🤖 fix: fence a child's stale goal-agent refusal and fix its recovery wording

[ThomasK33]

Summary

Fixes #5452. A sub-agent's fail-closed goal-agent check (#5402) awaited without fencing its dispatch. A refusal that went stale during the check could show an obsolete chat error and pause a goal that a newer attempt (or a just-resumed goal) owned. The refusal text also told the user to "select an available agent", which cannot help a child: the picker does not repin its agentType.

Implementation

1. Staleness fence (item 1). TaskService.sendChildGoalTurn builds a synchronous stillCurrent(goal) check. It holds while the captured attempt is still the task's current attempt and the goal keeps the same goalId, status and user-activation stamp (every user Resume restamps it).
   • The chat error: refuseUnavailableGoalTurnAgent takes a probe (it may now be async). The probe reads the goal under the goal file lock, then applies stillCurrent, so the attempt is checked after that await.
   • The pause: pauseForUnavailableAgent passes the fence to the write. A new internal SetGoalInput.stillCurrent is evaluated under the goal file lock next to expectedGoalId, and a refusal is a goal_conflict. reserveBudgetWrapupForRedispatch takes the same fence for a budget-limited goal.
   • A stale refusal therefore changes nothing. The stream end takes its normal path, which has its own attempt fences. Fail-closed behavior is unchanged for a current dispatch.
2. Child wording (item 2). resolveAgentForStream's goal-turn refusal for a sub-agent names the pinned agent to restore or enable before reactivating the task. It no longer asks the user to select an agent. It does not promise a goal resume, because a reported child cannot resume its goal yet (🤖 Sub-agent goals: follow-ups (parent spend ceiling, UI test, deleted child agent) #5411 item (c), deferred).

Validation

Test-first. Before the fix:

Expected: "active"
Received: "paused"
(fail) TaskService child goals > a refusal whose attempt was replaced during the check leaves the goal alone
Expected: false
Received: undefined
(fail) TaskService child goals > a refusal whose goal was paused and resumed during the check shows no chat error
error: expect(received).not.toContain(expected)
Expected to not contain: "select an available agent"
(fail) AIService.streamMessage … > a sub-agent's goal turn does not fall through to a later candidate or exec

All three pass after the fix. A further test, "an attempt replaced just before the refusal's pause write keeps its goal", fails if the write fence is removed.

Deferred

• A pause write racing an attempt rotation during its own file I/O, either cross-store or with two backends. This is fail-safe: the goal ends paused. It is tracked with 🤖 Sub-agent goals: follow-ups (parent spend ceiling, UI test, deleted child agent) #5411 (b): 🤖 Sub-agent goals: follow-ups (parent spend ceiling, UI test, deleted child agent) #5411 (comment)

Risks

Low. The change only drops a refusal whose dispatch is stale, and only on the child goal-turn path. The newer attempt or goal runs its own fail-closed check before any goal turn. The root goal dispatcher's synchronous probe is still evaluated in the same tick as its chat error, with no await (tested).

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-02T10:11:55.400873Z	2a83c22	New commits
🔒 Security Review	✅ Completed 2026-10-02T10:13:41.340641Z	2a83c22	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 90cfb0fe30

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 57afa72d14

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 200f6e9a15

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5c69b7d4f7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ThomasK33]

Readiness record (issue coordinator lane, workspace-63):

• Head: 2a83c22bfa79d34ab8f676c6deb1433949074147. All required checks pass.
• Assessments: 5 normal and 5 security review rounds, plus 1 final independent check, for 6 per category.
  • Normal-review findings by round: 3, 2, 1, 3 (two were duplicates, so 2 unique), then 0.
  • Every security review was clean.
  • All findings were fixed except one round-4 P1. That P1 is a pause write racing an attempt rotation during its own file I/O. It is fail-safe (the goal ends paused) and predates this PR. It is deferred to 🤖 Sub-agent goals: follow-ups (parent spend ceiling, UI test, deleted child agent) #5411 (b): 🤖 Sub-agent goals: follow-ups (parent spend ceiling, UI test, deleted child agent) #5411 (comment)
• Final independent check (a fresh clean-context agent, on this head): ready. It found the deferral acceptable.
• All threads are resolved.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high