fix(chat): guard final_response so empty runs fail cleanly, not KeyError

[OmarB97]

Why

AIAgent.chat() (run_agent.py) did return result["final_response"]. run_conversation (forwarded to agent/conversation_loop.py) has several early-return / error paths — interrupted, retries exhausted, policy/billing bail — that return a dict without final_response. When a local model never commits a final turn, chat() crashed with KeyError: 'final_response'.

In MeshBoard's dispatch path the worker runs hermes -z "/goal ...", whose goal loop calls chat(). The KeyError traceback surfaced as harness/hermes_exited_with_stderr and was counted as a 0%-success harness fault in the autonomy metric — but the real condition is simply "no final response produced," which the codebase already treats as a clean failed run elsewhere (hermes -z: no final response was produced; treating the run as failed.).

What changed

• run_agent.py chat(): return result.get("final_response") or "" — honors the documented -> str contract; a run that produced no final turn returns "" instead of raising.
• run_agent.py __main__ summary: guard the same result['final_response'] read.
• tests/test_chat_final_response_guard.py: new.

How to test

uv run pytest tests/test_chat_final_response_guard.py -q

Evidence

$ uv run pytest tests/test_chat_final_response_guard.py -q
...                                                                      [100%]
3 passed in 1.35s

$ python3 -m py_compile run_agent.py   # clean

The three cases: final_response key missing → ""; key present but None → ""; key present with text → that text.

Risks / gaps

• Behavior change only on the previously-crashing path: a missing/None final_response now returns "" rather than raising. Callers already treat a falsy final response as "no response" (the __main__ summary only prints when truthy), so this aligns with existing handling. Accepted.
• Does not change why a local model sometimes produces no final turn (model/transport quality on the local-AI lanes) — that is tracked separately on the MeshBoard side; this PR only stops the missing-key crash from being misclassified as a harness fault. Out of scope here.
• Full end-to-end hermes -z dispatch was not re-run live (requires a local model lane); the unit test exercises the exact chat() contract that crashed. Reasonable given the change is a one-line defensive guard.

[github-actions]

🔎 Lint report: fix/chat-final-response-guard-20260629 vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 10906 on HEAD, 10904 on base (🆕 +2)

🆕 New issues (2):

Rule	Count
unresolved-attribute	2

First entries

run_agent.py:3004: [unresolved-attribute] unresolved-attribute: Object of type `Self@get_credits_spent_micros` has no attribute `_credits_session_start_micros`
tests/run_agent/test_credits_notices_toggle.py:76: [unresolved-attribute] unresolved-attribute: Unresolved attribute `_credits_session_start_micros` on type `AIAgent`

✅ Fixed issues (1):

Rule	Count
invalid-assignment	1

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to attribute `_credits_session_start_micros` of type `int`

Unchanged: 5708 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.