fix(rescue): pass timeout: 600000 on the inner task Bash call

[ultsaza]

Problem

The /codex:rescue command dispatches to the codex:codex-rescue subagent, which makes one Bash call to codex-companion.mjs task. Neither the agent nor the codex-cli-runtime skill tells the model to pass a timeout, so the call inherits Claude Code's default Bash timeout of 120s. Codex rescues against non-trivial diffs routinely take several minutes; they get SIGTERM'd mid-turn at the 120s cap, leaving the companion job dangling in status: running.
The 120s default and 600s ceiling are documented in Claude Code's Bash tool reference under "Timeout: two minutes by default. Claude can request up to 10 minutes per command with the timeout parameter." Both are overridable via BASH_DEFAULT_TIMEOUT_MS / BASH_MAX_TIMEOUT_MS env vars, but passing timeout: 600000 explicitly works regardless of the user's BASH_DEFAULT_TIMEOUT_MS.

Reported in #122.

Fix

Add one bullet in each of two places, requiring timeout: 600000 (the Bash tool's 10-minute maximum) on the inner task invocation:

• plugins/codex/agents/codex-rescue.md (Forwarding rules)
• plugins/codex/skills/codex-cli-runtime/SKILL.md (Command selection)

Small rescues are unaffected; large-diff ones now have room to finish.

Why not the same approach as PR #239

PR #239 fixes the same 120s-cap symptom for /codex:review and /codex:adversarial-review by embedding a Bash({timeout: 600000}) block directly in those command templates. /codex:rescue cannot be fixed that way because its Bash call is constructed dynamically by the subagent, not declared in the command template — so the instruction has to live in the agent and the runtime skill that build the call.

Relationship to other open PRs

• PR fix: keep rescue task calls in foreground #214 (closes [BUG] codex:codex-rescue hangs in worktrees when run_in_background is used #198) touches the same two files but on different lines, addressing the orthogonal run_in_background problem. The two PRs merge cleanly in either order.
• PRs fix(runtime): prevent tracked jobs hanging forever on broker disconnect #184 / fix(runtime): add wall-clock timeouts to JSON-RPC request and captureTurn #302 / feat(codex): per-turn watchdog with TurnWatchdogError + exit 124 #312 add wall-clock timeouts at the captureTurn / RPC layer. They are defense-in-depth at a different layer and remain useful even with this PR landed, since they catch app-server-side hangs that a Bash-tool timeout cannot.

Test plan

• node --test tests/commands.test.mjs → 8/8 pass. The new contract assertions hold (Pass \timeout: 600000`...` matches in both the agent and the runtime skill).
• npm test (full suite): 82/86 pass. The same 4 failures are present on a pristine origin/main checkout — verified by temporarily replacing only the three files this PR modifies with their origin/main versions (git checkout origin/main -- <files>) and re-running npm test in the same environment. Identical failure set. The four are already documented as pre-existing in PR fix(runtime): add wall-clock timeouts to JSON-RPC request and captureTurn #302 and PR fix(runtime): prevent tracked jobs hanging forever on broker disconnect #184 bodies:
  • status shows phases, hints, and the latest finished job
  • status preserves adversarial review kind labels
  • result returns the stored output for the latest finished job by default
  • resolveStateDir uses a temp-backed per-workspace directory
• Manual: a /codex:rescue invocation over a ~30-file diff that previously died at the 120s cap completes end-to-end. (Not run from the session that prepared this PR.)

Closes #122