fix(rustyclawd): replace Bash-tool wall-clock cap with idle-liveness reaping (#2607)

[rysweet]

Summary

Resolves #2607 — no wall-clock timeout may kill in-flight agent/LLM work.

The #2607 audit swept every tokio::time::timeout / wait_timeout / recv_timeout on a child/agent/LLM path. Exactly one remaining wall-clock kill on local agentic work was found: the RustyClawd adapter's Bash tool. This PR replaces it with idle-liveness, mirroring the canonical meeting agent proxy (src/meeting_backend/agent_proxy.rs).

The Signal parity piece of #2607 (signal_agent_mode() → Meeting idle-liveness proxy) already landed on main in #2624; this PR closes out the remaining audit item.

The bug

The Bash arm wrapped child.wait_with_output() in tokio::time::timeout(120s, …) and spawned with ProcessSpawnConfig::default() (no isolation). A command still streaming output at 120 s was SIGKILLed mid-flight → ClientError::Timeout. Worse, because the child was neither process-group isolated nor signalled on future-drop, it was orphaned — left running detached while the tool call reported failure.

The fix (idle-liveness)

• Spawn isolated via ProcessSpawnConfig::with_isolation() (setsid): the child leads its own process group (pid == pgid) so the reaper can SIGKILL the whole subtree.
• Stream stdout/stderr over a channel; every chunk resets last_activity. A command that keeps producing output runs unbounded regardless of total runtime. Only sustained silence for the whole window reaps the child, killing the whole process group (numeric-PID libc::kill(-pgid, SIGKILL), shell-free signal policy) so no orphan survives. The error honestly identifies an IDLE reap, never a productive kill.
• New env knob SIMARD_RUSTYCLAWD_IDLE_LIVENESS_SECS (seconds) sets the window; 0 = fully unbounded escape hatch. Unset/malformed falls back to the model-supplied per-call timeout (ms, default 120 s), now reinterpreted as an idle window rather than a total budget.

Tests (the three #2607 acceptance scenarios + resolver)

• bash_producing_output_past_the_window_is_never_killed_2607 — a command emitting output past the window runs to completion, all output captured.
• bash_idle_child_is_reaped_with_honest_error_and_no_orphan_2607 — a genuinely idle child is reaped after the window with an honest "idle" error, and no PID retains the command's marker (whole process group reaped, no orphan).
• bash_env_zero_disables_idle_reaping_2607 — SIMARD_RUSTYCLAWD_IDLE_LIVENESS_SECS=0 disables reaping.
• Pure resolve_idle_window cases: positive override wins, 0 → None, unset/malformed → per-call fallback.

Audit — wall-clock caps on agent work (keep vs replace)

Location	Semantics	Verdict
base_type_rustyclawd/tool_executor.rs — Bash arm	(was) wall-clock 120 s on a tool child; killed productive work and orphaned it	REPLACE → idle-liveness (this PR)
meeting_backend/agent_proxy.rs — idle window	idle-liveness (resets per chunk; 0 = unbounded)	KEEP — canonical template
meeting_backend/agent_proxy.rs — recv_timeout(200 ms)	poll tick, not a turn cap	KEEP
signal_conversation/channel.rs — signal_agent_mode()	routes Signal turns through the Meeting idle-liveness proxy	KEEP (already merged, #2624)
engineer_loop/agent_spawn.rs — wait_with_output()	unbounded natural wait	KEEP — already compliant
agent_supervisor heartbeat staleness	idle-based (SIGTERM on stale heartbeat), not wall-clock	KEEP
terminal_session idle detection	idle-based, work-process aware	KEEP
meeting_backend/close_guard.rs	graceful post-turn meeting-close budget (shutdown, not a turn)	KEEP
copilot_task_submit/orchestration.rs	bounded polling of a remote GitHub agent job	KEEP — remote poll, not local agent I/O
operator_commands_dashboard/*	VM/SSH/tmux/websocket infra ops	KEEP — non-agentic infra
bridge_subprocess/subprocess.rs, update_check.rs	channel-framing / update-check reads	KEEP — non-agentic

Keeps are deliberate. The rule targets agent turns and LLM subprocesses; a remote-job poll, a channel-framing read, or a graceful-shutdown budget is legitimately bounded and stays bounded.

Docs

New reference docs/reference/rustyclawd-bash-tool-idle-liveness.md (behavior, configuration, mechanism, full audit, invariants), linked from docs/index.md and docs/reference/base-type-adapters.md. mkdocs build --strict passes.

Verification

• cargo test --lib base_type_rustyclawd::tool_executor — 24 passed.
• cargo clippy --release -- -D warnings and full --all-targets --all-features --locked clippy — clean.
• cargo fmt --all -- --check — clean.
• mkdocs build --strict — clean.
• Rebased onto latest origin/main; diff is exactly the 5 intended files.

Closes #2607

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[github-actions]

📊 Coverage Summary

Generated by cargo llvm-cov --workspace --summary-only (nightly, excluding test files)

Module	Lines	Covered	Coverage
Total	140496	116585	83.0%

_{Coverage data from CI run. Test files matching tests?/ are excluded from line counts.}