Suppress each raw commentary echo paired with a typed Codex item completion by protocol order, while preserving later raw-only notes and contributor-rewritten completion text.
Fixes#93296.
Thanks @Marvinthebored.
Verification:
- 95 focused projector tests
- full CI run 27593515603
- real behavior proof run 27593522821
- local and whole-branch autoreview clean
Co-authored-by: Peter Lindsey <peter@lindsey.jp>
Pinned session-extension registries now remain the owner even when empty, preventing later active registry churn from leaking agent-owned extensions into the gateway surface.
Inbound PDF/document text already flows to agents through the canonical
media-understanding pipeline (applyMediaUnderstanding -> extractFileBlocks),
but it inherited the OpenResponses input_file limits (5MB / 4 pages), so large
managed PDFs from channels/Control UI were skipped and locked-down agents saw
only an attachment marker.
- Size inbound file extraction from agents.defaults.mediaMaxMb (default 20MB,
cap 25MB) and pdfMaxPages (default 20, cap 150) via a new
resolveFileExtractionLimits; explicit gateway responses.files config still
wins per-field. (#90096)
- chat.send: let oversized (>5MB) managed inbound PDFs pass through sandbox
staging with their managed media path instead of a 4xx, so host-side
extraction reaches sandboxed agents without copying the file into every
sandbox; non-PDF oversize files are still rejected. (#90097)
Reuses the existing extraction/injection path; no parallel module or extra
prompt-injection sites.
When a child openclaw process is spawned via a backgrounded subshell that
exits before the new process reaches the stale-pid sweep, the new process
is reparented to the supervisor (PID 1 / launchd) and the ancestor walk
in getSelfAndAncestorPidsSync can no longer see the running gateway. The
running gateway then shows up on lsof as an unrelated sibling on the
port and gets SIGKILL'd by cleanStaleGatewayProcessesSync, recreating
the issue #68451 supervisor restart loop across a reparent boundary.
Real-world trigger: a user ~/.zshrc auto-start block
if ! pgrep -x openclaw-gateway >/dev/null; then
(openclaw gateway >/dev/null 2>&1 &)
fi
combined with codex per-turn `zsh -c "set -e; . shell_snapshot"` invocations
caused every chat turn on rh-bot to SIGKILL its launchd-managed gateway,
producing HTTP 000 errors and ~33 kill events captured by a forensic
launchd unified-log tracker before the zshrc was patched.
Fix: gateway-cli captures OPENCLAW_GATEWAY_SERVICE_PID from inherited env
BEFORE overwriting it with process.pid, then threads the captured PID
through cleanStaleGatewayProcessesSync into getSelfAndAncestorPidsSync's
exclusion set. The protection is opt-in per call site so existing
maintainer paths (openclaw update / openclaw doctor restart helpers) keep
their ability to terminate a running gateway intentionally.
The inherited-PID parser is strict positive-integer only: a malformed
inherited env value (`"123abc"`, `"123.4"`, `"0x7b"`, etc.) is rejected
rather than silently protecting PID 123 from cleanup and leaving the
stale listener alive. New focused unit tests cover the parser
contract.
Existing regression tests cover the reparent suicide-kill scenario and
the defensive ignore-non-positive-PID contract on the cleanup side.
Preserve rollback journaling for NFS and SMB-backed stores, refuse SSHFS after symlink-aware mount classification, and close Workboard database handles when filesystem policy rejects initialization.
Use transactionally consistent VACUUM INTO snapshots for every state-root SQLite database and exclude original journal sidecars so verified backups cannot restore torn plugin or memory state.
* test(qa): add smoke ci primary coverage evidence
* test(qa): remove overstated primary coverage claims
* test(qa): make release profile include smoke ci
* test(qa): trim taxonomy formatting churn
* test(qa): avoid hardcoded profile names in coverage test
* test(qa): make release profile cover taxonomy
* test(qa): type profile fixture all category flag
* test(qa): include channel delivery in smoke ci profile
