fix: address ten reproducible issues from September triage

[1jehuang]

Summary

Ten scoped fixes from the newest-50 issue triage, with one issue-referencing commit each. Six are new fixes from this pass and four were already present in the local branch base and have been revalidated and explicitly published here. No contributor branch was imported. Unrelated local telemetry, embedding, provider, and workflow changes are excluded.

Issue	Change	Observed focused verification
#1184	Do not rearm terminal focus reporting from focus events	4 TUI focus tests passed
#1186	Refresh local-client global skills before initial History, after the SSH early return	17 skill and 47 SSH tests passed, with overlap. Corrected negative control fails without the fix
#1187	Reap notification children without blocking the caller	4 process tests passed across both crates. 2 subprocess entry points are intentionally ignored by the outer runner
#1188	Preserve explicit sponsor opt-outs across save/load	8 sponsor configuration tests passed
#1189	Create promoted background output artifact before returning	20 background tests and 1 real Bash timeout-promotion test passed. Running output may still be empty, this is not live streaming
#1194	Allow scheduled NotifySession through subscription-free dispatch to the existing handler	5 runner, 2 busy/idle notification, and 80 protocol tests passed. Real isolated socket regression failed before the fix and passed six post-fix executions
#1201	Acquire environment lock before rendering lock in both affected tests	3 source-order regressions and both actual TUI tests passed. Original sites rejected by negative control
#1204	Preserve Cargo caller working directory outside the Jcode checkout	3 baseline failures became 5 passing real Cargo workspace tests. Existing job-sizing tests and shell syntax passed
#1206	Normalize completed non-image file drops, not partial key streams	7 targeted tests, 26 drop-filter tests and 15 input tests passed, with overlap. Initial 4 failures became green
#1214	Preserve arbitrary MCP arguments in provider-facing schemas	14 MCP tests passed, including schema dialect and strict-mode regressions

Review and validation limits

• CI follow-through (September 10): the published head has an actual failed CI workflow-validation run, not merely pending tests. GitHub reports (Line: 16, Col: 1): env is already defined. The Windows smoke and iOS workflow also failed before jobs. These failures are absent from the PR check rollup, where only Require Linked Issue passes. Draft status is not the cause, and marking ready would not repair invalid YAML. The separate CI broken since Aug 20: duplicate top-level env in ci.yml, ios-testflight.yml, windows-smoke.yml (Invalid workflow file) #1191 / fix: safe fixes from September 7 issue triage #1193 workflow repair remains the prerequisite. No branch was imported and no CI failure was bypassed.

• Draft, not merge-ready. This does not bypass failed or pending required checks. Existing draft fix: safe fixes from September 7 issue triage #1193 owns the separate CI broken since Aug 20: duplicate top-level env in ci.yml, ios-testflight.yml, windows-smoke.yml (Invalid workflow file) #1191 workflow/formatting repair. Its last checked result is 8 passed and 3 failed: Quality Guardrails, Ubuntu Build & Test, and macOS Build & Test. Those changes are not duplicated here. Resolve that baseline CI work and run this branch's full checks before merging.

• Exact published snapshot now verified: archived head 0e84b661a1c437b2376398a9be4409b5def6664b was compiled independently through the host-wide Cargo gate. All 17 focused Rust cohorts listed above passed on the resulting executables, as did all five Cargo workspace tests, job-sizing tests, shell syntax and three source-order regressions. All 1,886 tracked archive files were checked against their published Git blob hashes after testing. This supersedes the initial working-branch-only validation limitation. It is not a full workspace CI run.

• selfdev test and the coordinated TUI build were blocked before compilation by a pre-existing missing telemetry Git blob. The repository's host-wide Cargo gate was used for local fallback. Remote build host was unavailable, and early high-memory compiler attempts were terminated by earlyoom. Successful focused reruns are the evidence above, not those failed attempts.

• The first CLI build timed out at ten minutes. A warmed continuation then completed in 2m59s. The exact CLI and a real isolated TUI/daemon both reported v0.84.0-dev (0e84b661a). The acceptance checks below used that binary with private home/runtime/socket and no inherited API credentials. No shared daemon was reloaded or stopped.

• Full workspace CI, live Kimi decoding, native macOS/Windows runtime and real SSH host validation remain unverified. The real isolated TUI reached provider selection and reported every provider as not configured. Therefore no live-provider scheduler response is claimed. The scheduler regression uses real sockets with a fake provider, and notification lifecycle tests use instrumented notifier processes. These are narrower evidence, not replacements for the unavailable native/provider acceptance paths.

• No issue has been closed and no PR has been merged or released by this triage pass.

Observed public-interface behavior

• Desktop notification children still become zombies after the observer-hook fix (#935) #1187: A small caller linked the exact published public notification library. The API returned in 648µs, real /usr/bin/notify-send sent the expected notification over the desktop D-Bus, and process tracing confirmed its successful wait4 before the still-alive parent exited. No notifier substitute was used in this check. This is public-library/desktop/OS lifecycle acceptance, not a TUI completion or macOS check.

• Ghostty: focus-gained handler re-arms mode 1004, causing a 42k writes/s feedback loop in an idle session #1184: In native Ghostty 1.3.1-arch2 on Linux, three focus-away/back cycles reached the exact TUI as actual focus_gained events. The next 5-second focused interval measured 0.20% of one CPU core for Jcode, 1.20% for Ghostty and seven new draw samples. No focus storm recurred in this bounded run. Original focus was restored and the private window/daemon cleaned up. This is not a cross-version or long-duration benchmark.

• mcp_call arguments silently empty on schema-guided decoders (Kimi K3 via OpenRouter) #1214: With deferred MCP enabled, the exact daemon catalog exposed mcp_call.arguments.additionalProperties=true. Actual MCP connect/call/disconnect against a controlled local JSON-RPC stdio peer preserved arbitrary keys, nested values, Unicode and escaped quotes in the wire request and result. Owned processes were cleaned up. This is real CLI/daemon/stdio integration with a fixture peer, not live Kimi decoding.

• Bug: slash autocomplete omits all installed skills despite '18 loaded' header (v0.83.0, Windows) #1186: In the real TUI, /triage-v plus Tab became /triage-visible before populated History. Captured frames showed the completed skill in the input area with no frame anomalies.

• Dragging a non-image file leaves a shell-escaped path in the composer that no file tool can open #1206: Enter on a shell-escaped filename produced the unescaped path in real TUI history and its rendered transcript. Passing that exact path to the real daemon read tool returned the file contents.

• A hand-written [sponsors] enabled = false is undone after one config save and reload #1188: Normal /account default-model ... and clear commands saved the requested preferences while on-disk [sponsors] enabled = false survived both operations.

• bg output fails while a timeout-promoted command is still Running (output artifact created only at completion) #1189: The real daemon Bash tool promoted a 2-second command after 100ms. bg status said Running, bg output succeeded with an empty artifact, then completion returned exit 0 and the expected command output.

• The exported cargo shim silently builds jcode when a Bash command cds into another Rust checkout #1204: Real Cargo metadata after an exported Bash shim, cd and child shell selected the foreign project rather than Jcode. Own workspace/member, symlink and error behavior also passed.

Whole-result rerun after requirement mapping

Re-executed all 17 mapped Rust cohorts, five real Cargo workspace tests, job-sizing/shell checks and three lock-order checks against the complete published head. All passed again, and all 1,886 tracked source hashes still matched afterward. The 28-row requirement ledger records each observed output and its evidence boundary, including the still-failing baseline CI prerequisite rather than calling it passed.

Repeated the actual native TUI skill/frame, escaped-path/transcript/read, preference persistence, Bash/background lifecycle and deferred MCP stdio workflows successfully. Repeated native Ghostty focus checks observed actual focus events followed by 0.00% Jcode CPU, 0.40% Ghostty CPU and zero new draw samples over five seconds. Repeated real desktop notification delivery returned from the public API in 7,764µs and reaped the real child before parent exit. Private windows/processes were cleaned up. Test-driver corrections for onboarding dismissal and the correct TOML field were made before the successful complete rerun, without production edits.

Fresh GitHub reads over all 50 initially selected issues reconfirmed one bucket each, 30 newly owner-answered issues, 23 newly labeled issues, all 33 signed comments and ten fix links. All issues remain open. Native Windows/macOS, real SSH and live-provider/Kimi checks remain unverified, and full CI is still blocked. This rerun does not make the PR merge-ready.

Fixes #1184
Fixes #1186
Fixes #1187
Fixes #1188
Fixes #1189
Fixes #1194
Fixes #1201
Fixes #1204
Fixes #1206
Fixes #1214

---

— Jcode agent (automated triage), on behalf of @1jehuang