The ~820-line tail of free builders — make_frame/make_frame_csc/
make_frame_common, make_video_image, probe_rgb_direct, the H.265/AV1
parameter-set writers and the AV1 bit-writer — moves to a #[path] child
module, the amf_sys.rs shape: the child sees the parent's private items
(Frame and friends), so the split costs zero visibility churn. Six
parent-called items went pub(super); five stay child-private (dead_code
is per-item and each is used within the child). vulkan_video.rs drops
5,292 → 4,489 lines and the construction unsafe gets its own review
surface; steady-state encode logic stays in the parent.
⚠ Trap recorded for future child-module splits: inline `use super::X`
statements INSIDE moved fn bodies silently change meaning (super shifts
one level) — vk_av1_encode/vk_valve_rgb imports needed crate:: paths.
Proven on-glass, not just compiled: all 8 vulkan GPU smokes green under
the validation layers on the 780M post-split (H.265 + AV1, RGB-direct,
CSC, CPU paths — every moved constructor exercised).
nvenc_cuda.rs and qsv.rs are DECLINED the same treatment, with evidence
in the handoff doc: no equivalent self-contained seam — their candidate
regions are ~150-line loader/accessor clusters interleaved with the
encoders' own state types, and a thin-forwarder impl split is exactly
the churn a no-defect phase penalizes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Linux backend decision existed as open_video_backend's string match
plus five partial hand-copies (the zero-copy-plane gate, the pyro
advertisement gate, the software advertisement pin, the vulkan pref
ceiling, resolved_backend_is_gpu). Windows solved this long ago —
windows_resolved_backend() is consumed by its dispatch AND its mirrors,
with labels still stamped at the open sites. Linux now has the twin:
resolve_linux_backend (pure, lazy auto probe) + linux_resolved_backend
(config wrapper, unknown→auto exactly as every mirror's old `_` arm).
⚠ This deliberately DEVIATES from the audit's shadow-assertion
prescription, on both critics' findings: the shadow had no execution
venue (open_video had ZERO test call sites; shipped hosts are --release)
— unfalsifiable ceremony — and the file's own Windows half proves
dispatch-consumes-resolver is safe: the mgmt record is protected by the
label-from-the-open-site convention, not by resolver avoidance. One
consumed table beats two tables plus an inert cross-check.
Preservation riders from the critique, all applied:
- pref_ceiling KEEPS its cfg!(vulkan-encode) branch (the resolver is
feature-blind; dropping it re-creates advertise-then-die-at-open).
- linux_zero_copy_is_vaapi's Vulkan|Software arm preserves the old `_`
fallthrough EXACTLY — the vulkan-on-NVIDIA capture-plane mismatch and
the software twin are FILED in the design doc, not fixed here.
- resolved_backend_is_gpu splits as linux + not(any(windows, linux)) —
never a target list, so no exotic target loses the fn.
- The auto probe is lazy (impl FnOnce) — explicit prefs stay zero-probe
(/serverinfo polls through these mirrors), pinned by a panicking
closure in the resolver test.
New coverage: the full alias table pinned, and a GPU-free dispatch test
through the real open_video_backend (software arm, vendored openh264) —
its first test call site anywhere. Config-latch seam guarded loudly.
Cross-crate mirrors recorded out of scope in the design doc:
session_plan::resolve_encoder, gamestream/serverinfo.rs
base_codec_mode_support, capture.rs's "pyrowave" string.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The ~410-line hand-mirrored AMF vtable ABI was already isolated in an
inline `mod sys` — the remaining value of the audit's 'best split in
the crate' is the FILE boundary: amf.rs drops 3379 → 2965 lines and the
pure unsafe-FFI surface (25 unsafe fns, zero policy) is reviewable in
isolation, which is the crate's stated review goal.
Pure move: `#[path = "amf_sys.rs"] mod sys;` keeps the module name and
every `sys::` call site byte-identical; the module was self-contained
(only `use std::ffi::c_void`, no super:: references). The banner became
the file's module docs; contents de-indented one level — rustfmt-clean
on the first check, which is what proves the move byte-exact.
Gated on .173: all five clippy combos + the amf-qsv,qsv test leg (38
passed — the live AMF matrix on real VCN silicon among them).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The two direct-NVENC backends carried hand-copied twins of the same
loss-recovery decision: range validity, covering-range dedup, DPB
window, clamp — ~30 duplicated lines each. The decision now lives once
as nvenc_core::plan_range_recovery (the range half of WP7.2; the slot
half is enc/rfi.rs), pure and unit-tested; each backend keeps its
session gate, its unsafe per-timestamp driver loop, and its state
stores.
The step order is load-bearing and now pinned by tests: the covering
dedup runs with the UNCLAMPED last and BEFORE the DPB window (a covered
re-ask never touches the driver even when the range has since aged out
of the DPB), the boundary at next_ts - RFI_DPB is inclusive, and the
Invalidate carries the CLAMPED last — which is also what the caller
records in last_rfi_range, exactly as the inline code stored it. A
driver failure mid-loop still returns false with NO range recorded and
no anchor armed. Decline deliberately clears nothing (neither twin
touched pending_anchor on decline — same shape as Vulkan's non-clear,
opposite of AMF/QSV; do not harmonize).
The exact-cover → Covered test records EXISTING behavior including that
a covered range survives a forced IDR with zero driver calls — a
recorded fact, not an endorsement. RFI_DPB's import leaves both twins:
its only per-backend use was the arithmetic that moved.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
AMF (user-LTR bitfield), QSV (mfxExtRefListCtrl) and Vulkan Video (the
app-owned DPB slot table) each hand-implemented the same loss-recovery
decision: distrust every reference encoded at-or-after the loss start,
anchor on the newest one strictly older. Three copies had already
diverged once — the fecbec2d taint sweep reached AMF/QSV a commit before
the Vulkan backend was carved out, and Vulkan shipped without it. The
decision now lives once in enc/rfi.rs, pure and unit-tested (this path
had zero coverage and is the loss-recovery path); every mechanism —
how a force is applied, how distrust is persisted — stays in its
backend.
Behavior-preservation notes, each critic-verified against the shipped
code:
- Callers feed only currently-trusted references; taint (>= loss) and
anchor (< loss) predicates are disjoint, so pre-sweep-view pick ==
post-sweep-table pick on every input. Tie-break preserved (first
entry wins == the strict '>' all three used, ascending slot order).
- plan_slot_recovery delegates its pick half to pick_anchor, which
keeps both items live on every leg (pick_anchor's only external
caller is Linux+vulkan-encode; dead_code is an item lint) and makes
'anchor chosen from the taint snapshot' structural.
- The decline arms are deliberately NOT uniform and stay put: AMF/QSV
clear an un-consumed pending_force; Vulkan leaves pending_loss armed
(a stale arm re-resolves at frame-build into the healing IDR) — now
pinned in comments on both sides.
- Vulkan's frame-build site is a pick-only re-run by design (the arm
carries the loss start, not the slot); pick_recovery_slot is deleted
and its tests migrated 1:1 into the shared module.
Test-count deltas are expected: the rfi tests now run on every Windows
combo (amf.rs is featureless there) and on the Linux feature leg.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A defaulted Encoder method that TrackedEncoder doesn't forward silently
no-ops through the wrapper — the host loop only ever holds the wrapped
box, so the feature dies for every session with nothing in the logs. The
trap has bitten three times (set_wire_chunking's §4.4 chunking probe,
set_pipelined's LN3 escalation, applied_bitrate_bps's ABR truth), and
every Phase 7 consolidation that adds a trait method re-arms it.
Source-text parse of the trait and the forwarding impl (both top-level
rustfmt items: block ends at the first column-0 brace, method names sit
on 'fn '-prefixed lines), then set equality — the reverse direction is
already a compile error, so equality == completeness. Mutation-verified:
removing set_wire_chunking's forward fails naming exactly that method.
16/16 today.
Limit stated in the test doc: name-set equality only — a forward whose
body delegates to the WRONG inner method still passes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The HDR/4:4:4 leg (188edde2, shipped 0.14.0) taught the client's plane
ring to allocate R16_UNORM on a 10-bit session, but the
pyrowave_image_view structs handed to decode_gpu_buffer kept the
original SDR shape: image_format/view_format hardcoded R8_UNORM, chroma
extents hardcoded half-res. pyrowave wraps our VkImage under the
DECLARED format and vkCreateImageView's its storage view from it
(pyrowave_c.cpp WrappedViewBuffers::wrap) — an R8 view over a
non-MUTABLE R16 image is invalid Vulkan that release drivers execute
anyway: the view's addressing covers half the surface, so the iDWT's
8-bit codes fuse pairwise into 16-bit texels (structured tile-scrambled
garbage) and the never-written remainder samples as all-plane zeros
(saturated green). Field report 2026-07-25: RX 9070 XT host → 780M
client, total corruption on every 10-bit PyroWave session regardless of
bitrate/refresh, wire counters spotless; milder shapes of the same UB
on NVIDIA clients; Arc validation never saw it because it ran SDR
(R8 ring — declaration consistent by accident).
Declare what the ring really built: format from hdr16 (R16/R8), chroma
extents from chroma444 (full-/half-res) — exactly what upstream's own
pyrowave_image_get_image_view fills from the image itself. Host
encoders audited clean: Windows keys plane formats off hdr16 already,
Linux is 8-bit end to end.
The tray led its tooltip with "⚠ conflicting host: Sunshine" and (on Linux)
went NeedsAttention for a Sunshine that was merely *installed* — never
started, never listening, harmless. An always-on warning over a dormant
package is a false alarm on every poll, and the tray is the wrong surface
for a one-shot install-time observation anyway.
Removes the `conflicts` field from the tray's `Summary`, the tooltip prefix,
`has_conflicts()`, and the ksni attention branch that fed on it. The host
still detects and reports conflicts where that belongs — the startup
`punktfunk::detect` warning, the `detect-conflicts` subcommand, and the
`/api/v1/local/summary` field, which the tray must (and does, per the new
test) keep tolerating on the wire.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
WP4.5's device-selection half, closed as the observability intermediate after
TWO selection designs died in adversarial review:
- Attempt #1 (d26bcf05, withdrawn): match pf_gpu::selected_gpu(). Its Linux
auto arm answers "the NVIDIA GPU" whenever /dev/nvidiactl exists, moving
the encoder off the iGPU that can import the compositor's dmabufs on an
Intel-compositor + NVIDIA-present laptop — import failures feed the
process-wide raw-dmabuf latch, which never un-latches.
- Attempt #2 (this session, withdrawn before commit): anchor on the
PUNKTFUNK_RENDER_NODE-else-renderD128 node via VK_EXT_physical_device_drm.
Render minors are driver-BIND-ORDER artifacts, not display topology: on
the common AMD-iGPU + NVIDIA-display desktop, in-tree amdgpu binds before
out-of-tree nvidia, so the anchor deterministically picks the idle iGPU
while the compositor allocates on NVIDIA — the same latch, opposite
polarity, behind a success-looking log.
The correct oracle is evidence of which device ALLOCATED the capture
buffers — producer identity from the capture negotiation, threaded per
session into this open. Until that plumbing exists, selection stays
first-usable, both call sites still share one selector (pure over the
device list, so capture_modifiers and open_inner cannot diverge — including
across an in-place resize's re-open, which does not renegotiate capture),
and the open logs ONE greppable line: picked vendor/device, the anchor node
and its owner (DRM render major/minor, VK_EXT_pci_bus_info fallback), and
the console's selected GPU. A wrong-device session on a multi-GPU host used
to be completely invisible; a field report can now show it. No WARN arm on
purpose: the wrong-pick direction inverts between the laptop and desktop
topologies, so a mismatch is not evidence of a wrong pick, and a warning
that fires forever on healthy hosts teaches people to ignore warnings.
Decision recorded against the audit's framing: manual console GPU selection
stays unhonored by pyrowave on Linux (the Windows twin honors it) — honoring
console-mutable state without per-session threading is what made attempt #1
unsafe.
Verified on the 780M: the line resolves all three identities (1002:15bf x3,
DRM-props match live); full pyrowave on-glass suite green; selection
behaviour byte-for-byte unchanged.
WP4.5 (device half).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Second attempt at WP6.3; the first (ce543668) was withdrawn after its tight
CBR window measured as a 36x bandwidth regression (97% filler NALs). The
correction that unlocked this one: that measurement's BASELINE row was an
8-frame artifact. Under the shipped 1000ms/500ms CBR window a calm stream
overflows the CPB once the initial fill drains (~30 frames at 10Mbps/60fps)
and RADV then pads every frame to the exact rate share, forever. Measured on
the 780M (1280x720@60, 10Mbps, calm content): 64 frames = 97.5% filler;
300 frames = 5.63MB at 98.5% filler where this commit ships 83KB at 0%.
AV1: 99.6% -> 0%. The status quo was ~the full target bitrate of zeros on
every idle AMD/Intel Vulkan-encode desktop and Steam Deck — and stuffing to
exactly the target permanently satisfied the ABR calm brake
(actual >= 3/4 * current), the ratchet WP6.3's withdrawal feared from the
tight window, live in the shipped code all along.
The fix reads VkVideoEncodeCapabilitiesKHR::rateControlModes (previously
ignored — rateControlMode was hardcoded CBR with no capability check) and
installs VBR with average == max plus the house ~1-frame window
(vbv_window_ms, PUNKTFUNK_VBV_FRAMES-scaled) when the driver advertises VBR.
VBR permits underspend, the exact missing degree of freedom: Vulkan exposes
no filler-suppression control (AMF's filler_data=false / NVENC's default-off
have no VK equivalent), so the MODE is the only lever. CBR-only drivers keep
the loose window untouched — tightening it under CBR just starts the stuffing
30 frames earlier. Drivers advertising neither mode (ANV per current Mesa)
keep the pre-existing CBR install, now WARN-logged.
No pacing claim, deliberately: burst A/B on the 780M is byte-identical
between 1000ms CBR and 17ms VBR (max AU 1.19MB in both) — this firmware
ignores the window for QP decisions entirely. The payload is filler
elimination. PUNKTFUNK_VULKAN_RC=cbr|vbr is the field escape hatch and the
on-box A/B control (two withdrawn attempts bought that insurance).
Also on the same caps struct: maxBitrate is now read and clamps open +
retarget (RADV reports 1 Gbps — within 5% of the 4K120 ABR targets), and
applied_bitrate_bps() reports the encoder-side truth (pending-first, so the
session loop's read right after reconfigure_bitrate sees the clamp) — without
it a binding clamp would feed the ABR a phantom base, the trap the trait doc
names.
And the one-frame VUID-vkCmdBeginVideoCodingKHR-pBeginInfo-08254 violation
found in the withdrawal review: record_submit promoted a pending retarget
into self.bitrate BEFORE recording whenever first_frame was set, so after a
mid-stream reset() (which preserves the pending rate and rc_installed) the
begin-coding declaration named a rate the session had not installed — and the
two triggers, ABR retarget and the stall watchdog, correlate. Now the
declaration always names the session's current rate and the RESET install
carries the pending one via its own struct; promotion stays in
post_submit_bookkeeping. The extended validation-layer test reproduces the
retarget-then-reset coincidence: exactly one 08254 on the pre-fix build, zero
on this one (RADV PHOENIX, on glass).
Gates: docker amd64 legs green; Windows .173 seven legs (34 passed);
.25 full-suite parity vs origin/main (identical CUDA-only failures) + all 9
vulkan on-glass tests + validation layers clean.
WP6.3, plus WP7.1's ms-form half (vbv_window_ms).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PyroWave never checked frame dimensions against the session, and it applies no
alignment — `width`/`height` are the negotiated mode verbatim — so a mismatched
frame was encoded edge-smeared or cropped, silently, forever.
Every other Linux backend already refuses exactly this, with this shape, in
`submit`: libav-NVENC (`linux/mod.rs`), VAAPI (`vaapi.rs`) and openh264 (`sw.rs`)
all carry the same `ensure!`. PyroWave was the only one that didn't. That is the
justification; an earlier draft cited `vulkan_video.rs` instead, which is the
weaker precedent — it bails in its Dmabuf arms only, and its CSC path and CPU arm
have no dimension guard at all.
Mostly this is a wrong-picture bug and not a memory-safety one: `rgb2yuv.comp`
clamps every fetch with `min(p, textureSize - 1)` and the CPU arm uploads
`min(len, need)` into a session-sized image. But it also closes a narrow real
hazard that was not in the filing: `import_cached` keys on `(st_dev, st_ino)` and
returns the cached `VkImage` on a hit WITHOUT rechecking the extent, and unlike
the capture side it is never cleared on a renegotiation — so a dmabuf inode
recycled across a shrinking renegotiation would hand the encoder an image sized
for the old, larger allocation. This check closes that route.
⚠ Recorded at the code because it changes the failure mode, not just the
detection: a mismatch is NOT always transient. A compositor-initiated PipeWire
renegotiation updates the capturer's size in place and signals nothing the encode
loop reads, so it can be a permanent new steady state — and `reset()` reopens at
the same dimensions by construction, so the host's five-reset budget cannot
recover (~3.1 s of frozen stream, then the session ends) where before it would
have streamed on with a wrong picture. For a real mode change that trade is
clearly right; for a 16-row KWin mismatch it is not, and the proper fix is for
the host to classify this error as a PIPELINE rebuild rather than an encoder
reset. Filed, not done here.
The device-selection half of WP4.5 was written, reviewed and WITHDRAWN — see the
handoff doc. Matching the selected render GPU regresses the hybrid
Intel-compositor + NVIDIA-present topology this project has a live field report
for, because `selected_gpu()` answers NVIDIA whenever `/dev/nvidiactl` exists
regardless of where capture actually runs.
WP4.5 (dimension half).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The RGB-direct CPU-upload path allocated and zero-filled a whole padded frame on
every submit, filled it row by row, then memcpy'd the whole thing into the mapped
staging buffer. The zero-fill was entirely dead: the row loop writes every byte
of every row, and rows past the source re-copy the last source row. So the frame
paid for an allocation, a page-fault storm over fresh pages, a full zero-fill,
and two full-frame copies where one would do.
Map first, write the padded rows straight into the mapping. Nothing reads back
from the destination — the row source is always the caller's buffer, never the
staging memory — so writing into (write-combined) host memory costs nothing
extra, and `make_host_buffer` allocates HOST_COHERENT, so the writes need no
flush before the transfer reads them.
Measured on real RDNA3 780M silicon, release build, `submit` alone, 400 frames
x 3 rounds:
- **1920x1080** (rows padded to 1088): p50 2315 -> **1725 us** (-25%),
p99 2819 -> **1972 us** (-30%), and the spread (p99-min) tightens 739 -> 483 us.
- **1366x768 -> 1408x768** (BOTH axes padded, so the column tail loop runs):
p50 1013 -> **926 us** (-8%), p99 1303 -> **1110 us** (-15%). This is the one
case where the new code could have been slower — 4-byte stores straight into
write-combined memory — and it is not.
- **CONTROL, 1280x720** (64x16-aligned, so the branch is never entered):
p50 692 vs 692 us. No delta, which is what makes the two above attributable to
this change rather than to anything else on the branch.
The branch is reached by any RGB-direct session at a mode that is not 64x16
aligned, whenever capture delivers CPU frames (the default Linux capture path is
dmabuf and never enters it). 1080p qualifies, since 1080 aligns to 1088. ⚠ An
earlier draft of this message claimed "33 MB at 4K" — that is wrong: 3840 and
2160 are both already aligned, so 4K UHD never enters this branch at all. The
large case is an ultrawide like 3440x1440 -> 3456x1440, ~20 MB.
Three things this deliberately does NOT do:
- The extent guards stay scoped to the `pad` branch. The CSC path deliberately
supports a source SMALLER than the encode extent — its shader clamps at sample
time — so a guard hoisted above the branch would break it.
- Every fallible step stays above the map. An error raised between `map_memory`
and `unmap_memory` would strand the mapping for the life of the slot's staging
buffer, and the next frame's `map_memory` on it then violates
VUID-vkMapMemory-memory-00678. (The filed "two exits above the map leak Vulkan
objects" hazard is separately already gone: `47a23bec` moved that unwind into
`make_host_buffer`.)
- It adds guards rather than removing them: a zero source axis made `sh - 1`
underflow, and "cannot fail after the map" has to be true by construction.
Three corrections to an earlier draft, all found by review of that draft:
- The `dw*dh*4 == need` precondition was a `debug_assert!` placed BELOW the map.
That is wrong twice: it made the only check on the slice length vanish from the
builds that ship, and a fired assert would have unwound past `unmap_memory` —
the exact failure the bullet above says was designed out. It is now a real
checked `?` above the map.
- `need` was `(iw * ih * 4) as u64`: a u32 multiply widened after the fact, which
agreed with the usize slice length only up to ~32768x32768. The old code's
`min(need)` was a hard backstop against exactly that and the rewrite dropped
it. Now widened before the multiply, matching `read_slot`'s existing discipline.
- The extent guard now runs BEFORE the payload-length guard, so the usize
`sw * sh * 4` cannot overflow on a garbage frame header.
WP6.2(a).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`std::env::var` was on three per-frame paths. Measured on `.173` (Windows, 57
environment variables, 2M iterations): **121.9 ns** per call for the NVENC
in-flight cap and **114.9 ns** per call for the ffmpeg poll spin, against
**0.9 ns** once memoized. On Linux, 32 ns → 1.5 ns.
⚠ Those numbers deflate the filing, and that is worth recording: the audit
ranked this as a hot-path defect, but ~120 ns/frame is ~0.003% of a frame
budget. The fix is still right — it is free, and it takes a global environment
lock off the encode thread — but nobody should schedule it ahead of anything on
the strength of the "hot path" framing.
The severity RANKING was also inverted, and the measurement confirms why. The
site the audit called worst — `nvenc_cuda`'s backpressure loop condition — costs
a default session nothing, because the condition short-circuits on
`async_rt.is_some()` and the default session never engages the two-thread
retrieve. The one that actually pays every frame is Windows `submit`, which
consults `async_inflight_cap()` in BOTH arms of the ring-depth match, sync mode
included, where the result is then thrown away. Windows `poll` is the second
unconditional one, and the audit ranked it last.
Memoized inside each helper rather than latched into a session field. Nothing in
the workspace mutates these variables at runtime (enumerated: no `set_var` for
either key anywhere in first-party code, and the Windows service's arbitrary-key
`host.env` loader runs in the SCM supervisor, which re-execs the host as a child
and never opens an encoder itself). A field would instead change WHEN the value
is read — and the Windows `cap` composes the env with `input_ring_depth`, which
`set_input_ring_depth` may change after open, so freezing that half would
reintroduce the in-place-overwrite bug the ring term exists to prevent.
Two deliberate behaviour changes ride along on `PUNKTFUNK_FFWIN_POLL_MS`, so
"behaviour-preserving" describes the memoization only, not this whole commit:
- **A 1000 ms ceiling.** The reachable hazard was never the overflow — that
needed `ms >= 1.8e16` — it was a slipped digit: `=100000000` was a 27.7-hour
spin of the encode thread.
- **`.trim()`, now on all three parsers.** An earlier draft applied the house
rule (WP7.8) to one of the three, which left `PUNKTFUNK_NVENC_ASYNC=" 1 "`
working while `PUNKTFUNK_NVENC_ASYNC_DEPTH=" 6 "` silently fell back to 4 —
and memoization would have frozen that silent fallback for the process
lifetime. Reachable: the Windows `host.env` loader trims around `=` before
stripping quotes, so `=" 2 "` yields a value with inner spaces.
⚠ Correction to an earlier draft of this message, which asserted that the audit's
`saturating_mul` proposal would relocate an overflow panic into release builds.
That is FALSE and the code comment now says so: `Duration::from_micros(u64::MAX)`
is ~1.8e13 seconds, six orders of magnitude below `Duration`'s ceiling, so
`Instant + Duration` neither overflows nor panics (measured, with and without
debug assertions). `saturating_mul` is still wrong, for a different reason — it
sets a deadline ~584,000 years out on a spin that provably never produces the
owed AU, i.e. it wedges the encode thread permanently. A hang, not a panic.
WP6.1.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`encodeCodecConfig` is a C union, so the `hevcConfig` writes in the 4:4:4 branch are only
meaningful on an HEVC session — on H.264 or AV1 they reinterpret that codec's own config bytes.
The branch was gated purely on `chroma_444 && full_chroma_input` with no codec test, and stayed
non-UB only because `lib.rs` degrades 4:4:4 for non-HEVC codecs: a two-file invariant with
nothing asserting it, on the path BOTH direct-NVENC backends take.
The audit filed that much. What it did not note is that the same shape hides a second bug:
this is an `if`/`else if`, so a non-HEVC session that arrived with `chroma_444` set took the
HEVC branch and skipped the per-codec bit-depth arm entirely — ending up with neither HEVC
4:4:4 (wrong for it) nor its own 10-bit configuration (simply absent). AV1 would have lost
`pixelBitDepthMinus8`/`inputPixelBitDepthMinus8` silently. Non-HEVC now falls through to the
arm that knows what to do with it, and an unexpected request is logged rather than swallowed.
Adds two tests that need no GPU — `apply_low_latency_config` is pure config authoring, so an
AV1 session can assert it never receives the HEVC FREXT profile GUID (an INVALID_PARAM at
open) and that its own depth still lands, with an HEVC case guarding the good path.
⚠ `NV_ENC_CONFIG` must NOT be `mem::zeroed` in those tests: `frameFieldMode`/`mvPrecision` are
C enums whose discriminants start at 1, so all-zero is not a valid value and Rust's zero-init
check ABORTS the process (SIGABRT, caught by the Linux gate). They seed it from `Default` the
same way `build_config` does before overwriting from the driver preset. Worth knowing before
writing any further test against these SDK types.
Verified: Linux gate L1-L4 green (39 tests, the two new ones among them) and the full Windows
gate on .173 — 7 legs. Note the Windows test leg runs `--features qsv`, so these tests only
execute on the Linux leg; the Windows legs prove the change compiles in all five feature
combinations.
`ir_active` was `cfg.intra_refresh && set.co2.is_some()` — i.e. "we asked for it", not "we got
it". Both `Query` and `Init` can return MFX_WRN_INCOMPATIBLE_VIDEO_PARAM, a WARNING the open
path deliberately accepts, while dropping the intra-refresh wave on the floor.
That lie is not cosmetic. `ir_active` feeds `EncoderCaps::intra_refresh`, so the session
advertises gradual refresh to the client; the client then stops asking for the IDRs it would
otherwise request on packet loss, and a lost frame gets concealed instead of repaired. The
stream degrades exactly when recovery matters.
Now confirmed against the driver: a best-effort `GetVideoParam` with a CodingOption2 buffer
chained on, read back for `IntRefType`. Deliberately a SEPARATE query rather than a buffer
attached to the existing `BufferSizeInKB` call (which is what the audit finding suggested) —
that value backs every bitstream allocation, and a runtime that disliked the chained buffer
would take it down with the readback. A failure here costs only the verdict and resolves
conservatively (trust the request), so the worst case is the previous behaviour.
Verified on Intel UHD 750 (.42), which is where the Intel on-glass run originally caught this:
with PUNKTFUNK_INTRA_REFRESH=1, H.264 now logs "silently dropped intra-refresh
(GetVideoParam reports IntRefType=0)" and advertises it OFF, while H.265 on the same GPU
stays ON — it is genuinely active there. Per-codec, so it could never have been decided
statically. 8 live QSV tests pass.
Gated on .173: clippy -D warnings at nvenc,amf-qsv,qsv (host + pf-encode --all-targets),
amf-qsv without qsv, qsv alone, no-features, cargo test --features qsv, rustfmt — all green.
qsv.rs is Windows-only so the Linux legs do not compile it.
Closes WP3.2(c), the last Phase 3 item that was hardware-blocked.
Phase 5's Linux half (audit WP5.1 + WP5.4), each item shaped by the
review that rejected the obvious fix:
Dmabuf import unwind (vk_util): every failure after create_image leaked
the VkImage, and the dup'd dmabuf fd leaked as a raw i32. The sharp edge
is that a SUCCESSFUL vkAllocateMemory transfers fd ownership to Vulkan
(vkFreeMemory closes it), so the naive close-on-error is a double close
that clobbers whatever unrelated descriptor recycled the number. The dup
now lives in an OwnedFd released exactly in the allocate-success arm;
every other path drops it once, and bind/view failures free image+memory.
PyroWave open unwind: open_inner had ~20 fallible steps that each leaked
everything before them (instance, device, pyrowave objects, the whole
CSC pipeline). Rather than a parallel teardown guard — whose reviewed
hazards were a null-unsafe pyrowave_encoder_destroy and a drifting
duplicate of Drop — Self is now constructed right after create_device
with every later resource null, and the existing Drop (wait-idle first,
pw_enc null-guarded, delete-nullptr and VK_NULL_HANDLE destroys are
no-ops) is the single unwind path for error and normal teardown alike.
The ensure_cpu_rgb staging twins (create/allocate/bind, both backends)
and the RGB-direct make_view pair get the same discipline via a shared
make_host_buffer. Observed on hardware: 32 forced import failures, zero
fd drift (the new import_failure_leaks_no_fds smoke on RADV).
24-bpp CPU service (WP5.4): pixel_to_vk had no mapping for the packed
Rgb/Bgr the PipeWire portal negotiates, so a session committed to a path
the backend could not serve and died at its first frame. The filed
open-gate was rejected as a half-mirror — the dmabuf axis is keyed by
fourcc at submit, unknowable at open — so instead the CPU axis is
SERVED: a 3-to-4 expand at the staging upload (normalize_cpu_rgb, the
CPU twin of WP1.4's swscale expand), order-preserving for the CSC
samplers and BGRA-forced for the RGB-direct encode source, whose session
pictureFormat is B8G8R8A8 — the on-glass run caught R-first sources
violating VUID-vkCmdEncodeVideoKHR-pEncodeInfo-08207, a mismatch that
predates this change for plain Rgbx CPU sources. The dmabuf axis feeds
pf-zerocopy's raw-dmabuf degrade latch (3efbe416) from both Vulkan
import caches — deterministic refusals flip capture to CPU delivery,
which now serves every format capture produces; transient OOM and
native-NV12 sessions are excluded.
RECORDING-state hygiene: the fallible recording prefix (cursor prep,
import, staging) now resets the command buffer on error instead of
leaving it RECORDING for the next begin to trip
VUID-vkBeginCommandBuffer-commandBuffer-00049; PyroWave's submit-level
blanket reset is replaced by the same scoped shape, closing its
fence-timeout hole where a reset hit a PENDING buffer
(VUID-vkResetCommandBuffer-commandBuffer-00045).
Verified: docker linux/amd64 4-leg gate green; RADV 780M on-glass under
the validation layers at 1920x1080 — all 10 smokes pass (including the
new 24-bpp CSC + RGB-direct legs and both channel orders), decoded
colour truth checked with ffmpeg, and the only remaining validation
messages are the two documented VALVE-extension layer gaps plus a
pre-existing pyrowave shader/feature mismatch now filed in the planning
doc.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
import_plane closed the shared NT handle on every pyrowave_image_create
failure, believing pyrowave only consumes it on success. The vendored
truth was messier: Granite's allocator closed by-reference handles
UNCONDITIONALLY after the first vkAllocateMemory — success AND failure —
while failures before the allocator (validation, vkCreateImage, no
memory type) left the handle open, and both classes surface as the same
error code. So the call site could not know whether to close, and an
allocate-stage failure double-closed a possibly-recycled handle value
(audit WP5.3). The filed fix (DuplicateHandle + close the original
unconditionally) traded the double close for a guaranteed leak on every
pre-allocate failure and left the ambiguity in place.
Patch 0006 fixes the callee instead: the allocator never closes a
caller's handle, and pyrowave_image_create consumes it exactly at its
success return — which is what pyrowave.h ("take ownership and close the
HANDLE on import") documents, and what Granite's semaphore import
already did, so import_fence was correct as written and the two paths
now share one contract. The close-on-failure in import_plane is thereby
correct on EVERY path, including a vkBindImageMemory failure after a
successful allocate. Bonus fix recorded in the patch: the allocator's
block-recycling retry loop used to re-run vkAllocateMemory with
import_info still pointing at the just-closed handle.
Both-platform gates green; pyrowave_win_smoke (.173, 34/34) re-validates
the live import path against the rebuilt vendored C++.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three Windows NVENC teardown/accounting defects (audit WP5.2, marked
BLOCKING — the obvious fixes were reviewed and rejected before this):
Event-handle leak: init_session pushed each completion event to
self.events only AFTER nvEncRegisterAsyncEvent, so a failed registration
leaked the Win32 handle. Push first — the init-failure path already runs
teardown, which closes everything in the list (unregistering the one
never-registered event is a harmless error return).
Wedged drain: teardown drains the retrieve thread's backlog, and each
queued job could burn its full 5 s completion wait — cap x 5 s on the
encode thread, paid again on every rebuild of the stall-recovery ladder.
The rejected fix (a shutdown flag) abandoned the backlog on EVERY
teardown, turning the routine drain into destroy-while-encoding. Instead
the wedge evidence stays where it lives: after one full-budget timeout in
the retrieve loop, later jobs wait a 250 ms slice; a success resets the
latch. Routine teardown is byte-identical to before — nothing is ever
abandoned — and a first timeout is already encoder-fatal, so short
slices behind an established wedge change only how long teardown blocks.
Session-unit accounting: LIVE_SESSION_UNITS was refunded even when
destroy_encoder FAILED, drifting the budget low and over-admitting
parallel displays. The rejected fix (never refund on failure) let one
transient wedge episode poison admission until a host restart. Now the
refund follows PROOF: destroy succeeded, or its status says the driver
holds no session (device gone/TDR). Ambiguous failures park the handle —
units still charged, fail-closed, pinning the D3D11 device the driver
session references — and init_session retries the destroy while zero
sessions are live, under a gate that serializes every session open
against the reap so a recycled handle address can never alias a live
session. The charge lasts exactly as long as the driver keeps refusing
the destroy, which is the definition of still-leaked. Admission stays a
lock-free atomic load.
Both-platform gates green (docker linux/amd64 4 legs; .173 all 7 legs).
The retry-destroy-on-a-failed-handle assumption is documented at the
reap: the SDK defines no retry semantics either way.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A dmabuf import the GPU driver refuses is refused identically on every
retry, but the only recovery above it was the encode-stall ladder: five
in-place encoder rebuilds, then the video session ends. So a host whose
driver will not take what its compositor allocates lost every session
on its first frame, and every reconnect repeated it — while the very
same host streamed fine with `PUNKTFUNK_ZEROCOPY=0`. The software knew
how to run that machine and never chose to.
Latch it, exactly as the sibling CUDA-import path already does after
repeated worker deaths: three consecutive import failures with no frame
in between disable the raw-dmabuf passthrough for the host process, and
capture negotiates CPU frames from the next session on. Three sits
below the encoder's rebuild budget, so the latch is set before the
session it doomed ends — one bad session, then a working (if slower)
host, with a log line saying which and why instead of an operator
having to find an environment variable.
Only the two stages that ARE the import are counted — the buffersrc
push of the DRM-PRIME descriptor and the buffersink pull where `hwmap`
maps it into a VA surface. `avcodec_send_frame` is deliberately left
out: that one is the encoder stalling, which the in-place rebuild
exists to recover, and taking zero-copy away permanently over a
transient fault would be a bad trade.
The latch lives in pf-zerocopy because it is the leaf both sides can
see — the capture→encode edge is one-way by design, so pf-capture
cannot ask pf-encode anything.
The zero-copy DRM-PRIME descriptor declared `objects[0].size = 0`, on
the belief that ffmpeg would work the real size out. It does not: both
of its import paths pass the value straight through to libva —
`prime_desc.objects[i].size` on the PRIME_2 path, `buffer_desc.data_size`
on the legacy fallback it tries next — so every VA driver we have ever
handed a dmabuf to was told the backing object was empty and left to
derive the size itself.
The drivers this path has run on (radeonsi, modern Intel iHD) derive it
correctly, which is why nobody noticed. A Gen9 Intel host does not get
that far: `vaCreateSurfaces` answers VA_STATUS_ERROR_ALLOCATION_FAILED
on the first frame and every frame after it, `av_buffersink_get_frame`
returns EIO, and five in-place encoder rebuilds later the video session
is over. That host cannot stream at all with zero-copy on.
`lseek(SEEK_END)` is the standard dma-buf size query and the same one
this tree's Vulkan bridge already performs on these very fds. A kernel
that refuses it leaves the old 0 rather than costing a frame that might
still have encoded.
Whether this alone fixes that host is unconfirmed — the descriptor was
wrong either way, and it is the first thing the driver reads.
The EGL importer — the head of the CUDA/NVENC zero-copy path — opened
`/dev/dri/renderD128` and hoped. On a single-GPU host that is the
NVIDIA node and it always worked. On a hybrid laptop the iGPU is bound
first, so renderD128 is Intel: we built a GBM device on Mesa, asked it
for a pbuffer-capable OpenGL config, got none (Mesa's GBM platform
advertises only window-capable ones — gbm_surface is the point of that
platform there), and reported
zerocopy worker init failed: no EGL config for OpenGL
on a machine whose NVIDIA EGL stack was demonstrably fine, naming
neither the device nor the reason. `PUNKTFUNK_RENDER_NODE` was no
escape either: this path never read it.
Pick the node by what it is — the first `/dev/dri/renderD*` whose
sysfs PCI vendor is 0x10de, the same identification the crate's Vulkan
bridge already uses for its physical device. A local scan rather than a
`pf_gpu` dependency, because this crate is a leaf whose worker is its
own process, and `pf_gpu::linux_render_node` answers a different
question anyway: it follows the operator's GPU *preference*, which may
legitimately name the iGPU. What is needed here is not which GPU to use
but where CUDA is. `PUNKTFUNK_ZEROCOPY_RENDER_NODE` overrides, and no
NVIDIA node found keeps the historical guess — sysfs may just not be
mounted, and the CUDA context is what properly fails on a host with no
NVIDIA at all.
Then stop the config query being able to fail this way: retry without
the surface-type constraint, which is honest because we never create an
EGLSurface — `eglMakeCurrent` runs surfaceless. And name the node in
every error and in the ready line, so the next hybrid host diagnoses
itself.
`punktfunk-scripting` with nothing to run pinned a full CPU core
indefinitely — in exactly the state its own systemd unit documents as
inert ("the runner does nothing until you add scripts or install
plugins"). A field report on a 7.7 GB laptop had it at 99.9% CPU for
two hours straight, `strace` showing a bare `clock_gettime` loop and no
other syscall at all, which starved the compositor badly enough to
blank the physical display.
Nothing in the runner was looping. `Effect.runFork` registers nothing
with the event loop — effect's keep-alive timer lives in
`Runtime.makeRunMain`, which runner-cli deliberately doesn't use
because its shutdown is the bespoke two-signal one — and a bun process
whose only pending work is an unresolved promise busy-spins rather than
blocking. With units running there is normally a socket or timer to
park on; with none there is nothing at all, so the no-op state was the
one state that span.
Hold one idle handle for the runner's lifetime and drop it when the
unit tree ends on its own. Measured on the packaged bun (1.3.14), three
seconds idle with empty scripts/plugins dirs: 3098 ms of CPU before,
280 ms after (bun's startup, then nothing).
The regression is a property of the process rather than of any function
in it, so the test spawns the real CLI and reads the child's CPU time,
asserting it also logged and stayed alive — a crashed runner burns no
CPU either and must not pass.
Found by running the smokes at 1920x1080 under the validation layers while chasing the three
items left over from WP4.2. AV1 forbids the encode source's `codedExtent` differing from the
sequence header without `FRAME_SIZE_OVERRIDE`
(VUID-vkCmdEncodeVideoKHR-flags-10324), or from the reference slots without
`MOTION_VECTOR_SCALING` (`-10325`). RADV PHOENIX advertises NEITHER — and RGB-direct is the
default on EFC hosts with true-extent the default at unaligned modes, so plain 1080p AV1 was
tripping both on every frame: source 1920x1080 against an app-aligned 1920x1088 header and
DPB. Measured 16 violations per 8-frame run; the CSC path had none.
The fix is to make all three agree at the RENDER size rather than the aligned one — an
unpadded coded size is valid on this hardware, so the coded frame simply IS the visible
frame:
- the AV1 sequence header follows the render size when true-extent is active (joining
native NV12, which already authors true-size headers for the same reason);
- the DPB setup and reference slots carry `src_extent` instead of the aligned `ext2d`.
`src_extent` already collapses to `ext2d` whenever true-extent is off, so every other
configuration is untouched;
- `render_and_frame_size_different` now compares render against the DECLARED source extent
instead of the aligned size, or true-extent would have claimed a mismatch that no longer
exists.
Fixing only the header is not enough and is actively misleading: it clears -10324 and
immediately exposes -10325, because the mismatch has moved to the reference slots rather than
gone. Both had to move together.
This keeps the EFC fast path. Two alternatives were implemented, measured and rejected on the
way here: falling back to the compute CSC costs the zero-copy the B2 work existed to deliver,
and routing to the padded-copy staging trades these two VUIDs for
VUID-VkImageCreateInfo-pNext-06811 — `pad_img`'s extra TRANSFER_SRC usage is not in the
profile-advertised set, measured 8x per session on HEVC-padded too, so that is a pre-existing
defect of the padded path and not somewhere to route a default session.
HEVC is deliberately untouched: it has no equivalent constraint (its crop rides the
conformance window), it measures zero violations, and its aligned-SPS path is the validated
one.
On RADV PHOENIX (780M, Mesa 26.0.4) the AV1 stream now decodes as coded 1920x1080 / render
1920x1080 — genuinely unpadded, where before the alignment rows were encoded and cropped
back out. The CSC path still reports coded 1920x1088 / render 1920x1080, which is correct for
it. All four `vulkan_smoke*` pass at 256x256 and 1920x1080.
Two validation errors remain on the RGB-direct path and are NOT ours, now with evidence
rather than assumption: `VUID-VkImageViewCreateInfo-image-08336` uses the PROFILE-BLIND format
query, so it cannot see that RGB conversion legalises BGRA as an encode source — the
profile-aware query used by -06811 accepts the very same image; and
`VUID-VkQueryPoolCreateInfo-pNext-pNext` rejects
`VkVideoEncodeProfileRgbConversionInfoVALVE`, which the VALVE extension REQUIRES for profile
identity, and the layer diagnoses itself as "a struct from an extension added to a later
version of the Vulkan header".
Verified: canonical Linux gate (docker linux/amd64) L1-L4 green; on-glass on RADV PHOENIX
under `VK_LOADER_LAYERS_ENABLE='*validation*'`, with the bitstreams read back through
libdav1d/trace_headers.
Four items, gated on the RTX box (.173, all 7 Windows legs) and the Linux docker gate, with
the QSV HDR one proven by an A/B on real Intel silicon (.42, UHD 750).
WP4.1 — Windows NVENC sticky state. Both halves were the same mistake: a single field doing
duty as both "what the session negotiated" and "what this session can actually do", so the
first downgrade destroyed the negotiation.
- `query_caps` clears `self.hdr` on a GPU without 10-bit encode, but `submit` re-derives the
requested HDR state from every frame's pixel format and compared it against that cleared
value. On such a GPU a P010 capturer therefore reported "HDR changed" on EVERY FRAME and
tore down and rebuilt the whole encode session each time. The trigger now compares against
`hdr_requested`, and the downgrade is latched in `hdr_unsupported` so it is remembered
(and warned once) instead of rediscovered per init.
- One subsampled-YUV frame cleared `chroma_444` permanently, so 4:4:4 never returned when
the capturer went back to RGB. The negotiation now lives in `chroma_444_requested` and the
effective value is recomputed at every init; the per-session downgrade still happens and
is still reported honestly through `caps()`, it just is not destructive any more.
WP4.3(a) — QSV HDR mastering luminance was divided by 10,000. The old comment justified it as
"VPL wants whole cd/m²", which is true of the VIDEO-PROCESSING unit but not the encode path:
the runtime passes the value straight into the ITU-T H.265 Annex D mastering-display SEI,
whose unit is 0.0001 cd/m². Verified end to end on Intel UHD 750 by dumping the bitstream and
reading SEI 137 with trace_headers, before and after:
before: max_display_mastering_luminance = 1000 (0.1 nits)
after: max_display_mastering_luminance = 10000000 (1000 nits)
with `min` (500 = 0.05 nits) and both content-light-level fields unchanged. That asymmetry —
a correct min beside a 10,000x-low max — was the original tell.
WP4.3(b) — `reconfigure_bitrate` never re-read `BufferSizeInKB`, so an ABR step-up could hand
the runtime an AU buffer sized for the old, lower bitrate. Fixed in both places it was broken:
re-read the driver's answer via `GetVideoParam` after the `Reset` (mirroring `init_encode`),
AND stop `take_bs` recycling pooled buffers without checking their capacity — the pool would
otherwise keep serving short buffers even once `bs_bytes` was correct.
WP7.8 — the `PUNKTFUNK_QSV_FFMPEG` half, finally landable now that a Windows gate exists (it
sits behind `#[cfg(feature = "qsv")]`, so nothing in a Linux-only matrix compiles it — it was
deliberately held back for exactly that reason). Trimmed like its siblings, and kept
case-insensitive: the knob already accepted `TRUE`, and the bare house `matches!` would have
silently dropped that spelling.
Phase 4 of the audit is now complete, and WP7.8 with it.
Verified: Windows gate on .173 — clippy -D warnings at nvenc,amf-qsv,qsv (host + pf-encode
--all-targets), amf-qsv without qsv, qsv alone, no-features, `cargo test --features qsv`
(34 passed), rustfmt. Linux docker gate L1-L4 green. On-glass: the QSV A/B above on .42.
Owed: NVENC sticky-state on-glass on .173 (needs a live 10-bit-capable capture toggle).
WP4.2 of the pf-encode audit, validated on RADV PHOENIX (780M, Mesa 26.0.4) under the Vulkan
validation layers rather than by inspection alone. Three of the five reproduce with the
driver naming the violation; one turned out milder than filed and is recorded as such.
1. `ensure_cpu_rgb` cached the staging image on FORMAT ONLY while sizing it to the SOURCE
frame, and the copy uses the CURRENT frame's extent — so a same-format source-size
increase copied past the allocation. Now keyed on (format, width, height). This is the
memory-safety one: on hardware it was 8 x
VUID-vkCmdCopyBufferToImage-imageSubresource-07971 ("extent.width (512) exceeds
imageSubresource width extent (128)") and `submit` returned Ok every single time, so
nothing upstream could notice. Zero after the fix. Note the trigger for anyone re-testing:
the image is cached PER RING SLOT, so the ring must wrap before a slot sees a larger frame
— two differently-sized submits in a row land on different slots and look fine.
2. The RGB-direct CPU padding loop only ever grows a source into the aligned extent; a source
LARGER than the encode extent panicked on the row `copy_from_slice`. Now refused by name,
matching what the DMA-BUF arms already do, so a bad frame takes the encoder-rebuild path
instead of unwinding out of the encode thread with an index message.
3. AV1 wrote `render_width/height_minus_1` but never set
`render_and_frame_size_different`, and AV1 ignores the render size unless that flag says it
differs — so every mode needing 64x16 alignment presented the padding. libdav1d on the
encoded stream now reports `size 1920x1088 ... render 1920x1080`; before, the flag being
unset meant render defaulted to the coded 1088, i.e. 8 rows of duplicated edge pixels
shipped to the client.
4. `read_slot` read the PERF timestamp pool whenever the pool merely EXISTED, but the
padded-RGB path's CPU-upload arm records its own command buffer and writes no timestamps —
reading an unreset query with WAIT is undefined. Now gated on a per-slot `ts_written`,
because "a pool exists" was being used as proof of "the pool was written". SEVERITY
CORRECTION vs the audit: this is not a hang in practice. It needs PUNKTFUNK_PERF *and*
RGB-direct *and* the non-default PUNKTFUNK_VULKAN_RGB_TRUE_EXTENT=0 *and* an unaligned mode
*and* CPU capture, and RADV then fails the read rather than blocking. Another driver may
block, hence the fix.
5. `reset()` re-arms `first_frame`, which was also what decided whether begin-video-coding
declares the rate-control state. But a reset does not rebuild the session, so the CBR
installed earlier is still current when the next frame opens its coding scope: the
declaration was omitted exactly when it was required. Split into `rc_installed`, which
survives reset. On hardware this was
VUID-vkCmdBeginVideoCodingKHR-pBeginInfo-08253 ("no VkVideoEncodeRateControlInfoKHR ... but
the currently set mode is CBR"); zero after.
Also fixes two pre-existing AV1 spec violations that were firing on EVERY frame of a shipped
path and are not in the audit at all — found only because the validation layers were on, and
confirmed untouched by this branch before being attributed as pre-existing:
`constantQIndex` must be 0 unless rate control is DISABLED (VUID-...-10320; this session is
always CBR, so the driver owns Q), and `pStdPictureInfo->pSegmentation` must be NULL
(VUID-...-10350; it pointed at a zeroed struct, and leaving it unset means the same
"segmentation disabled" while being spec-correct).
Items 1 and 5 keep their reproductions as `#[ignore]`d hardware tests, documented as
meaningful only under `VK_LOADER_LAYERS_ENABLE='*validation*'` since the violations are
invisible to the API's return values.
STILL OPEN, filed not fixed (three more pre-existing validation errors the cleanup exposed):
VUID-VkQueryPoolCreateInfo-pNext-pNext on the AV1 and RGB paths, and 2 x
VUID-VkImageViewCreateInfo-image-08336 (the RGB-direct encode-src view's format features lack
VIDEO_ENCODE_INPUT) — the latter may be a validation-layer gap around the VALVE extension
rather than our bug, and wants real analysis rather than a quick patch.
Verified: canonical Linux gate (docker linux/amd64) fmt + clippy --all-targets at default and
at nvenc,vulkan-encode,pyrowave + both test legs; and on RADV PHOENIX all four `vulkan_smoke*`
tests pass at 256x256 and at 1920x1080 with zero validation errors on the paths touched here.
Both come straight out of the first compositor-less bring-up (P3), on a
two-GPU box:
PUNKTFUNK_DRM_CARD=<n> pins SDL's KMSDRM device index. SDL enumerates
/dev/dri/card* and takes the first it can open, which is regularly the
wrong one: it chose the card a live compositor already held DRM master on
and died at swapchain creation, while the idle card with the connected
display sat unused. Kept an explicit operator choice rather than
auto-detection — deciding "is this card already mastered" needs the very
ioctl that taking master IS, so any in-process guess would be fragile.
The swapchain error now carries what to actually check. "vkCreateSwapchainKHR:
Initialization of an object has failed" is useless to someone bringing up a
kiosk; on the kmsdrm backend it now names the pinned card and lists the
three real causes in order (no connected connector / another DRM master /
NVIDIA). Empty on every other backend, where it would be noise.
The NVIDIA note is measured, not guessed: on NVIDIA proprietary + kmsdrm,
Vulkan enumerates the GPU AND the display (VK_KHR_display reports the
connected HDMI connector) and still fails — as root, with
nvidia_drm.modeset=Y, on a card no compositor was using. Not permissions,
not DRM master; their direct-display path wants the display leased via
vkAcquireDrmDisplayEXT and SDL's kmsdrm surface path does not do that.
Plan: punktfunk-planning design/embedded-arm64-client.md §P3
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two things were x86-specific, both now expressed properly instead of
hardcoded:
* PKG_CONFIG_PATH carries the runtime's multiarch directory.
flatpak-builder does NOT shell-expand `env` values, so ${FLATPAK_ARCH}
would be taken literally — a build-options.arch override supplies the
aarch64 string and inherits the rest.
* The prebuilt Skia archive is per-target and pinned by sha256. Two
`type: file` sources discriminated by only-arches now share one
dest-filename, so SKIA_BINARIES_URL stays a single literal path.
Upstream publishes the aarch64 archive under the same skia commit hash
and the same resolved-feature key, so the two stay in lockstep on a
skia-safe bump.
build-flatpak.sh gains ARCH (default: this machine's), passes --arch to
both flatpak-builder and build-bundle, and arch-suffixes the bundle name so
an x86_64 and an aarch64 build can coexist in dist/ instead of the second
silently overwriting the first. CI composes its own published filename, so
nothing downstream changes.
The aarch64 Skia sha256 was verified by downloading the published archive,
not taken from the API. No aarch64 flatpak has been built — flatpak-builder
builds in a sandbox for the target arch, which needs an arm64 machine.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
arch=('x86_64' 'aarch64'), and on aarch64 `pkgname` drops punktfunk-host
so makepkg never enters the host's build or package path; build() takes a
client-only cargo invocation without the NVENC/Vulkan-encode features.
Dropping the host from pkgname rather than giving package_punktfunk-host a
per-package arch is what makes the skip unambiguous, and it mirrors how
PF_WITH_WEB already extends that array.
The host stays x86-only because its encode stack (NVENC/QSV/AMF) is.
Verified by sourcing the PKGBUILD under both CARCH values: x86_64 yields
`punktfunk-host punktfunk-client`, aarch64 yields `punktfunk-client`. NOT
verified by a real build — there is no official arm64 Arch container, and
the README says so rather than implying it is tested.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The blocker was never ExclusiveArch — it was that %build builds host +
tray + client together and %install lays down the host unconditionally,
with no way to express "client only". `%bcond_without host` (default ON,
so an x86_64 build is unchanged) now gates the host binary, the tray, the
headless-session data, the firewalld services, the main %post and the bare
%files section. Omitting %files for the MAIN package is the load-bearing
part: it is what stops rpm emitting an empty `punktfunk` alongside
punktfunk-client.
build-rpm.sh exposes it as PF_WITHOUT_HOST=1 and skips the libcuda stub
regeneration and the libcuda leak check, neither of which means anything
without the host.
Verified two ways. `rpmspec -P` shows the default expansion still carrying
3 host install lines, the tray build and one bare %files, while
--without host carries none of those and one %files client. And a real
rpmbuild in a native aarch64 Fedora 43 container produced
punktfunk-client-*.aarch64.rpm with correct aarch64 sonames
(libc.so.6(GLIBC_2.17), libSDL3.so.0, libavcodec.so.61) and no main
package.
Not a cross-compile: %build runs cargo for the build machine's arch, so
this wants an arm64 builder. No CI leg yet — deliberate.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Adds a `rust-arm64` job: clippy -D warnings for the client crates at
--target aarch64-unknown-linux-gnu, plus a --no-default-features session
build so the minimal embedded configuration keeps compiling in its own
right rather than only as a subset of the default one. Runs in the cross
image on the ordinary amd64 runner; client crates are listed explicitly
because --workspace would drag in the x86-only host encode stack.
This is a regression guard, not an artifact leg (deb.yml ships those).
c_char signedness cuts both ways and neither direction is visible from an
x86-only CI: hardcoding i8 fails to COMPILE on ARM, while casting to u8
fails the LINT on ARM. Two such defects were already found this way.
Also documents the arm64 client .deb recipe in packaging/debian/README.md.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`c_char` is i8 on x86_64 and u8 on aarch64, so `cert_pem_out as *mut u8`
is a REQUIRED conversion on one target and a no-op on the other — where
clippy::unnecessary_cast then denies it. `.cast::<u8>()` is correct and
lint-clean on both.
Caught by the new aarch64 clippy leg on its first run. A sweep of the
remaining `as *mut u8` / `as *const u8` sites in the client crates found no
other c_char-derived casts; the rest convert from c_void or u16.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Adds punktfunk-rust-ci-arm64cross (the rust-ci toolchain + an Ubuntu ports
arm64 multiarch sysroot) and a deb.yml job that cross-builds the client
package on the ordinary amd64 runner — no arm64 runner in the fleet and
none needed. Client only: the Linux host encodes with NVENC/QSV/AMF, all
x86. The apt registry keys pool entries by the .deb's own arch, so an arm64
box picks the package up with no client-side configuration.
Three things this had to get right, each of which cost a failed build:
* rustup target add must run against the toolchain rust-toolchain.toml
PINS, not the image's default `stable` — otherwise the pinned toolchain
has no aarch64 std and the build dies ~230 crates in on "can't find
crate for core". Copying the pin file in also pre-downloads the
toolchain every CI job currently re-fetches on first use.
* ffmpeg-sys-next compiles a probe it intends to RUN, so it forces
.target(HOST) while still passing the TARGET's include paths; the arm64
dir then shadows the host's own libc headers and the x86 compiler dies
in bits/math-vector.h on NEON/SVE types. Prepending the host dir does
NOT help — GCC drops a -I that duplicates a system directory, keeping
its original later position — so ci/pf-host-cc strips target include
dirs from host compiles instead.
* the job hard-fails on a non-AArch64 session binary, rather than
shipping an amd64 payload under an arm64 package name.
skia needs no special handling: the aarch64-linux-gnu textlayout+vulkan
prebuilt resolves, so arm64 ships the FULL client, OSD included — unlike
Windows ARM64, which still builds --no-default-features.
The cross image builds in its own docker.yml job (needs: build-push) since
it is FROM punktfunk-rust-ci:latest and must not race the entry that
publishes that base. rpm/Arch/Flatpak stay x86_64 for now — their builders
have no arm64 sysroot, which is its own piece of work.
Plan: punktfunk-planning design/embedded-arm64-client.md
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`punktfunk-session --pair <PIN> --connect host[:port]` runs the SPAKE2
ceremony with no window and no toolkit, prints the same
`paired <addr>:<port> fp=<hex>` line as `punktfunk-client --pair`, and
exits. Until now the PIN ceremony lived only in the GTK shell or the Skia
console, so enrolling an embedded/kiosk client meant installing a desktop
on it — or copying the identity store by hand.
Dispatches above every graphics call (the machine may have no display) and
is in the `--no-default-features` build: enrolling must never be the reason
a minimal image pulls in Skia. `--name` defaults to the hostname instead of
the desktop path's hardcoded "Steam Deck".
`forget_placeholder` and `device_name` move into pf_client_core::trust so
the two binaries share one implementation rather than the session
re-deriving them.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`char` is signed on x86_64 but UNSIGNED on aarch64, so the hardcoded
`Vec<*const i8>` compiles on every desktop target and then fails to match
ash's `&[*const c_char]` when cross-compiling the client for ARM.
Found by the aarch64 cross-build spike (punktfunk-planning:
embedded-arm64-client.md) — the only portability defect in the client
stack; nothing else in the client crates assumes an architecture.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
openh264 tops out at level 5.2 — 3840x2160 landscape or 2160x3840 portrait — and enforces
that ceiling inside `reinit`, which the crate calls on the FIRST ENCODE rather than at
encoder construction. So an oversized mode built a perfectly healthy-looking encoder and
then failed every single submit: the session connects, negotiates, and never delivers a
frame, with the real reason buried in a per-frame error rather than at the open.
`validate_dimensions` does not cover this. It is keyed on the codec, and H.264 legitimately
reaches 4096 on every hardware backend — this ceiling belongs to the software backend alone,
which is exactly the path a GPU-less host falls back to.
Rejects at open instead, mirroring the rule from the openh264 version we actually ship
(0.9.3) rather than from its docs — including the orientation-aware shape, since a naive
per-axis `w <= 3840 && h <= 2160` would wrongly refuse a legal 2160x3840 portrait session.
Three tests: the accepted modes in both orientations, the modes `validate_dimensions` lets
through but openh264 cannot serve, and that `open` itself refuses rather than deferring to
submit. They cost nothing to run — the guard fires before openh264 is initialised at all.
Verified with the canonical Linux gate (docker linux/amd64): fmt, clippy --all-targets at
default and at nvenc,vulkan-encode,pyrowave, and the pf-encode test leg (37 passed).
The knob tested `v == "0"` for the disable case and treated everything else that was set as
a force-enable. So a trailing space — the kind a `.env` line or a shell script leaves behind
— made `PUNKTFUNK_VULKAN_RGB_DIRECT=0 ` mean the exact opposite of what the operator wrote,
forcing the RGB-direct encode source on. On a cursor-blend session that is not a subtle
difference: the EFC front-end cannot blend, so the pointer disappears from the stream. An
empty value and any typo did the same thing.
Now parsed like every sibling knob in the crate — `matches!(v.trim(), "1"|"true"|"yes"|"on")`
and the matching false spellings — so unrecognised input falls back to the default instead
of being read as a force-on. `=0` and `=1` keep working exactly as documented.
Splits the parse into a pure function so the accepted spellings are covered by tests. That
matters more than it looks: the env-var read itself is untestable in a parallel test binary
(the process environment is shared), which is precisely why this knob's behaviour had no
coverage and the inverted case survived.
Verified with the canonical Linux gate (docker linux/amd64): fmt, clippy --all-targets at
default and at nvenc,vulkan-encode,pyrowave, and the pf-encode test leg (37 passed).
Every capability probe drops libav's log level to AV_LOG_FATAL around an encoder open it
expects to fail, then restores what was there before. That level is one process-global int
and the probes genuinely overlap — they are reached both from `/serverinfo` and from session
bring-up. Two overlapping save/restore pairs interleave as get(INFO) → get(FATAL) →
set(INFO) → set(FATAL), and the process is then pinned at AV_LOG_FATAL permanently: every
later libav diagnostic silently dropped, including the ones you want when a stream fails to
open later on.
Replaces the four hand-rolled save/restore pairs with one RAII guard over a single shared
mutex. The audit filed two sites; there are four — the NVENC 4:4:4 and 10-bit probes in
`linux/mod.rs` and both VAAPI probes, which is why the lock has to be shared across the two
modules rather than kept local to either.
Being RAII also closes a smaller hole the old shape had: the restore was a statement after
the open, so any early return added to those functions later would have leaked the quiet
level. Now it cannot.
The guard is poison-tolerant — a probe that panicked mid-window has already restored the
level through `Drop`, so refusing the lock forever afterwards would be strictly worse than
proceeding. It is not re-entrant, which is safe here because no probe body reaches another
probe (they only call encoder-open helpers); that invariant is written down at the type.
Serialising the probes costs nothing measurable: they are process-once behind their caches
and each already pays for a real encoder open.
Verified with the canonical Linux gate (docker linux/amd64): fmt, clippy --all-targets at
default and at nvenc,vulkan-encode,pyrowave, and the pf-encode test leg (37 passed).
`vk_av1_encode.rs` and `vk_valve_rgb.rs` are hand-copied `#[repr(C)]` structs handed to the
driver through raw `p_next` chains. Nothing in the type system relates them to the C
definitions any more, so an edit that inserts, drops, widens or re-pads a field is not a
compile error — it is the driver reading our bytes at the wrong offsets, silently. The
sibling vendored ABI in `amf.rs` has carried assertions for exactly this reason; these two
had none at all.
Adds size, alignment and per-field offset assertions for all 19 vendored structs. They are
`const` rather than `#[cfg(test)]` (the shape `amf.rs` uses) so they hold in every build
including the shipped one, and on any target the modules compile for.
Assertions cannot catch a swap of two same-typed fields — the offsets are unchanged — so
the field order was diffed field-by-field against the authoritative headers while writing
them: `vulkan_core.h` and `vk_video/vulkan_video_codec_av1std_encode.h` from Vulkan-Headers
`main` as of 2026-07-25. That diff covered every struct, every `ST_*` structure-type value,
every flag bit and both enum groups, and found no drift — the vendored copies are faithful.
The one remaining hand-copied table the compiler still cannot see is the bitfield member
order inside the three `*Flags` words, where a wrong index means the driver reads
`use_superres` where we meant `render_and_frame_size_different`. Three tests pin those to
the header by listing the members in C declaration order and asserting the Nth setter
writes bit N, with the trailing `reserved` field checked too — a dropped member shifts
`reserved` down and fails.
Verified with the canonical Linux gate (docker linux/amd64): fmt, clippy `--all-targets`
default and `nvenc,vulkan-encode,pyrowave`, and both test legs.
`NV_ENC_ERR_INVALID_VERSION` is how the driver reports two opposite failures, and we only
ever explained one of them. The message told the operator to update the NVIDIA driver or
reboot — correct for a genuine header/kernel-module skew, and actively misleading for the
field case behind it: a host that streams once per boot and then fails every later session
at the caps probe, until the *process* restarts. A version skew is static; it cannot come
and go inside one process, so that advice cost the reporter a reboot per stream.
Split the two on the only fact that tells them apart: whether a session has already opened
in this process. `nvenc_status` gains a `SESSION_OPENED` latch set right after every
successful `open_encode_session_ex` — both backends' caps probe and real open, plus the
Windows availability probe. No session yet, the version word really is in question and the
existing skew advice stands. A session already opened, and the kernel module demonstrably
accepted this build's version word, so the message now names per-process driver state and
points at the cheap fix (restart the host service, no reboot) plus a request for the log.
The load-time gate cannot serve as this discriminator: `NvEncodeAPIGetMaxSupportedVersion`
is a pure userspace query, so the classic "updated the driver, didn't reboot" skew sails
through it and only fails later at the open. Only a session that actually opened proves the
kernel module agreed.
The split lives in a pure `invalid_version(bool)` so both halves are unit-tested without
touching the process-wide latch. This is a diagnosis change only — it does not fix the
underlying field bug, whose root cause is still open.
Verified on Linux (192.168.1.25): clippy `-p pf-encode --features nvenc` clean, `cargo test
-p pf-encode --features nvenc --lib` 15 passed, rustfmt clean.
`33121ece` added a block explicitly labelled "TEMP KWin composite probe (DROP BEFORE
MERGE)" to prove the cursor blend was dispatching. It merged, and has been running on
every blended frame since: an atomic fetch_add per frame plus a `tracing::info!` with
seven computed fields every 512 frames, on the submit hot path of the default-on
Linux direct-NVENC backend.
The signal it existed for is already covered — the failure arm above it warns once
with the dispatch error, and the `else` arm warns when an overlay arrives with no
blend at all. What is deleted is only the success-path telemetry.
Verified `-D warnings --all-targets` + tests on Linux (default; shipped
nvenc,vulkan-encode,pyrowave), and `clippy --features vulkan-encode,pyrowave` on real
AMD RDNA3 hardware.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
With `PUNKTFUNK_ENCODER=vulkan`, `open_video_backend`'s vulkan arm bails outright for
anything that is not HEVC/AV1 ("the Vulkan Video encoder supports HEVC + AV1; the
session negotiated {codec:?}"). But `host_wire_caps` for that same pref fell through
to the VAAPI probe / static superset, which includes H.264 — so the host advertised
H.264, a client could negotiate it, and the session died at encoder open.
The pref now contributes a CEILING that is intersected with the device probe, never a
replacement for it. That distinction is the whole fix: pinning a static HEVC|AV1
would have ADDED AV1 on the AMD/Intel hosts whose probe currently withholds it
(pre-RDNA3, pre-Arc), re-creating this very bug for a different codec. Intersecting
can only ever narrow.
Without the `vulkan-encode` feature the pref cannot open anything at all — that arm
bails with "requires a build with --features vulkan-encode" — so the ceiling is
empty there rather than optimistic.
Costs nothing on the handshake path: a `&str` match plus a const-folded `cfg!`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`LP_MODE` cached the resolved VAAPI entrypoint (full-feature `EncSlice` vs low-power
`EncSliceLP`/VDEnc) in a process-global array indexed by codec alone. That is not a
staleness bug, it is a session-killer, because the cache is load-bearing: once a mode
is latched the open tries exactly ONE mode and the `[false, true]` fallback is gone.
Latch low-power on an Intel Arc (Gen12+ removed the full-feature entrypoints, so it
is the only one that works there), then switch the web-console GPU preference to an
AMD dGPU. `render_node()` follows that preference, so every VAAPI open now goes to
radeonsi passing `low_power=1`, which it rejects — with no full-feature retry, for
the process lifetime. `probe_can_encode` reports all-false (so the advertisement
silently falls back to the static superset) AND the session's own encoder open fails.
Now keyed on (render node, codec, bit depth):
- The render node because the entrypoint is a property of the DEVICE libva opens,
and it is literally what `render_node()` hands libva — so key and device cannot
describe different GPUs. Deliberately not `pf_gpu::selection_key()`, which can
name a different adapter than the node actually opened.
- The bit depth because Main10 and 8-bit can resolve to different entrypoints on
the same device; one shared slot let an 8-bit answer pin the 10-bit open, i.e.
HDR under-advertisement.
Verified `-D warnings --all-targets` + tests on Linux (default; shipped
nvenc,vulkan-encode,pyrowave). The multi-GPU switch itself needs a box with two
VAAPI-capable adapters — owed on-glass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`base_codec_mode_support` fell through to the static superset (H.264|HEVC|AV1) on a
GPU-less host, so Moonlight negotiated HEVC or AV1 and the session then died at
encoder open with "the software encoder emits H.264 only" — openh264 encodes nothing
else. The native plane's twin of this function, `pf_encode::Codec::host_wire_caps`,
has gated on exactly this since it was written; this one never did.
Deliberately a local gate rather than delegating wholesale to `host_wire_caps()`.
Delegation is the drift-proof shape and was the first design, but on Windows it
re-runs the DXGI adapter enumeration several times per `/serverinfo` GET — the probe
helpers each sample it uncached — and this endpoint is polled by every client. The
software case is a plain config read, so it costs nothing here.
Recorded in a comment rather than fixed: the static `MaxLumaPixelsHEVC` in the
serverinfo XML still advertises an HEVC limit even when the mask now drops HEVC.
Harmless (Moonlight gates capability on the mask, not the limit) but it is a second,
now-inconsistent advertisement.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`flush` looked dead — the only caller in the host is the `spike` dev subcommand — so
an audit sweep filed it as "wire it in or delete it". Both are wrong, and without
this note the next sweep will re-file it.
Wiring it in is refuted by control flow: both encode loops reach their exit only
AFTER the transport is gone (client disconnected, or the session stopped), so the AUs
a flush would recover have nowhere to go. And flush is the one call on this trait
that can BLOCK on a wedged encoder, on exactly the teardown path a stopped session
needs to finish promptly — the Linux direct-SDK NVENC retrieve-thread join is untimed,
so flushing there could hang a session that is already ending.
Deleting it is refuted by real consumers: `spike` encodes a FINITE clip and wants the
tail, and the `#[ignore]`d hardware smoke tests across the backends assert the drain
contract on real GPUs. Those are finite-stream users; a live session is not one.
Doc only, no behaviour change.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`open_video`'s `cursor_blend` argument was a request with no answer: lib.rs did
`let _ = cursor_blend;` and only three backends ever read `CapturedFrame::cursor`.
So a session could ask for a composited pointer, get a backend that silently
discards it, and stream with no mouse cursor and nothing in the logs. Two
separately-confirmed audit findings — the VAAPI dmabuf path and the libav-NVENC CUDA
path — are symptoms of that one hole.
`EncoderCaps::blends_cursor` makes it a fact each backend states. The four exhaustive
`EncoderCaps { .. }` constructors mean adding the field is a compile error until every
backend answers, which is the enforcement mechanism for future backends rather than a
side effect. Vulkan Video answers from its ACTUAL configured source rather than
statically: only the CSC path composites (`prep_cursor` feeds the compute shader),
while the RGB-direct/EFC front-end and the native-NV12 source have no compositing
stage at all and merely warn once that the pointer is being dropped.
`open_video` warns when a session asked for blending and the opened backend cannot
deliver it. A warning is deliberately all it does: `open_video` cannot re-plan
capture, so refusing would trade a missing pointer for a dead session. The host owns
`plan.cursor_blend` and is the only layer that can fall back to capturer-side
compositing — this gives it something to base that on.
Enforcement is NOT included. The reviewed design proposed refusing the client's
host-composite flip to keep the client drawing its own pointer, but `CursorRenderMode`
is client->host only: there is no host->client counterpart, so refusing yields no
pointer at all — the same failure it claimed to prevent.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`set_input_ring_depth` is a DEFAULTED trait method, so a caller that forgets it
fails silently — and one did. The GameStream loop opens an encoder
(gamestream/stream.rs:671 and the rebuild at :831) and never called it, while the
Windows IDD-push capturer declares a ring of 2 and `async_inflight_cap()` defaults
to 4. With PUNKTFUNK_NVENC_ASYNC=1 a Moonlight session therefore pipelined four
encodes against a two-texture ring, letting the capturer rotate a texture out from
under a live encode: torn or mixed frames, never an error — precisely the corruption
the cap exists to prevent, and precisely what the trait doc warns "fails silently and
intermittently".
Guarded at the single point of CONSUMPTION rather than by plumbing the setter into
every loop: `input_ring_depth` has exactly one reader in the workspace, so one
fail-safe covers every caller including ones not yet written, whereas fixing N call
sites only fixes the N someone remembered. An unconfigured ring is now treated as
the shallowest any capturer here declares, so the unconfigured path degrades to less
pipelining — a latency cost, not corruption.
The two GameStream sites also pass the REAL depth, because the fail-safe is a floor,
not a substitute: `idd_depth` is configurable and a deeper ring is free pipelining
the fallback would forfeit.
Default (sync) sessions are unaffected — the cap is only read while the async
retrieve thread exists, and that is opt-in.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Escalate-and-hold's missing half. The contention escalation (capture depth,
then the NVENC pipelined retrieve) was permanent: one sustained overrun — even
one CAUSED by the ABR overdrive's rebuild storms — cost the session its
depth-1 latency and its sub-frame streaming forever, and `encode_us` reported
queue depth instead of ASIC time for the rest of the session.
- The leaky bucket now keeps scoring after escalation. A sustained
every-frame-on-cadence run (~5 s at 120 fps) winds back one stage in reverse
order: pipelined retrieve first (its rebuild restores the IO-stream binding
and sub-frame chunked streaming), then capture depth back to 1. Attempts are
paced by an exponential backoff (1 → 5 → 25 min, capped) — a workload that
truly needs the escalation converges to keeping it, but never a permanent
latch.
- NVENC (Linux) implements `set_pipelined(false)`: a `want_sync` latch handled
at the same drained safe point as the engage side (`maybe_disengage_async`
mirrors `maybe_engage_async`); the lazy sync re-init re-arms everything and
opens on an IDR. The stream loop polls until the switch lands, then re-runs
the escalation warmup so the wind-back's own stall can't re-escalate it.
`PUNKTFUNK_NVENC_ASYNC=1` (operator-pinned async) refuses the wind-back;
the trait doc now specifies the two-way contract.
- While escalated, `cadence_degraded` stays latched (bitrate climbs refused)
even with the bucket drained: the headroom is spent, and climbs resuming
mid-escalation would saw against it and starve the clean run.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Host half of the §ABR-overdrive fix. The stream loop now reads
`Encoder::applied_bitrate_bps()` after every bitrate apply and stores THAT into
`bitrate_kbps`/`live_bitrate` — the send pacer, web console, mgmt registry and
control-task acks all track what the ASIC really targets instead of the
requested rate (the pre-fix ack promised 1.01 Gbps while the encoder ran
794 Mbps, and the client controller climbed from the phantom base forever).
- A short apply teaches `encoder_ceiling_kbps` (shared atomic): the stream loop
pre-clamps incoming requests to it and SKIPS the apply when nothing would
change — ending the reconfigure-reject → full-rebuild(~0.6 s + IDR) storm —
and the control task resolves future SetBitrate acks against it, so the
client learns the ceiling through the existing ack path (no wire change; old
clients converge too).
- `cadence_degraded` (shared flag, leaky-bucket level ≥ 10 or an escalated
session): while set, the control task resolves climbs to the current applied
rate — on a fat LAN no network signal ever stops a climb the encoder can't
serve, and past the compute knee more bits only deepen the cadence miss.
Descents always pass; they're the cure.
- Escalation-warmup hygiene: an ABR rebuild stall (~70 missed deadlines at
120 fps, 3.5× the escalate threshold) and the backlog scored against a
heavier rate no longer feed the latency escalation — rebuilds reset the
leaky bucket and re-run the warmup; in-place down-steps clear the bucket.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two client-side halves of the §ABR-overdrive fix (the 4K120 sessions that
climbed to 1 Gbps against a 794 Mbps encoder and a 8.33 ms budget):
- Short-ack cap learning: the host now resolves a climb it can't serve to what
the encoder actually runs at (its codec-level ceiling, or the current rate
while encode is behind cadence). Two consecutive identical short acks latch
that value as a host cap the climb logic folds into its ceiling — one short
ack stays a transient (a failed rebuild also acks short once). The cap is
mode-scoped (cleared on an accepted mode switch, tracked via a mode
generation counter the control task bumps) and re-probes one step after ~60 s
parked clean, so a heavy-scene refusal can't quietly cap the whole session.
- Host-encode-latency down-driver: the per-AU 0xCF `encode_us` the host
already ships (and the overlay already draws) now feeds the controller
through its own window accumulator — the overlay channel is lossy and
embedder-drained, so the ABR gets a dedicated mirror of the decode-latency
path. Baseline-relative like the decode signal (an escalated host reports
encode_us inflated by ~a frame of queue depth; an absolute budget threshold
would read permanently red), with the baseline rebased after our own
decreases so one backoff doesn't train-fire into the floor. This is the only
signal that can push an already-too-high rate back under the encoder's
compute knee — host climb refusal stops the climb, but nothing else descends
on a clean LAN.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three encode-time fixes from the 4K120 field analysis (sessions pinned ~107 of
120 fps, 14-15 ms reported encode):
- Ceiling truth: the codec-level bitrate clamp (binary search at session open)
now records its result in a process-lifetime advisory cache keyed by
GPU/config, so an ABR overshoot opens — or in-place reconfigures — straight
AT the ceiling instead of re-running the ~6-open search (and a full rebuild
+ IDR) on every overshoot. New `Encoder::applied_bitrate_bps()` exposes the
post-clamp rate so the session loop can stop pacing/acking a phantom
requested rate (consumed host-side in a follow-up).
- Clamp-search hygiene: only NVENC parameter/caps rejections steer the search
now (`NvCallError` keeps the raw status downcastable); a transient failure
(busy engine, session limit, OOM, driver skew) propagates instead of
shrinking into — and now caching — a bogus ceiling. The floor fallback also
records the split mode it actually opened with, so a later reconfigure
re-presents the real session params.
- Split-frame selection: one shared `resolve_split_mode` replaces the two
byte-identical direct-SDK copies; the force-2-way threshold moves to
`SPLIT_FORCE_PIXEL_RATE` (950 Mpix/s, shared with the libav path) because
4K120 = 995,328,000 px/s missed the old `> 1e9` gate by 0.47% and stayed on
AUTO — which never engages at 2160 px height, leaving the second NVENC
engine idle in exactly the mode the threshold existed for. The Linux
session-ready info! line now carries the final split mode (journals are
INFO+; this was undiagnosable from user logs).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
09aa2db3 turned on `-p pf-encode --all-targets -D warnings` for Windows and removed
the crate-wide `allow(dead_code)`; together those surfaced two genuinely-dead items
in `nvenc,amf-qsv,qsv` — the combination the installer ships — and CI went red on my
own step.
- `WinVendor::Amf`: native AMF replaced the libavcodec AMF path in production, so
the only remaining CONSTRUCTOR is the `#[cfg(feature = "amf-qsv")]` latency A/B
in amf.rs — test code, so the lib target constructs it nowhere. The module
header already says this machinery is kept deliberately for that measurement, so
it gets a targeted `#[allow(dead_code)]` naming the reason rather than deletion.
- `probe_can_encode`: `lib.rs`'s only caller is under
`cfg(all(not(feature = "qsv"), feature = "amf-qsv"))`, because with the native
VPL backend compiled in `qsv::probe_can_encode` answers instead. Gated
`#[cfg(not(feature = "qsv"))]` to match its call site.
Both are real findings the blanket allow had been hiding; neither is a behaviour
change.
WHY THIS ESCAPED MY PRE-PUSH CHECKS, since that is the useful part: the Windows
runner I verify on has no FFmpeg dev tree, so `amf-qsv` could not build there and
`ffmpeg_win.rs` was compiled by NOTHING in my matrix — the same class of blind spot
this whole audit is about, reproduced by me. CI's own FFmpeg turns out to be cached
on that box at C:\Users\Public\ffmpeg, so the verification recipe now sets
FFMPEG_DIR and covers `nvenc,amf-qsv,qsv` (the CI command and my new step) plus
`amf-qsv` without `qsv` — the combo that still uses `probe_can_encode`. All three
clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The tray sat at "Idle" through an entire stream, and its Windows context menu
could come up without its main action. Two separate causes.
`GET /api/v1/local/summary` reported `video_streaming`/`audio_streaming` straight
from `AppState.streaming` — flags only the GameStream media pipeline raises. The
native punktfunk/1 plane is the DEFAULT (GameStream is opt-in) and never touches
them, so every native session read as idle: idle icon, and a tooltip that printed
that session's own resolution next to the word "idle" (the `session` field was
already native-aware). This is the blind spot `/status` was fixed for — see the
`session_status` module doc — which the summary kept. Both flags now OR in a live
native session, and the tray additionally treats a reported session as streaming,
so a new tray reads an older host correctly too.
The menu built "Open web console" (and the pairing entry) only while a live
loopback probe of the console had just succeeded. A console still starting, or an
SSR render slower than the 2 s probe timeout, therefore deleted the tray's
most-wanted action outright — indistinguishable from a tray that never had one,
with left-clicking the icon as the only, undiscoverable, way in. The entry is now
always present and the default item; a failed probe changes its label to
"(not responding)" instead of hiding it, and takes two consecutive misses to say
so, since one timeout is not "down".
While here: a "Release kept display…" entry when displays are held (the summary
field's own doc promised a one-click release that did not exist), and entries
deep-link to /pairing and /displays instead of all landing on the dashboard. The
Linux SNI menu mirrors all of it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`linux_native_nv12_ok` is the verdict the host threads into capture negotiation
(session_plan -> OutputFormat::nv12_native -> ZeroCopyPolicy::native_nv12_session
-> pf-capture's prefer_native_nv12). Its doc claimed "the backend must be eligible
to open", but it asked only three static questions — codec is H265/AV1, an env
default, and a pref denylist — and never whether this GPU has an encode queue at
all. It could not: vulkan_video.rs exposed no probe.
That verdict is uniquely load-bearing. Once the producer has been asked for
two-plane NV12 there is NO fallback: open_video deliberately makes a failed Vulkan
open FATAL for an NV12 capture rather than degrading to libav VAAPI, because VAAPI
would import that buffer as packed RGB and stream silent garbage. So on a gamescope
host whose Mesa lacks Vulkan HEVC encode the session died at its first frame, while
the same host with PUNKTFUNK_PIPEWIRE_NV12=0 streamed fine — and the comment
promising such a device "degrades gracefully to the old backend rather than
breaking the stream" was stale on every gamescope host.
Two changes:
1. The gate. The third conjunct was a denylist of the EXPLICIT prefs that skip
Vulkan Video ("nvenc"|"nvidia"|"cuda"|"pyrowave"), which silently missed the one
that matters: the DEFAULT encoder_pref is "", and "" resolves to auto, which on
an NVIDIA box opens NVENC. A stock NVIDIA host passed this gate. It now consults
`linux_zero_copy_is_vaapi`, which layers the pref on top of the same auto
decision open_video makes — what the note on `linux_auto_is_vaapi` says a
capability probe must use, and what the downstream consumer of this very verdict
already uses. This alone was worth fixing; a probe added behind the old gate
would have opened a Vulkan instance on every NVIDIA handshake.
2. The probe. `vulkan_video::probe_encode_support` runs the FIRST check open_inner
performs and hard-fails on — the physical-device + encode-queue-family scan for
this codec's op — and nothing more, so it is provably no stricter than the open
and can never talk a working host out of the fast path. The scan is now a shared
`find_encode_device` used by BOTH, because a probe that mirrors a dispatch goes
stale the first time the dispatch grows a case (the failure open_video's
backend-label note already records). Cost: one instance plus physical-device
queries — no logical device, no video session, no VRAM — cached per (selected
GPU, codec) in the can_encode_10bit idiom, with the probe run outside the lock.
Per codec, not once: codec_op_for selects a different queue-family bit for AV1,
and HEVC-encode-without-AV1-encode is the common VCN/ANV configuration.
Later stages (create_device, create_video_session, the capability query) can still
fail for reasons the probe does not model. Those are harmless: the capture format
is only committed once the probe says yes, and everything after keeps the ordinary
packed-RGB negotiation.
Verified `-D warnings --all-targets` + tests on Linux (default; shipped
nvenc+vulkan-encode+pyrowave). The behaviour needs an on-glass check on the bazzite
RADV box — including the negative case, which `RADV_DEBUG=novideo` reproduces.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`submit_cpu`'s RGB24/BGR24 → rgb0/bgr0 expand ran `w*h` iterations per frame, each
building two bounds-checked sub-slices for a 3-byte copy plus a pad-byte store — a
shape LLVM will not vectorise into the byte shuffle it actually is. At 3840x2160
that is 8.3M iterations on the encode thread, per frame.
It is not an edge case. The module header says the portal commonly negotiates
packed 24-bit RGB, and pf-capture offers `VideoFormat::RGB` first because wlroots
commonly fixates it — so this was the mainstream CPU path, running roughly an order
of magnitude slower than the 4-bpp sibling branch (a plain row memcpy) purely
because of how it was written.
swscale's packed-RGB expanders are SIMD, the sibling VAAPI backend already routes
RGB24/BGR24 through them, and this file already owned the `sws_getContext` /
`sws_freeContext` lifecycle — so this is net subtractive rather than new machinery:
the existing CSC condition widens to include `expand`, and the hand-written loop
goes away. No new field, no second context, nothing added to `Drop`: `expand` is
false whenever `want_444` (see the `nvenc_pixel`/`expand` binding), so the three
users are mutually exclusive and one context serves whichever applies. The `expand`
field itself is gone with its only reader.
`sws_setColorspaceDetails` is now applied ONLY for the 4:4:4/HDR users. The expand
is a pure byte shuffle — NVENC does the RGB→YUV itself downstream — and handing it
a matrix and range would have silently range-converted every packed-RGB session,
which is exactly what the module header promises does not happen.
Verified `-D warnings --all-targets` + tests on Linux (default; shipped
nvenc+vulkan-encode+pyrowave). The pixel path itself is unverifiable without an
NVIDIA box: `nvenc_hdr10_smoke` and friends are #[ignore]d, so the channel order
and the pad byte want an on-glass check on the CachyOS 5070 Ti.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`sws_getContext` hands back a raw pointer whose only `sws_freeContext` lives in
`Drop for NvencEncoder` — and `Drop` needs a CONSTRUCTED `Self`, which does not
exist on either of `open`'s early returns: the intra-refresh-unsupported path
(which recurses into `Self::open`) and the plain error return. Both sit between
the context's creation and the `Ok(NvencEncoder { … })` that would give it an
owner, so every failed open leaked one.
That is not a once-per-process wart. `open_nvenc_probed` walks an EINVAL bitrate
ladder — requested rate, then the codec-level cap, then ×3/4 down to 50 Mb/s —
re-entering `open` at each step, so a GPU that refuses the requested bitrate leaks
a context per step (up to ~10). The intra-refresh retry adds one more.
Fixed structurally rather than with a guard: the block depends only on values known
before the encoder open (`want_444`, `want_hdr10`, `cuda`, `format`, `nvenc_pixel`,
the dimensions, `full_range_444`), so moving it BELOW `video.open_with(opts)` — to
just above the struct construction, with nothing fallible in between — makes the
leak unrepresentable instead of merely handled. No behaviour change: same context,
same colourspace details, same field.
Found while designing the WP1.4 swscale expand, which would have promoted this from
a 4:4:4/HDR-only leak (the only sessions that build a context today) to one on every
packed-RGB session. Fixing it first is the prerequisite for that change.
Verified `-D warnings --all-targets` + tests on Linux (default; shipped
nvenc+vulkan-encode+pyrowave).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
All three Windows backends derived it the same wrong way —
`bit_depth >= 10 || matches!(format, P010 | Rgb10a2)` at ffmpeg_win.rs:154,
amf.rs:1275 and qsv.rs:777 — so the NEGOTIATED depth could force a 10-bit encoder
over an 8-bit capture.
That combination is not hypothetical. A client advertises 10-bit, the handshake
negotiates bit_depth=10, and then enabling advanced colour on the IDD virtual
display fails — at which point the capturer says exactly that and delivers 8-bit
NV12 anyway ("10-bit HDR was negotiated but enabling advanced color on the virtual
display FAILED — encoding 8-bit SDR"). Every vendor then lost the session, each in
its own way:
- native AMF and native QSV derived `expected = P010`, saw Nv12, and bail!'d at
open;
- the libavcodec path accepted the open, built a P010 encoder, and then failed
EVERY submit forever, because its per-frame check recomputes the depth from the
frame and never matches. reset() could not help: the rebuild re-derived the
same wrong depth from the same stored bit_depth. The host burned
MAX_ENCODER_RESETS and ended the session.
The last one is the worst of the three because it IS the fallback: native QSV
correctly refuses this input at open, and lib.rs then falls back to the ffmpeg path
"for robustness" — which accepted it and died per frame instead.
Since the duplication was the bug, the fix is one shared `ten_bit_input()` in
codec.rs that all three call, and it follows the delivered pixels. That also keeps
the stream HONEST rather than merely alive: the depth selects the colour signalling
(BT.2020 PQ vs BT.709) and the staging surface format, so an 8-bit capture now
produces an 8-bit stream that says it is SDR — which is what the capturer already
reported it was sending. It warns when the negotiated depth is discarded.
The negotiated depth stays an upper bound. The session LABEL may still claim HDR;
that mismatch lives in the negotiation, not in the encoder, and is not addressed
here.
`is_10bit_format` keeps its (now format-only) definition and is used by
submit_d3d11's per-frame check, so the predicate the encoder was built from and the
one it re-checks per frame cannot drift. That check can now only fire on a genuine
mid-stream depth change.
Verified `--all-targets -D warnings` on Windows (no features / pyrowave / qsv /
nvenc,qsv; 31 tests) and Linux (default; shipped nvenc+vulkan-encode+pyrowave).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The vendored rate controller packs the 32x32 block index into the low 16 bits of
`RDOperation::block_offset_saving` (pyrowave-sys patches/0002). Past u16::MAX the
index collides with the `saving` field, the resolve over-credits, and the emitted
payload can overshoot the buffer `pyrowave_encoder_packetize` writes into — whose
only bounds check is an `assert` that the Release (NDEBUG) vendored build compiles
out. There is no Rust-side guard behind it.
All three callers of `pyrowave_mode_fits_rdo` hardcoded 4:4:4, leaving 4:2:0
unchecked — but 4:2:0 overflows too, just later: 8192x6144 is 73728 blocks and
8192x8192 is 98304, while `Codec::PyroWave.max_dimension()` permits 8192 per axis,
so both are reachable from a client-requested `mode=WxHxFPS`. Worse, the host's
only use of the helper (native/handshake.rs) is a 4:4:4 -> 4:2:0 downgrade, so an
oversized mode was actively routed INTO the unguarded branch.
The cap now lives in `validate_dimensions`, checked against 4:2:0 — the most
permissive chroma, so a mode that cannot fit there fits no PyroWave session at any
chroma. That is the single chokepoint both the negotiator (handshake.rs:180) and
`open_video_backend` already run. Both encoder opens additionally check the chroma
actually being opened: the 4:4:4 half, plus defence in depth for the
`PUNKTFUNK_ENCODER=pyrowave` lab override.
Not dormant: punktfunk-host has `default = ["pyrowave"]` and no build passes
`--no-default-features`, so these backends ship in every host binary.
Tests pin the arithmetic (8192x6144 = 73728, 8192x8192 = 98304, 7680x4320 still
fits) and assert the cap does not leak to H.265/AV1, which have no such controller.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
pf-encode's GPU backends are off by default, so almost none of them were under
`-D warnings`:
- ci.yml lints/tests with default features, which do cover VAAPI, libav-NVENC
and — via punktfunk-host's `default = ["pyrowave"]` — the PyroWave backends,
but not `nvenc` or `vulkan-encode`.
- deb.yml builds those two, but with `cargo build`, where warnings are not errors.
That left enc/linux/nvenc_cuda.rs, enc/linux/vulkan_video.rs and the vendored
vk_av1_encode / vk_valve_rgb bindings — ~8,150 lines carrying ~70 `unsafe` blocks —
never linted anywhere, so the crate's own
`#![deny(clippy::undocumented_unsafe_blocks)]`, its stated unsafe-proof gate, was
never actually enforced on them.
Linux gains a clippy+test leg at the SHIPPED feature set: deb.yml builds
`punktfunk-host/nvenc,punktfunk-host/vulkan-encode` WITHOUT
`--no-default-features`, so the .deb carries pyrowave too and that combination is
what deserves the lint. GPU-free — the hardware tests are `#[ignore]`d and
NVENC/CUDA dlopen their entry points, so the test binary links with no driver.
Windows gains a separate `-p pf-encode --all-targets` lint. The existing lint is
`-p punktfunk-host`, which never builds pf-encode's test targets — the blind spot
that let the Linux twin's tests rot. It must be clippy rather than `cargo test`:
on MSVC nvidia-video-codec-sdk link-imports NvEncodeAPICreateInstance /
NvEncodeAPIGetMaxSupportedVersion, so a test binary cannot link without the
driver's import lib. clippy type-checks without linking; ci.yml runs the tests.
`--all-targets` is load-bearing, not decoration: without it the feature-gated
`#[cfg(test)]` modules are never compiled at all.
Also widens windows-host.yml's paths filter, which listed `crates/pf-encode/**`
but none of the crates it compiles against (pf-frame, pf-gpu, pf-zerocopy,
pf-host-config, pf-capture, ...), so a change reaching the Windows host through
one of those triggered no Windows build.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Inherited from the pre-extraction host crate root as scaffolding for backend
paths defined ahead of the build that used them. A census across every feature
combination on both platforms (flip it to `warn`, rebuild) found it was hiding
exactly two items — so it bought nothing while blinding the crate to future rot:
- `vaapi::fourcc` — superseded by `pf_frame::drm_fourcc`; no call sites left.
- `vulkan_video::open_opts` — test-only (the smoke tests use it to pass the
RGB-direct request explicitly rather than through the env), now `#[cfg(test)]`.
Removing it surfaced a real latent defect the Linux census could not see:
`amf.rs`'s `percentile` / `drive_and_measure` helpers are used only by the
`#[cfg(feature = "amf-qsv")]` latency A/B benchmark, so a `--features nvenc,qsv`
build compiled the helpers with their caller gated out. They now carry the same
gate as their caller.
Verified `--all-targets -D warnings` on Linux (no features; shipped
nvenc+vulkan-encode+pyrowave) and Windows (no features; pyrowave; qsv; nvenc,qsv).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
All ten `NvencCudaEncoder::open` call sites in the `#[cfg(test)]` module passed 9
arguments to a 10-parameter function: `cursor_blend` was added to `open` and the
tests were never updated. The module has been uncompilable ever since —
`cargo clippy -p pf-encode --features nvenc --all-targets` fails with 10x E0061.
Nothing in CI noticed because no job compiles this crate's feature-gated test
targets: ci.yml lints/tests with default features (which do not include `nvenc`),
and windows-host.yml lints `-p punktfunk-host`, which builds pf-encode as a
dependency and so never builds its test targets. The CI commit below closes that
gap; this has to land first or that leg starts red.
`false` is what these tests exercised before `cursor_blend` existed: every frame
they build sets `cursor: None`, and there is no cursor-blend test.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Field report (Nobara, 0.19.2): entering Game Mode mid-stream left the monitor
on and the stream mirroring at the desktop's resolution. Root cause: the
DM-flavor takeover needs privilege to stop plasmalogin, the polkit rule is a
docs-only manual step nobody installs, so every managed entry silently
degraded to ATTACH — and with a physical display connected the attach path
mirrors the box's own session at its own mode.
Ship the privilege with the packages instead: a root helper
(libexec/punktfunk/pf-dm-helper, verbs stop|restore) behind its own polkit
action (io.unom.punktfunk.dm-helper, allow_any — the same mechanism Nobara's
os-session-select uses, and the helper derives the DM unit from the
display-manager.service symlink itself so callers never name a unit across
the privilege boundary). The host tries the plain system-bus verbs first
(root / operator rule still take precedence), then pkexec's the helper; the
restore paths (idle restore + in-stream desktop-switch honor) use the same
ladder so a takeover that needed the helper can always be undone by it.
Packaged in rpm/deb/arch (Arch under /usr/lib/punktfunk with the policy's
exec.path annotation rewritten; the host probes both layouts). Nix is left
out deliberately: store paths can't match the probe, NixOS keeps the manual
rule. Docs updated — the rule snippet stays as the scoped alternative.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A fresh install left PunktfunkWeb registered but not running: `web setup` waited
only for the mgmt token before firing `schtasks /run`, while `web-run.cmd` also
requires the host identity cert — and the host writes the token during argument
parsing but `cert.pem` only after the pure-Rust RSA-2048 keygen inside `serve`.
The launcher lost that race, exited 1, and since the task carried no trigger but
boot (Task Scheduler does not reliably restart on a non-zero exit code) the
console stayed down until the next reboot, with the installer still reporting
"web console set up + started".
- `web setup` gates on cert.pem (written last) as well as the token, 90 s budget.
- After `schtasks /run`, poll for the :47992 listener and retry before giving up;
warn honestly instead of claiming a start that did not happen.
- `web-run.cmd` (installed + dev) waits in-process for the token + cert (~5 min)
rather than exiting 1 and hoping restart-on-failure retries.
- Register the task with a logon trigger alongside boot, falling back to the
boot-only XML if a Task Scheduler build rejects it.
- Linux had the same defect: punktfunk-web.service's Restart=on-failure gave up
permanently after systemd's default 5-starts-in-10 s limit. StartLimitIntervalSec=0.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An all-fixes patch on top of 0.19.1, all host-side:
- KDE Plasma display takeover moved off kscreen-doctor to in-process
kde_output_management_v2 (86d5d666), with the >60 Hz custom-mode install
ported too (64511af4) — a KWin host whose kscreen-doctor wedges now disables
its physicals and hands over the streamed output in ~2.4 ms instead of hanging
~26 s and never taking over.
- The gamescope managed takeover is display-manager-flavor-aware (e35dad52) so
switching a non-SDDM Linux desktop (Nobara / plasmalogin) into Game Mode no
longer start-limits the display manager into a permanent black screen, and
in-stream 'Switch to Desktop' is honored under it (a3e21d92).
- Windows display isolation anchors the kept sources at the desktop origin
(9b241d9d) so an origin-less desktop no longer fails 0x57, with restore
guaranteed never to leave the desktop all-dark.
- Windows pen/touch/absolute-mouse map over the streamed output's rect rather
than the whole virtual desktop (087e7d04), so absolute input lands correctly
when a physical monitor is kept on beside the streamed display.
Nothing moved on the wire (protocol 2) or at the C ABI (13), so no embedder
rebuild and 0.18/0.19 peers keep mixing freely. The Windows virtual-display
driver protocol is unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pen, touch, and absolute mouse arrive normalized to the STREAMED display's
frame, but pointer_windows::to_screen and sendinput's MouseMoveAbs mapped
them over the whole virtual desktop — correct only when the virtual display
is the sole active display (Exclusive topology). In Extend — a physical
monitor kept on beside the virtual output, or an Exclusive isolate degraded
to the 0x57 keep-physicals fallback — the streamed output sits at a non-zero
origin, so every sample landed shifted and mis-scaled. The pen exposed it
first (field report 2026-07-24): a stylus is strictly absolute, with no
closed-loop correction onto the target like a cursor.
New pf-inject::stream_target (Windows): the host publishes the streamed
output's CCD target id at capture bring-up (one central site —
capture_virtual_output covers the native and GameStream planes); mapping
sites resolve its current desktop rect through pf-win-display's
source_desktop_rect — the same resolver the cursor-readback poller uses, so
inject and readback always agree — TTL-cached (250 ms) because a
group-layout re-arrange moves a live output's origin mid-session. No target
set / never resolved falls back to the whole virtual desktop (the historical
mapping, still right for Exclusive topology and devtest).
One change in to_screen covers pen + touch; MouseMoveAbs converts the same
desktop pixel into the 0..65535 MOUSEEVENTF_VIRTUALDESK coordinate, closing
the identical latent absolute-mouse bug (apollo-comparison open item #14/#30).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Follow-up to the previous commit: the only KWin topology op still shelling out to
kscreen-doctor was the >60 Hz custom-mode INSTALL (`set_custom_refresh`). Port it
to `kde_output_management_v2` so the whole KWin path is kscreen-free on modern KWin.
`kwin_output_mgmt::set_custom_mode` builds a one-entry `kde_mode_list_v2`
(set_resolution / set_refresh_rate / set_reduced_blanking=full / add_mode), applies
it via `kde_output_configuration_v2.set_custom_modes` (since 18), waits for KWin to
generate the mode (its CVT generator may align the width down — matched with the
same height-exact / width-within-8 / refresh-within-1Hz gate as before), then
selects it — which changes the output size and drives the sacrificial-birth stream
renegotiation. `kwin::create`'s want_high branch tries this first and only resolves
a kscreen id + calls `set_custom_refresh` if it returns None (pre-6.6 KWin without
set_custom_modes, or the compositor not answering) — so a >60 Hz session on a box
where kscreen-doctor wedges no longer eats a 5 s resolve timeout either.
Bonus: `set_custom_modes` REPLACES the custom list, so reconnects are idempotent —
no more one-custom-mode-per-connect growth of the user's display list.
Verified: cargo test -p pf-vdisplay (73 pass), clippy -D warnings clean, fmt clean.
The set_custom_modes round-trip is proven live on KWin 6.6.4: apply=applied, and a
1648x928@75 request generates 1648x928@74.901 (CVT fractional refresh, caught by the
1 Hz tolerance). The mode SELECT + full >60 Hz stream is still owed on-glass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A KDE host on Nobara stopped disabling its physical screens and creating the
virtual output the moment it updated: streaming still came up, but bring-up took
~26 s and the streamed output never became the desktop (`also_disabled=[]`).
Root cause is not our topology logic — it's that every `kscreen-doctor` call on
the reporter's session wedges. kscreen-doctor drives libkscreen, which (per
setup) waits on the kscreen KDED module over D-Bus; when that layer is stuck it
blocks in its own connect and never returns, so all five topology queries hit
their 5 s budget and got killed (host log). Reproduced live on a Nobara / KWin
6.6.4 box: `kscreen-doctor -j` there times out at 8 s, every time.
But the compositor's OWN Wayland is fully responsive on that same session — the
host just created a virtual output over it via zkde_screencast. So drive the
topology (resolve our output, take primary, disable the physical/bootstrap
outputs, capture their modes, re-enable on teardown, position) directly over
`kde_output_management_v2` + `kde_output_device_v2` instead of shelling out.
On that same Nobara box the in-process path binds management (v19) and
enumerates the outputs in 2.4 ms — ~3400x faster than the 8 s hang, and immune
to whatever wedges the standalone tool.
- vendor kde-output-management-v2 / kde-output-device-v2 (KWin advertises mgmt
v19, device v20); generate client bindings inline (the interdependent-protocol
module split from the wayland-protocols crate; needs the `bitflags` dep for the
device protocol's bitfield enums).
- new `kwin_output_mgmt`: bounded enumerate-then-apply over one Wayland
connection; every wait is time-bounded so a genuinely wedged compositor
degrades to `handled = false` and the old kscreen-doctor path still runs.
- `kwin::create` topology + `apply_position` prefer the in-process path (address
our output by its stable device UUID, supersede-robust) with kscreen-doctor as
the fallback. The 60 Hz path now makes ZERO kscreen-doctor calls; only the
>60 Hz custom-mode install still shells out (its in-process port is a follow-up).
Verified on Linux: cargo test -p pf-vdisplay (73 pass), clippy --all-targets
-D warnings clean, fmt clean, punktfunk-host checks clean. In-process enumeration
proven live against KWin 6.6.4 (the reporter's env); the disable/apply path is
not yet on-glass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
First on-glass test of the DM-flavor takeover confirmed the black-screen
fix, but exposed that the managed model could not switch back: Steam's
in-stream session-select is a silent no-op while the display manager is
stopped (every config-write branch in os-session-select requires the DM
to be running), so the capture loss it causes just relaunched game mode.
The user pass of the switch script does leave one durable trace — the
~/.config/steamos-session-select sentinel, written before any DM check.
The managed launch now baselines its mtime; a capture loss with the
sentinel advanced reads as the switch request and is honored by
replaying the switch with the DM up: restore the display manager, run
the distro's own os-session-select (its internal pkexec is authorized
allow_any by the distro policy, so it works from the host's sessionless
context), then stop the autologin game-mode unit so Relogin lands in
the newly selected desktop. The capture-loss re-detection then follows
KWin as it comes up; a 120 s post-honor grace stops the rebuild loop
from racing the booting desktop back into game mode (superseded early
if the box's own game-mode unit reappears — the desktop→game leg stays
fast). The baseline is re-recorded on crash-restore so a pre-existing
sentinel never reads as a fresh request. Every verb in the sequence was
live-validated on the Nobara repro VM.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Field report (Nobara 43 HTPC): switching the host to Steam Game Mode
mid-stream permanently black-screened the box. Live-proven root cause on
a Nobara repro VM: our takeover masks the box's gamescope-session-plus
unit, plasmalogin's Relogin=true then fails its session Exec repeatedly
and trips systemd's start limit within ~1 s — the display manager dies,
and our restore verb (unmask + user start) cannot bring a seatless
gamescope back. Only 'reset-failed + restart' of the DM recovers.
The takeover is now display-manager-flavor-aware:
- SDDM (Bazzite/SteamOS) keeps the proven mask+SIGKILL path unchanged.
- plasmalogin/unknown DMs never mask: with privilege (root or an
operator polkit rule scoped to the DM unit — documented) the host
stops the DM for the stream and restores it with reset-failed +
restart (PUNKTFUNK_RECOVER_SESSION_CMD as fallback), recorded in the
persisted takeover state so a host crash still restores; without
privilege the managed takeover degrades to ATTACH and mirrors the
box's live Game Mode instead of destabilizing the seat. Both legs of
the privileged cycle live-verified on the repro VM (headless managed
session works with zero login sessions; render nodes are 0666).
A loaded-but-inactive leftover instance never triggers the DM stop.
Companion fixes from the same triage:
- ensure_box_gamescope_mode gains the attach-only rebuild-probe guard
both managed paths already had (stale post-capture-loss detection
restarted the box's unit), and no longer re-modes a box that drives a
physical display — attach mirrors on-glass; re-mode is the
headless-box model.
- Capture-loss rebuilds targeting gamescope get a 100 s budget: the
40 s budget expired inside the first 45 s Steam-cold-start launch
attempt, a guaranteed single-shot failure.
- A PUNKTFUNK_COMPOSITOR pin now WARNs once per capture loss when the
live session no longer matches it (the pin disables
session-following — the reporter's original stream-death trigger).
- A managed session that took nothing over (client gamescope pin beside
a live desktop) is stopped on disconnect instead of being orphaned
forever.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Round 3 of the field-reported exclusive-topology 0x57, with the reporter's
retest logs finally separating the variable: the SAME isolate call converged
rc=0 whenever a kept member already sat at (0,0), and failed 0x57 on every
shape — doomed-path-carried AND keep-only escalation — whenever the doomed
physical held the origin. A committable config must still contain a primary
(a source pinned exactly at (0,0)); deactivating the display that held it
while the kept virtual stays at its EXTEND offset supplies an origin-less
desktop, which Windows rejects wholesale. Translate the kept sources rigidly
so the top-left-most lands on the origin; sets already covering (0,0) are
untouched so plain re-commits stay byte-identical.
Also from the same logs:
- restore_displays_ccd now guarantees the desk is never left all-dark: the
saved snapshot can be unappliable (0x64a — it pinned a virtual target
incarnation removed before teardown) or apply cleanly yet re-light nothing
(snapshotted while an earlier failed teardown had the physicals off). If no
external physical is active after the apply while one is connected, fall
back to the database EXTEND preset. Internal panels don't count — a closed
lid must not be forced back on.
- the final isolate failure now names the surviving targets instead of
asserting 'a non-virtual display stayed active' — the logs showed a sibling
VIRTUAL display as the survivor (linger-expiry shrink), and the wording sent
triage the wrong way.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
An all-fixes patch on top of 0.19.0. The headline is session teardown — the
native stop flag made enforceable and bounded (ac3dc432) and GameStream ending
on the control peer's ENet disconnect (41fa25c4), so a session can no longer
outlive the client that started it on either plane.
Also: the Windows input-desktop pair (display writes b5fa878b + pen/touch
injection 24a24734 both surviving a UAC prompt), the Apple display-sleep guard
(bc5f6a38), the gamescope corner-parked cursor (4b2d2d1e), the KWin custom-mode
selection (3a33a694), the tvOS glob warnings (cc8cccbe), and the Linux
starter/env docs revamp (66172753).
Nothing moved on the wire (protocol 2) or at the C ABI (13), so no embedder
rebuild and 0.18/0.19 peers keep mixing freely.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
`punktfunk-core` emitted `unused import: super::super::*` for
client/pump/handshake.rs and client/pump/input_task.rs — but only when
built for a tvOS target, where the cfg set leaves the glob contributing
nothing. Nobody saw them because CI never builds tvOS: that needs
BUILD_TVOS=1 and the Tier-3 build-std nightly path, and the Apple
workflow only runs swift build/test (macOS).
Both files also `use super::*`, i.e. the parent `pump` module, which
itself globs `super::*` (`client`) unconditionally — so every name stays
reachable by that path and the inner glob was pure duplication. Their
sibling datagram_task.rs already imports only `use super::*`, so this
also makes the module consistent. control_task.rs and data.rs keep
theirs: those genuinely resolve names through it and never warned.
Verified warning-free AND compiling on every slice the xcframework
ships — aarch64/x86_64-apple-darwin, aarch64-apple-ios,
aarch64-apple-ios-sim, aarch64-apple-tvos, aarch64-apple-tvos-sim,
x86_64-apple-tvos — plus Linux with and without the `quic` feature. A
full xcframework rebuild now reports 0 warnings (was 6: two per tvOS
slice), and swift build, swift test (170 tests), and xcodebuild for the
macOS, iOS and tvOS schemes all pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Native sessions could survive long after the client was gone, in two
independent ways.
HOST: `stop` was only advisory. The one thing that ends a session is the
stream thread returning — every teardown (conn.close, the joins, and the
RAII drops of the session permit, admission entry and stream marker)
sits after that await, and nothing forced the issue. The encode loop
checks `stop` between iterations, so any unbounded call INSIDE one never
reaches the check, and one stuck syscall became a permanent zombie: it
held its semaphore slot (four of those and the host stops accepting
QUIC entirely), its admission entry (a later client gets "host busy"
forever), and not even the console's Stop button could clear it — that
button sets this same flag.
* Bound the wait: once the session has been told to stop, the thread
gets STREAM_STOP_GRACE (90 s, well past the 40 s capture-rebuild
budget) to return, then teardown runs anyway. The thread is detached,
not killed — Rust can't cancel a blocking thread — so it keeps its
capturer/encoder until the stuck call returns, but the session's slot
and admission entry come back and the host keeps serving. It logs at
ERROR as the host wedge it is.
* Bound the audio/input joins too — the last unbounded await in
teardown.
* Take the session permit AFTER the QUIC handshake instead of before
`accept()`, so a host at its concurrency cap still accepts and the
waiting client sees a live path instead of a silent dial timeout.
* Bound the compositor helpers that caused the wedge in the first
place: new pf-vdisplay `proc::{status_within, output_within}` kill a
child that outlives its budget. `kscreen-doctor` is a Wayland client
of the very compositor it configures, so against a wedged KWin it
never returned; same for systemctl/dbus against a stuck session bus.
CLIENTS: the connection was never closed, so the host was right to keep
the session — it still had a live, keep-alive-answering peer.
* Android: backgrounding did no teardown at all, and Android doesn't
suspend the process, so the worker kept answering keep-alives until
the OS reclaimed it (on a TV box, never). End the session on ON_STOP,
via the existing onDispose path; a plain close, not a quit, so the
host lingers the display for a fast return.
* Apple: the .background arm was iOS-only AND gated on an opt-in that
defaults off, so backgrounding did nothing — while the `audio`
background mode kept the app (and its connection) alive indefinitely.
Act unconditionally, and cover tvOS.
* Core: `conn.close()` only queues the frame, and run_pump is the body
of a block_on whose runtime is dropped the instant it returns, so the
driver could never put it on the wire — a deliberate quit reached the
host as silence (8 s idle timeout, no quit code, and the linger meant
for an unwanted disconnect). Carry the endpoint out of the handshake
and flush with wait_idle(), the same discipline the pairing and probe
paths already use.
Linux check/clippy/tests green: 262 host, 71 pf-vdisplay (incl. new
bounded-process tests), 231 core.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
GameStream sessions outlived their client: the only complete teardowns
were the explicit ones (RTSP TEARDOWN, nvhttp /cancel, mgmt DELETE
/session), and the only automatic detector was a media-UDP send error —
which needs an ICMP port-unreachable, so a true vanish (Wi-Fi drop,
sleep, power off, crash) left video+audio encoding into the void
forever, and even a plain Moonlight quit (which sends neither TEARDOWN
nor /cancel) leaked the session. The stale state then cascaded: a
lingering launch 503-blocked a different client under
mode_conflict=reject, and streaming=true made a reconnect's PLAY take
its "stream already running" branch — no new threads, old threads still
aimed at the dead endpoint, the reconnect got no media.
ENet already detects all of this — the control peer's reliable-ping
timeout (or clean disconnect) fires Event::Disconnect within ~5-30 s —
but the handler only reset input state. Wire the real teardown into it:
* AppState::end_session — THE compat-plane session teardown: stops both
media threads (their flags), clears launch + negotiated stream config;
idempotent. /cancel and mgmt stop_session now share it.
* control.rs Disconnect → end_session. Gated on the TRACKED session
peer, and Connect only tracks a peer from the /launch owner's IP (the
same source-IP bind the RTSP/media plane uses), so an unauthenticated
LAN peer connect+disconnect can't end a live session, and a fast
reconnect's stale-peer timeout can't kill its successor.
* Client-unreachable UDP send errors now end the whole session via an
OnSessionLost callback (built at PLAY) instead of stopping only the
plane that noticed — audio no longer keeps streaming after video
detects the dead client, and vice versa.
Linux check/clippy/tests green (53 gamestream tests incl. the new
end_session regression test).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A KDE host streaming to a phone-shaped client fell back to 60 Hz even though the
120 Hz mode was sitting right there in KDE's display list, one manual click away.
KWin generates every custom mode's timing with libxcvt, whose first step is
`hdisplay_rnd = hdisplay - (hdisplay % 8)`. A 2868x1320@120 request (iPhone 16
Pro Max) therefore becomes 2864x1320@119.92 — 4 px narrower, fractional refresh.
We then asked kscreen-doctor for `mode.2868x1320@120`; its `findMode` matches a
mode's id or its own `WxH@qRound(Hz)` name, so that string matched nothing, the
select silently no-op'd, the output stayed on its sacrificial birth mode, and
`size_applied` came back false → "KWin rejected the custom mode" → 60 Hz.
Widths like 1920/2560/3840 are all multiples of 8, which is why only clients with
phone-shaped panels ever hit this.
Resolve the mode out of the output's OWN list instead and address it by kscreen
mode id: exact height, width at most one cell narrower than asked, refresh within
1 Hz (which excludes the native 60 Hz entry). `set_custom_refresh` now returns the
whole achieved mode, and `create` reports that as the output's `preferred_mode` —
so the capturer's renegotiation gate waits for the size KWin will actually deliver
and the encoder opens against it, rather than starving on the requested one.
Also skip `addCustomMode` when a usable mode is already installed: kscreen-doctor
APPENDS to the list and KWin persists it per output name, so the old code grew the
user's display list by one entry per connect.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Same field report as the display-write fix, other half of the symptom: with a
UAC consent prompt up — one the user could SEE in the stream, because capture
already renders the secure desktop (326d6e17) — pen and touch did nothing,
while mouse and keyboard kept working.
The split was exactly which paths knew the input desktop can move. sendinput.rs
reattaches, so mouse/keyboard reached the prompt. pointer_windows.rs had no
desktop handling at all, so every InjectSyntheticPointerInput came back
ERROR_ACCESS_DENIED:
22:20:31 virtual pen created (PT_PEN)
22:20:31 pen inject failed error=Zugriff verweigert (0x80070005)
22:20:42 touch inject failed error=Zugriff verweigert (0x80070005) contacts=1
Measured on glass before fixing, to find out what it actually takes:
device on Default, thread on Default -> 0x80070005 (the field failure)
device on Default, thread on INPUT -> OK
device on INPUT, thread on INPUT -> OK
The middle row is load-bearing: the synthetic pointer device is NOT
desktop-affine, so rebinding the thread suffices and the device is never
destroyed and recreated across a desktop switch — which would have dropped
in-flight contacts and the pen's in-range state mid-stroke.
Injection now retries once bound to the input desktop. The binding is scoped,
not persistent like sendinput's: inject_pen/inject_touch_frame run on TWO
threads (the caller's apply_batch and the refresh threads), and the batch caller
is a shared task thread that must not be left parked on a Winlogon desktop that
disappears when the prompt is dismissed.
The first-failure WARN now carries the rejected sample (flags, pen flags/mask,
pressure, rotation, tilt, position). A 0x80070057 INVALID_PARAMETER was seen
once BEFORE any prompt existed and is still unexplained; the ranges all look
sound on inspection (coordinates clamped to the virtual screen, roll/azimuth
u16 so the modulo cannot go negative, tilt bounded 0..=90), so catching the
actual offending sample is the way to find it.
Verified on glass: with a consent prompt up, pen and touch now reach it — zero
inject failures where every prior session failed immediately. A UAC prompt can
now be dismissed from an iPad with the Pencil.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Field-reported 2026-07-23: a consent prompt left up on an unattended host made
the box unreachable. Every connect black-screened and failed after ~60 s, and
the host had to be reached over SSH to clear the prompt.
Windows refuses ChangeDisplaySettingsEx/SetDisplayConfig issued from a thread
that is not on the desktop currently receiving input. While UAC (or the lock /
logon screen) is up that desktop is Winlogon, and the host — which the service
launches explicitly onto WinSta0\Default — got DISP_CHANGE_FAILED /
ERROR_ACCESS_DENIED for every write. A session starting in that window could
never set its virtual display's mode, so the capturer sized its ring to a stale
mode, every composed frame was dropped for a size mismatch, and bring-up burned
8 retries before giving up.
Measured on glass (RTX box, SYSTEM host in console session 2, real consent
prompt up, virtual display active):
INPUT desktop = Winlogon
UNBOUND CDS_TEST -> -1 (DISP_CHANGE_FAILED)
UNBOUND SDC_VALIDATE -> 0x5 (ERROR_ACCESS_DENIED)
BOUND CDS_TEST -> 0 (DISP_CHANGE_SUCCESSFUL)
BOUND SDC_VALIDATE -> 0x0 (ERROR_SUCCESS)
New input_desktop.rs mirrors pf-inject's sendinput.rs retry model: issue the
write, and only when it fails the way a wrong-desktop write fails, rebind this
thread to the current input desktop and retry once. A working write is never
touched, so the normal path is unchanged. The retry predicate stays narrow
(ERROR_ACCESS_DENIED only) so the unrelated 0x57 exclusive-mode topology bug is
not re-issued and mis-attributed. Unlike sendinput's dedicated injector thread,
the binding here is SCOPED — a shared display-write thread left on a Winlogon
desktop that is later destroyed would fail every subsequent write, which is the
wedge this removes.
Applied to all eight SetDisplayConfig sites and both ChangeDisplaySettingsExW
calls in set_active_mode. The isolate site now decides its supplied config once,
outside the write, so a retry re-issues the identical config rather than logging
its escalation twice.
A save is logged, because a silent one is indistinguishable in a field log from
a write that never needed saving — which made the first on-glass verification of
this change inconclusive.
Both ACCESS_DENIED diagnostics now ask which cause applies instead of naming
only the disconnected-RDP one: that phrasing sent this investigation chasing a
phantom RDP session on a host that was already service-launched in the console
session, while a consent prompt was the actual cause.
Verified on glass: connect with a consent prompt up now reaches first frame in
3.0 s (was a black screen then "Connection failed" after ~60 s), with six
"retried bound to it and it applied desktop=Winlogon" lines and rc=0x0
throughout.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
gamescope hides its pointer by WARPING the X pointer to the root's bottom-right corner pixel;
it does not swap in a transparent X cursor, so XFixesGetCursorImage keeps handing back the last
opaque arrow. The XFixes source picked its display by "whichever pointer moved", and a parked
pointer never moves again — so it froze on the parked Xwayland and composited that arrow at
(w-1, h-1): a sliver of cursor welded to the corner of the stream for the rest of the session
while the real pointer went undrawn. Reported on-glass in every game (a game grabs the pointer,
so the hide is permanent): "part of cursor shows up on bottom right ... isn't where it really is".
Follow GAMESCOPE_CURSOR_VISIBLE_FEEDBACK instead — gamescope publishes it on every nested
Xwayland root: 1 on the server whose pointer it is drawing, 0 on the others and 0 on all of them
once the pointer is hidden. It answers both questions correctly for a STATIC pointer, which
motion cannot: which Xwayland owns it, and whether to draw it at all. Read at connect, re-read
on its root PropertyNotify (event drain now discriminates CursorNotify vs PropertyNotify) with a
250 ms resync as the self-heal. A pointer gamescope draws nowhere is published visible:false,
not dropped — the encode loop overwrites the frame's overlay from this slot and strips invisible
ones, so a None would leave the last visible overlay standing on repeat frames. Honouring the
atom also gives the stream gamescope's own idle auto-hide, which this source never had.
The pointer-motion heuristic stays as the fallback for a gamescope that publishes no verdict
(logged at session start via cursor_feedback=false), so nothing regresses to a cursorless stream.
Measured on a live 1920x1080 Steam Gaming Mode session (RTX 5070 Ti, gamescope c31743d+,
--xwayland-count 2 --hide-cursor-delay 3000): idle 3 s => pointer (1919,1079) + feedback 0;
real evdev motion => pointer live + feedback 1. End-to-end against that session the source now
publishes visible=false while parked and tracks the live pointer within one 250 ms sample of
motion. cargo clippy -p pf-capture --lib -D warnings and the 5 new pick_active tests are green
at cargo 1.96.1 (CI parity); cargo fmt --all --check clean.
A stream is not user activity to the OS, and controller input does not feed
the HID idle timer on any Apple platform — so a gamepad-only session reliably
idles the panel out from under the user mid-play. A keyboard/mouse capture
session masks the bug because those events are real local HID. The Android
client has held FLAG_KEEP_SCREEN_ON while streaming all along; the Apple
clients held nothing.
DisplaySleepGuard is acquired in beginStreaming and released at the top of
disconnect, so it is scoped to the session — every teardown path (user quit,
sessionEnded, the backgrounded keep-alive timeout) funnels through disconnect.
iOS/iPadOS/tvOS: UIApplication.isIdleTimerDisabled. App-wide and ignored
while backgrounded, which is what the audio-only keep-alive wants.
macOS: ProcessInfo.beginActivity(.userInitiated, .idleDisplaySleepDisabled),
the Foundation wrapper over IOKit power assertions. That defers display
sleep but NOT the screen saver, which runs off the HID idle timer
independently — so a 30 s IOPMAssertionDeclareUserActivity heartbeat runs
alongside it. Intended side effect: an idle-lock that follows the screen
saver is deferred for the session too.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Field triage (Nobara, Discord): the kde.md starter host.env told desktop
users to set PUNKTFUNK_COMPOSITOR=kwin, which PINS the backend — detect()
short-circuits and the capture-loss rebuild never re-detects — so a
mid-stream switch to Game Mode killed the stream instead of following it.
A follow-up hardcoded XDG_RUNTIME_DIR=/run/user/1000 anchor broke PipeWire
for any non-1000 uid (pw audio connect: Creation failed).
Revamp across every starter/example/reference:
- Desktop starters (kde/gnome/hyprland/sway) shrink to
PUNKTFUNK_VIDEO_SOURCE=virtual + an explicit warning that pinning
disables session-following; forcing a backend is CI/appliance-only.
- host.env.example: rewritten around auto-detection; anchors demoted to a
commented ssh/cron-only block with the uid trap spelled out; the
gamescope ATTACH/MANAGED knobs documented (previously missing);
case-sensitivity called out.
- packaging/bazzite/host.env + README: drop the uid-1000 anchors (a
systemctl --user service inherits/derives them); README's stale
PUNKTFUNK_COMPOSITOR=gamescope-era template synced to the real one.
- packaging/kde/host.env: loud APPLIANCE-ONLY header (it pins on purpose).
- configuration.md: session-anchors section inverted to "leave unset",
compositor row states the pin consequence, case-sensitivity note.
- troubleshooting.md: new "session fails right after editing host.env"
section (case, wrong-uid anchors, stale pin, restart-to-apply).
- gamescope.md/bazzite.md: attach/managed descriptions match current
behavior (managed is the infra-detected default; attach re-modes a
box-owned session to the client's resolution).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Release notes were written for the people who build Punktfunk, not the people who use it —
dense with protocol/ABI/type names that confused even technical readers. Rewrite v0.19.0 in a
benefit-first, plain-language voice (New/Improved/Fixed) with all internal terms removed from the
body and every protocol/ABI/embedder detail moved to a single bottom "Under the hood" section.
docs/releases/README.md now codifies this voice as the format spec for all future notes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The v0.19.0 deb + decky release-attach steps source gitea-release.sh under `sh` (dash) and
died with "Bad substitution" at _release_notes_path's `${BASH_SOURCE[0]:-$0}` — a bash array
subscript dash rejects. So those legs never attached their assets to the release (the bash
legs — apple, android, rpm — were fine, which is why the body still seeded and the DMG/.ipa
attached). CI always sources this from the repo root, so resolve the notes as
docs/releases/<tag>.md relative to CWD and drop $BASH_SOURCE entirely. Verified under dash.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Headline: the whole pen/tablet input stack (P0–P5) — a pressure-sensitive stylus plane
across core wire, Linux, GameStream/Moonlight, Windows, iPad Pencil and Android. Plus
touch-injection fixes, cursor-DPI scaling, three Windows-host display/cursor fixes, a
gamepad same-PID conflict fix, and the Android mouse/keyboard (TV) regressions.
Release notes authored in-repo at docs/releases/v0.19.0.md (first release on the new flow;
CI seeds the release body from it). cargo update --workspace syncs the lock (versions only).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The pen P0 work added a new embeddable C entry point (punktfunk_connection_send_pen)
but left ABI_VERSION at 12 — the value 0.18.0 shipped. Per the project's convention
(0.18.0 bumped to 12 for the additive cursor channel), additive C-surface growth bumps
the ABI so embedders can gate on the new symbol. Regenerated include/punktfunk_core.h
via cbindgen (cargo build -p punktfunk-core); WIRE_VERSION is unchanged (pen is
capability-gated on HOST_CAP_PEN).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The iOS leg archives with development signing and re-signs for App Store distribution at
export time, but the existing export uses destination=upload (straight to TestFlight) and
leaves no .ipa on disk. Add an export step that re-exports the SAME archive with
destination=export to produce an App Store distribution-signed .ipa, then:
* attach it as the `punktfunk-ios-ipa` run artifact on every build (upload-artifact@v3,
since Gitea's backend rejects @v4), and
* attach it to the unified Gitea release on vX.Y.Z tags (alongside the DMG), via the
shared ensure_release/upsert_asset helpers.
Same gate as the archive; warn+skip (never fails the best-effort iOS leg) if the archive is
absent. Note: an App Store-signed .ipa installs only via TestFlight/App Store, not by direct
sideload — it is a release/archival artifact.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The hex dumps from the on-glass session show textbook SS_TOUCH packets with
pressureOrDistance = 0x7fc00000 (NaN): VoidLink encodes 'unknown finger
pressure' as NaN (iPad fingers have no force sensor), and the anti-forgery
finiteness gate rejected the entire packet — silently disabling the touch
plane while pen (real Pencil force, always finite) worked. Coordinates keep
the strict gate (they feed injector scaling); pressureOrDistance now
sanitizes non-finite to 0.0, the spec's own 'unknown', which the pen path
already maps to full-scale ink on contact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
An on-glass 'touch does nothing' currently leaves NO trace: log the first
SS_TOUCH per session at info (proves the client sends + we decode), and warn
when a packet carrying a pointer magic fails the body parse (layout mismatch)
instead of vanishing into the unknown-magic drop.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two Android TV M&K regressions from the c90343c2 overhaul, both user-reported:
1. Mouse back/forward buttons doubled as Android navigation while streaming.
The BUTTON_BACK/FORWARD motion edges already go over the wire as X1/X2, but
Android's input reader ALSO synthesizes KEYCODE_BACK/FORWARD key events from
those buttons unconditionally (stamped SOURCE_MOUSE, no FLAG_FALLBACK) — and
the guard only swallowed the FLAG_FALLBACK variant, so the duplicate reached
the BackHandler and yanked the user out of the stream. Swallow mouse-sourced
BACK/FORWARD too; a remote/keyboard BACK is never mouse-sourced and still
exits.
2. The soft keyboard popped up on physical-keyboard typing. KeyCaptureView (the
always-focused IME/capture anchor) unconditionally claimed to be a text
editor with a live editable InputConnection, so the TV IME counted input as
active and showed its UI the moment hardware keys arrived. Gate the editor
identity on imeShown (the user actually summoning the keyboard via gesture /
remote toggle), with restartInput on both edges so the gate is re-read, and
an onKeyPreIme BACK hook so an IME self-dismissal can't leave the gate
stale.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Field bug (Steam Deck client, exclusive topology): every isolate retry failed
SetDisplayConfig 0x57 with the physical panel left lit, while iPhone sessions
worked — their virtual monitor's connected-set already had a sole-display
topology persisted, so Windows deactivated the physical on hotplug and the
isolate was a no-op re-commit (deactivated 0). The Deck's monitor identity had
no such entry, forcing the live-deactivate path: supplying the doomed path
(inactive, modes unpinned) plus its orphaned mode entries is rejected 0x57 by
some driver/OS validation combos, so the retry loop could never converge — the
prior mode-index unpin (5e088af7) was necessary but not sufficient on this box.
Attempt 1 keeps the shipped shape; once a survivor is seen, later attempts
supply ONLY the keep paths with the mode table rebuilt to just their referenced
entries (indexes remapped, clone-shared entries deduped). The same host applies
exactly this keep-only shape rc=0 whenever the DB has pre-isolated the virtual
display, so the escalated shape is field-validated on the failing machine. The
final attempt also drops SDC_FORCE_MODE_ENUMERATION — a real path removal
drives COMMIT_MODES on its own.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On-glass report (iPad Moonlight vs the Windows host): pen inked, native touch
did nothing. Cause: POINTER_INFO.pointerId was the client's raw wire id —
Moonlight sends arbitrary large pointerIds, and synthetic-pointer injection
rejects them (Apollo compacts ids into dense slots for exactly this reason);
the failure was logged at trace, i.e. invisibly. Wire ids now map to the
lowest free slot for the contact's lifetime, and the FIRST injection failure
of a device logs at WARN (pen too) so an inert input plane is never silent
again.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Android leg of design/pen-tablet-input.md §7: against a HOST_CAP_PEN host,
stylus/eraser pointers split out of BOTH touch models (passthrough + gesture)
into StylusStream — state-full samples with pressure, AXIS_TILT, azimuth from
AXIS_ORIENTATION (Android's 0 = away-from-user IS wire north), AXIS_DISTANCE
hover, both stylus barrel buttons, the eraser tool, and historical (coalesced)
samples batched oldest-first. Kotlin heartbeats ≤100ms per the wire contract.
JNI: nativeHostSupportsPen + nativeSendPen (flat 10-float stride, sentinels
<0). No barrel-roll axis exists on Android — roll stays unknown here.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The flagship leg of design/pen-tablet-input.md §7: against a HOST_CAP_PEN host
the Pencil splits out of the finger path (independent of touch-input mode) into
PencilStream — state-full samples with force→pressure, altitude→polar tilt,
azimuth (+90° to the wire's north-clockwise convention), Pencil Pro rollAngle→
barrel roll (17.5+), coalesced touches batched ≤8/send for full 240Hz fidelity,
and Pencil hover (zOffset>0 distinguishes it from trackpad hover) as in-range
samples with distance. Squeeze holds barrel 1, double-tap clicks barrel 2 (the
Pencil has no hardware buttons/eraser — these are how host apps get those
affordances). Implements the ≤100ms heartbeat wire contract (80ms timer) so a
stationary held stroke survives the host's 200ms dead-client failsafe. Toward a
pen-less host nothing changes (Pencil stays a finger).
Also fixes hostSupportsCursor testing the STALE 0x04 bit — HOST_CAP_CURSOR
moved to 0x08 when TEXT_INPUT claimed 0x04; the old test would mistake a
text-input host (e.g. Windows) for a cursor grant and double-render the pointer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Release notes now live in the repo at docs/releases/vX.Y.Z.md, authored during the
version bump (before the tag). ensure_release (gitea-release.sh + .ps1 twin) seeds the
Gitea release body from that file in the create POST, so the release is born WITH its
notes instead of being created empty and PATCHed afterward. canary/rc have no file ->
empty body, unchanged.
New manual `announce` workflow (workflow_dispatch, tag input) re-syncs the notes file
over the live release and posts a violet embed (title, notes lead-in, release-page link)
to the Discord #releases channel via the DISCORD_RELEASE_WEBHOOK secret. Stable-only:
a -rc tag is refused unless allow_prerelease=true. Pressing "go" is the quality gate, so
a half-built or failed release is never announced.
docs/releases/README.md documents the ritual; TEMPLATE.md is the skeleton.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Parity with the Apple client (35c2c145): the SDL/Linux presenter built the
host-forwarded cursor as an OS cursor at its native bitmap size, so the
on-screen pointer tracked the HOST's display scaling — a 4K/high-DPI virtual
display forwards a 96 px bitmap that then rendered huge on the client.
SDL cursors are fixed-size from their surface (no draw-time scaling), so
cache the shapes RAW and resample each by the live video-fit factor
(min(window_px / mode) — the same aspect-fit `finger_to_content` uses) at
install time, rebuilding when the serial OR the fit changes (a resize). The
resample area-averages on premultiplied alpha so a transparent pixel's
colour can't bleed into the fringe. The pointer now stays sized to the
streamed desktop at any host scaling.
Not yet eyeballed on a Deck/Linux box — SDL's HiDPI cursor sizing is
version-dependent; validate on glass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Windows leg of design/pen-tablet-input.md §6, following Apollo's recipe:
a per-session PT_PEN device (pressure rescaled 0..1024, polar tilt→tiltX/Y,
barrel roll on rotation — Windows Ink renders Pencil Pro roll natively, barrel
button, eraser via INVERTED/ERASER flags, hover) with a 40ms refresh thread
against the ~100ms synthetic-pointer staleness auto-lift, plus a PT_TOUCH
device closing the historical SendInput wire-touch no-op (full active-contact
frames, per-id DOWN/UP edges, self-healing lost-DOWN synthesis). pen_supported
now probes PT_PEN creation on Windows (1809+), which lights up HOST_CAP_PEN
and the GameStream featureFlags there — Moonlight iPad + Pencil can ink on a
Windows host. Frame grouping mirrors the Linux uinput backend.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Moonlight leg of design/pen-tablet-input.md §4: DESCRIBE now advertises
SS_FF_PEN_TOUCH_EVENTS wherever pen injection exists (Linux+uinput, same gate
as HOST_CAP_PEN — elsewhere the flag stays 0 and Moonlight keeps client-side
mouse emulation exactly as today). SS_PEN packets merge over last-known state
into state-full PenSamples (BUTTON_ONLY per spec carries no position; unknown
contact pressure 0.0 inks at full scale; UP keeps proximity only for clients
that demonstrated hover) and drive the same PenTracker→VirtualPen chain as the
native plane — Moonlight iOS + Apple Pencil now exercises the full injection
path end to end. SS_TOUCH forwards as wire touch on a synthetic 65535² surface
with CANCEL_ALL replayed per tracked contact. No stroke timeout on this plane:
ENet is ordered/reliable and Moonlight doesn't heartbeat a stationary pen.
Packet layouts pinned against moonlight-common-c Input.h/Limelight.h upstream.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Linux injection leg of design/pen-tablet-input.md: 0xCC/0x05 pen batches now
route through the per-session PenTracker into a lazily-created 'Punktfunk Pen'
uinput tablet (BTN_TOOL_PEN/RUBBER, pressure, tilt-from-polar, ABS_Z barrel
roll, hover distance, INPUT_PROP_DIRECT) — compositors pick it up via libinput
and hand apps zwp_tablet_v2 with full fidelity. The host now advertises
HOST_CAP_PEN when /dev/uinput is accessible (PUNKTFUNK_PEN=0 kill-switch);
transitions group into SYN frames so proximity-enter carries its position.
Stroke failsafe: clients heartbeat ≤100ms while in range (documented wire
contract — capture APIs are silent for a stationary pen); 200ms of silence
force-releases. 'punktfunk-host pen-test' draws a pressure-ramped sine stroke
through the real tracker→uinput chain, no client needed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Field report (iPad via GameStream, 2026-07-23): cursor missing from the
stream. The session's fresh target 257 reported cursor_excluded=false
(DECLARED_TARGETS is keyed per target and only 259 had declared), so the
host skipped forced-composite — but DWM's pointer exclusion after an
IddCx hardware-cursor declare reaches EVERY later monitor of the adapter,
not just the declaring target (proven on-glass: declare on 259 under
WUDFHost 2192, then 257's frames streamed pointer-free with CURSOR_SHOWING
set and SM_MOUSEPRESENT true). GameStream/Moonlight clients cannot draw a
forwarded cursor, so those sessions were silently cursor-less.
ADD replies now report any-declare-this-WUDFHost-life (the state's real
scope — it dies with the adapter reset, same as before); per-target ids
stay only as the dbglog audit trail. Channel-less sessions on a declared
adapter then hit the existing forced-composite gate (GDI poller + blend)
and carry the pointer in-video. Host code unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Mac client drew the host-forwarded cursor bitmap at its native pixel
size (pixel-as-points), so the on-screen pointer tracked the HOST's display
scaling: a 4K/high-DPI virtual display renders a 96 px pointer, which then
appeared huge on the client — and on retina every pointer was already
2x-inflated and upscaled-blurry, since the backing scale was ignored.
Cache the shape raw and rebuild the NSCursor at the live video-fit factor
(fit.width / mode.width — the same AVMakeRect fit the input mapping uses)
on each apply, re-fitting on resize / retina moves. The pointer now renders
at its true size relative to the streamed desktop at any host scaling:
crisp and 1:1 with the video at 100%, and proportional (not ballooning)
when the host sits at 200-300% — including a deliberately raised scale.
iOS/tvOS (UIPointerInteraction) and Android (no forwarded-bitmap path)
render the pointer differently and need no change; the SDL/Linux client has
the same native-size issue and will follow.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
0.18.0 regression: UAC consent and Winlogon (lock/logon) stopped appearing
in streams. The cursor channel's IddCx hardware-cursor declare — re-issued
by the driver on every swap-chain assign — keeps the path out of the OS's
software-cursor mode, which is the only mode the secure desktop renders
through; DWM then never presents UAC/Winlogon into our swap-chain and the
stream repeats the last normal-desktop frame for the whole interaction.
The GDI cursor poller now classifies the input desktop on its reattach
cadence (UOI_NAME != Default = secure, 250 ms cadence instead of 2 s), and
the capturer edge-triggers the existing-but-unused proto-v6
IOCTL_SET_CURSOR_FORWARD flip: OFF at secure entry (the driver stops its
per-assign re-declare; the host facade forces the same-mode re-commit that
actualises the software-cursor default, under the vdisplay manager lock via
the new force_recommit()), back ON at dismissal for channel sessions.
Channel-session open resets the driver's persisted desired state to ON so
a session that died mid-secure-desktop can't leave the next one adopting
UNdeclared (the §8.6 cross-session composite trap); orderly teardown does
the same. The flip closure is built for every Windows session — a
channel-less session can reuse a driver monitor whose earlier-session
cursor worker still re-declares, and NOT_FOUND from never-declared targets
is ignored.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Steam-conflict gate matched any physical Valve `28DE` HID device, so a
Steam Machine with a physical Steam Controller 2 (`28DE:1302`) plugged in
degraded a client's requested virtual Steam Deck (`28DE:1205`) to a DualSense
— the reporter's controller passthrough came up as the wrong pad.
Steam Input drives *distinct* Steam controllers side by side; only two
*identical* `28DE` devices confuse it (the Deck-as-host case: built-in
`28DE:1205` + a second virtual Deck, observed live on Bazzite). Narrow the
gate to key on the exact VID+PID instead of the `28DE` vendor alone:
- `steam_backend_product()` maps each virtual Steam backend to the PID it
enumerates as (Deck 1205, classic SC 1102, SC2 1302, SC2 Puck 1304).
- `physical_steam_controller_present()` -> `physical_steam_product_present(pid)`
matches `:28DE:{PID}`, keeping the FVPF-serial / `vhci_hcd` / `/virtual/`
exclusions so our own virtual pads never count.
- `degrade_steam_on_conflict` degrades only on a same-PID duplicate and logs
the conflicting identity. `PUNKTFUNK_STEAM_FORCE=1` override unchanged.
A physical SC2 no longer blocks a virtual Deck (and vice versa); a genuine
same-identity duplicate (physical Deck + virtual Deck, or physical SC2 +
SC2 passthrough) still degrades. Adds a unit test for the PID mapping.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The pen plane from design/pen-tablet-input.md, protocol side only (the P1 uinput
tablet injector will consume it): 0xCC kind 0x05 carries batches of state-full
PenSamples (pressure, polar tilt + azimuth, barrel roll, hover distance, eraser
tool, barrel buttons); PenTracker diffs samples into injector transitions so a
lost datagram self-heals; HOST_CAP_PEN (0x10) gates sending and stays
unadvertised until a real backend exists. C ABI: PunktfunkPenSample +
punktfunk_connection_send_pen, documented in docs/embedding-the-c-abi.md §8.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Steam's "Add a Non-Steam Game to My Library" entries carry no `appmanifest`, so
`scan_manifests` can't see them and they never appear in the streamed library.
Read each Steam account's binary `userdata/<id>/config/shortcuts.vdf` directly
(a binary KeyValues parse) to surface the user's own custom entries, plus their
grid artwork from `userdata/<id>/config/grid`. Best-effort: an unreadable or
absent shortcuts.vdf contributes nothing, and hidden shortcuts / duplicate
appids are dropped. Verified: clippy -D warnings + compile green on Linux.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The visible "Punktfunk" shortcut's Steam appId is cached in SteamUI
localStorage, which lives in Steam's CEF profile — so it survives a
plugin uninstall/reinstall. Deleting the shortcut left the key dangling;
on the next mount ensure*Shortcut recalled the dead appId, took the reuse
branch, and ran SetShortcut* against a non-existent id (silent no-ops).
AddShortcut was never reached, so the entry never came back. The same
stale-reuse also made the hidden stream shortcut RunGame a dead gameid.
Verify the remembered appId still maps to a live shortcut via
appStore.GetAppOverviewByAppID before reusing it; a confident miss falls
through to AddShortcut. When appStore is unavailable, assume it exists so
a false negative can't spawn a duplicate library entry.
Also add a "Recreate library shortcut" recovery button to the QAM About
section (recreateShortcuts): drops any dead cached keys and re-ensures.
Covers deleting the shortcut without reinstalling, where no mount fires
the self-heal.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Every flatpak Tooling failure on 2026-07-22 (4 runs, incl. two AFTER the
runner fleet got its local dnsmasq cache) died in flatpak remote-add with an
instant 'Could not resolve hostname' ×10 while dnf in the SAME container
resolved fine. The one thing only this leg does is force glibc onto TCP DNS
(use-vc) — a 2026-07-11 mitigation for the embedded resolver dropping UDP
under fleet load. With job containers now resolving against the on-box
dnsmasq cache (daemon.json dns → docker0), UDP is reliable again and the TCP
path through the same chain is the flaky one under concurrency. Remove the
hack; keep the nss-resolve drop and retry.sh as the backstop.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The cursor-channel sweep grew AddReply by the cursor_excluded u32 tail (24
bytes, legacy prefix 20) and bumped PROTOCOL_VERSION to 6, but left two lib
tests behind: the AddReply byte-layout roundtrip still built the legacy
initializer (a compile error that turned the rust CI leg red on main), and
the version-coherence assertion still pinned v5 (hidden behind that compile
error). Cover the new tail bytes + the legacy-size constant, and widen the
compat-window note to v4–v6 (floor stays 3).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
bun-audit went red mid-day on a fresh advisory wave (13 findings across two
snapshots of a moving DB): node-tar DoS trio+1 (critical), sharp/libvips CVEs,
fast-uri host confusion, undici ×7, brace-expansion/linkify-it/js-yaml/postcss
DoS-class, immutable, dompurify. All transitive pins in the management-console
tree; none ship in the streaming stack.
Every fix version is in-range for its consumers, so: pin overrides for the
single-major packages (tar/dompurify/linkify-it/sharp/fast-uri/immutable/
undici/postcss/js-yaml — bun ignores range-scoped override keys, so plain keys)
and a full lockfile re-resolve, which also lifts brace-expansion 2.1.1→2.1.2
in place while its 5.x line (minimatch@10) stays untouched. : No
vulnerabilities found.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The KWin/Phase-A/B commits were built but never clippy-checked (that distrobox
had no clippy component), so they left TEMP on-glass probes and lint debt in the
tree. With clippy now runnable (fedora rust 1.96.1 = CI parity):
- drop the `fec424ee`/`8cff30d5` TEMP probes: the `update_cursor_meta` SPA_META
diagnostic logs (also un-detaches the `// SAFETY:` comment from its `unsafe`
block → fixes `undocumented_unsafe_blocks`) and the KWin composite-arm probe in
the encode loop.
- `#[allow(clippy::too_many_arguments)]` on `spawn_pipewire` (8 params since the
KWin leg added `expect_exact_dims`; mirrors `from_virtual_output`).
clippy `-p pf-capture -p pf-vdisplay -p punktfunk-host --locked --features
nvenc,vulkan-encode -- -D warnings` is now green. (--all-targets additionally
trips a pre-existing env mismatch: the fedora libspa binding lacks
`SPA_VIDEO_TRANSFER_SMPTE2084`, referenced only by a `#[cfg(test)]` guard-test —
not a code issue; CI's pinned pipewire has it.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A declared IddCx hardware cursor is IRREVOCABLE for its OS target's life
(§8.6), and the sticky exclusion survives monitor REMOVE→ADD because each
client gets a STABLE target id — so once any desktop-mode session declared,
every later pure-capture session on that target streamed a cursor-less
desktop: DWM excluded the pointer, no channel forwarded it, no blend drew it
(the exact no-regression gap §8.6's per-session cap gate cannot see).
Driver: track every successful SetupHardwareCursor per target
(DECLARED_TARGETS — scoped to the WUDFHost's life, exactly the sticky
state's scope) and report it in a new AddReply::cursor_excluded tail field
(dual-size discipline, both skews degrade cleanly; no proto bump).
Host: the flag rides AddedMonitor → Monitor → WinCaptureTarget; a session
WITHOUT the cursor channel on a flagged target forces composite mode in the
IDD-push capturer — GDI poller + blend for the session's life, pinned on
(set_cursor_forward cannot clear it: with no client drawing, un-compositing
would erase the pointer entirely).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
gamescope excludes its pointer from the PipeWire node it feeds us and can't embed
one either (`set_hw_cursor` is inert), so every gamescope stream was cursorless.
Read the pointer from gamescope's nested Xwayland instead — XFixesGetCursorImage
for shape/hotspot/visibility, core QueryPointer for position — and publish a
CursorOverlay into the capturer's existing `cursor_live` slot, so the encoder
blend composites it into the video exactly like the SPA_META_Cursor path.
- pf-capture/src/linux/xfixes_cursor.rs (new): the XFixes reader. Connects to
EVERY nested Xwayland (Gaming Mode runs one per --xwayland-count) and follows
the focused one each tick (the display whose pointer moves), reading that
display's own cursor shape. Un-premultiplies ARGB -> straight RGBA. Drop stops
the thread.
- Capturer::attach_gamescope_cursor + the PortalCapturer override spawn it into
the same `cursor_live` slot; pf_vdisplay::gamescope_xwayland_cursor_targets
discovers the (DISPLAY, XAUTHORITY) pairs via the GAMESCOPE_WAYLAND_DISPLAY scan.
- host: SessionPlan.gamescope_cursor (set from the compositor at both resolve
sites AND on a mid-stream Desktop->Gaming switch); the blend gate now builds the
encoder blend for gamescope; a sibling composite arm attaches capturer.cursor()
per tick for capture-mode clients (no channel needed). native NV12 is disabled
for these sessions — that encode path can't blend the cursor (it assumes
gamescope embeds the pointer), so we capture RGB and route to the proven CUDA
VkSlotBlend / compute-CSC blend.
- the pipewire thread no longer clobbers `cursor_live` with None on a buffer that
carries no SPA_META_Cursor (gamescope) — that raced the XFixes writer and
strobed the composited pointer on/off.
On-glass (home-bazzite-2, RTX 5070 Ti, Gaming Mode): cursor visible + steady in
Big Picture and CS2 menus, follows focus between the two Xwaylands, hidden in-game.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Since the cursor-channel work, every Linux virtual output was created in
metadata pointer mode — making ALL sessions depend on host-side cursor
compositing, including the ones that can never use the channel (Moonlight/
GameStream, legacy clients, capture-mode starts). Those sessions paid the
blend bring-up per session and, whenever a visible cursor was composited,
the loss of NVENC's stream-ordered submit — for a strictly worse cursor
than the compositor's own.
Mirror the Windows no-regression gate: the already-wired per-session
set_hw_cursor(cursor_forward) now drives the Linux backends too. A
cursor-channel session gets metadata (shapes forwarded, composite flip
blends host-side — today's validated path, unchanged); every other session
gets the pointer compositor-EMBEDDED at creation (KWin zkde pointer=2,
Mutter cursor-mode=1, wlroots/hyprland portal CursorMode::Embedded — their
pre-channel default; the portal pair also gains the Metadata arm for
channel sessions, wired but untested on-glass). SessionPlan.cursor_blend
narrows to cursor-forward sessions and rides into the direct-SDK NVENC
open, which skips the Vulkan slot-blend bring-up entirely when off —
embedded sessions ring on plain CUDA surfaces and pay zero cursor cost,
per-session or per-frame.
The keep-alive registry's reuse key grows the created pointer mode (new
VirtualDisplay::hw_cursor getter): a kept embedded display has no cursor
metadata for a channel session to forward, and a kept metadata display
would leave a channel-less session with no pointer in its frames.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A vendored PTX blob is JIT'd against the driver's ISA ceiling, so the
cursor-blend module silently dies on drivers older than the generating
toolkit (CUDA_ERROR_UNSUPPORTED_PTX_VERSION/INVALID_PTX, 222/218 — the
KWin leg's invisible composite cursor on driver 595/CUDA 13.2 vs a
CUDA 13.3 blob). SPIR-V has no such coupling, and the in-tree precedent
already exists twice (vulkan_video's CSC blend, VkBridge's exportable
OPAQUE_FD → cuImportExternalMemory bridge).
New pf_zerocopy::vkslot::VkSlotBlend: the direct-SDK NVENC encoder now
allocates its input ring as exportable Vulkan buffers CUDA-imports (same
contiguous InputSurface layouts, pitch = row bytes rounded to 256), and
the cursor composite is a compute dispatch over the cursor's rectangle
(cursor_blend.comp, vendored .spv; spec-constant selects ARGB/NV12/YUV444;
BT.709 limited, matching the retired .cu). The surface SSBO is uint[] with
every invocation owning whole words — no 8-bit-storage device dependency.
Cursor-bearing frames force the existing CPU-synced submit path so the
CUDA copy → Vulkan dispatch (fence-waited) → NVENC encode ordering is
CPU-established; cursorless frames keep the stream-ordered fast path
untouched. Any bring-up/alloc/registration failure falls back wholesale
to plain pitched CUDA surfaces (never a mixed or short ring): sessions
always encode, composite mode just loses the cursor, warned once.
cursor_blend.cu / cursor_blend.ptx and the CursorBlend PTX loader are
deleted.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rate-limited journal logging bracketing the silent segment of the Linux
composite path: what overlay (if any) the encode loop's composite arm
hands the encoder, and whether the NVENC cursor-blend kernel launches and
with what geometry. On-glass (KWin leg) the blend module loads yet the
composite cursor stays invisible — these two probes attribute it to
no-overlay / stripped-en-route / kernel-draws-nothing. The module-loaded
INFO line is permanent (success must be as attributable as failure);
everything else in this commit is temporary.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
KWin's ScreenCastStream builds its PipeWire format offer — including the
maxFramerate cap it actively throttles delivery to — ONCE at stream
creation, when the virtual output sits at its hardcoded birth 60 Hz. A
kscreen custom-mode change afterwards updates the OUTPUT (readback says
240) but never the offer, so every consumer negotiates max=60 and the
stream delivers 60 (pw-dump-verified on KWin 6.6.4). The only path that
rebuilds the offer is the stream's own resize handling: a source SIZE
change while recording re-runs buildFormats — picking up the output's
current refresh — and renegotiates the live stream.
So above 60 Hz, birth the output at a sacrificial height (+16), then
install + select the real WxH@hz custom mode: the first frame recorded
after the consumer connects triggers KWin's resize path, which renegotiates
the live stream to WxH@hz. The capturer holds (requeues) buffers until the
negotiated size matches — new expect_exact_dims plumbing VirtualOutput →
registry (fresh creates only) → open_virtual_output → a self-disarming
gate in the process callback — bounded by a 3 s deadline that accepts the
producer's dims rather than wedging the session into the first-frame
retry loop. set_custom_refresh reads back the full active mode; a size
reject (pre-6.6 KWin) recreates plain at the real size @60.
Every kscreen operation now addresses the output by its NUMERIC id,
resolved by matching the managed name-prefix AND the just-created birth
size (newest id wins): a mode-switch supersede reuses the per-slot output
NAME (deliberately, for KWin's per-name config persistence) while the
superseded sibling is still alive, and name addressing hit the FIRST
match = the OLD output — on-glass that resized the live session's display
out from under it (wrong-res/black), read the old output back as 'mode
applied', and set the old output primary while the replacement starved at
its birth mode.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The vendored cursor_blend.ptx was regenerated with the CUDA 13.3 toolkit,
which stamps '.version 9.3' — and a driver whose JIT predates that ISA
refuses the whole module with CUDA_ERROR_UNSUPPORTED_PTX_VERSION (222),
silently killing host-side cursor compositing (on-glass: composite-mode
cursor invisible on driver 595.58, KWin leg). The kernels use only baseline
arithmetic, so hand-lower the version directive to 8.0 (CUDA 12.0), which
every supported Turing+ driver JITs. Both files now warn that an nvcc
regeneration re-stamps the version and must be re-lowered.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The frame-attached overlay is the pointer position at the last damage
frame; repeats re-encoding a static desktop froze the blended cursor
between redraws — on-glass the composite-mode cursor stuttered while
window drags (constant damage) were smooth. Refresh the repeat's overlay
from Capturer::cursor each tick so pointer-only motion re-blends at tick
rate — the bandwidth the pre-channel embedded mode already paid. Linux
only: the Windows capturer composites internally and its frames must
never carry an overlay a blend path would double-draw.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The forwarder read only frame.cursor — the overlay attached to the LAST
EMITTED frame. Mutter's pointer-only buffers update the capture-side cursor
state but are skipped as frames, so between damage frames the forwarder
re-sent a stale snapshot: on-glass the client cursor appeared exactly when
clicking or crossing hover targets (damage) and froze/vanished otherwise —
the same pointer-only-motion gap the Windows IddCx channel fills, which is
why the encode tick already prefers the capturer's LIVE cursor. Publish the
overlay from every dequeued buffer (frames or not) into a shared slot and
implement the Capturer::cursor hook on PortalCapturer.
Host-composite mode on Linux still updates the blended pointer only on
damage frames (no re-encode of a static desktop on cursor motion) — the
regen-on-cursor-change the Windows path has is a follow-up.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Mutter only rewrites a buffer's SPA_META_Cursor region when the cursor
changed; recycled buffers between damage frames carry a stale id-0 meta.
Treating id 0 as a hidden pointer flickered the client cursor off between
hovers on-glass — per the SPA contract id 0 is 'invalid/no cursor info',
so keep the last-known state (OBS's consumer does the same). A genuinely
hidden pointer stops producing updates; the M3 relative-mode hint keeps
its Windows CURSOR_SUPPRESSED source.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Mutter offers SPA_META_Cursor with a FIXED size pod —
SPA_POD_Int(CURSOR_META_SIZE(384, 384)) in meta-screen-cast-stream-src.c —
while our request capped the range at meta_size(256, 256). The intersection
is empty, so the Meta param silently failed to negotiate and no buffer ever
carried the meta region: the entire Linux cursor pipeline (forward AND
blend) was blind on GNOME, proven by an on-glass probe counting 32k+
buffers with zero metas. KWin's offer fits inside 256², which is why the
same consumer code passed there. Raise the max to 1024² headroom (the
negotiated allocation follows the producer's value, not our max) and align
the bitmap parse guard.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rate-limited journal logging of SPA_META_Cursor arrival/absence, reported
id/position/bitmap offset, and bitmap acceptance — the Linux cursor
pipeline is otherwise blind end-to-end during Mutter RecordVirtual
metadata bring-up on .21.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The libei absolute device advertises one region per logical monitor, and
both MouseMoveAbs and touch mapped into regions().first() — whichever
output the compositor announced first. Next to a physical monitor (on-glass:
GNOME with a dummy HDMI beside the virtual primary) that put the pointer and
every click on the WRONG output: the seat cursor never entered the streamed
monitor, so neither embedded nor metadata cursor capture could ever see it,
and the cursor channel had nothing to forward. Pick the region whose logical
size matches the streamed mode (the wire flags already carry it); fall back
to first() for the single-monitor case. Region geometry now rides the
device-RESUMED log line for diagnosability; matching the screencast
mapping_id instead of the size is the follow-up for same-sized-monitor
ambiguity.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Same trap as the Mutter backend: zkde_screencast pointer mode 2 (embedded)
never delivers SPA_META_Cursor, starving both the cursor channel and the
encoder blend. Mode 4 (metadata) feeds both.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The resize rebuild creates the new virtual display BEFORE retiring the old
one (create-before-drop), so the registry saw a live same-backend sibling —
its own dying predecessor — and told the backend it was not first in group.
Mutter then skipped the Primary/Exclusive apply ("joining an existing
display group — extending") and the retiring owner took the topology with
it: every resize silently demoted the virtual output to an extended,
shell-less desktop. Thread the superseded pool gen through acquire so the
replacement establishes topology, and stop counting kept (Lingering/Pinned)
entries as demoting siblings — no session owns them, so there is no live
desktop to clobber.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
RecordVirtual was created with cursor-mode EMBEDDED, which never delivers
cursor metadata — the cursor channel had nothing to forward (client-drawn
cursor invisible) and the always-built encoder blend had nothing to
composite (host-drawn cursor invisible too). Metadata mode feeds both:
non-forwarding sessions get the blended pointer (same pixels Mutter would
have embedded), forwarding sessions strip the overlay and send shape/state.
The capturer already negotiates the meta on every stream (meta_param).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
faad4b57 shipped the field, shader scale, and plumbing but the
scripted edit inserting the sdr_white_level_scale() refresh in the
scratch-rebuild arm matched nothing and silently no-op'd (no assert) —
sdr_white_scale stayed at its 1.0 init, i.e. 80 nits, i.e. exactly the
old dark cursor. Now queried on every scratch rebuild, with an info
log of queried/applied so an HDR session shows what it uses.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- The polled position is also the composite-blend position: at 16 ms a
240 fps session reused a stale position for ~4 consecutive frames and
the composited pointer visibly stuttered against the video. 4 ms
out-paces every session rate; a tick is one GetCursorInfo syscall.
- Log the queried DISPLAYCONFIG_SDR_WHITE_LEVEL scale on each blend-
scratch rebuild so an HDR session shows what is actually applied
(the virtual display's own SDR-brightness setting is the suspect
when the cursor still reads dark).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
sRGB→linear alone lands cursor-white at 80 nits (scRGB 1.0) while DWM
composes the surrounding SDR desktop at the user's SDR-brightness
setting (~200+ nits by default) — the cursor read visibly dark on HDR.
Scale by DISPLAYCONFIG_SDR_WHITE_LEVEL (new
win_display::sdr_white_level_scale, queried on blend-scratch rebuilds —
the CCD query stays off the per-frame path).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Root cause, proven on-glass across five driver builds: a declared IddCx
hardware cursor is IRREVOCABLE. There is no un-declare DDI, the
empty-caps re-setup is rejected INVALID_PARAMETER, and after a
successful same-mode re-commit with the driver's re-declare provably
suppressed (sticky per-target flag, zero re-setups logged) DWM still
never composites the software cursor back into the monitor's frames.
Every DWM-based composite flip is therefore unfixable.
The composite (capture) model is now implemented in OUR pipeline:
- the driver keeps its hardware cursor declared for the session's whole
life (the state that works) — frames stay pointer-free always;
- in composite mode try_consume routes the conversion through a blend
scratch: slot copy + one alpha-blended quad of the GDI poller's shape
at its polled position (CursorBlendPass — fullscreen-triangle VS with
viewport placement, straight-alpha PS, sRGB→linear for FP16/HDR
rings), covering all four convert paths (NV12, 4:4:4 copy, P010,
PyroWave) GPU-side under the slot's keyed mutex;
- pointer-only motion produces no driver publish (declared hw cursor),
so try_consume REGENERATES from the last slot whenever the polled
cursor state changes — the cursor moves on a static desktop at tick
rate, without faking freshness for the stall/death watchdogs;
- set_cursor_forward becomes purely host-internal state; the
SET_CURSOR_FORWARD IOCTL machinery is no longer used by the host
(driver keeps it, dormant) and its sender plumbing is removed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
On-glass: 're-setup on swap-chain assign -> 0x0' fired AFTER 'enable=0
stored' — duplicate/successor monitor entries (re-arrival churn) kept
their default-true flag and re-declared the hardware cursor right after
the composite flip's re-commit, undoing it.
The desired state now lives in a per-target static
(CURSOR_FORWARD_DESIRED): every arrival inherits it (no generation can
resurrect a declare the session turned off), and the flip stamps EVERY
live entry matching the target, declaring against whichever has the
live worker.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The empty-caps un-declare is REJECTED by IddCx (STATUS_INVALID_PARAMETER,
on-glass driver 9.9.0722.1407) — there is no un-declare DDI. The
composite flip now works with the only lever the OS gives us: every
mode COMMIT reverts the path to the software cursor, and the driver
skips its per-commit re-declare while cursor_forward_on is off. So:
- driver: disable stores the flag only (no DDI call); enable still
declares immediately against the live worker's event.
- host capturer: after a successful disable flip, force a same-mode
re-commit (win_display::force_mode_reenumeration) — DWM composites
the pointer from the very next commit; the swap-chain flap is the
class the driver's preserved-publisher machinery already rides out.
- state now survives re-arrivals end-to-end: the capturer caches the
APPLIED state (cleared on every channel (re)delivery, which
re-created driver entries need — their flag defaults to declared),
and the stream loop re-applies every tick instead of edge-gating
(steady-state cost: one Option compare).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
On-glass (.173, match-window session): every window-size convergence
re-creates the driver-side monitor (re-arrival resize / sibling-session
slot churn), destroying the cursor worker the channel was delivered to —
the flip then died NOT_FOUND (no entry with target_id ∧ worker) and the
declared state was lost entirely.
Driver: set_cursor_forward becomes STATE, not an edge on one monitor
generation — find by target_id ∧ hw_cursor, store cursor_forward_on
even without a live worker (it steers the next delivery/re-setup);
declare/un-declare only when a worker exists. Channel delivery honors
the flag: setup_and_spawn(declare=false) adopts + spawns WITHOUT
declaring (composite mode), so a later enable-flip declares against the
worker's event.
Host: retain the SET_CURSOR_CHANNEL sender; extract
deliver_cursor_channel and RE-deliver the surviving section on every
ring recreate and — with one retry — when a flip IOCTL fails
(the re-arrived entry gets a fresh worker, declared per its flag).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Belongs to 493f4fae (which shipped only the crates/ side — this
packaging/ half was left unstaged, so the deployed driver silently
lacked the IOCTL and every flip died NOT_FOUND at dispatch):
IOCTL_SET_CURSOR_FORWARD dispatch + monitor::set_cursor_forward
(cursor_forward_on flag, resetup gate) + cursor_worker::
unsetup_hardware_cursor (empty-caps un-declare experiment).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
PunktfunkConnection.setCursorRender(clientDraws:) over the new
punktfunk_connection_set_cursor_render export, driven by one
edge-detected reconciler in StreamView (clientDraws = captured &&
desktopMouse) called from every transition — the ⌃⌥⇧M chord,
engage/release, and session start. Capture model and released now get
the host-composited pointer back in the video (full fidelity); the
desktop model keeps the local NSCursor path. xcframework rebuilt v12.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The cursor channel excluded the host pointer from the video for the
whole session, but the client only draws it under the DESKTOP mouse
model — flipping ⌃⌥⇧M to capture mid-stream left the session
pointer-blind (user-found). The fix makes render ownership LIVE state:
- wire: CursorRenderMode 0x51 (client→host, control stream) —
client_draws: true = desktop model (host excludes + forwards),
false = capture model / released (host composites, full fidelity —
DWM on Windows incl. real XOR inversion, encoder blend on Linux).
Sessions start client_draws=true (the pre-message behavior).
- host: control task stores the flag; the encode loop edge-detects it —
forwarding + frame.cursor strip while the client draws, quiet
forwarder + overlay-into-blend while composited. SessionPlan now
grants blend CAPABILITY wherever the capture has a pointer (dropping
the !cursor_forward gate) so the composite side needs no encoder
rebuild; per-tick frame.cursor decides what is drawn.
- Windows: Capturer::set_cursor_forward → retained
IOCTL_SET_CURSOR_FORWARD sender (proto v6) → driver (un)declares the
IddCx hardware cursor on the LIVE monitor. Un-declare re-issues the
setup with EMPTY caps (candidate mechanism — the DDI has no
documented un-setup); resetup-on-commit is skipped while off, so a
mode commit's software-cursor default is the fallback path. Failure
against a pre-v6 driver logs and keeps declared-at-ADD behavior.
- SDL client: one edge-detected reconciler at the cursor pump site —
desktop-active = client draws; capture model or released = host
composites. Covers the chord, the M3 auto-flip, and engage/release.
- C ABI v12: punktfunk_connection_set_cursor_render (additive; wire
version unchanged — pre-§8 hosts ignore the unknown message type).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
cursorChannelActive/hostCursors/cursorState stayed latched across
sessions — a next session against a host WITHOUT the cursor cap would
wear the previous session's stale shapes via resetCursorRects.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The M2c redesign (design/remote-desktop-sweep.md §8): the IddCx hardware-
cursor query is alpha-only by design — IDDCX_CURSOR_SHAPE_TYPE has no
monochrome value, the OS pre-converts monochrome to masked-color, and
masked-color delivery is dead code on modern builds (proven on-glass at
every ColorXorCursorSupport level). The driver keeps its hardware cursor
declared (XOR FULL) purely so DWM excludes every cursor type from the
IDD frame; the SHAPE now comes from a GDI poller in the capture host —
which runs as SYSTEM inside the interactive session, so no helper
process is needed.
CursorPoller (idd_push/cursor_poll.rs, the RustDesk/WebRTC/OBS pattern):
~60 Hz GetCursorInfo on a dedicated thread; rasterise via
CopyIcon→GetIconInfo→GetDIBits only when the HCURSOR value changes;
monochrome via the WebRTC truth table with invert pixels as opaque black
plus a white outline; AND-mask alpha for alpha-less color cursors;
CURSOR_SUPPRESSED = hidden; per-output visibility via the target's
desktop rect; per-monitor-V2 DPI awareness thread-scoped; input-desktop
reattach on failure + 2 s cadence (GetCursorInfo on a stale desktop
succeeds with stale data). Spawned iff the cursor channel was delivered;
a live poller is the sole overlay source (no serial-namespace mixing) —
the driver shm read remains as fallback if the poller dies.
Not chosen: DXGI Desktop Duplication GetFramePointerShape —
PointerPosition.Visible goes stale under injected-only input on current
Win11 (Sunshine #5293, exactly our topology), it burns one of the
session's four duplication slots, and IDD-output metadata has
conflicting field reports.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Durable results of the M2c on-glass bring-up (.173, 2026-07-22):
- The base IddCxMonitorQueryHardwareCursor DDI slot is stubbed to
STATUS_NOT_SUPPORTED on WDK 26100 — add the ...QueryHardwareCursor3
wrapper (wdk-iddcx) and drain it instead; v3 X/Y are only meaningful
when PositionValid, so a position-invalid tick keeps the prior position.
- Hardware-cursor setup is per-mode-commit: the OS silently reverts to a
software cursor on every mode commit, so re-issue the setup on each
swap-chain assignment (monitor::resetup_cursor, called from
assign_swap_chain) or the query fails NOT_SUPPORTED forever.
- One caps definition for initial setup and re-setup, resting at XOR
FULL: the query delivers ONLY alpha shapes in every configuration
(all three ColorXorCursorSupport levels, event-driven and 30 Hz
polled) — masked/monochrome cursors never arrive. FULL keeps the
frame cursor-free for ALL cursor types; the full-fidelity shape comes
from the session-side cursor source in the host
(design/remote-desktop-sweep.md §8).
- Poll the query at ~30 Hz instead of pure event-wait: masked cursors
never fire the data event, and the seqlock's position/visibility
should stay fresh regardless.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
On-glass (Mac ↔ Windows M2c host): the host cursor's VISIBILITY is not a
usable 'a game grabbed the mouse' signal — Windows hides the pointer for
ordinary desktop activity (clicking, typing). The per-frame forwarder
turned those transients into relative_hint=true, so the client flipped
desktop→capture→desktop, which (a) warped the cursor to view-centre and
(b) flushed held buttons via releaseAll — a spurious mouse-up ~200 ms
into every press that broke window dragging/resizing.
Disable the auto-flip; the mouse model stays user-driven (⌃⌥⇧M). The
hint still rides the wire, unused, for a future M3 that keys off a REAL
pointer-lock signal (host-side ClipCursor/raw-input detection) instead
of visibility.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The FIRST session's Welcome runs before any backend construction
(vdisplay::open happens at display prep, after the Welcome), so the
capability probe's vdm() expect panicked the very first handshake of a
fresh service — the client saw a dead connect and auto-wake kicked in.
init() is idempotent and the driver facade is a unit struct, so the
probe now initialises the manager itself. Found on-glass (.173, first
Mac-client connect), fix deployed there.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Mac client now exercises the full cursor channel against capable
hosts: desktop-mode sessions advertise CLIENT_CAP_CURSOR, the host
stops compositing the pointer, and StreamView draws the forwarded
shapes as the real NSCursor — plus the M3 host-driven model flip.
- ABI v11: punktfunk_connect_ex9 (adds client_caps; ex7/ex8 pass 0 —
Android untouched), PunktfunkCursorShape/State + the two
next_cursor_* poll fns (audio-style borrow contract), and
PUNKTFUNK_CLIENT_CAP_CURSOR. Fixed in passing: the first insertion
split next_host_timing's cfg/no_mangle attributes off the fn, which
silently dropped the symbol from the header AND dylib — caught by
the Swift build, reattached with its docs.
- PunktfunkConnection: clientCaps connect param, hostSupportsCursor,
CursorShapeEvent/CursorStateEvent + nextCursorShape/nextCursorState
(one cursor thread drains both planes; cursorLock joins the close()
ladder).
- SessionModel: desktop-mode sessions (DefaultsKey.mouseMode) connect
with the cursor cap on macOS.
- StreamView: shape cache (serial → NSCursor, straight-alpha RGBA via
CGImage), resetCursorRects wears the HOST shape while the desktop
model is engaged (hidden host pointer ⇒ invisible), latest-wins
state pump, and the M3 auto-flip — edge-triggered on relative_hint,
⌃⌥⇧M sets an override latch cleared by the next host edge, leaving
relative warps the pointer to the host position via the inverse
letterbox mapping (CGWarpMouseCursorPosition, CursorCapture's
coordinate convention).
Verified: swift build + full test suites green (xcframework rebuilt at
ABI v11, signed); core 218 tests + clippy -D warnings on Linux (.21).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Brings the cursor channel to Windows hosts. The pf-vdisplay driver
declares an IddCx hardware cursor for sessions that negotiated
cursor-forward — DWM then EXCLUDES the pointer from the IDD frame and
delivers shape/position out-of-band, into the same CursorOverlay →
forwarder → wire → client pipeline the Linux portal path uses.
- pf-driver-proto v5 (additive, host floor stays 3): AddRequest's spare
tail becomes hw_cursor (same size/offsets); IOCTL_SET_CURSOR_CHANNEL
delivers a host-created CursorShm section (64-byte seqlock header +
256² shape buffer, layout pinned + tested). No event crosses the
boundary — the host polls at encode-tick pace.
- driver: wdk-iddcx grows the two cursor DDI wrappers; a per-monitor
cursor worker (event wait → QueryHardwareCursor → seqlock publish)
starts only when BOTH the ADD asked and the channel arrived, so a
failed delivery leaves DWM compositing as today. Shape bytes ship raw
(BGRA/masked + pitch); the host converts.
- host: the section rides the existing sealed-channel broker (least-
privilege dup, remote reap on failure); IddPushCapturer::cursor()
seqlock-reads → CursorOverlay (BGRA→RGBA, masked-color approximation,
desktop→frame origin shift, per-shape conversion cache). New
Capturer::cursor() trait hook — the encode loop prefers it over the
frame-attached overlay because hardware-cursor moves produce NO new
frame on a static desktop. hw_cursor survives the re-arrival resize
(carried on the manager's Monitor).
- negotiation: cursor_forward grows the Windows arm (client cap ∧
driver proto ≥ 5, probed once via the control device); SessionPlan
carries cursor_forward → OutputFormat.hw_cursor.
- drive-by: wdk-probe's two pre-existing same-type casts (clippy) and
pf-vdisplay's stale spike-test refs (crate::win_display moved to
pf-win-display; tracing-subscriber was never a dep) repaired.
Verified: proto tests on Mac AND MSVC (15/15 incl. the CursorShm
layout pin); clippy -D warnings for proto/frame/capture/vdisplay/host
on Linux (.21) and native Windows (.173); the DRIVER workspace clippy
-D warnings green against the real WDK 10.0.26100 bindgen (DDI names,
enum variants and IDARG layouts all bind). On-box driver deploy +
on-glass validation follow.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The full Parsec model: launching a game from a desktop session flips the
client into captured relative automatically, and back — no chords.
- Capture overlay now distinguishes 'hidden' from 'no cursor yet':
CursorOverlay grows a visible flag, overlay() returns Some whenever a
bitmap is known (the encode loop strips invisible overlays after
forwarding so no blend path ever draws one; the CPU composite guarded
on visibility already).
- Host forwarder maps it onto the reserved wire bit: visible ⇒ VISIBLE,
hidden-but-known ⇒ RELATIVE_HINT (an app grabbed/hid the pointer),
never any hint before the first bitmap — a cold start can't flip a
desktop session into capture.
- Presenter: edge-triggered auto-flip on hint changes via the new
Capture::set_desktop (chord semantics untouched); a manual ⌃⌥⇧M sets
an override latch that holds until the HOST's intent next changes, so
the hint never fights the user. Leaving relative warps the local
cursor to the host's last pointer position through the new
content_to_window inverse-letterbox mapping (unit-tested against
finger_to_content) — the hand-back is seamless.
Verified on .21: fmt + clippy -D warnings + tests (presenter 20 incl.
the inverse-mapping roundtrip, host 245).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The host cursor stops riding the video and becomes a real OS cursor on
the client (the Parsec/RDP model): pointer feel no longer pays the
capture→encode→network→decode→present round trip.
Wire (M2a):
- Hello grows a client_caps trailing byte (CLIENT_CAP_CURSOR) after the
fixed display_hdr block — presence disambiguated by remaining length,
which caps the post-HDR tail at 27 bytes (documented); Welcome answers
HOST_CAP_CURSOR (capable-and-asked, the 444/clipboard precedent).
- CursorShape (0x50, control stream): serial + dims + hotspot + straight
RGBA, ≤120px/side so the u16 frame always fits (128² would overshoot);
client caches by serial — re-showing a known shape costs 14 bytes, not
a bitmap (RDP pointer-cache for free).
- CursorState (0xD0 datagram): serial + visible/relative_hint flags +
position, sent once per encode-loop tick — latest-wins, self-healing
under loss, no refresh timer. relative_hint is reserved for M3.
- Client core: two new planes (control-task + datagram-task arms) →
next_cursor_shape/next_cursor_state; connect() grows client_caps
(C ABI passes 0 until the v11 cursor poll fns exist).
Host (M2b, Linux portal only):
- handshake::cursor_forward is THE predicate (client asked ∧ Linux ∧
compositor ≠ gamescope) — Welcome bit and session wiring both read it.
- SessionPlan.cursor_blend goes false for a forwarding session; the
encode loop ticks a CursorForwarder every iteration: shape-serial diff
→ control-task bridge (mirrors probe_result), state datagram → conn.
- CursorOverlay/capture CursorState carry the hotspot through
(nearest-neighbor downscale backstop for XL cursors, unit-tested).
Presenter:
- CursorChannel drains both planes per loop iteration; shapes become
SDL color cursors (from_surface + hotspot), applied while the desktop
mouse model is engaged; visibility follows the host; capture/released
hands back the system cursor. Sessions advertise the cap when they
START in desktop mode.
Verified on .21: fmt + clippy -D warnings (7 crates) + tests green
(core 218 incl. new wire roundtrips, host 245 incl. e2e + forwarder
downscale tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Folds the parked client-side-cursor machinery (cursorMode auto/always/
never, hidden while disabled) into the cross-client mouse model:
MouseInputMode capture|desktop under DefaultsKey.mouseMode, picked in
Settings ▸ Keyboard & mouse (macOS), resolved at session start and
gated off on gamescope hosts (relative-only EIS).
Desktop model = the un-neutered absolute path with the SDL cursor
policy: pointer never disassociated (enters/leaves the stream freely),
monitor forwards letterboxed absolute positions, and the local cursor
hides only while over the view via an invisible-cursor rect (the
host's composited cursor is the one you see; AppKit manages the rect,
so no hide/unhide balancing). ⌘⇧C becomes ⌃⌥⇧M — the same chord as
the SDL clients — flipping the model live with an atomic
release/re-engage.
Verified: swift build + full test suites green (macOS arm64).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New physical-mouse model beside capture: Desktop leaves the pointer
uncaptured and sends absolute positions through the letterbox
(MouseMoveAbs — every host injector already consumes it). Capture
stays the default and the game model.
- pf-client-core: MouseMode (capture|desktop) persisted in Settings,
default capture so existing stores are unchanged.
- pf-presenter: Capture grows a desktop model — latest-wins pending
abs position coalesced per loop iteration (same 1000 Hz discipline
as relative), flushed before clicks/keys/wheel so they land where
the cursor is; Ctrl+Alt+Shift+M flips the model live; local cursor
stays hidden over the window (the host's composited cursor is the
one you see until the M2 cursor channel); Windows keyboard grab
only engages for capture (a desktop stream is something you
Alt-Tab away from).
- gamescope gating: its EIS is relative-only, so desktop mode is
pinned off there (resolved_compositor), with a log note.
- Settings surfaces: GTK row (dynamic caption), WinUI combo,
console-UI row + step test, capture-hint line.
Verified: fmt + clippy -D warnings + tests (33/40/19) on Linux .21;
clippy -D warnings for all five crates incl. punktfunk-client-windows
native on .173.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 0.17.2 placeholder underestimated this cycle: Android mouse & keyboard
(physical mouse + capture, TV remote-as-pointer, IME text, clipboard), the new
committed-text input plane (InputKind::TextInput / HOST_CAP_TEXT_INPUT), and
the ChaCha20-Poly1305 session cipher are feature work — a minor, not a patch.
Wire/driver/C-ABI markers all unchanged (2 / 4 / 10): every addition is
negotiated or ignored by older peers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The android leg failed twice on 2026-07-22 with sdkmanager's "Error on
ZipFile unknown archive" — once in our NDK/build-tools step, once inside
setup-android's default package set. Root cause is the shared runner fleet
dropping packets under parallel-job load (the same class retry.sh documents
for the flatpak leg): sdkmanager unzips while streaming, so a truncated
stream reads as a corrupt archive. A solo re-download on the same runner
verified clean.
- setup-android: install only platform-tools — the action's default legacy
`tools` package drags in the ~250 MB emulator this job never uses, which
was the largest and flakiest download in the whole leg.
- NDK/build-tools/CMake: wrap sdkmanager in retry.sh (4 attempts, linear
backoff) — a failed attempt leaves no partial package, so re-invoking is
safe.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
punktfunk negotiates HDR correctly (the stream is BT.2020 PQ 10-bit) but never
told the client to switch its display into HDR picture mode, so HDR streams
stayed in SDR on the TV. Sunshine sends an async control-stream message
(IDX_HDR_MODE, type 0x010e, carrying SS_HDR_METADATA) shortly after the stream
starts; Moonlight only flips the TV into HDR on that cue. We sent exactly one
host→client control message (rumble 0x010B) and never this one. (aurora-tv PR
#53 worked around it client-side by inferring HDR from the negotiated format;
this fixes it host-side so stock Moonlight clients switch too.)
Once the control stream is live and the session negotiated HDR, send a one-shot
0x010e message (enable + 26-byte SS_HDR_METADATA from generic HDR10 defaults),
latched per connection and re-armed on Disconnect. The rumble sequence counter
is unified into a single monotonic `host_seq` shared by both message types —
the GCM nonce derives from `seq`, so a separate per-type counter would reuse
(key, nonce) pairs in the host direction.
The BT.2020 PQ VUI already rides the bitstream (the encoder derives it from the
P010 capture format), so this is purely the missing "switch now" signal; no
encoder change needed. Wire layout locked by unit tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
rustls-webpki 0.103 accepts only X.509 v3 certs and returns
`UnsupportedCertVersion` for anything older, aborting the mTLS handshake at
the `CertificateVerify` step before it ever checks the signature. The
moonlight-embedded client family (moonlight-embedded, aurora-tv, …) still mint
self-signed **v2** certs with no keyUsage extension, so pairing fails against
punktfunk while it works against Sunshine — whose OpenSSL verify callback never
inspects the cert version. (aurora-tv PR #53 worked around this client-side by
switching to v3 + keyUsage; this fixes it host-side for stock clients too.)
The cert's X.509 version and extensions carry no security weight here: the
client cert is self-signed and pinned by SHA-256 (`peer_is_paired`), and the PIN
pairing ceremony is the real proof of identity. The property that DOES matter —
that the peer holds the private key for the cert it presented — is still
enforced: when webpki rejects the cert we re-run exactly that `CertificateVerify`
check with a version-agnostic parser (x509-parser SPKI + the same `rsa` verify
`pairing::verify256` uses). This is a full cryptographic check, never a bypass:
a bad signature still fails, and non-RSA / unsupported schemes fall through to
webpki's original error. Matches Sunshine's leniency without loosening the
pinned trust model.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The desktop clients' clipboard protocol, adopted on Android (text-only v1):
opt-in ClipControl at stream start when the host advertises HOST_CAP_CLIPBOARD
and the new "Shared clipboard" setting (default on) allows.
Device → host: local copies (primary-clip listener + a probe at start) are
announced as lazy format-list offers; the text crosses only when the host
actually pastes (FetchRequest → served from the live clipboard). Host →
device: a host copy's offer is fetched eagerly and lands in the system
clipboard (Android has no practical lazy-paste provider), with an echo guard
so the resulting clip-changed callback doesn't bounce it back as a new offer.
Native side: session/clipboard.rs JNI shims over NativeClient's clip_*
surface; events cross to Kotlin as compact strings from a blocking
nativeNextClip poll on a dedicated thread (the nativeNextRumble pattern),
joined before the session handle is freed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The VK key-event vocabulary cannot express text an input method COMMITS
(autocorrect, gesture typing, non-Latin scripts, emoji). Add a first-class
text event and negotiate it:
- punktfunk-core: InputKind::TextInput (= 15) carries one Unicode scalar per
event in `code`; HOST_CAP_TEXT_INPUT (0x04) in Welcome::host_caps.
- Host advertises the cap only where the session's inject backend can type
text: Windows SendInput (KEYEVENTF_UNICODE, surrogate-pair aware) and the
Linux wlroots backend — a dedicated second zwp_virtual_keyboard whose xkb
keymap grows Unicode keysyms on demand (the wtype model), so keymap
re-uploads never disturb the main device's layout/modifier state. The
KWin-fake-input/libei/gamescope backends can only press layout keycodes, so
those sessions don't set the bit and clients keep the VK fallback.
- GameStream plane: Moonlight's UTF-8 text packet (MAGIC_UTF8, previously
recognized-and-dropped) now decodes to the same TextInput events.
- Android: KeyCaptureView picks a real editable InputConnection when the host
has the cap — the IME runs its full machinery, mirrored to the host live via
common-prefix diffs of the composition (backspaces + new suffix), with
setComposingRegion adopting committed text so autocorrect-revert flows diff
instead of retyping; newline→Enter, deleteSurroundingText→Backspace/Delete.
Older hosts keep the TYPE_NULL raw-key path unchanged.
- keymap: media VKs (0xB0-0xB3) → evdev so the Android media keys land on
Linux hosts too.
Verified: punktfunk-core + host gamestream + pf-inject tests green on Linux
(Ubuntu box), clippy clean; Android app+native builds.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Phase A — hardware mouse: new MouseForwarder routes every SOURCE_MOUSE event
while streaming. Uncaptured (default): hover/drag positions forward as absolute
cursor moves, wheel with fractional accumulation (high-res wheels), all five
buttons incl. back/forward as X1/X2; the local cursor is hidden over the stream
(TYPE_NULL pointer icon) so the host's composited cursor is the visible one.
Captured ("Capture pointer for games" setting, or Ctrl+Alt+Shift+Q live): the
OS pointer is grabbed via requestPointerCapture on the (repurposed) key-capture
view and raw relative deltas forward as mouse-look; held buttons are tracked
and flushed on capture loss/teardown, and the engaging click is swallowed
(desktop parity). Mouse clicks are intercepted before Compose so they can never
be misread as trackpad taps; a mouse back-button's FALLBACK BACK is dropped so
it can't yank the user out of the stream.
Phase B — keyboard: media/consumer keys (play/pause/next/prev/stop) now map to
their VKs instead of falling to the system; TV-remote SELECT maps to Enter.
Phase D — Android TV: RemotePointer turns a D-pad-only remote into a pointer
(Siri-remote analogue): hold SELECT ~0.8s toggles, D-pad glides the host cursor
with ramped acceleration (Choreographer-paced, diagonal-normalized), SELECT
taps click, play/pause right-clicks, holding it toggles the on-screen keyboard,
BACK leaves pointer mode; an overlay hint teaches the vocabulary.
Settings: "Capture pointer for games" + "Invert scroll direction" (applies
to the wheel and two-finger touch scrolling).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Field report: with Exclusive Topology, the physical monitor sometimes stayed
lit — every isolate attempt re-committed and failed SetDisplayConfig with
0x57 ERROR_INVALID_PARAMETER. Clearing DISPLAYCONFIG_PATH_ACTIVE alone leaves
the deactivated path's source/target modeInfoIdx pinned to the queried mode
entries, and per the SetDisplayConfig contract a path being turned off must
have BOTH mode indexes set to DISPLAYCONFIG_PATH_MODE_IDX_INVALID; some
driver/topology combinations (notably when the doomed path shares a source
with the kept virtual path) reject the supplied config outright, so the
retry loop could never converge.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Remote INPUT resets the compositor's idle timers, but a passive
(video-only) viewer sends none — observed live: a SteamOS Game-Mode
host s2idled mid-day and dropped off the network (in a VM with GPU
passthrough it never woke again). Refcounted across planes: the native
LiveSessionGuard and the GameStream video worker each hold a share;
first hold takes the logind sleep:idle block inhibitor on a dedicated
plain thread (zbus blocking must not run on a tokio worker), last drop
releases. Best-effort: no logind → warn once and stream on. No-op off
Linux (macOS/Windows manage their own assertions).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- rebuild-check.sh + punktfunk-rebuild-check.service (enabled, ordered
Before=punktfunk-host): ldd-probes the host binary at session start —
milliseconds when healthy, a full update.sh rebuild only when a
SteamOS update actually broke its library links. update.sh restarts
go --no-block so the check → update.sh → restart chain can't deadlock
against the unit ordering. update.sh retrofits the unit.
- installer summary: web console is https (the unit serves TLS).
- docs: steamos-host.md (runner in the build step, keep-list + auto
rebuild = updates survive hands-free), gamescope.md (Gaming Mode
touch = single-finger pointer, exact taps, no multi-touch),
plugins.mdx (SteamOS runner note), scripts/steamdeck/README.md
(rebuild-check, runner payload, keep list, honest packaging
trade-off + the CI-prebuilt future direction).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The raw-client-pixels fallback is exact only while the managed session
runs at the client's mode. When they diverge — foreign-gamescope
attach at its own resolution, supersample/under-render, transitions —
raw pixels drift or land out of range. The EIS relay file now carries
the session's current output size as a second "WxH" line (from
current_gamescope_output_size()); the injector scales normalized
client positions into it, keeping raw pixels only as the last resort.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The previous commit's no-region fallback never fired: gamescope DOES
advertise a region — (0,0,INT32_MAX,INT32_MAX), meaning "coordinates
are raw". Mapping normalized positions into it explodes a center tap
to x≈1e9, clamped by gamescope to the far corner — the cursor pinned
at (1279,799) and every degraded-touch tap landing there. Treat a
region as an output geometry only when plausibly sized (≤16384 px);
otherwise emit raw client pixels.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
gamescope's "Gamescope Virtual Input" advertises pointer_abs but no
region, so every MouseMoveAbs (and the degraded-touch moves built on
it) was silently dropped (emitted=false) — clicks then landed at a
stale cursor position. With no region, the managed session runs at the
client's mode, so client pixels are output pixels: emit them raw.
Also: log region count/dims at device-resume, and add
PUNKTFUNK_INPUT_TEST_ABS=WxH to `input-test` (corners + center, 1s
apart) so the degraded-touch path is verifiable with xdotool alone.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
gamescope's EIS ("Gamescope Virtual Input") advertises pointer/
pointer_abs/button but never a touchscreen — so in Game Mode every
remote TouchDown/Move/Up was dropped on the floor (Desktop Mode works
via KwinFakeInput, which does touch). Headless KWin's libei has the
same gap. Instead of dropping: degrade to a single-finger absolute
pointer — down = abs-move + left press, move = abs-move, up = release —
synthesized through the normal Mouse* inject paths so region mapping,
held-state tracking, and release_all apply. First finger drives the
pointer; later fingers are ignored (a pinch degrades to a drag). A
touchscreen device appearing later takes over on the next touch —
checked per event, never latched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Verified live on a SteamOS 3.8.14→3.8.16 update: the A/B partset switch
rebuilds /etc and drops everything not on Valve's keep list — the udev
rule (uhid access: virtual pads silently degrade to Xbox 360), the
vhci-hcd autoload (native Deck pad transport), and the UDP buffer sysctl
(back to 208 KB → stutter). Valve's sanctioned extension point is a
drop-in under /etc/atomic-update.conf.d/ — itself on the stock keep
list, so it self-preserves. install.sh §4 and update.sh now register
scripts/punktfunk-atomic-keep.conf there.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- scripts/steamdeck/install.sh (§2b) + update.sh: build and install the
scripting runner user-scoped (~/.local wrapper + pinned bun + bundle) —
SteamOS's read-only /usr can't take the .deb layout, so the console's
plugin store reported "the plugin runner isn't installed" on every
SteamOS host. update.sh retrofits it onto existing installs.
- plugins.rs runner discovery: check the ~/.local layout after the /usr
ones; mention the SteamOS path in the not-installed error.
- update.sh: define warn() — it was used but never defined, so under
`set -e` the first warn aborted the update before services restarted.
- pf-client-core gamepad: a Steam Input virtual pad forwarded on a real
Deck (the only-pad case) now declares the DECK kind in its per-pad
arrival instead of the wrapper's Xbox 360 identity — the session-level
auto_pref already resolved SteamDeck, but the arrival overrides it on
current hosts.
- reject tests: cover SetupFailed; move the foreign-code probe off 0x68.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three fixes for the "most connects simply fail" trap on a SteamOS host
with no physical display (VM testbox, panel-less mini PC):
- restore path: skip the restore-to-physical-panel when no DRM connector
reports a connected display — removing the headless drop-in and
restarting gamescope-session.target on such a box just crash-loops
gamescope and strands every later connect on "no usable compositor".
Keep the headless session (and the takeover state) instead.
- connect path: when no live graphical session exists but the MANAGED
gamescope infra is present (SteamOS gamescope-session / Bazzite
session-plus), route to the gamescope takeover — which rebuilds the
session at the client's mode — instead of failing the connect.
- error delivery: a session-setup failure used to just drop the QUIC
connection, reaching the client as "control stream finished mid-frame"
(indistinguishable from transport trouble). Close with a typed
setup-failed code (0x68, RejectedSetupFailed = -29) carrying the error
text, and give the client handshake/pairing error paths a short grace
for the CONNECTION_CLOSE to beat the stream error it caused.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
pyrowave-sys's build.rs runs cmake to build the vendored PyroWave C library,
but cmake wasn't in the distrobox apt list — so `cargo build -p punktfunk-host`
dies on every SteamOS host with "is cmake not installed?". Add it.
Also move sudo acquisition from step 4 (after the ~15-min build) into preflight
and keep the timestamp warm across the build. Before, a non-interactive run
(`ssh host 'bash install.sh'`) silently skipped the UDP-buffer/udev/vhci-hcd/
input-group tuning AND enable-linger — a "successful" host with no gamepad
passthrough, lossy streaming, and no headless persistence (dies on logout).
Now sudo is prompted once up front (authorize and walk away), and a no-TTY /
no-password run fails loudly before the build instead of silently at the end.
Validated end-to-end on SteamOS 3.8 in a VM with Radeon 780M passthrough:
full host + web build, all planes listening, VAAPI render node, headless
auto-start via linger.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Lifts the ~100 Mbps decrypt ceiling on clients without hardware AES — the
armv7 soft-AES targets (webOS TVs), where AES-128-GCM resolves to fixsliced
software AES + software GHASH (~50-100 cpb) while ChaCha20-Poly1305's ARX
construction runs ~10-17 cpb portable, a 4-7x lift that PyroWave-on-TV needs
(design/chacha20-session-cipher.md).
Phase 1 (core crypto, no wire change): SessionKey merges cipher choice and
key material (invalid combinations unrepresentable, zeroize + redacted-Debug
discipline kept); SessionCrypto dispatches both aead-0.5 ciphers per call —
the salt||seq nonce scheme, per-direction salts, seq-as-AAD and replay
window carry over verbatim (same 96-bit nonce / 16-byte tag, const-asserted).
Config.key becomes SessionKey; validate's zero-key rejection follows the
active variant. The C ABI keeps its fixed 16-byte key mapped to AES — no
ABI_VERSION bump.
Phase 2 (negotiation): VIDEO_CAP_CHACHA20 (0x40) — support-plus-request in
one bit, the VIDEO_CAP_444 precedent. Welcome grows cipher@68 +
key_chacha@69..101, emitted only when non-zero so an AES session's Welcome
stays byte-identical to the pre-cipher form; decode is fail-closed (short
key or unknown id -> Err, never a silent AES fallback). No WIRE_VERSION
bump; downgrade resistance inherited from the pinned-TLS control channel.
Phase 3 (host): grant only when the client advertised the bit and the
PUNKTFUNK_CHACHA20 kill-switch (default on, documented) allows; fresh
32-byte per-session key from the same RNG discipline, legacy key field
stays independently random; resolved cipher logged at session start.
Verification: seal/open suites parameterized over both ciphers + a
cross-cipher tamper case; Welcome roundtrip/truncation/fail-closed tests;
ChaCha lossy-loopback soak (loss/replay is cipher-independent); bench
gains _chacha20 series (AES ids unchanged for CI history) — host-side
sealing line-rate-trivial on both x86 (~640 MiB/s) and Apple Silicon
(~535 MiB/s). punktfunk-probe drives the interop matrix via
PUNKTFUNK_CLIENT_CHACHA20=1 and logs the negotiated cipher.
Phase 4 (pf-webos pin bump + unconditional cap bit) follows the next
core release.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The CI workflow assembled the store zip by hand and silently drifted from
scripts/package.sh when 60d46530 added assets/ and controller_config/ — every
published build (canary and stable through 0.17.1) was missing both, so the
Steam shortcut a fresh install creates had no images and no controller layout
(shortcut_art() found nothing to send; apply_controller_config failed with
template-missing). CI now stages through scripts/package.sh — the plugin file
list lives in one place — and a backstop verifies the runtime-loaded pieces are
in the zip before publishing.
Frontend hardening for installs poisoned by those builds: applyArtwork stamped
its per-appId ART_VERSION marker even when zero assets were applied, so
shipping the files alone would never heal existing shortcuts — only mark done
when something actually landed, and bump ART_VERSION 2→3 to force one re-apply.
Same guard for the controller-config marker (ok with zero account configset
dirs is not done), plus one deferred artwork retry for the fresh-AddShortcut
registration race setShortcutHidden already defers around.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The PUNKTFUNK_PIN_CLOCKS clock pin (AMD power_dpm_force_performance_level=high / NVIDIA nvmlDeviceSetGpuLockedClocks) was armed once at host start and released only at host exit, so the box's clocks stayed pinned the whole time the host ran — even with no client connected.
Move the pin behind a box-wide refcount (gpuclocks::session_pin()) shared across both streaming planes (native + GameStream): it arms on the first live client and releases on the last disconnect. on_host_start() now only installs the process-scoped NVIDIA P2-cap application profile. No behavior change when PUNKTFUNK_PIN_CLOCKS is unset.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Launching from the Decky plugin against a sleeping host fired the
Wake-on-LAN packet and dialed immediately — but the dial was a single
quinn attempt that gave up after the transport's 8 s idle window, far
shorter than a suspend-to-RAM resume. The host woke to find the client
already gone (its retransmitted Initials show up host-side as
"QUIC accept failed error=timed out"), the stream shortcut exited, and
the launch looked cancelled.
- punktfunk-core: dial in a loop until the embedder's connect budget is
spent — short 3 s attempts keep the Initial cadence dense, so the
connect lands within ~a second of the host's network returning. Only
silence is retried: pin mismatches, typed rejections, and any real
answer surface immediately, and the shutdown flag stops the loop.
- decky: stretch the budget to 75 s when a magic packet actually went
out (PF_CONNECT_TIMEOUT → --connect-timeout via the wrapper), so the
retry window covers the resume; an awake host still connects in <1 s.
- host: a clean close before the control handshake is a reachability
probe (hosts-page online pips), not a failed session — log it at
debug instead of WARN "session ended with error", which buried the
real failures in the reporter's log.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The system-tuning block (UDP buffers, gamepad udev rule, vhci-hcd autoload,
input group) was gated on `sudo -n true` — passwordless sudo — so on a stock
SteamOS box (the deck account needs a password) the whole block was silently
skipped: /dev/uhid stayed root-only so the gamepad degraded to an Xbox 360 pad,
and the UDP buffers stayed at the 416 KB default. Video was unaffected because
Game Mode/gamescope needs none of it, which masked the skip on-glass.
- install.sh + update.sh: acquire sudo interactively (`sudo -v` prompt on a TTY)
instead of requiring passwordless; skip only when there's no TTY or auth fails,
and then print the exact manual block plus a `passwd` hint — a stock SteamOS
'deck' account has no password, so sudo can't work until one is set.
- update.sh: also retrofit the UDP-buffer sysctl for older/skipped installs, and
loudly nag to reboot when the input group was just added (a --user restart does
not pick up the new group; only a fresh login does).
- steamos-host.md: note this step prompts for the sudo password.
Root is unavoidable: the udev rule, input group, and vhci-hcd module are all
kernel operations with no user-space alternative.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A fresh SteamOS host install failed on-glass two ways: KWin denied Desktop-mode
capture (zkde_screencast_unstable_v1) so every session died in the pipeline build,
and the native Steam Deck pad degraded to an Xbox 360 controller. Both stem from
setup that only a fresh login applies — the host user-service keeps the
pre-install group set and KWin reads its .desktop grant at session start — plus a
missing vhci-hcd autoload the pad's USB/IP transport needs (the deb/rpm/arch
packages ship it; the Deck installer did not).
- install.sh: install punktfunk-modules.conf (+ modprobe vhci-hcd now) for the
native Deck controller transport, seed the KDE RemoteDesktop grant
(kde-authorized) for Desktop-mode input, and print a loud "reboot before
streaming" notice when a relogin is actually required (input group added or the
.desktop grant first installed).
- update.sh: retrofit the udev rule, vhci-hcd autoload, input group, and grant so
existing installs pick them up on the next update without a reinstall.
- steamos-host.md: document the first-install reboot and the native controller
passthrough requirements (input group + vhci-hcd + client-side Steam Input Off).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The VirtualHere DualSense recipe linked the `virtualhere-dualsense.ts` SDK
example as a deployable recipe but skipped everything needed to run it: that
VirtualHere is a server (couch) + client (host) pair, and that the example —
like every example — imports `../src/index.js`, which only resolves inside the
SDK repo. A user copying it out had no way to know they must
`bun add @punktfunk/host` and swap that import.
- automation.md: rewrite the recipe into a full walkthrough — the two-sided
VirtualHere setup, the `-t` verbs, the zero-code hooks version, and a new
scripted section with exact deploy steps, env vars, and running it as a
service (its own SIGTERM handler makes `systemctl stop` release the pad).
- sdk/README.md: say how to run an example in-repo vs deployed on a host.
- virtualhere-dualsense.ts: header note on the import swap + service setup,
since that file is where the doc link lands.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Task 1 (latency): the transactional present now commits from the RENDER
thread inside an explicit CATransaction + flush() instead of hopping to
main. The present harness (2026-07-21, 240 Hz Studio, full-size window)
measured the main hop batching presents at runloop-iteration rate when
main is busy — an active implicit transaction NESTS the explicit one —
which is the field's presents=55 @ fps=240 / display_p50 18.6 ms.
Off-main commits measured immune to main churn: glass p50 ~10 ms,
cadence a clean 4.17 ms at 240. flush() is load-bearing: without it the
render thread's own implicit transaction (drawableSize/colour writes)
swallows the explicit commit and NOTHING reaches glass — the harness
reproduced the exact freeze (every present dropped). Legacy main-hop
kept as PUNKTFUNK_TXN_PRESENT=main for field A/B. New pf-windowed
Console line (subsystem io.unom.punktfunk, category presenter)
decomposes the issue side per second.
Task 1 lever D: the f407f418 IOSurface contents-swap path is
resurrected format-aware as WindowedPresentMode.surface — bgra8 SDR /
rgba16Float+PQ-tagged HDR pool, swap committed off-main from the
completion handler; EDR anchored by the metal layer underneath.
Env-only prototype (PUNKTFUNK_WINDOWED_PRESENT=surface) until the HDR
composite is eyeballed on glass.
Task 2 (setting): punktfunk.windowedSafePresent (default ON =
transactional mitigation; OFF = the fast async path whose panic the
saga is about — the caption says so in plain words). Rows in the touch
Settings (Presentation) and gamepad settings, macOS-only.
PUNKTFUNK_WINDOWED_PRESENT=async|transaction|surface overrides for dev
A/B. maximumDrawableCount stays 3 — the harness measured 2 starving the
render loop (172/s) and worsening glass p50.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Swift `Int` is 64-bit (Foundation infers %lld); the stats mirror's format
string used %d (32-bit C int) for fps/presents/lost. macOS 26's strict
String(format:) validator rejects the mismatch and drops the whole line
(the error also cascades onto the float args, mis-blaming floor_p50). Use
%lld for the three Int fields. Pre-existing, unrelated to the DCP work;
surfaced while reading presenter logs during the panic fix.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The windowed-macOS "mismatched swapID's" @UnifiedPipeline.cpp kernel
panic is the CAMetalLayer ASYNC image queue (commandBuffer.present,
mandatory with displaySyncEnabled=false) diverging from WindowServer's
compositor on a high-refresh composited session — it survived glass
pacing (PyroWave, 2026-07-18) and hit windowed HEVC on the same 240 Hz
Mac Studio (2026-07-21), so it's the async queue itself, at any codec
or pacing.
f407f418's mitigation routed windowed PyroWave through a BGRA8 IOSurface
layer.contents pool, which cannot carry HDR — extending it to HEVC would
have tone-mapped windowed HDR down to SDR. Instead, present the drawable
through a Core Animation transaction (CAMetalLayer.presentsWithTransaction)
for every windowed session: commit, waitUntilScheduled, then drawable.present()
inside a CATransaction, so the swap commits with the layer tree and stays
in lockstep with the compositor instead of racing it. The full
rgba16Float / PQ / EDR render path is untouched — windowed HDR is real
HDR again. Fullscreen keeps the async path (direct scanout, lowest latency,
no panic there).
Removes the IOSurface surface pool, its surfaceLayer sibling, and the
macOS windowed PQ->SDR tone-map (tvOS keeps its no-headroom tone-map).
Net -150 lines. swift build + 169 tests green. Needs on-glass soak on
the 240 Hz Mac Studio (kernel race — only real hardware confirms).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Flip both zero-copy levers from opt-in to default:
- The capture negotiation now PREFERS gamescope's producer-side NV12 pod by
default (PUNKTFUNK_PIPEWIRE_NV12=0 restores the packed-RGB negotiation).
Codec-aware gating rides a new OutputFormat::nv12_native ->
ZeroCopyPolicy::native_nv12_session edge resolved by the host facade from
the session plan (pf_encode::linux_native_nv12_ok): only H265/AV1 sessions
whose backend can open the raw Vulkan Video encoder ever see the NV12 pod
-- an H264/Moonlight session (libav VAAPI, which would misread the
two-plane buffer) keeps today's BGRx negotiation, as do the
GameStream-resolve and portal-mirror paths, PyroWave, and NVENC prefs.
- RGB-direct's unaligned modes default to the true-extent direct import
(PUNKTFUNK_VULKAN_RGB_TRUE_EXTENT=0 restores the padded-copy staging).
Guarded-tested on Van Gogh with the kernel journal watched: clean, and at
5.38 ms p50 the fastest 1080p encode path measured on that hardware. The
EFC only exists on Mesa >= 26, where the codedExtent-driven session_init
padding is guaranteed (verified back to Mesa 24.2).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Opt-in alternative to RGB-direct's padded-copy staging at unaligned modes:
direct-import the visible-size capture and declare the TRUE-SIZE source
codedExtent. RADV has derived the VCN session_init firmware padding from
srcPictureResource.codedExtent since Mesa 24.2, so the EFC is told the source
lacks the alignment rows and the hardware edge-extends them internally --
which also reframes the 2026-07-20 field GPU reset: that crash passed the
ALIGNED extent as the source codedExtent (firmware padding zero) over a
visible-size buffer. Off by default until a guarded live test proves the EFC
front-end honors the padding like the YUV fetch path does; the padded-copy
staging remains the shipped behavior.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Drop the padded-copy staging for producer-native NV12 and direct-import the
visible-size buffer at unaligned modes (1080p) too. Safe by the driver's own
contract rather than by allocation luck: native sessions now author the H265
SPS / AV1 sequence header at the RENDER size and pass the matching codedExtent
on every picture resource. RADV rounds the bitstream SPS up itself (with a
conformance window -- radv_video_patch_encode_session_parameters, per the
VK_KHR_video_encode_h265 proposal's "implementations may override" clause) and
programs the VCN session_init with the true extent plus nonzero firmware
padding, so the hardware edge-extends the alignment rows internally and never
fetches past the source's real extent -- the exact mechanism VAAPI/radeonsi has
always used to encode 1080p. The driver-emitted header NALs already carry the
patched SPS, so the wire format is unchanged.
The CSC and RGB-direct paths keep the app-aligned convention (their sources
genuinely cover the aligned extent); RGB padded-copy staging is untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Software decode uploads swscale RGBA with no CSC/tonemap pass. On the new
desktop-compositor HDR10 surfaces (GNOME 48 / Plasma 6 + Mesa >= 25.1 offer
ST2084 even on SDR desktops) that sRGB-encoded content was composed as PQ —
the field-reported psychedelic picture (Fedora clients decode HEVC in
software because stock Mesa strips it; reproduced + fix live-validated on
GNOME Wayland: mode-0 soup -> HDR→SDR washed-out-but-correct).
The CPU lane keeps the SDR swapchain and its known untonemapped-PQ gap; a
real PQ→sRGB pass for CPU frames is the follow-up.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
SDL_APP_ID = io.unom.Punktfunk — compositors match it against the shipped
.desktop for the window icon; without it KWin/mutter show the generic Wayland
icon (field-noticed on the Deck). Linux analog of win32 AppUserModelID.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
GNOME 48 / Plasma 6 with Mesa >= 25.1 now offer HDR10_ST2084 surfaces even on
SDR desktops, silently flipping PQ streams into the mode-0 passthrough lane —
previously unreachable on desktop compositors and never validated there. The
hatch pins the shader tonemap for diagnosis and as the field escape.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The native QSV encoder only attached mfxExtVideoSignalInfo on HDR sessions —
an SDR stream carried no colour description at all (NVENC writes 709-limited
unconditionally; qsv.rs now mirrors it).
Diagnostics for the field-reported blue/magenta Intel-host colours:
- hdr-p010-selftest takes an optional WxH + GPU vendor (intel|nvidia|amd) and
prints which adapter it tested — 64x64 on the default GPU proved nothing on
dual-GPU boxes, and the field heights (1080/1400) are not 16-aligned.
- pf-capture: ignored live test running the converter at 1920x1080 pinned to
the Intel adapter.
- pf-encode: qsv_live_p010_1080_colorbars_dump — known P010 bars through the
UNALIGNED-height ingest copy (1080 src -> align16 1088 pool, the seam no
640x480 test exercises) to Main10 HEVC, dumped for off-box decode checks.
- dxgi.rs: drop the stale 'falls back to the R10 path' claim (no such path).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
With PUNKTFUNK_PIPEWIRE_NV12=1 (bring-up gate), the PipeWire negotiation
offers an NV12 LINEAR DMA-BUF pod (BT.709 limited pinned MANDATORY) ahead
of the BGRx one, and gamescope's producer-side RGB->YUV pass replaces the
host CSC entirely: the encoder imports the two-plane buffer as a profiled
VIDEO_ENCODE_SRC image and the VCN encodes it directly. Contributed
measurements: encode p99 2.9 ms (from ~4-4.5 ms via EFC RGB-direct),
60 fps capture, 0 send drops.
Hardening on top of the contributed patch:
- unaligned modes (1080p!) stage through a padded aligned NV12 copy (edge
rows/columns duplicated, transfer-only) instead of direct-importing the
visible-size buffer -- a direct import would make the VCN read past the
producer allocation, the exact OOB class behind the 2026-07-20 field GPU
reset; the encode extents return to the aligned coded extent everywhere
- the UV plane layout honors the producer's plane-1 chunk (offset/stride)
when the SPA buffer carries one (same-BO verified by inode), with the
contiguous-plane contract as fallback
- PyroWave sessions are excluded from the gate (their Vulkan compute CSC
ingests packed RGB), and a native-NV12 session that resolves to libav
VAAPI (H264 codec, PUNKTFUNK_VULKAN_ENCODE=0, feature off, or a failed
Vulkan open) refuses at open instead of streaming garbage chroma
- pad staging images carry TRANSFER_SRC (the width-padding pass self-copies
the staging image -- previously missing on 1366-wide modes)
- metadata-cursor one-shot warn (parity with RGB-direct) and a padded-NV12
PUNKTFUNK_PERF split label; the padded RGB copy gets its timestamps too
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
settle_desktop_portal pushes the live desktop's WAYLAND_DISPLAY into the
systemd --user manager (import-environment) so a re-activated portal inherits
the right session — but the import PERSISTS after that desktop dies. Every
later user unit then inherits the stale socket, including
gamescope-session.target: gamescope sees WAYLAND_DISPLAY, runs NESTED, and
aborts with "Failed to connect to wayland socket: wayland-0" — observed live
on a Deck, where Game Mode could not start at all (autologin crash-looped)
until the var was unset by hand.
Two-layer fix, mirroring the protection launch_session already gives its
transient unit:
- observe_session_instance: when a desktop compositor instance goes away,
unset WAYLAND_DISPLAY/DISPLAY from the manager env (a desktop bounce is
harmless — the next portal settle re-imports).
- write_steamos_dropin: UnsetEnvironment=DISPLAY WAYLAND_DISPLAY on the
takeover drop-in, unit-scoped belt-and-suspenders for host-initiated
restarts.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Observed live on a Deck host: the user switched Game Mode → Desktop, KDE came
up in 0.8 s — and the capture-loss rebuild's FIRST detection ran inside that
gap, still said Gaming, and its gamescope re-acquire restarted
gamescope-session.target. On SteamOS that steals the seat: the user was
yanked out of the KDE session they had just chosen, the two session managers
fought (job canceled / dependency failed), no gamescope node appeared within
30 s, and the stream died at the 40 s rebuild budget.
A rebuild probe acting on possibly-stale detection must never stop, relaunch,
or take over box sessions. New pf_vdisplay::rebuild_probe_scope (RAII,
counted): while held, the gamescope managed/takeover create paths only attach
to a live node and fail fast otherwise — no stop_autologin_sessions, no
session-plus relaunch, no gamescope-session.target restart. The capture-loss
loop holds it for the first 4 s after a loss (the detection-ambiguity
window); after that, destructive rebuilds are allowed again, so a genuine
switch INTO Game Mode still gets its headless takeover. Probe failures are
instant, so the loop now paces itself at ~2 Hz instead of spinning.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two session-transition stalls found live on a SteamOS Deck host, one cause:
the pipeline retry loop couldn't tell a wait-it-out failure from a
retry-now-and-it-works one.
- Gamescope first-frame race: a PipeWire stream connected while gamescope
re-initializes its headless takeover negotiates a format, reaches
Streaming, and never receives a buffer — while a fresh connect delivers
within ~0.5 s. Every gamescope bring-up ate the full 10 s first-frame
budget on attempt 1 (17 s bring-ups; KWin: 1.2 s). The retry loop's FIRST
attempt now waits 2.5 s (Capturer::next_frame_within), so the reconnect
that fixes the race happens at ~3 s. Later attempts keep the patient 10 s
— the documented 30-60 s Big Picture cold start still fits the budget.
- Capture-loss rebuild vs session switch: the rebuild loop re-detects the
active session between build_pipeline_with_retry calls, but each call
burned 8 attempts (~13 s) against a compositor that no longer exists (a
Desktop→Gaming switch spent 13 of its 27 s retrying gone-KWin). The
capture-loss path now passes max_attempts=2, turning the outer loop into
~1 s detect-and-retry cycles that follow the box to the new session.
The resize path's direct build_pipeline call keeps the default budget (no
retry wrapper there to absorb an early bail).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The SteamOS scripts built the host featureless, so PlatformAuto could never
pick the Vulkan Video backend (default-on since 84329205) and every session
silently ran libav VAAPI — unlike the deb/arch/rpm/sysext builds, which all
pass the feature. Three gaps, found live on a Deck host:
- install.sh/update.sh: cargo build with --features punktfunk-host/vulkan-encode
(matches the packaged builds; pure-Rust ash, no new system dep).
- host.env: RADV_PERFTEST=video_encode — Van Gogh RADV still gates
VK_KHR_video_encode_* behind it, so without it the backend can't open and
falls back to VAAPI. Harmless where encode is exposed by default.
- io.unom.Punktfunk.Host.desktop (Exec rewritten to this install's binary):
KWin only grants zkde_screencast_unstable_v1/org_kde_kwin_fake_input to an
exe a .desktop authorizes — without it Desktop-Mode capture fails outright
and a mid-stream Game↔Desktop switch strands the stream on the old backend.
update.sh retrofits the last two idempotently so existing installs get them
without a reinstall.
Validated on a SteamOS 3.8.14 Deck (Van Gogh): Vulkan HEVC opened on both the
KWin and gamescope paths at 5120x1440 — in-place ABR reconfigure (no IDR, no
encoder rebuild) vs the VAAPI path's full teardown per ABR step, and Desktop
capture came up without a re-login.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
`<CardContent className="p-0">` never removed the padding it looked like it
removed: tailwind-merge resolves conflicts within a variant, so `p-0` cancels
`p-4` but leaves `sm:p-6` standing. Every call site that used it nests a
CardHeader inside, which brings its own `sm:p-6` — so at >=640px the title sat
24px further in than its neighbours'. Visible on the Plugins page as 'Installed
plugins' not lining up with 'Plugin runner'.
Give CardContent a `flush` prop that omits the padding outright, and use it at
the three call sites (Store, Pairing, Stats) rather than fixing the symptom
three times and leaving the footgun in place.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The guard checked the package NAME shape, and `@punktfunk/plugin-kit` — the
framework every kit-built plugin depends on — matches `@scope/plugin-*`
exactly. Windows on-glass accepted an uninstall of it; bun no-ops removing a
non-dependency so nothing broke, but the store was offering to pull a library
out from under the plugins using it.
Require membership in the top-level dependency list, the same authority that
already keeps plugin-kit out of the installed listing. Same blind spot, two
call sites — the listing was fixed during implementation, this one wasn't.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two things looked wrong in the store:
The cards had no space above the icon. `CardContent` zeroes its own top padding
(`pt-0`/`sm:pt-0`) because it normally sits under a `CardHeader` that supplies
it — but these cards have no header, so the top padding simply vanished while
the other three sides kept `p-card`. `p-card` does not undo it: `card` is a
custom `--spacing-*` token, which tailwind-merge does not recognise as a
spacing value and so never dedupes against `pt-0`, leaving the longhand to win.
Both header-less cards in this section now restore it explicitly, at both
breakpoints.
Platform chips read "windows" / "linux". Those are the catalog's platform
IDENTIFIERS (the index validator pins `linux | windows | macos`), which is right
for the wire and wrong on screen; the card now maps them to display names —
Linux, Windows, macOS. Proper nouns, so deliberately not routed through i18n.
Adds Store/StoreCard stories covering the padding, the platform labels, and the
installed / update / incompatible / blocked / external states, so the card can
be eyeballed without a host or a catalog.
Gates recorded (nvenc-subframe-slice-output.md Phase 3): loss-harness curve
identical streamed-vs-whole; adversarial security review clean; live .21
3-leg A/B — streamed 0 corrupted frames, p99 8527→5363 µs vs sliced-whole,
first_slice_us ~500 µs of a ~1250 µs encode reaching the send thread; wire
bytes rate-governed under CBR + 1-frame VBV (the ~1-2 % slice-header cost is
absorbed by rate control, not added to the wire). GameStream: plane has zero
code diff (own RTP packetizer; shares only untouched send_pacing); a live
Moonlight re-test joins the standing owed Moonlight item.
- resolve_slices(codec, default): env 1..=32 wins (1 = the NEW explicit
single-slice escape), else the backend default — 4 on Linux direct-NVENC,
1 elsewhere. resolve_subframe(default_on): PUNKTFUNK_NVENC_SUBFRAME
tri-state (0 = never, 1 = force, unset = backend default).
- Linux nvenc_cuda: resolves both ONCE in query_caps — the sub-frame default
is gated on the GPU's SUBFRAME_READBACK cap so an unsupporting GPU never
has enableSubFrameWrite forced into its init params — and feeds the same
resolved values to build_config, build_init_params (open AND in-place
reconfigure) and the chunked-poll latch.
- Windows: env-only as before (async path untouched, byte-identical config
for unset env); LowLatencyConfig.slices + the explicit subframe param
replace the shared env reads.
- Tests: chunked e2e now runs at the DEFAULTS (no knobs); the fallback test
becomes the escape test (SLICES=1 disarms; SUBFRAME=0 disarms while
4-slice encode continues on the plain poll path).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Slipped through 421e8112: the on-box clippy runs covered punktfunk-core and
punktfunk-host but not the tools crates; CI's workspace-wide pass caught it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The bootstrap key generated during implementation was lost, and its
replacement was pasted into the wrong repository's secret store — Gitea
secrets neither cross repos nor read back, so neither is usable. The private
half of this one lives in punktfunk-plugin-index's own INDEX_SIGNING_KEY,
which is the only thing that signs the catalog.
Nothing has shipped pinning any of these, so this is a straight replacement
of slot 0; the second slot stays reserved for a genuine rotation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
B2 default-on (design/vulkan-rgb-direct-encode.md): wherever the probe
passes, Vulkan Video sessions now take the EFC RGB source by default —
direct import on aligned modes, padded-copy staging on unaligned ones
(1080p). PUNKTFUNK_VULKAN_RGB_DIRECT becomes the override: =0 disables,
=1 forces, unset = the default below.
The one session class that must NOT default on: cursor-as-metadata
captures (every non-gamescope compositor), where the CSC shader's blend
IS the visible pointer — the EFC cannot composite, and defaulting there
would silently drop the cursor from the stream. The hint rides the
existing plumbing:
- SessionPlan gains cursor_blend, resolved once where the compositor is
known (gamescope embeds the pointer itself → false; kwin/mutter/
wlroots/hyprland → true), and shows up in the logged plan line.
- open_video/open_video_backend thread it through (native pump: all
three encoder-open sites read plan.cursor_blend; GameStream monitor
capture: true — it negotiates metadata cursor; spike: false).
- VulkanVideoEncoder::open resolves: env override, else ON iff the
session never hands us cursor bitmaps. The warn-once for a cursor on
an RGB session (forced via =1) stays.
Verified on-hw box (Linux): pf-encode + punktfunk-host compile, clippy
clean, unit suite green. The GPU paths themselves are unchanged from the
smoke-validated 96e19986 — this commit only changes which sessions
select them by default.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Review verdict: no exploitable memory-safety / amplification / splice /
resurrection issue; nonce order and two-lane seal byte-identical for the
legacy path. Actionable findings applied:
- explicit per-block data_shards cross-check next to the recovery_shards
one (defense-in-depth — the geometry invariants make it unreachable
today, but have_data indexing assumes it)
- partial delivery skips an UNPINNED streamed frame (frame_bytes still 0)
instead of truncating its max-sized buffer to an empty "partial"
- regression tests for the two untested load-bearing guards: the
final-first out-of-range-sentinel drop (the exact-sized-buffer guard)
and second-final-with-different-totals rejection
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Phase-2 gate: more, smaller wire units must not regress FEC recovery.
Adds a GF16 streamed column (three chunk pushes + finish per AU — sentinel
blocks then real totals). Measured: identical to whole-AU at every loss
rate (50/50 through 1/6, the same 25%-budget cliff).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The probe's reassembler is the shared core one, and the probe is the tool
that measures the streamed-AU overlap win — advertise the cap so a host
with chunked encode streams to it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The wire half of sub-frame slice output (latency §7 LN1, planning
design/nvenc-subframe-slice-output.md Phase 2): toward a client that
advertises the new cap bit (0x20), a chunked-poll encoder session ships each
AU's completed FEC blocks while the tail of the frame is still encoding —
the AU's last packet leaves the host as the encode finishes instead of
after it.
Wire semantics (negotiated; zero change for anyone else):
- Non-final blocks ride SENTINEL headers: block_count = 0 (a value no legacy
sender emits) + frame_bytes = 0 + exactly max_data_per_block data shards,
so the receiver's shard-offset formula needs no total.
- The final block's headers carry the real frame_bytes/block_count (+
FLAG_EOF) and RETRO-VALIDATE the whole frame: totals under which a
received sentinel block is out of range or not full-K kill the frame
wholesale (no spliced delivery) and the index can't be resurrected.
- Firewall: sentinels are bounded by the negotiated limits (full-K exactly,
never the last block the limits allow, no total to lie about); the exact
derived-geometry check runs unchanged on every non-sentinel packet and
retroactively at pinning. Sentinel opens commit a max_frame_bytes buffer,
bounded by the existing IN_FLIGHT_BUF_FACTOR budget (amplification test).
- Order-agnostic like legacy: a reversed frame (final block first) opens
legacy-shaped and still accepts its sentinels against the pinned totals.
- Small/empty streamed AUs degenerate to byte-identical legacy headers.
Host: Packetizer::{begin,push,finish}_streamed seal full-K blocks (data +
parity per block) as chunks arrive; Session::seal_streamed_* share the
pooled-wire + two-lane seal machinery via the new seal_run; the send thread
paces each flush under the frame's existing deadline (pace_sealed split out
of paced_submit) and runs the whole-AU accounting at the last chunk; the
encode pump forwards poll_chunk output as ChunkMsg when the client has the
cap AND the encoder chunks (re-queried per AU — an escalation falls back
seamlessly). Probes never run mid-AU. PUNKTFUNK_STREAMED_AU=0 = host escape
hatch. Client core ORs the cap into Hello (the shared reassembler carries
the support). Sampled first_slice_us vs encode_us PERF log measures the
overlap; the 0xCF stage-field extension stays a follow-up.
Core tests: streamed round-trip (clean/loss/reorder/duplicate, both orders),
sentinel firewall bounds, lying-final wholesale kill + no-resurrect,
open-amplification budget, header-shape pins. Gates still owed before
default-on: security review pass, loss-harness curve, GameStream smoke
(plane untouched structurally), bitrate A/B.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The encoder half of sub-frame slice output (latency §7 LN1, planning
design/nvenc-subframe-slice-output.md Phase 1): with PUNKTFUNK_NVENC_SLICES=N +
PUNKTFUNK_NVENC_SUBFRAME=1 on a sync depth-1 session, Encoder::poll_chunk hands
the in-flight AU out as slice-boundary chunks read through doNotWait sub-frame
locks while the tail is still encoding — the readback loop the on-hw probe
validated (~200 µs slice spacing on the 5070 Ti), productionized.
- codec.rs: AuChunk (AU metadata on the first chunk, `last` closes the AU;
chunks concatenate to exactly the poll() bytes) + supports_chunked_poll /
poll_chunk trait surface. Default impl wraps poll() as one self-closing
chunk, so a chunk-driven consumer works against every backend.
- nvenc_cuda: chunked readback cut at slice boundaries only (bitstream size at
n reported slices = end of slice n, Annex-B contiguous); completion is NEVER
numSlices alone — one finishing BLOCKING lock is the authority and the wedge
watchdog, so the final chunk blocks exactly like sync poll (the depth-1 pump
contract; 6dc195f9 bug class). Keyframe on early chunks is the submit-time
IDR prediction (exact under P-only/infinite-GOP), cross-checked at finish.
Debug builds shadow-check emitted chunks against the finished AU prefix.
Mutually exclusive with pipelined retrieve (gated off when async_rt exists,
dropped by the escalation rebuild); composes with stream-ordered submit.
- nvenc_core: slices_env/subframe_requested shared parses so the config author
and the chunked-poll arming can't disagree.
- TrackedEncoder forwards both new methods (the set_wire_chunking trap class).
Host loop untouched — Phase 2 (VIDEO_CAP_STREAMED_AU seal/send) consumes this.
On-hw: nvenc_cuda_chunked_poll_end_to_end + nvenc_cuda_chunked_poll_fallback_whole_au.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Every session encoder is boxed in TrackedEncoder (open_video), and the
wrapper never forwarded Encoder::set_pipelined — so the host loop's
contention escalation (ae673158) hit the trait default, which returns
false ("backend can't pipeline, stop asking"), and the pipelined-retrieve
stage could never engage adaptively. Only the explicit
PUNKTFUNK_NVENC_ASYNC=1 open-time path worked. The exact defaulted-method
trap the wrapper's own comment documents for set_wire_chunking.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Console: a Plugins section (Browse / Installed / Sources) on a static nav
entry, with install friction proportional to trust — a plain confirm for a
verified entry, a warning naming the curator for an external one, and a
danger dialog that makes you retype the spec for a raw package. Tier badges
are permanent and follow the plugin onto its own UI page.
Index: unom/punktfunk-plugin-index published and served from Gitea's raw
endpoint (real HTTPS, byte-exact, no vhost to stand up) — merge to main is
publish, which resolves the design's open hosting question.
Four things only running it could find:
- runner discovery matched @punktfunk/plugin-* only, so a third-party
scoped plugin (which D8 requires) would install and never run
- ...and that convention also matches @punktfunk/plugin-kit, a plugin's
own framework: it listed as installed and would have been imported as a
unit. Both now key off the plugins dir's top-level dependencies, with an
emptied dependency list meaning 'nothing installed' rather than falling
back to the naming convention
- the store must not pass new flags to the runner: the scripting package
ships separately and an older one reads an unknown flag's value as a
package name. The host writes the bunfig scope mapping itself
- ureq reports only >= 400 as Err, so a conditional request's 304 arrives
as Ok with an empty body — handled as an error it made every refresh
after the first verify a signature over zero bytes and sit stale
Also: the console's first Tabs use exposed an @unom/ui theme gap that
rendered inactive tabs invisible (caught in a browser pass, not by types).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The index is the verification gate: a catalog entry pins one exact version
plus that version's tarball integrity hash, so 'verified on every release'
is a property of the data rather than a promise about process. Nothing can
express 'track latest' for a catalogued plugin.
- store/index.rs signed index parse + ed25519 verify (ring), validate-and-drop
per entry, semver minHost/advisory ranges
- store/sources.rs built-in unom source (compiled-in URL + two key slots for
rotation) + operator sources in plugin-sources.json
- store/catalog.rs https fetch with size/timeout/redirect caps, signature before
parse, last-good disk cache (stale-but-usable when offline)
- store/jobs.rs single-flight install/uninstall: registry-integrity preflight
against the pin, spawn the runner CLI with live log capture,
post-install version check with rollback, provenance record,
runner restart (discovery is startup-only)
- store/manifest.rs install provenance; absence means CLI-installed
- mgmt/store.rs 12 routes under /api/v1/store, denied to the plugin token
(a plugin that can install plugins is an escalation primitive)
Also generalizes runner discovery and listInstalled from @punktfunk/plugin-*
to ANY scope's plugin-*: catalog entries must be scoped so the scope can map
to that entry's registry, so a third-party plugin necessarily arrives under
its own scope and would otherwise install but never run.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Lifts 3aacec53's alignment gate into a mode select: an aligned mode keeps
the true zero-copy direct import; an unaligned one (1080p!) now blits the
visible frame into a per-slot ALIGNED BGRA staging image and duplicates
the edge rows/columns into the 64x16 padding — transfer-only regions in
one vkCmdCopyImage (1080p: visible + 8 row regions; width padding adds a
second self-copy pass in GENERAL), no compute shader. The encode reads
the staging image, never the capture buffer, so the EFC can never read
past a producer allocation (the field GPU hang). Still one ~8 MB copy vs
the CSC path's ~17 MB + dispatch + plane copies. Verdict line:
active(padded-copy). The staging import drops the video profile entirely
(TRANSFER_SRC only).
CPU-payload paths made honest on the way (they were the smoke baseline
AND the software-capture fallback):
- rgb mode: the staging upload is padded CPU-side (edge duplication) so
the aligned encode-src is fully defined;
- CSC mode: the sampled image is now SOURCE-sized with a matching copy
extent — the old aligned-size image + tightly-packed buffer sheared
rows and left garbage rows at unaligned modes (black-bar artifacts;
YMIN=16 in every smoke frame), which also masked as a 24 dB PSNR
'regression' against the (correct) padded output.
On-glass (780M, host Mesa 26.0.4): all four smokes pass at 256x256
(direct) AND 250x250 (padded); padded-EFC frames decode perfectly
uniform (YMIN==YMAX) at the exact 709-narrow values (79/148/60 for the
first three fills); CSC-vs-padded PSNR 49.9 dB avg after the baseline
fix. clippy clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Field report (commit 25624443, RDNA2/VCN3 host, 1080p): repeated VCN0 VM
protection faults from punktfunk-host, vcn_enc_0.0 ring timeouts, two ring
resets, then a failed reset escalating to a full MODE1 GPU reset with VRAM
loss. Root cause is B1's RGB-direct source: the session's coded extent is
64x16-aligned (1920x1088) but the captured dmabuf is only the real mode
size (1920x1080) — the VCN EFC reads the 8 padding rows past the end of
the buffer (8 x 7680 B = 61 KB, matching the faulting page span exactly).
The CSC shader absorbs alignment by clamping reads and duplicating edges;
RGB-direct has no such stage. Every prior validation surface happened to
be aligned (256x256 smoke, 1440p live box) — 1080p is the common mode
that is not. The stall watchdog's rebuild-and-refault loop then turned a
per-session fault into the GPU death spiral.
Two layers:
- Open-time gate: RGB-direct engages only when the real mode equals the
aligned coded extent (720p/1440p/4K yes, 1080p no); otherwise the CSC
path runs and the verdict line says why (unaligned-mode). The padded-
copy variant that re-enables 1080p is design-doc B2 work.
- Submit-time check: an RGB-direct frame that does not cover the coded
extent errors out (encoder-rebuild path) instead of importing an
undersized source.
On-glass (780M, host Mesa 26.0.4): all four smokes pass aligned; at
PF_SMOKE_W/H=250 the gate refuses RGB-direct and the rgb smokes soft-skip.
clippy clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
CardContent is shadcn's header-adjacent variant (`p-4 pt-0 sm:p-6 sm:pt-0`),
but the four Live-status tiles use it with no CardHeader. tailwind-merge lets
the call site's unprefixed `p-4` cancel the base `pt-0`, yet nothing cancels
`sm:pt-0` — so the tiles lost their top padding at >=640px only, and the
content sat against the top edge. Restores the top inset at the sm breakpoint
and lets the content fill the stretched grid cell so it stays centred when a
sibling tile is taller.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The RGB→NV12 compute CSC shares the SMs with the game; request an
elevated global-priority queue (VK_KHR/EXT_global_priority) so the
driver can schedule it ahead. PUNKTFUNK_VK_QUEUE_PRIORITY =
off|high|realtime (default realtime), downgrading REALTIME→HIGH→none
on NOT_PERMITTED/INITIALIZATION_FAILED — a refused class never fails
the bridge. Mirrors PyroWave's ac0e7332 (same honest caveat: WDDM
didn't move; the Linux driver may — contended-tail lever, correct and
harmless either way).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PUNKTFUNK_NVENC_ASYNC gains a real tri-state: 1 = always (as before, now
documented as ~+1 tick at depth-1), 0 = never (vetoes escalation), unset =
ADAPTIVE — off until the session loop's cadence-overrun detector escalates.
The host loop's adaptive-depth leaky bucket grows a second stage: once the
capturer's depth is maxed (Linux portal is permanently depth-1), it asks
the encoder for pipelined retrieve via the new Encoder::set_pipelined hook
(asked exactly once; default impl declines, Windows untouched).
nvenc_cuda engages at a safe point via a clean session rebuild WITHOUT the
IO-stream binding: with input==output stream bound, later stream work
waits on prior encode completions and would serialize a pipelined session
— stream-ordered submit and two-thread retrieve are mutually exclusive.
The ordered gate now also requires async_rt absence (belt-and-braces for
the runtime switch). Re-open's first frame is the standard session IDR.
On-hardware test: escalate mid-session → retrieve thread live, binding
gone, all AUs deliver, first post-escalation AU is the IDR.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
design/vulkan-rgb-direct-encode.md B1: when the probe passes AND
PUNKTFUNK_VULKAN_RGB_DIRECT=1 (default OFF until the B2 on-glass A/B),
the session opens with pictureFormat=B8G8R8A8 and the captured RGB frame
becomes the encode source — the VCN EFC front-end does the 709-narrow CSC
inline during encode. Deleted from the per-frame path: the compute CSC
dispatch, both NV12 plane copies, the semaphore hop and one queue submit
(dmabuf frames are ONE encode-queue submit; ~17 MB/frame of GPU traffic
gone). The DPB stays NV12; RFI/RC/quality-level machinery is untouched.
Shape:
- probe_rgb_direct now returns the chroma-siting bits to create the
session with (midpoint preferred, else cosited-even per axis); the open
verdict line gains "active" / "available(off; ...)" states.
- RgbProfileStack: the rgb-chained video profile rebuilt on the stack per
profiled-image creation after open (profile identity is by value) —
dmabuf imports become profiled VIDEO_ENCODE_SRC images (import cache
unchanged), the CPU staging image likewise (concurrent encode+compute).
- record_submit_rgb: steps 2–4 twin (dmabuf: single submit; CPU: staging
copy on the compute queue, semaphore-ordered); shared step-1/bookkeeping.
- begin_encode_cmd takes a SrcAcquire (CSC general / fresh-import FOREIGN
QFOT / cached visibility-only / staging TRANSFER_DST) so both paths
share the encode recording.
- RGB-direct frames skip the CSC per-slot resources entirely (make_frame
split into csc/common halves); cursor bitmaps warn once (EFC cannot
composite; gamescope — the flagship — embeds the cursor itself).
On-glass (780M RADV PHOENIX, host Mesa 26.0.4): all four smokes pass
(vulkan_smoke{,_av1,_rgb,_rgb_av1}); the EFC-encoded H265 stream decodes
clean and matches the CSC-encoded stream at 49.9 dB average PSNR (min
48.9) — within the design's ±1-code-value tolerance. (Raw-OBU ffmpeg
probing of the tiny AV1 dumps fails identically for BOTH paths —
pre-existing dump quirk, not a stream defect.) check+clippy+full unit
suite green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
servePluginUi's Bun.serve leaves idleTimeout at Bun's 10s default, so the old
15s keepalive could never arrive — an idle status feed was closed before its
first ping, which is exactly what a live host showed. Default is now 5s.
Also adds the regression suite that should have caught it: the original test
only covered a finite, self-driving stream. The new one exercises the
production shape (PubSub-backed, published after the response opens) plus the
keepalive — with a reader that does not re-enter read(), which is what made the
first version of this test lie.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@unom/ui rings cards ring-2 ring-accent, sized for its own subtle accent; in
this palette --accent is the brand violet, so every card shouted. The console
already softens the same ring in its Card wrapper — plugin UIs now inherit that
instead of each wrapping AnimatedCard themselves.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PUNKTFUNK_NVENC_SLICES=N (2..=32, default off) splits H.264/HEVC frames
into N slices (sliceMode 3); PUNKTFUNK_NVENC_SUBFRAME=1 (default off)
arms enableSubFrameWrite + reportSliceOffsets on sync sessions only.
Both experimental groundwork for sub-frame slice output (plan §7 LN1).
nvenc_cuda_subframe_slice_probe (on-hardware, ignored) answers the LN1
go/no-go: spins lock_bitstream(doNotWait) against an in-flight frame and
prints the (t_us, status, numSlices, bytes) timeline — incremental slice
availability and its spacing, or all-at-completion.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On-glass B0 result from the 780M (RADV PHOENIX, host Mesa 26.0.4): the
probe returned no-709-narrow-midpoint because RADV advertises
xChromaOffsets = COSITED_EVEN only — the VCN EFC does canonical H.26x
left-cosited x-chroma, while the requirement was written to bit-match our
2x2-average (midpoint) shader. The delta is a half-pel chroma-x phase,
imperceptible and unsignalled in our bitstream; EFC's siting is arguably
the more correct one. Accept either siting bit per axis (model + range
stay exact: 709 narrow); B1 will pass the preferred available bit.
With this the 780M probes rgb_direct=available; both GPU smoke tests
(H265 + AV1) pass on that box against host Mesa 26.0.4 — the first
real-RADV validation of the poll fix, the quality-level control, and B0.
(Container mesa alone can't run the H265 smoke: Fedora ships RADV with
H264/H265 encode compiled out — AV1 only. Use the host ICD via
VK_ICD_FILENAMES on Atomic boxes.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@unom/ui reads --secondary as muted TEXT; this palette (like the console's) uses
it as a dark SURFACE, so EmptyState captions, badge text and table headers
rendered dark-on-dark. Tailwind derives both utilities from one token, so the
text lane is re-pointed at --muted-foreground here — once, for every plugin,
instead of in each plugin's stylesheet.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Bind the session's IO streams to the encode thread's high-priority copy
stream: in sync-retrieve depth-1 use the per-frame input copy and cursor
blend now enqueue with NO cuStreamSynchronize and encode_picture orders
after them on the stream. Same stream both directions, so the encode's
completion is inserted into the stream and later work (the next frame's
copy into a reused ring slot) waits for it.
Soundness gate: the fast path engages only when pending is empty (true
depth-1 usage) — every prior encode was drained by a blocking poll, and
the caller holds the frame payload across the matching poll (contract now
documented on Encoder::submit; both host loops already comply). Pipelined
callers and PUNKTFUNK_NVENC_ASYNC mode keep the blocking copies.
True zero-copy input registration (registering the worker-owned IPC
buffer directly) stays the LN2 v2 follow-up — it needs a contiguous
worker-pool NV12 layout and a registration<->IPC-mapping lifetime tie.
PUNKTFUNK_NVENC_STREAM_ORDERED=0 restores the old blocking behavior.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Latency plan §7 LN0 (Linux/NVIDIA encode follow-on):
- sampled PUNKTFUNK_PERF submit split (copy/blend/map/pic) in nvenc_cuda —
the host loop's submit_us folds all four together; the D2D input copy is
the LN2 zero-copy target and now measurable on its own
- sampled blocking lock_bitstream timing on pf-nvenc-out — in two-thread
mode the host loop's wait_us wraps a non-blocking poll, so the real
encode wait was measured by no timer
- caps probe + log SUPPORT_SUBFRAME_READBACK / SUPPORT_DYNAMIC_SLICE_MODE
(LN1 sub-frame slice-output prerequisites, fleet visibility)
- explicit zeroReorderDelay=1 in the shared low-latency config (P-only +
no lookahead has no reordering anyway; pins the bit against preset or
driver drift; shared with the Windows backend)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
First phase of design/vulkan-rgb-direct-encode.md (punktfunk-planning): make
the captured BGRx dmabuf the direct encode source with the VCN EFC front-end
doing the 709-narrow CSC — deleting the per-frame compute CSC, both plane
copies, the semaphore hop and one queue submit.
B0 changes nothing about the encode path. It vendors the extension surface
(vk_valve_rgb.rs — ash 0.38 predates it; same rationale and style as the
vk_av1_encode module) and probes at open whether this host qualifies:
extension present (Mesa >= 26.0 + EFC hardware) → feature bit → conversion
caps cover the compute shader's exact math (709 / narrow / midpoint both
axes) → encode-src format set offers B8G8R8A8 with DRM-modifier tiling.
The verdict is one INFO line (rgb_direct=available | first missing
requirement) — the field telemetry that decides where B1 can default on.
Verified: cargo check + clippy clean + unit tests green (Linux box); the
probe short-circuits to no-ext on non-RADV drivers.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Field follow-up to 6dc195f9 (the AMD host whose H265 encode sits at ~5 ms):
the remaining time is the VCN ASIC preset, and we never chose one.
1. Explicit encode quality level. RADV only emits a VCN preset op when the
app issues ENCODE_QUALITY_LEVEL — we never did, so every session ran the
firmware's default preset. Install the resolved level on the first
frame's RESET control (chained ahead of the CBR state, which satisfies
the spec's quality-change-carries-rate-control rule) and bake the same
level into the session-parameters object (the spec requires the match).
Default 0 = the fastest tier the driver exposes (RADV: SPEED, except
H265 on pre-RDNA4 which the driver pins to BALANCE); clamped to the
profile's maxQualityLevels. PUNKTFUNK_VULKAN_QUALITY=0..3 overrides for
quality-biased setups. Logged at open so field logs show the tier.
2. Persistent bitstream mapping. read_slot vkMapMemory/vkUnmapMemory'd the
host-visible bitstream buffer every frame; map each ring slot once at
build instead (coherent memory; vkFreeMemory implicitly unmaps).
3. PUNKTFUNK_PERF CSC split. GPU timestamps bracket the compute batch
(import barriers + cursor prep + CSC dispatch + plane copies) and a
sampled csc_us line logs every ~2 s — separating shader/copy time from
the ASIC encode inside the pump's wait_us, the VAAPI submit-split
analog for this backend. Off (and unrecorded) unless PUNKTFUNK_PERF.
Verified: cargo check + clippy + DPB/RPS unit tests green (Linux box).
The GPU smoke tests still need a RADV host (NVIDIA driver fails session
open on the build box, pre-existing).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The session pump's depth-1 loop (capture → submit → poll) treats a poll()
None as "the backend holds the frame internally, re-poll next tick" —
true only of the libav AMF/QSV wrappers, whose codecs genuinely hold ~2
frames. The Vulkan Video backend probed its slot fence with
get_fence_status instead, so the AU of every frame — finished on the
ASIC in ~5 ms — sat unharvested until the tick AFTER the next frame was
submitted: encode_us read as one full frame period (~17 ms at 60 Hz vs
VAAPI's ~5.3 ms on the same VCN, the AMD field report) and every frame
shipped a frame period late — one frame of avoidable glass-to-glass
latency on the path that exists to beat libav VAAPI.
Wait the oldest in-flight slot's fence in poll() instead, matching the
sync NVENC backend's blocking lock_bitstream and the documented
Capturer::pipeline_depth contract ("capture → submit → poll-blocks").
Bounded by ENCODE_FENCE_TIMEOUT_NS like enqueue()'s backpressure wait,
for the same reason: poll runs on the thread the stall recovery runs on,
so a wedged GPU must surface as an error, not park it. None now only
means "nothing submitted". Covers H265 and AV1 (shared poll).
Verified: cargo check + DPB/RPS unit tests green (home-worker-5); the
GPU smoke tests fail at session open on that box's NVIDIA driver
identically on unmodified origin/main — pre-existing, needs the RADV
box for on-glass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Plugin contracts share ProviderEntry/Artwork/LaunchSpec/PrepStep with their
UIs; the kit root stays node-only (reconcile keeps the HostClient transport).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`web/project.inlang/settings.json` pointed `modules` at
https://cdn.jsdelivr.net/npm/@inlang/plugin-message-format@4/dist/index.js —
inlang's stock `paraglide-js init` scaffolding, which nobody had a reason to
keep. It made `bun run codegen` need the network at build time, so in a
sandbox that has none (the Nix build; any air-gapped CI) the plugin import
failed and paraglide compiled ZERO messages.
That failure is silent by design: the inlang SDK logs a PluginImportError as a
WARNING, paraglide still prints "Successfully compiled" and exits 0, vite
bundles the empty message set, and the console only dies at request time inside
renderToReadableStream because every `m.foo()` is undefined. Reported from a
NixOS host running punktfunk-web-server.
The plugin is a normal npm package, so depend on it like one: add
@inlang/plugin-message-format as a devDependency and reference it by relative
path. `loadProjectFromDirectory` (which both `paraglide-js compile` and the
vite plugin use) imports non-`http` modules straight off disk, and the shipped
dist/index.js is a self-contained bundle with no imports of its own, so it
loads from the data: URL exactly as the CDN copy did. Compiles 268 messages
offline, with no project.inlang/cache round-trip.
Add tools/check-i18n.mjs, run at the end of `codegen`, so this can never ship
silently again: it fails the build on a remote `modules` entry, a module that
does not resolve, or a compiled message count below what messages/en.json
defines. Verified it catches all three (remote URL, missing plugin, corrupt
plugin that compiles to 0 messages) and passes a healthy build.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The ~80% of every plugin that was copy-pasted (rom-manager ↔ playnite),
extracted as one Effect-v4 package:
- runtime: definePluginKit — async-main boundary hiding a ManagedRuntime
(two-effect-instances discipline), signal-driven interruption with a
bounded shutdown grace
- config: Schema-driven raw round-trip (defaults only in the Schema via
withDecodingDefaultKey + encodingStrategy omit; file stays authored-shape),
atomic writes, world-writable refusal, changes stream
- cache-store: disposable derived state, corrupt→empty, write-through
- reconcile: kit-owned provider wire schemas + typed client over the
skew-safe untyped pf.request seam
- sync-engine: generic poll/watch/debounce/single-flight-coalesce/
fingerprint-skip engine with a status PubSub (the SSE feed)
- ui-server + sse: effect/unstable/httpapi behind servePluginUi with
core-only env layers (validated by the phase-0 spikes); raw SSE route
(httpapi has no event-stream media type)
- cli: minimal command dispatcher reusing the plugin layer graph
(deliberately not effect/unstable/cli — it needs platform packages)
- react subpath: plugin router (path→hash→fallback deep-link restore +
pf-ui:navigate bridge), ResultGate, sseAtom, resolvePluginBase
- theme.css: the console's violet identity packaged for plugin UIs
18 bun tests incl. the two phase-0 spikes; publish workflow mirrors
sdk-publish (tag plugin-kit-v*; 0.1.0 published manually).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Manual publish (no sdk-v tag — the version was published directly so the
plugin-kit and rewritten rom-manager can consume pluginStateDir/pluginIngestDir
from the registry; the previously published 0.1.1 predates the security-tiers
merge and lacks those exports).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two spikes validating the riskiest seams of the plugin-kit design before
implementation:
- HttpApi (effect/unstable/httpapi) served on Bun behind the SDK's
servePluginUi using ONLY effect-core layers (Etag.layerWeak, Path.layer,
HttpPlatform.layer + FileSystem.layerNoop) — no platform package. Auth,
schema validation, and static fallthrough all verified end-to-end.
- Browser client prefix strategy for the console proxy: both
transformClient+prependUrl AND baseUrl preserve the /plugin-ui/<id>
path prefix. Nested Schema.withDecodingDefaultKey defaults confirmed
(Effect-valued defaults; encodingStrategy "omit" restores raw shape
on encode).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Second and final batch from the 2026-07-20 core-sweep lows:
client:
- connect() timeout now sets `quit` before shutdown, so a handshake that
completes after the deadline closes with QUIT_CLOSE_CODE instead of
leaving the host lingering (virtual display up) for a reconnect that
never comes.
- probe_result: saturating_add on the wire-supplied wire_packets +
send_dropped counters (debug-build overflow panic / release wrap).
- the standing-latency bleed no longer sets flush_in_window: that flag
is the ABR's SEVERE (×0.7) verdict, and the bleed fires only on
provably loss-free windows the controller itself scores as fine.
clipboard:
- fetch_cancels pruned on every new fetch (was: one dead oneshot per
paste for the session).
- serve chunks gated on a parked waiter + capped at CLIP_FETCH_CAP with
an Error event (was: unbounded silent accumulation under any req_id).
- serve_inbound park bounded by FETCH_STALL_SECS + send.stopped() (was:
an unanswered FetchRequest parked the task, waiter, and bi-stream
forever; ~100 of them exhaust the connection's bidi budget).
C ABI (ABI_VERSION 9 → 10, header regenerated, additive only):
- new punktfunk_connection_clock_offset_now_ns — the LIVE re-synced
offset (Swift/Kotlin latency math read the frozen connect-time value
~40ms wrong after a wall-clock step).
- to_config: checked u64→usize narrowing of max_frame_bytes (32-bit
armeabi truncated >4GiB to a plausible residue).
- host_poll_input: no &mut held across the embedder callback (re-entry
aliased it — UB under noalias); mid-drain callback clears now stick.
- next_audio_pcm: DTX (empty) payloads skipped — decode synthesized
120ms of concealment per 5ms slot and grew the playout ring forever.
- next_clipboard releases the parked payload on an empty poll (a one-off
50MiB paste stayed resident all session).
- frames_dropped / wants_decode_latency write their documented 0/false
defaults before the NULL-handle check.
- gamepad constant docs match pick_gamepad() reality (DualSenseEdge/
SwitchPro landed; DualSense/DS4 honored on Windows UMDF too).
FEC:
- gf8 reconstruct/reconstruct_into reuse the (k,m) codec cache like
encode_into (was: fresh 230×200 generator + decode inversion per
lossy block on the pump thread).
- vendored fec-rs: the 8 safe wrap_mul_slice shims assert equal lengths
(x86 SIMD callees bound stores on input.len() — safe-code OOB write);
ReconstructShard's safety contract gains the len()==get().len()
clause and reconstruct_internal sizes raw slices from the slice.
transport/GTK:
- Linux GSO super-buffer capped at the real UDP payload ceiling (65487)
not 65535 — seg sizes ≥1024 could EMSGSIZE and latch GSO off
process-wide, blamed on the network.
- GTK settings dialog no longer rewrites an unlisted-but-valid stored
gamepad preference to "auto" on close.
Already fixed on main (verified stale, skipped): the reassembler FEC
ceiling, wants_decode_latency's third term, request_probe rollback +
the generalized probe watchdog.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
From the 2026-07-20 punktfunk-core quality sweep's adjudicated lows
(~/punktfunk-sweeps/punktfunk-core-2026-07-20.json), each with a
regression test:
- session: the two-lane seal's dead-worker fallback dropped the frame's
back half and returned Ok with half an access unit; the corpse lane
was also never respawned. The send-failure arm now reclaims the tail
(single-lane seals the WHOLE frame) and both failure arms drop the
lane so the next large frame respawns it. A recv-side death now
surfaces as an error instead of silent truncation.
- reassemble: Reassembler::reset() left pending_partial parked, so a
pre-flush stale partial survived Session::flush_backlog and was
delivered as the first "frame" after a jump-to-live.
- caps: resolve_codec echoed a non-conformant multi-bit preferred byte
verbatim (downstream from_wire folds it to HEVC — possibly outside
the shared set). It now isolates one bit of the intersection.
- endpoint: stream_transport_idle only floored the value; an absurd
operator-supplied idle timeout blew past QUIC's VarInt ms ceiling and
panicked host startup through the expect. Clamped to 1s..1h.
- pairing: PairRequest::encode cut the device name at a raw byte-64
boundary, splitting multi-byte UTF-8 (host showed U+FFFD forever);
it now shares Hello's char-boundary truncate_to.
The frozen-FEC-ceiling finding (reassemble.rs:109) was already fixed on
main by the sweep's high-severity commit — skipped as stale.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- build.rs: drop the stale `rerun-if-changed=src/client.rs` (the file
became client/ in W7; the path never fires).
- lib.rs: the crate doc's module list named 6 of ~19 modules — add the
missing ones (config/error/stats/input/reject/reanchor/render_scale/
audio/wol + the quic-gated control-plane group), without intra-doc
links to feature-gated modules.
- quic/mod.rs: the module doc still described the deleted `msgs`
module; list the actual post-split module set.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
client/pump.rs was a single 1192-line async fn stacking six concerns.
It is now a 190-line orchestrator (same spawn order, same channel
wiring) over five focused modules:
- pump/handshake.rs — connect + Hello/Welcome/Start + skew handshake +
hole punch + Session construction (HandshakeOut)
- pump/input_task.rs — the gamepad snapshot/removal/arrival re-send
state machine + passthrough input forwarding
- pump/control_task.rs— the control-stream select loop (renegotiation,
probes, bitrate acks, clock re-sync, clip
metadata), bundled as ControlTask
- pump/datagram_task.rs — the datagram tag demux + rumble reorder gate
- pump/data.rs — the blocking data-plane pump (loss reports, ABR
+ capacity probe, jump-to-live detectors,
standing-latency bleed), bundled as DataPump
Every body moved verbatim (dedent + arg-struct destructures aliasing
the original binding names); no per-frame indirection added — the pump
loop, its locals, and the hot-path shape are byte-equivalent. The
Ctrl/DataPump arg structs exist only to stay under too_many_arguments.
196 lib tests pass; clippy clean on --features quic AND
--no-default-features (--all-targets); include/punktfunk_core.h
byte-identical; fmt clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
session.rs (1203) sharpens down to the two hot-path state machines +
lifecycle (887): ReplayWindow + seq_of + their tests -> session/replay.rs;
the PUNKTFUNK_PERF trio PumpPerf/SealPerf/TimedCoder -> session/perf.rs;
the Phase-1.5 lane machinery SealLane/SealJob/seal_wire_slice/
TWO_LANE_MIN_PACKETS -> session/seal.rs. Facade pattern (session.rs stays
the parent file); pub use keeps session::{PumpPerf,SealPerf} stable and
lib.rs re-exports are untouched. Pure code motion + pub(super) bumps —
seal_frame_inner/poll_frame/poll_input bodies unchanged; the
wire-equivalence tests stay co-located with the seal path they pin.
196 lib tests pass, clippy --features quic --all-targets clean, fmt clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
quic/tests.rs (1813 lines, 43 tests) was the W7 split's leftover: the
source moved into handshake/caps/control/clock/pairing/pake/datagram/
endpoint/clipstream/io but every test stayed in one monolithic file.
Each test now lives in a #[cfg(test)] mod tests at the foot of the
module it exercises, verbatim. The two CompositorPref/GamepadPref
wire/name tests moved to config.rs (where those enums live), so they
now also run under --no-default-features. The clip_loopback and
ctrl_framing integration mods share connect_pair via a cfg(test)-only
quic/test_util.rs.
Test-only motion: 196 lib tests pass unchanged on macOS, clippy
--features quic --all-targets clean, include/punktfunk_core.h
byte-identical.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`ci.yml`'s clippy gate has been failing on main since the pre-0.16.0 sweep landed, and
because clippy stops at the first crate it can't compile, the visible error was only
ever the first of four. `ci.yml` is not tag-triggered, so v0.16.0 cut and shipped over
a red main; none of this reaches the release artifacts (the two E0308s are in a test
binary and a dev tool, and the rest are lints) but the gate has been blind since.
Fixed, in the order clippy surfaced them:
- clients/probe failed to COMPILE (E0308, x2). `810d918d` moved `clock_sync` onto the
resumable `io::MsgReader` but only updated the client pump, leaving the probe passing
a bare `&mut RecvStream`. The probe now wraps the control stream in a `MsgReader` at
`open_bi` and threads that everywhere — Welcome, the --remode and --bitrate watchers,
and the speed-test result read — which is also what the refactor was for: those reads
sit behind timeouts and `select!`, exactly where a straddling frame would desync the
stream for the rest of the run.
- pf-capture: `SPA_META_Cursor as u32` is a `u32 -> u32` no-op
(`clippy::unnecessary_cast`). Line 1274 already passes the same constant uncast, so
the type is not in question.
- pf-inject: `noop as usize` on the SIGUSR1 wake handler added in `986402f7` trips
`clippy::function_casts_as_integer`; goes via `*const ()` as the lint asks. Same value
in the `usize`-typed `sa_sigaction` slot.
- punktfunk-core: the `ctrl_framing` test module's `use super::*` is unused, which
`-D warnings` promotes to an error.
Verified with CI's own commands on a Linux box (this is all Linux-gated code, so a Mac
cannot check it): `cargo clippy --workspace --all-targets --locked -- -D warnings`
finishes clean, `cargo build --workspace --locked` succeeds, and
`cargo test --workspace --locked` exits 0 with no failures — including punktfunk-core's
196-test suite, which covers the control-stream framing the probe change touches.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The flathub remote-add has now burned its entire retry budget and failed the job on
three consecutive tag builds — v0.15.0, the v0.15.0 re-point, and v0.16.0 — each one
needing a manual re-run to land the bundle.
The cause was already root-caused here on 2026-07-11 (see the Tooling step's comment):
home-runner-1's Docker embedded resolver at 127.0.0.11 drops UDP lookups while the
shared multi-org fleet is saturated. Nothing retransmits a dropped datagram, so the
lookup simply times out. The fix then was to widen retry.sh to 10 attempts (~9 min),
which comfortably outlasts a main push's ~8-workflow fan-out — but a TAG push starts
13 workflows at once, and that burst now outlives the whole budget. Retrying harder
is chasing the symptom; the transport is the problem.
`options use-vc` moves resolution to TCP, where the kernel retransmits and a query
cannot be silently lost under load. Same resolver and search path — only the
transport changes — so internal names still resolve exactly as before. Deliberately
no additional nameservers: a public resolver in the list could answer an internal
name (git.unom.io) with NXDOMAIN. retry.sh stays as the backstop for real upstream
blips.
Applied non-fatally, because Docker bind-mounts /etc/resolv.conf and may present it
read-only: a DNS tuning that cannot be applied must not be the thing that fails a
release build. If the write is refused we warn and stay on UDP, i.e. exactly today's
behaviour.
Scoped to the flatpak workflow, where the failure is actually evidenced, rather than
applied fleet-wide on suspicion.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
v0.16.0's Apple leg failed at the xcframework build with "No space left on
device (os error 28)": the runner's boot volume was at 100% with 152 MB free.
None of the four `xcodebuild archive` calls passed -derivedDataPath, so each one
used the default ~/Library/Developer/Xcode/DerivedData/<name>-<hash> — and that
hash is derived from the PROJECT'S ABSOLUTE PATH. act_runner rotates its
workspace (~/.cache/act/<hash>/hostexecutor), so every rotation looked like a new
project to Xcode and minted a fresh ~760 MB tree. Nothing ever collects those:
they live outside the workspace, so act's own cleanup never sees them. 31 had
accumulated in three days (17 -> 20 July), which with the 12 GB shared
ModuleCache came to 32 GB — on a 228 GB volume already 95% full.
Pin all four archives to one path. The tree is now REUSED rather than multiplied,
which also keeps the module cache warm instead of rebuilding it per run. The new
step additionally prunes anything week-stale left in the default root, covering
both the legacy per-path trees and any other job that lands there.
Cleared by hand on home-mac-mini-1 to unblock the release (152 MB -> 31 GB free);
this is the change that stops it coming back. Note the same class of problem bit
the Windows runner the same night from the other direction — its disk-cleanup task
purges C:\t and deleted files out from under ISCC mid-pack — which is worth its
own look and is NOT addressed here.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
MINOR, not the 0.15.1 originally planned: the 24 commits since v0.15.0 carry 7
features, and they are not incidental — they change the plugin runner's privilege
model. It stops holding the console's full-admin token (scoped plugin-token lane),
the Windows runner drops from SYSTEM to LocalService, public-registry plugin
installs are gated behind an explicit flag, and the systemd user unit is sandboxed.
Operators have workflow-visible changes to absorb (a plugin genuinely needing the
admin surface must now set PUNKTFUNK_MGMT_TOKEN; installing outside the @punktfunk
scope needs --allow-public-registry). A patch label would understate all of that,
and sdk-publish pushes this version straight to consumers.
The other 12 fixes are largely the pf-encode/zerocopy/capture/inject quality sweep
landing: teardown deadlocks, error-path leaks, a GL->CUDA copy race, an odd-height
NV12 overrun, a cursor-meta OOB read, plus the two loss-recovery gaps 0.15.0 left
open (Vulkan Video never got the LTR taint sweep; QSV's sweep skipped its modal
single-swept-slot case).
`WIRE_VERSION` stays 2 and the driver protocol stays v4 — pairings, clients and
installed drivers keep working. The C ABI moves 8 -> 9 (PunktfunkFrame grows
`received_ns`, so receipt is stamped at the session boundary rather than at the
embedder's pull); the gitignored Apple xcframework needs a rebuild for that to
reach the Apple clients. No OpenAPI path changes (45 before and after) — the new
plugin-token lane is auth-layer only — so api/openapi.json moves info.version only.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
There is one `ProbeState` per session and no correlation id, and two independent
requesters share it: the pump's startup link-capacity probe and the embedder's
`NativeClient::request_probe` (the shipped "Test connection" in the Windows GUI
and the Linux GTK app). The startup path had a busy check, a watchdog and a
window rebase; the embedder path had none of them, and the two collide by
default — both shipped speed tests call `request_probe` on the statement right
after `connect()`, and the pump's probe fires 2 s later, inside that burst.
Four defects, one cluster:
- The pump overwrote the shared slot unconditionally. Mid-burst it drops
`base_packets`/`base_bytes`, the pump re-snapshots them against the host's
full-burst denominator, and the user's speed test reports roughly an order of
magnitude low; if the result lands first instead, the reset wipes `done` and
the embedder's poll loop never sees its own measurement. Now it defers and
retries rather than stealing the slot.
- An unanswered embedder probe never timed out. `probe_active` gates the entire
report tick — LossReport, the ABR window feed, the standing-latency ladder and
a pending ClockResync all live inside it. A host that ignores ProbeRequest is
an anticipated configuration (the startup path was given a 6 s timeout for
exactly that) so the embedder path could latch `active` forever and silently
switch off every adaptation mechanism for the rest of the session. Now a
watchdog covers a probe of either origin.
- `request_probe` latched `active` before `try_send`, so a full or closed ctrl
channel returned `Closed` to the caller while leaving the session wedged in
the state above. It now rolls back, mirroring the startup path.
- Only the startup path rebased the ABR window past the burst. Probe filler is
counted into `bytes_received` for every accepted datagram but never reaches
the decoder, and the tick is suppressed for the whole burst, so the first
post-burst window read the burst rate as `actual_kbps`. That poisons
`proven_kbps`, a monotone high-water mark that is never decayed, which
disables the x1.5 evidence-gated climb guard for the session and authorizes a
climb into a rate the decoder has never carried. The rebase now happens on the
falling edge of any probe, and covers the loss/packet anchors too so the first
post-burst LossReport isn't divided by a filler-inflated packet count.
Also: `wants_decode_latency` advertised on two of the three terms the pump
actually requires to arm ABR, omitting `resolved_bitrate_kbps > 0`, so against
an old host that reports no rate an embedder fed decode latency to a controller
that never runs.
No wire or ABI change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
`io::read_msg` frames a message with two `quinn::RecvStream::read_exact` calls,
and quinn documents `read_exact` as explicitly NOT cancel-safe: the bytes it has
already taken out of the stream live only in the future's own buffer and nothing
puts them back on drop. Both long-lived control loops drive that read from a
`tokio::select!` arm — the client pump alongside `ctrl_rx.recv()` and the resync
tick, the host alongside probe/reconfig/clip-offer channels — and neither uses
`biased;`, so any sibling that becomes ready ends the iteration and drops a
partially-progressed read. `clock_sync` has the same shape via
`tokio::time::timeout`, which can fire mid-frame before the session even starts.
A control frame only has to straddle two wakeups for this to bite: a ClipOffer
carries up to 16 kinds x 128 bytes of MIME, ~2 KB, which exceeds one QUIC packet
and is subject to the pacer; any frame whose second half is lost or reordered
does it too. Losing the consumed length prefix misaligns the stream permanently
— the next read takes two payload bytes as a length, so Reconfigured,
ProbeResult, BitrateChanged, ClockEcho and ClipState all decode as garbage and
are silently dropped, and a bogus length up to 64 KiB parks the read forever.
Mode switches, adaptive bitrate, mid-stream clock resync and clipboard are dead
for the rest of the session; only a reconnect recovers, and the log shows at
most one `warn!`.
Add `io::MsgReader`, which keeps the frame in progress in the reader rather than
the future and reads via quinn's cancel-safe `read`, and switch the three
cancelling sites to it (client control loop, host control loop, clock_sync).
The sequential handshake/pairing callers keep the plain `read_msg`, whose doc
comment now states the constraint. No wire bytes and no ABI change — only how
the same length-prefixed frames are assembled.
Tests: a frame split across two wakeups with the read cancelled in between must
resume and leave the following frame correctly framed (confirmed to fail — it
hangs on the desynced stream — against the old behavior), plus a zero-length
frame round-trip.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Four defects from the punktfunk-core quality sweep, all in the data plane.
transport/udp: the hole-punch adopted the source address of ANY datagram whose
first 8 bytes matched PUNCH_MAGIC — a fixed public constant with no key, nonce
or session id — and the authenticated QUIC peer was passed only as the
no-punch fallback, so it was never used to validate. Hole-punch is the default
bring-up path (it is skipped only for a fixed --data-port), and the data socket
is an OS ephemeral, so spraying the ephemeral range during the 2500 ms punch
wait let anyone steal the video plane: the legitimate client is then filtered
out by the connect() and gets nothing, while QUIC stays healthy so no reconnect
fires. With a spoofed source the same 8 bytes aim a full-rate stream at a third
party. Take the authenticated peer IP and require the punch to match it — only
the PORT is in question (that is what a NAT remaps and what the punch exists to
discover); the client dials the same host IP as its QUIC connection, so a NAT
presents one source IP for both planes. Also budget each read from the REMAINING
window, so off-peer datagrams cannot stretch the wait past punch_timeout.
transport/udp: the punch keepalive treated every send error as fatal and broke
out of its loop permanently and silently. It holds a clone of the connected,
non-blocking data socket — exactly the socket whose transient conditions this
module defines and documents in is_transient_io. One ENOBUFS from a full wlan tx
queue or an ENETUNREACH during an AP roam killed the only thing holding the NAT
mapping open; the path recovers, video keeps flowing, and the stream dies later
when the idle timer expires the mapping during a static scene. Route it through
is_transient_io like every other send site in the file.
packet: adaptive FEC moved fec_percent live (host bands it 1..=50 while Welcome
advertises 10) but the receiver's per-block acceptance ceiling was computed once
from the negotiated percentage and never re-derived. Once FEC ramped, every
packet of a maximal block failed `total > max_total_shards`, the block never
accumulated a shard, the frame aged out, and the resulting loss drove FEC higher
still — large frames wedged at 100% loss exactly when FEC was meant to rescue
the link. Fixed on both sides, because hosts and clients update independently:
the sender clamps per-block parity to the ceiling the peer negotiated, and the
receiver sizes that ceiling from the whole clamp range rather than a stale
snapshot of it.
No wire bytes and no C ABI signature change; WIRE_VERSION and ABI_VERSION are
unchanged. Regression tests cover all three (the punch tests were confirmed to
fail without the fix).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The LocalService runner can't traverse the interactive user's profile the way
the old SYSTEM runner could — so a plugin can no longer read a file an app
running as *you* produced (the acute case: the Playnite exporter's library
JSON under %APPDATA%). Confirmed on-glass: as LocalService the glob finds
nothing and the profile is un-traversable.
Add the inverse of plugin-state: <config_dir>\ingest, granted BUILTIN\Users
Modify by plugins enable (disable reverts to inherited Users:RX). An
interactive-user app drops ingest\<plugin>\… and the de-privileged runner
reads it there — the one Users-writable carve-out in the otherwise
Users-read-only tree. SDK exports pluginIngestDir(name) to resolve it; on
Linux the systemd --user runner owns the config dir so same-user producers
write there with no grant.
Accepted tradeoff: the inbox is writable by any local user (trusted-single-
user model; it feeds only a LocalService runner). Consumers must treat ingest
data as lower trust than their own state.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A plain `punktfunk-host plugins add …` printed the host's startup banner and
— on Windows — ran the win32u GPU-preference hook, whose DPI-awareness probe
emits a scary `SetProcessDpiAwarenessContext … "access denied"` WARN. None of
it is relevant to a package-manager command, and the WARN reads like a failure.
Gate both behind is_management_cli(): plugins / openapi / library /
detect-conflicts / driver / web / service-management verbs skip the banner and
the DXGI hook. `service run` (the SCM-launched host) is explicitly NOT
lightweight, so the hybrid-GPU ACCESS_LOST fix it depends on stays intact.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The LocalService runner cannot write anywhere under %ProgramData%\punktfunk
(the config dir is Users-read-only), so a state-writing plugin's saveCache /
config-edit / first-run mkdir all fail EPERM — proven on-glass (rom-manager
only looked fine because its state dir was pre-created by an admin run and a
0-title reconcile skipped the write).
Add the one writable grant the model was missing, keeping the split crisp —
code dirs RX+WA, secrets R, and now a dedicated state root RW:
- plugins enable / build-scripting.ps1: create %ProgramData%\punktfunk plugin-state and grant LocalService (OI)(CI)(M); disable revokes. Users stay
read-only, so another non-admin still can't tamper with a plugin's launch
templates.
- SDK: export pluginStateDir(name) -> <config_dir>/plugin-state/<name>. Same
path on Linux (the systemd --user runner owns the config dir, writable with
no grant), so plugins use one branch-free helper.
Plugins must persist under pluginStateDir(), not straight under the config dir.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- plugins: the Windows runner task runs as LocalService now, and
public-registry names need --allow-public-registry
- automation + SDK README: connect()'s zero-config credential is the
scoped plugin token; pairing administration and hook registration
need an explicit PUNKTFUNK_MGMT_TOKEN opt-in
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Supply chain: resolvePackage() used to pass any punktfunk-plugin-* or
foreign-scope name straight to bun add — a typo or a squatted look-alike
on npmjs.org would install operator-privileged code. Only the @punktfunk
scope (pinned to the Gitea registry by the bunfig scope map) resolves by
default now; anything else throws with an explanation unless
--allow-public-registry is passed, and even then installs print a loud
warning. Removal never gates — uninstalling stays safe regardless of
origin.
systemd (user unit): free hardening for well-behaved plugins —
NoNewPrivileges, PrivateTmp, ProtectSystem=strict with ReadWritePaths=%h
(plugin state and download dirs under $HOME keep working), and
RestrictAddressFamilies=AF_UNIX AF_INET AF_INET6. Plugins writing
outside $HOME, and distros that restrict unprivileged user namespaces
for user units, are handled via a documented systemctl --user edit
drop-in.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
fileIsSafe() returned true unconditionally on win32 ('ACL check is a
follow-up'), leaving the best-effort directory DACL as the only guard on
what the runner imports and executes. The follow-up: before importing a
unit, read its SDDL (Get-Acl via the full-System32-path powershell) and
refuse loudly — exactly like the Unix mode-bits path — unless the owner
is SYSTEM/Administrators/TrustedInstaller (or the account the runner
itself runs as, mirroring Unix's 'your own file is fine') and no other
principal holds a write-capable allow ACE (write/append data, EA/attrs,
DELETE, WRITE_DAC, WRITE_OWNER, generic write/all).
The SDDL verdict is a pure exported function with platform-independent
tests; unknown ACE shapes and unreadable ACLs fail closed. Inherit-only
ACEs (templates for children) and deny ACEs don't trip it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The PunktfunkScripting scheduled task ran operator-installed plugin code
as SYSTEM with -RunLevel Highest — any plugin defect was a full
compromise of the box. The principal is now NT AUTHORITY\LocalService
(minimal privileges, no password to manage, exists at boot, loopback
networking works), registered without RunLevel:
- installer: New-ScheduledTaskPrincipal -UserId 'LocalService'
- plugins enable: converges the principal idempotently (migrating tasks
an older installer or a dev box registered as SYSTEM) BEFORE starting,
then grants LocalService read — via icacls, full-System32-path — on
exactly the two SYSTEM/Admins-DACL'd files the runner's connect()
needs: the scoped plugin-token and the TLS-pin cert.pem. Never
mgmt-token. plugins disable revokes the grants; plugins status now
prints the task principal so the migration is verifiable.
- build-scripting.ps1 mirrors the convergence + grants on dev deploys.
The usage text also mentions the new --allow-public-registry gate that
lands with the supply-chain commit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The scripting runner used to hold the console's full-admin mgmt-token — a
plugin defect could rewrite hooks.json (arbitrary command execution) or
administer pairing. The host now mints a second persisted secret,
plugin-token (PUNKTFUNK_PLUGIN_TOKEN), and require_auth grows a third lane
for it: loopback-confined like the admin token, but plugin_may_access
carves out /hooks (read AND write), everything under /pair, /native/pair,
/native/pending, client unpair DELETEs, and other plugins' ui-credential.
Everything a plugin legitimately does (status/library/events/sessions,
its own UI lease) is untouched.
The SDK's zero-config connect() now prefers PUNKTFUNK_PLUGIN_TOKEN /
plugin-token over mgmt-token, so plugins pick the scoped credential up
automatically; a script that genuinely needs the admin surface sets
PUNKTFUNK_MGMT_TOKEN explicitly. Old hosts without a plugin-token fall
back to mgmt-token unchanged. No OpenAPI change: the lane is auth-layer
only.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`punktfunk-host plugins add/remove/list` forward to the bun runner, which on
Windows is resolved relative to the running exe (<exe-dir>\bun\bun.exe +
<exe-dir>\scripting\runner-cli.js). Only the installer ever laid that payload
down, so on a deploy-host.ps1 dev box — where the service runs out of
target\release — both paths are absent and the CLI bails with "the plugin
runner isn't installed". deploy-all.ps1 was host + web console only.
Add build-scripting.ps1: it mirrors CI (bun install --frozen-lockfile
--ignore-scripts + bun build src/runner-cli.ts --target=bun, gated on the same
`attempt=` sentinel that proves the dynamic plugin import stayed a runtime
import), then lays the bundle + scripting-run.cmd + the shared bun next to every
host exe it finds — the built one and whatever the PunktfunkHost service runs —
so it is correct on a dev checkout or an installed {app}. It stops a running
PunktfunkScripting first (a live runner holds bun.exe open) and does not
silently enable the opt-in task (-EnableTask to do so).
deploy-all.ps1 is now host -> web console -> runner, always, so the host binary
and the runner bundle never drift apart.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The two-pair investigation (wired Mac clients stuck at a rock-steady
~18-19 ms "network" that survived the load ending and cleared only on
reconnect) exposed two structural gaps, one of measurement and one of
recovery:
- Receipt was stamped at the hand-off PULL (Swift nextAU, pf-client-core,
Android decode loops), not at reassembly completion — so any client-side
standing state between the reassembler and the pull read as NETWORK
latency, undiagnosable from the HUD. ABI v9: `PunktfunkFrame`/`Frame`
grow `received_ns`, stamped by `Session::poll_frame` as the AU crosses
the session boundary. Every embedder now uses the core stamp; the Apple
client keeps the pull instant as `AccessUnit.pulledNs` and shows the
receipt→pull wait as its own "client queue" term (detailed HUD tier from
2 ms + a `queue_p50` stats-log field). Decode stages keep their pull
anchor on all platforms, so no historical stage shifts meaning.
- The jump-to-live detectors deliberately ignore anything under 6 queued
frames / 400 ms behind — so a small, constant, loss-free elevation (a
sub-frame standing backlog, or a stale clock offset after a wall-clock
step/slew) is carried for the rest of the session. New third detector
(`StandingLatency`, unit-tested ladder): window-MIN one-way delay
≥ 10 ms above the session floor with zero loss for ~4.5 s escalates
gently — a free clock re-sync first (an applied re-sync re-bases the
floor), then at most 3 flush+keyframe bleeds sharing the jump-to-live
cooldown, then a loud disarm naming what it means. Loss windows reset
the run: congestion belongs to FEC/ABR, not this detector.
Also: mid-stream re-sync apply/discard logs debug→info — they are the
forensic trail for the stale-offset case and were invisible in the field.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Both MSIX matrix jobs failed at the v0.15.0 tag, for two unrelated reasons, both
introduced by this release. 0.13.0 and 0.14.0 shipped both packages; 0.15.0 shipped
neither.
ARM64 failed to BUILD. The PyroWave Windows un-gating (1ef0229b) moved `pyrowave-sys`
in pf-client-core out of `cfg(target_os = "linux")` into a general optional dependency
that pf-client-core's own `default = ["pyrowave"]` turns on. The ARM64 leg builds
`--no-default-features` precisely to skip it, but that only disables defaults for the
two packages named with `-p` — every internal edge kept inheriting them:
clients/windows -> pf-client-core, clients/session -> pf-client-core, clients/session
-> pf-presenter, and pf-presenter's own `default = ["pyrowave"]` -> pf-client-core.
So the vendored Granite C++ compiled on ARM64 and stopped where it always does:
muglm.cpp reaching for _MM_TRANSPOSE4_PS/_mm_storeu_ps, then
`simd.hpp:103 #error "Implement me."`.
Those three internal edges now pass `default-features = false`; the feature is enabled
only through an explicit chain (clients/session's `pyrowave` -> pf-client-core/pyrowave
+ pf-presenter/pyrowave, and pf-presenter's `pyrowave` -> pf-client-core/pyrowave), so
`--no-default-features` finally means what it says. Verified by feature resolution
rather than assertion — `cargo tree -i pyrowave-sys`:
aarch64-pc-windows-msvc, --no-default-features -> not in the graph (was: present)
x86_64-pc-windows-msvc, default features -> present
x86_64-unknown-linux-gnu, default features -> present
PyroWave is NOT dropped from the Windows client. Decode has always run in the spawned
punktfunk-session binary, whose own default keeps the feature on, and unification turns
it back on for the shared pf-client-core wherever that binary is in the build (x64,
Linux). The shell only ever offered "pyrowave" as a codec preference string — it holds
no decoder and never called one, so linking the C++ into it was dead weight even on x64.
x64 built fine and failed at PACKAGING: `makepri new` rejected the manifest with
PRI191 "Appx manifest not found or is invalid" / malformed comment syntax. The console
tile added in 2508b720 documented its flag inside an XML comment, and XML forbids `--`
anywhere in a comment body, so the literal flag spelling made the whole manifest
unparseable. Reworded, with a note to keep the next person from reintroducing it.
Confirmed by parsing both revisions after template substitution: the tagged manifest
fails at line 63 col 9, this one parses.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two medium findings from the round-1 sweep, each applied to both siblings.
- create_swdevice stack-allocated the SwCreateCtx that the async PnP completion
callback writes through (result + up to 127 u16 of instance id) and then
SetEvents. The wait is bounded at 10s, so on a wedged-PnP timeout the callback
can still be PENDING: the frame is popped, the input thread reuses that stack,
and a late callback corrupts it and SetEvents an already-closed (possibly
recycled) handle. The context is now heap-allocated and reclaimed only where
the callback provably ran; on the timeout path the box is deliberately leaked
and the event left open, so a late write always targets live memory. Costs a
one-off ~264 B + one HANDLE on that rare path. Applied to the DualSense path
and its XUSB sibling in gamepad_windows.rs.
- Ds4WinPad::open swallowed a create_swdevice failure into a WARN and returned
Ok with no devnode. PadSlots::ensure then stored Some(pad) AND called
gate.on_success(), so the slot short-circuited on is_some() forever and the
capped-backoff retry that exists precisely to self-heal a transient PnP failure
never ran — the game saw no controller for the rest of the session unless the
client unplugged the pad. Now propagates, matching the XUSB sibling. Same fix
applied to steam_deck_windows.rs.
Windows .173: pf-inject 53/0. Linux .21: pf-inject 74/0 (8 ignored).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two medium findings from the round-1 sweep.
- The `.process` callback dequeued a buffer INSIDE its `catch_unwind`, and every
requeue site was inside too. `newest` is a raw pointer with no Drop, so any
caught panic (update_cursor_meta / consume_frame) unwound past all three
requeues and permanently stranded that buffer. Once the stream's fixed pool
drained, `dequeue_raw_buffer` returned null every call and capture silently
wedged while still reporting negotiated/active — defeating the very
catch_unwind that was meant to keep a panic survivable. The drain loop now runs
OUTSIDE the catch (dequeue/queue are non-panicking C FFI pointer ops) and
`newest` is requeued exactly once after it, on every path: normal,
corrupted-skip, or caught panic.
- `recreate_ring` invalidated `video_conv` and `hdr_p010_conv` but not
`pyro_conv`. That converter is mode-baked — BgraToYuvPlanes selects entirely
different shaders and output formats for SDR (8-bit BT.709 → R8/R8G8) vs HDR
(scRGB→PQ BT.2020 → R16/R16G16) — and `ensure_pyro_conv` only builds when None,
so a display_hdr flip reused the stale SDR converter against a freshly
HDR-formatted pyro ring, corrupting every frame. Reachable at the documented
"Downgrade point D": a PyroWave session with client_10bit=true that opens on a
box where HDR can't enable, then flips once the display comes up.
Linux .21: pf-capture 1/0. Windows .173: `cargo check -p pf-capture` clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Seven medium findings from the round-1 sweep, all in pf-zerocopy. Adjudicated
against source; all zero-copy-preserving (no GPU→CPU→GPU roundtrip introduced).
- import_src (vulkan.rs) leaked a VkBuffer + VkDeviceMemory + dup'd fd on every
fallible step before the success-only src_cache.insert. A failed import is
survived by the worker and RETRIED by the caller every frame, so this leaked
per frame for the worker's whole lifetime. Now owns the dup fd via OwnedFd
(closes on early return until Vulkan consumes it at allocate_memory) and
destroys the buffer + frees memory on each error path.
- ensure_dst destroyed the old exportable buffer and nulled self.dst BEFORE
building the replacement, so a failed rebuild both dropped the working buffer
and leaked every object the partial rebuild created (raw ash handles, no Drop;
VkBridge::drop only frees the live self.dst). Now builds fully, unwinds each
error locally, and swaps only on success.
- The submit→wait→reset sequence in import_linear_nv12 / import_linear `?`-ed out
of wait_for_fences BEFORE reset_fences on a TIMEOUT/DEVICE_LOST, leaving the
shared self.cmd PENDING and self.fence IN-USE while the caller retries on the
same bridge (and ensure_dst later destroys dst.buffer assuming nothing is in
flight). Now drains the GPU (device_wait_idle) and resets the fence before
propagating — fixing the reported ensure_dst UAF at its source.
- copy_pitched_nv12_to_buffer writes height.div_ceil(2) chroma rows into a UV
plane sized at height/2, so an odd height overruns by one uv_pitch row (OOB
device write / CUDA_ERROR_ILLEGAL_ADDRESS). Added the even-dimension guard to
import_linear_nv12, matching Nv12Blit/Yuv444Blit.
- sweep_reaper only reaped already-exited workers, so a worker wedged inside a
driver call lingered forever holding its CUcontext + BufferPool (hundreds of MB
VRAM). Now force-kills a child parked past REAPER_KILL_DEADLINE (20s).
Compile + tests green on Linux .21 (RTX 5070 Ti): pf-zerocopy 17/0. The error
paths themselves are not fault-injected; the fixes are structural.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The three high-severity defects from the round-1 sweep of pf-inject /
pf-zerocopy / pf-capture (adjudicated against source — all 7 reported criticals
in these crates downgraded; these were the real highs).
- pf-inject steam_gadget: `SteamDeckGadget::drop` set `running=false` then joined
the control thread, which spends steady state parked in a blocking, no-timeout
`EVENT_FETCH` ioctl that only tests `running` at its loop top. The flag never
reaches it, closing the fd can't wake an in-flight ioctl (the syscall holds a
file reference, and the fd is shared via Arc by the very threads being joined),
so the join hung — and it runs on the session input thread via
`PadSlots::sweep`, driven by the client's `active_mask`, so a remote peer
clearing its pad bit could freeze all session input. Now wakes the parked
threads with SIGUSR1 (no-op, non-SA_RESTART handler → the ioctl returns EINTR
and the loop exits), retried until each reports done and bounded (~1s).
- pf-zerocopy cuda: the GL→CUDA "sync point" was never established for the copy.
`cuGraphicsMapResources`/`UnmapResources` were issued on the NULL stream, but
the D2D copy runs on `copy_stream()`, a `CU_STREAM_NON_BLOCKING` stream exempt
from implicit NULL-stream ordering — and the GL de-tile/CSC that produced the
texture ends with only `glFlush` (no fence). So the copy could race ahead of
the not-yet-retired GL draw: intermittent stale/torn frames under GPU load, on
the default NVIDIA capture→encode path. Map, copy, and unmap now share
`copy_stream()`, so map's device-side guarantee orders the GL work before the
copy. Zero-copy preserved (no GPU→CPU→GPU roundtrip).
- pf-capture cursor meta: `update_cursor_meta` trusted three producer-written
fields (bitmap_offset, pixel offset, stride) with no bound against the metadata
region, driving OOB pointer arithmetic and an oversized `from_raw_parts` — an
OOB read that SIGSEGVs inside the PipeWire `.process` callback (uncatchable by
the surrounding `catch_unwind`). Switched to `spa_buffer_find_meta` to obtain
the region's real `size` and validate every offset with checked arithmetic
before each deref/slice, mirroring the fd-length guard the main frame path
already applies.
Compile + existing tests green on Linux .21 (real RTX 5070 Ti): pf-inject 74/0,
pf-zerocopy 17/0, pf-capture 1/0. The gadget deadlock path only executes on a
SteamOS host with raw_gadget/dummy_hcd (not reproducible on the CachyOS box), so
that fix is reasoned + compile-verified, not runtime-exercised.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
VulkanVideoEncoder::open_inner creates ~20 Vulkan objects across ~15
fallible steps, but all cleanup lived in the encoder's Drop — which only
runs once the value exists at the final Ok(Self). Any earlier ?/bail!
leaked everything built so far (a VkDevice + GPU memory per retried
open, and this backend is the default encode path on AMD/Intel Linux
hosts where open can fail transiently).
Factor the entire teardown sequence — unchanged — into a VkTeardown
guard whose Drop destroys any prefix of the build (vkDestroy*/vkFree*
are defined no-ops on VK_NULL_HANDLE): open_inner mirrors each object
into the guard as it is created and disarms it only at Ok(Self); the
encoder's own Drop rebuilds one from its fields, so both paths share
one sequence and cannot drift. make_frame now builds in place into a
guard-parked null Frame so a mid-build failure unwinds its partial
handles too, and make_video_image / vk_util::make_plain_image (also
used by the PyroWave backend) / build_parameters_h265 no longer leak
their own partially-created objects on failure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Completes the partial fix from the previous commit. The Windows PyroWave backend
caches its imported plane images keyed on the D3D11 texture's COM address, and
holds no reference on that texture — so once the capturer recreates its ring,
those addresses can be handed straight back out by the allocator and a
pointer-keyed cache hit returns an image bound to a texture that no longer
exists. Adding the extent to the key ruled out same-address-different-size
aliasing, but a recycle at identical dimensions still aliased.
The capturer already tracks exactly the value needed: `generation`, bumped on
every ring recreate. Plumbed it onto `PyroFrameShare` and the encoder now
flushes every cached import when it changes, which makes cache identity
independent of allocator behaviour rather than a bet against pointer reuse.
Validated on the RTX box: `pyrowave_win_smoke` (forced with `--ignored`, the
only test that actually exercises this path on real hardware) passes all ten
configurations — 1024²/720p/1080p/1440p across SDR/HDR and 4:2:0/4:4:4 — with
correct decoded chroma means, so the steady-state cache-hit path still works.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Windows direct-NVENC backend registers and encodes the capturer's textures
IN PLACE (no CopyResource), so how deep it may pipeline is a property of the
CAPTURER, not of the encoder. It was bounded only by `async_inflight_cap()` —
`PUNKTFUNK_NVENC_ASYNC_DEPTH`, default 4, clamped to the output-bitstream pool
— which consults nothing about the capturer, while the comment at the
backpressure loop claimed it "keep[s] in-flight depth within the capturer's
texture ring". It never did.
The IDD-push capturer rotates `OUT_RING = 3` per delivered frame with no regard
for encode completion (its own invariant note says OUT_RING(3) > max
pipeline_depth(2)). With the default async depth of 4 the encoder can therefore
still be reading a texture the capturer has already handed out again and
overwritten: torn or mixed frames. It is visual corruption rather than UB, so it
fails silently and intermittently — the worst shape to diagnose from a field
report.
Adds `Encoder::set_input_ring_depth`, reported from `Capturer::pipeline_depth`,
and bounds the async backpressure loop by `min(async_inflight_cap(), depth)`.
For IDD-push that yields 2, matching the capturer's stated contract; backends
that copy their input, or are synchronous, ignore it.
Wired at ALL THREE encoder-creation sites (initial open, stall/resize rebuild,
ABR rebuild) and forwarded through `TrackedEncoder` — this crate has a
documented trap where an unforwarded defaulted trait method silently no-ops
through that wrapper, which has already bitten the direct-NVENC work once and
the wire-chunking probe once.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Five of the nine medium findings from the pf-encode sweep. The remaining four
need cross-crate plumbing or an unwind refactor and are deliberately left out.
- vulkan_video `enqueue` waited on the backpressure fence with `u64::MAX`. That
wait runs ON the host encode thread — the same thread the stall watchdog's
`reset()` would run on — so a wedged GPU parked the one thread that could
recover the session: no error, no reset, and teardown blocking on the join.
This is the DEFAULT encode path for AMD/Intel Linux hosts (both shipped build
recipes enable `vulkan-encode` and `vulkan_encode_enabled()` defaults true).
Now bounded by ENCODE_FENCE_TIMEOUT_NS with expiry surfaced as an error.
- vulkan_video `import_cached` evicted a cached dmabuf import and destroyed its
image/view/memory with no fence wait, while up to `ring_depth - 1` submitted
frames may still reference it — a GPU-side use-after-free. `Drop` and `reset`
both idle first; this was the one unguarded destroy. Now idles before the
eviction loop, guarded on the length test so the steady state pays nothing.
- vulkan_video `read_slot` never asked for the encode's operation status, so a
FAILED encode was indistinguishable from a successful one and its feedback
was read as if it described real bitstream. Now requests WITH_STATUS_KHR and
refuses anything that is not COMPLETE.
- linux/pyrowave `encode_frame` opens its recording window early and has six
fallible steps inside it; every one returned with `cmd` still RECORDING, and
nothing repaired it (one `begin_command_buffer` in the file, and neither
`reset()` nor `Drop` touches `cmd`), so the next frame called `begin` on a
recording buffer — invalid usage. `submit` now resets the buffer on error;
legal on all these paths since the pool carries RESET_COMMAND_BUFFER and the
buffer is not pending.
- windows/pyrowave `encode_frame` ignored `frame.width`/`height` and imported
planes at the encoder's configured extent, so a ring recreate at a new mode
(the IDD capturer does this autonomously on a confirmed descriptor change)
read the planes under a stale VkImageCreateInfo. Added the size guard its QSV
and AMF siblings already carry, and keyed the plane cache on
(address, width, height) so a recycled COM address cannot resurrect an import
of a different size. NOTE: a recycle at the SAME size is still theoretically
possible; the complete fix keys on the capturer's ring generation and needs
that plumbed onto `PyroFrameShare`.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three defects from the pf-encode sweep, each adjudicated against source.
- Vulkan Video never received fecbec2d's taint sweep (it was carved out one
commit later). `pick_recovery_slot` accepts any resident slot whose wire is
below the CURRENT loss start, but "resident and older than this loss" is not
"the client decoded it": after an earlier loss [a,b] recovered at wire r,
everything in [a, r-1] is undecodable at the client — the lost frames plus
every frame that predicted through the gap — and those wires stay eligible
until the 8-slot ring rolls them out. A later loss could therefore anchor on
one and ship it tagged `recovery_anchor`, which is the client's definitive
re-anchor signal (punktfunk-core/src/reanchor.rs): the host lifts the client's
post-loss freeze onto a picture built from a reference it never had. Swept
before anchor selection, matching AMF/QSV.
`slot_wire` is blanked and `slot_poc` deliberately is NOT: `slot_poc` feeds
`build_h265_rps_s0`, which must keep naming every physically-resident DPB
picture or a conforming decoder evicts them and the anchor then references a
picture the client already dropped.
- QSV's sweep was incomplete, and in its MODAL case. `ltr_slots` mirrors the
hardware DPB, but nulling an entry issues no VPL call — the frame stays marked
long-term until that LongTermIdx is re-marked or an IDR flushes it (amf.rs
states this verbatim). The rejection loop iterates the post-sweep mirror and
only rejects `Some` slots, so it silently skipped the single entry the sweep
exists to distrust, leaving the recovery frame free to predict from it. With
NUM_LTR_SLOTS=2 the "exactly one slot swept" case is the common one, and the
two existing tests cover only the both-survive and both-swept cases. Taint is
now recorded in `ltr_tainted` with the FrameOrder left in place, so anchor
selection and the queued-force guard skip it while the rejection list still
names it.
- `read_slot` built a slice from the driver-reported (offset, bytes-written)
encode feedback with no validation against `bs_size`, so a driver reporting a
range outside the bitstream buffer produced an out-of-bounds read shipped
straight onto the wire. Checked in u64 (so the add cannot wrap) before
`map_memory`, so the error path has no unmap to unwind.
Adds `taint_sweep_excludes_slots_from_an_earlier_loss` covering the two-loss
case the existing single-loss test does not reach.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
MINOR, not patch: 44 commits since v0.14.0 split 21 features / 19 fixes, and
/api/v1 gained the per-scanner library-toggle endpoints (additive, already
regenerated into api/openapi.json by c2bba134). sdk-publish pushes this version
to consumers, so a patch bump would understate a new API surface to anyone
pinning ~0.14.
Headline work is the desktop clients drawing level with the Apple revamp: the
Windows client picked up settings parity, a findable console UI in the header,
the shared clipboard (with a per-host toggle), PyroWave decode in the codec
picker, and D3D11VA-first decode + HDR pass-through on Intel; the Linux GTK4
client got the same category-map settings rebuild. Apple landed the intent-based
presenter rebuild and the Dynamic Island redesign. Also: the punktfunk-host
plugins CLI, per-scanner library toggles in the console, and PyroWave raw-dmabuf
zero-copy capture on the Linux NVIDIA host.
Notable fixes: LTR-RFI loss recovery under sustained loss, two encode teardown
memory-safety holes, the audio first-open retry that was leaving sessions
silent, and the GameStream stream-marker announcement on the compat plane.
Every workspace crate is on version.workspace = true, so this stayed a one-line
bump plus the lock sync. (fec-rs, pf-driver-proto, usbip-sim, the Windows driver
crates and pf-vkhdr-layer are deliberately versioned independently and stay put.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
All four found in the pf-encode quality sweep and verified against source.
- NVENC partial-init leak (BOTH platforms, high): `init_session` publishes
`self.encoder` — and on Windows charges LIVE_SESSION_UNITS — *before* its
remaining fallible steps (bitstream buffers; on Linux also the input-surface
alloc and `register_resource`). A failure there left a live session with
`inited == false`, and every guard on the re-init path keys off `inited`, so
the next submit skipped teardown and overwrote `self.encoder`: the session
leaked permanently toward the driver's per-process cap, and its budget units
never returned, progressively starving parallel-display admission. `teardown`
already keys off `encoder.is_null()` rather than `inited`, so it cleans up
exactly this half-built state — it just was never called. Now invoked on the
`init_session` error path on both platforms.
- `can_encode_10bit` asked the wrong backend (medium): it resolved via
`linux_auto_is_vaapi`, which ignores `encoder_pref`, while `can_encode_444`
and `open_video` honour it. On a host that forces a backend (e.g.
`encoder_pref = "vaapi"` on an NVIDIA box) the probe answered for NVENC while
the session opened VAAPI, so the negotiated bit depth — and the HDR/SDR colour
label derived from it — described a backend that never ran. Now uses the same
`linux_zero_copy_is_vaapi` mirror, and `linux_auto_is_vaapi` carries a warning
that it resolves the `auto` case only and is not a dispatch mirror.
- Linux software arm ignored SW_BITRATE_CEIL (low): the Windows arm clamped
openh264 to 100 Mbps, the Linux arm passed the full negotiated rate. The
constant is now module-scope so both arms share one value.
- QSV/AMF env-parity (low): `PUNKTFUNK_IR_PERIOD_FRAMES` was a no-op on QSV
despite the comment claiming parity with AMF, and `PUNKTFUNK_NO_QSV_LTR` /
`PUNKTFUNK_INTRA_REFRESH` had dropped AMF's `trim()` and `yes`/`on` spellings,
so a value with stray whitespace silently did nothing on Intel while the same
value worked on AMD.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-19 20:47:55 +02:00
402 changed files with 53470 additions and 9627 deletions
assert_eq!((slot,wire),(3,3),"newest clean survivor is wire 3");
// Encoding resumes after recovery; wires 8..11 refill the swept slots and are clean. A
// later loss at wire 10 legitimately anchors on wire 9 — the sweep must not over-reject.
wires[4]=8;
wires[5]=9;
wires[6]=10;
wires[7]=11;
letplan=plan_slot_recovery(&view(&wires),10);
assert_eq!(plan.anchor,Some((5,9)),"wire 9 is post-recovery, clean");
apply(&mutwires,plan.tainted);
// A loss covering every live wire leaves nothing clean → decline, caller serves an IDR.
letmutall=[5i64,6,7,8,9,10,11,12];
letplan=plan_slot_recovery(&view(&all),5);
assert_eq!(plan.tainted,0b1111_1111);
assert_eq!(plan.anchor,None);
apply(&mutall,plan.tainted);
assert_eq!(pick_anchor(&view(&all),5),None);
}
}
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
