Fix Codex subagent final result preservation

[SmallSpider0]

Summary

Fix Codex subagent result handling so completed subagent cards preserve the actual final assistant message instead of sometimes showing only a terminal status such as done.

Problem

In some Codex multi-agent flows, a child agent could complete successfully but the parent-side result would only contain a completion marker (done / completed) rather than the child agent's final message.

This could happen when:

• wait_agent returned a completed/done status without a message
• child terminal events and child message events arrived in a non-ideal order
• later child trace/tool events arrived after the card had already reached a terminal state

Changes

• Track each child Codex runtime's latest final agent_message
• Use that cached final message as fallback when wait_agent reports completed/done without a message
• Normalize done, complete, and completed agent states consistently
• Preserve structured completed-state payloads instead of collapsing them to bare completed status
• Prevent terminal child agent cards from regressing back to running due to late child events
• Allow explicit parent operations like send_input / resume_agent to remain visible without letting late child tool events overwrite terminal state
• Reset cached final message and terminal state on child task_started
• Preserve late child messages as completed result updates instead of writing/running updates

Tests

cd cli
bun run test -- src/codex/utils/appServerEventConverter.test.ts src/codex/codexRemoteLauncher.test.ts
bun run typecheck

Also smoke-tested against a deployed local HAPI build with real Codex subagents:

• spawned two subagents
• waited for both
• verified wait_agent result contained full final messages, not only done

Notes

This is a clean resubmission from the latest upstream main, replacing the closed PR #599 without force-pushing the old branch.

[github-actions]

Findings

• [Major] Empty status message suppresses structured output — statusObjectFromAgentState now preserves completed agent states such as { status: 'completed', message: '', output: { value: 42 } }, but the launcher reads record.message with nullish coalescing, so an empty string is treated as the final result and the structured output is dropped. That reintroduces the blank/terminal-only completed card for this added case. Evidence cli/src/codex/codexRemoteLauncher.ts:962.
  Suggested fix:

  const extractAgentStatusMessage = (record: Record<string, unknown>): unknown => {
      const message = asString(record.message);
      if (message) return message;
      if (record.output !== undefined && record.output !== null) return record.output;
      if (record.result !== undefined && record.result !== null) return record.result;
      if (record.finalMessage !== undefined && record.finalMessage !== null) return record.finalMessage;
      return record.final_message;
  };

Summary

• Review mode: initial
• One issue found: completed child-agent states with blank message still lose structured output in the launcher.

Testing

• Not run (automation)

HAPI Bot

[github-actions]

Findings

• [Minor] Reset the new harness flag between tests — emitChildWaitStructuredOutput is enabled by the new structured-output test, but afterEach resets the neighboring child-event flags without resetting this one. Because the harness object is shared, every later test now runs with agentsStates carrying message: '' and output: { value: 42 }, so the suite becomes order-dependent and can mask regressions in the normal child-message path. Evidence cli/src/codex/codexRemoteLauncher.test.ts:1078.
  Suggested fix:

  harness.emitChildDoneStatusWithoutMessage = false;
  harness.emitChildWaitStructuredOutput = false;
  harness.emitChildTaskCompleteBeforeMessage = false;

Summary

• Review mode: follow-up after new commits
• One test-isolation issue found in the follow-up changes. The prior structured-output finding appears addressed in the launcher path.

Testing

• Not run (automation; bun is unavailable in this runner: /bin/bash: bun: command not found)

HAPI Bot

[github-actions]

Findings

• No findings.

Summary

• Review mode: follow-up after new commits
• No issues found in the latest PR diff. The prior HAPI Bot test-isolation finding is fixed in the current head.
• Residual risk: Codex app-server event ordering is hard to exhaustively validate from the diff alone.

Testing

• Not run (automation/review only; PR code not executed)

HAPI Bot