Missing automated test for new subagent fallback branch

The new isSubagent branch in resolveFallbackRetryPrompt is not covered by any automated test. The existing agent.test.ts suite never inspects the prompt argument forwarded to runCliAgent, so the case isFallbackRetry=true && isSubagentLane=true is only validated by the manual test plan in the PR description.

Because this is the only meaningful code path changed by the fix, a regression here could go undetected. Consider adding a test (either by exporting the function for unit-testing, or by asserting on the prompt captured by the runCliAgentSpy when lane is set to AGENT_LANE_SUBAGENT and the fallback run is triggered):

it("preserves body on model fallback retry for subagent lane", async () => {
  // configure runCliAgent to throw FailoverError on first call, succeed on second
  runCliAgentSpy
    .mockRejectedValueOnce(new FailoverError("timeout", "timeout"))
    .mockResolvedValue(/* ... */);

  await agentCommand({ body: "Do the task", lane: AGENT_LANE_SUBAGENT, ... });

  // second call (fallback) must still carry the original task, not the generic prompt
  const secondCallPrompt = runCliAgentSpy.mock.calls[1]![0].prompt;
  expect(secondCallPrompt).toBe("Do the task");
});

This is a comment left during a code review.
Path: src/commands/agent.ts
Line: 127-144

Comment:
**Missing automated test for new subagent fallback branch**

The new `isSubagent` branch in `resolveFallbackRetryPrompt` is not covered by any automated test. The existing `agent.test.ts` suite never inspects the `prompt` argument forwarded to `runCliAgent`, so the case `isFallbackRetry=true && isSubagentLane=true` is only validated by the manual test plan in the PR description.

Because this is the only meaningful code path changed by the fix, a regression here could go undetected. Consider adding a test (either by exporting the function for unit-testing, or by asserting on the `prompt` captured by the `runCliAgentSpy` when `lane` is set to `AGENT_LANE_SUBAGENT` and the fallback run is triggered):

```ts
it("preserves body on model fallback retry for subagent lane", async () => {
  // configure runCliAgent to throw FailoverError on first call, succeed on second
  runCliAgentSpy
    .mockRejectedValueOnce(new FailoverError("timeout", "timeout"))
    .mockResolvedValue(/* ... */);

  await agentCommand({ body: "Do the task", lane: AGENT_LANE_SUBAGENT, ... });

  // second call (fallback) must still carry the original task, not the generic prompt
  const secondCallPrompt = runCliAgentSpy.mock.calls[1]![0].prompt;
  expect(secondCallPrompt).toBe("Do the task");
});
```

How can I resolve this? If you propose a fix, please make it concise.

Limit subagent body replay to first fallback turn

resolveFallbackRetryPrompt now returns the full body for every fallback attempt when isSubagent is true, but the subagent lane is used beyond initial spawn (persistent follow-ups/steer reruns), not only for first-task bootstrapping. In cases where the first model attempt already recorded the user turn before failing (the same duplicate-turn scenario guarded by fallbackAttemptIndex in agentCommandInternal), this change re-injects the same instruction on retry and can repeat side-effecting tool actions; the body replay should be scoped to initial subagent task bootstrapping rather than all subagent retries.

Useful? React with 👍 / 👎.

Detect initial subagent spawn independently of session freshness

isInitialSubagentSpawn is derived from isSubagentLane && isNewSession, but real sessions_spawn runs pre-create the child session before invoking agent (they call sessions.patch first in src/agents/subagent-spawn.ts, and applySessionsPatchToStore creates a fresh entry with sessionId/updatedAt in src/gateway/sessions-patch.ts). That makes resolveSession report isNewSession = false, so fallback retries still use the generic continuation prompt and drop the original [Subagent Task] text when the first model attempt times out. In other words, this condition misses the production spawn path and leaves the reported fallback-context loss unresolved.

Useful? React with 👍 / 👎.

Identify subagent first turn without freshness gating

resolveFallbackRetryPrompt now preserves the original body only when isInitialSubagentSpawn is true, but this flag is computed as isSubagentLane && isNewSession; in the real sessions_spawn flow, src/agents/subagent-spawn.ts calls sessions.patch before invoking agent, and applySessionsPatchToStore creates a fresh session entry, so resolveSession reports isNewSession = false even on the first subagent task. That means fallback retries still switch to the generic continuation prompt and drop [Subagent Task] text in the production spawn path this change was meant to fix.

Useful? React with 👍 / 👎.