Greptile Summary

This PR fixes a bug in resolveLiveSessionModelSelection() where cross-provider fallback candidates were incorrectly treated as user-requested live model switches for agent sessions. The root cause was that when an agentId was present, the function used the agent's configured default model as the session baseline instead of the current attempt's provider/model — causing hasDifferentLiveSessionModelSelection to fire a LiveSessionModelSwitchError on every fallback attempt.

Changes:

• live-model-switch.ts: Removes the resolveDefaultModelForAgent call and uses params.defaultProvider/params.defaultModel directly as the baseline. The call site in pi-embedded-runner/run.ts already passes the current attempt's provider and model for these fields, so no change was needed there.
• live-model-switch.test.ts: Removes the now-unnecessary resolveDefaultModelForAgent mock and adds a dedicated test for the no-override fallback path.
• model-fallback.run-embedded.e2e.test.ts: Adds an e2e regression test covering cross-provider fallback for agent sessions without explicit session store overrides (agentId: "main").

Confidence Score: 5/5

Safe to merge — targeted, well-understood fix with matching unit and e2e regression coverage.

The change is minimal and surgical: one function call removed, two field accesses simplified to use the already-correct caller-supplied values. The call site in run.ts already passes the current attempt's provider/model, so the fix is properly integrated. New unit test and e2e regression test both cover the exact scenario described in the root cause. No P0 or P1 issues found.

No files require special attention.

_{Reviews (1): Last reviewed commit: "fix(agents): preserve fallback candidate..." | Re-trigger Greptile}

Follow-up validation after local deploy:

• rebuilt the patched repo and restarted the local gateway on build 0a9bfa8
• exercised the previously failing Telegram/live-session path directly against session 27c82782-4e2a-40a3-a626-f7b94b796154
• command used:

  openclaw agent --agent main --session-id 27c82782-4e2a-40a3-a626-f7b94b796154 --message 'Reply with POST_DEPLOY_TELEGRAM_FALLBACK_OK and nothing else.' --json\n  ```\n- result completed on `openai-codex/gpt-5.4` with payload `POST_DEPLOY_TELEGRAM_FALLBACK_OK`\n- fresh logs for run `f2eded73-809c-4d58-a0ff-17d6316a343f` showed Anthropic `429` on the primary attempt, then `model_fallback_decision` `candidate_succeeded` for `openai-codex/gpt-5.4`\n- importantly, there were no new `live session model switch detected before attempt` lines for that run\n\nSo the follow-up commit fixes the remaining Telegram/live-session regression that still reproduced after the first patch.

Staff Engineer Review: APPROVED

PR: fix(agents): preserve fallback candidates in live sessions
Author: goldmar

Root Cause Analysis: CORRECT

The fix properly addresses the root cause. resolveLiveSessionModelSelection() was using resolveDefaultModelForAgent() as the baseline whenever an agentId existed. During fallback retries, the runner compared the retry candidate (e.g., openai:gpt-5.4) against the agent-default baseline (e.g., anthropic:claude-opus-4-6) and threw LiveSessionModelSwitchError — treating internal retry candidates as user-initiated live switches.

The fix removes resolveDefaultModelForAgent() entirely and uses the passed-in defaultProvider/defaultModel (the current attempt's provider/model) as the baseline. Cross-provider fallback now works because the baseline moves with each retry attempt.

Code Quality: SOUND

normalizeSelectedAuthProfileSource — Clean utility. Correctly normalizes authProfileOverrideSource to auto | user | undefined. No edge cases missed.

resolvePersistedAuthProfileSelection — Correctly filters out auto-selected profiles during fallback. The compaction metadata heuristic is a sound backwards-compat approach for distinguishing legacy data.

hasDifferentLiveSessionModelSelection — The new comparison logic handles all edge cases:

• Same profile + different source → detected as change
• Missing → present → detected as change
• Present → missing → detected as change
• Both undefined → not a change

Return value consistency — The final return is cleaner: always returns authProfileIdSource directly. resolvePersistedAuthProfileSelection already handles the undefined case, so behavior is preserved.

Test Quality: STRONG

Unit tests cover all auth profile source edge cases including the tricky legacy/stale compaction metadata scenarios. The E2E tests directly validate the bug:

• "falls back across providers for agent sessions without explicit model overrides" — the core fix
• "ignores auto-selected persisted auth profiles when crossing providers during fallback" — regression coverage

Mock cleanup is correct: resolveDefaultModelForAgentMock was properly removed since that function is no longer called.

Minor Observation (non-blocking)

The compaction-count heuristic is a reasonable backwards-compat inference, but it's not obvious from the call site. A brief code comment would help future readers. Not worth blocking over.

VERDICT: APPROVED

The fix is correct, the tests are thorough, and the auth profile source handling is more rigorous than before. No blockers.

@joshavant @tyler6204 I addressed the follow-up review comments and the current checks are looking healthy again. This should be ready for another look when you have a minute.

Looks like this was independently addressed in bd89e07 on main — closing this one. Thanks for the quick fix!