fix(agent): chat() surfaces error instead of KeyError on no-final_response result

[OmarB97]

What does this PR do?

AIAgent.chat() raised KeyError: 'final_response' whenever run_conversation
returned via one of its error/failure paths. Those paths (API error after max
retries, billing/credits exhaustion, policy/toolguard halt, interrupts) return a
result dict shaped like {messages, completed, api_calls, error, failed} with
no final_response key, so return result["final_response"] crashed —
masking the real error behind a KeyError traceback.

This makes chat() surface the error message and degrade gracefully instead of
crashing.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• run_agent.py — AIAgent.chat() now returns
  result.get("final_response") or result.get("error") or "" instead of
  result["final_response"], so error/failure results surface their error
  string rather than raising KeyError.
• run_agent.py — guard the CLI conversation-summary printer with
  result.get('final_response') (same omitted-key hazard).
• tests/test_chat_final_response_resilience.py — regression tests covering the
  error path (no final_response → returns error), the happy path, and the
  empty case.

How to Test

1. pytest tests/test_chat_final_response_resilience.py → 3 passed.
2. Before this change, any chat() call whose underlying conversation hit an
   error path (e.g. an API error after retries) raised
   KeyError: 'final_response' at run_agent.py return result["final_response"].
   After, it returns the error string and the caller can handle/log it.

Notes

Surfaced in practice when a non-interactive hermes -z oneshot run hit a
provider error path: the launcher recorded a KeyError: 'final_response'
traceback instead of the underlying error.

[liuhao1024]

Verified the fix is correct and the bug is real.

Confirmed: run_conversation() has 6+ error return paths that omit final_response (e.g., line 1471 in agent/conversation_loop.py: "Invalid API response after max retries" with "failed": True but no final_response key). The current chat() code (return result["final_response"]) raises KeyError on all of them.

The or chain result.get("final_response") or result.get("error") or "" is the right approach:

• .get() avoids KeyError
• Falls back to error message so the caller sees what went wrong
• Empty string as final fallback when neither key exists

Tests cover all three branches correctly. The main() function fix (result['final_response'] → result.get('final_response')) is consistent — same crash path, same fix.

One minor note: the or chain means a genuinely empty-string final_response ("") would fall through to error. This is unlikely in practice (agents don't produce empty responses), but worth being aware of.

[OmarB97]

Thanks for verifying, @liuhao1024 — appreciate the independent confirmation of the error paths. The branch is conflict-free and ready for maintainer review whenever someone gets a chance.

[teknium1]

Thanks for the focused regression report and tests. The reported failure mode is already fixed on current main.

• Automated hermes-sweeper review verified commit 053424c4865db0e8cc6ef9a8c2f49bf8882afd45 (fix(agent): preserve final_response on failure returns) is on current main and shipped in v2026.7.1.
• agent/conversation_loop.py:1568-1575 now returns _final_response under both final_response and error for retry exhaustion; the same superseding fix covered the other terminal failure branches.
• tests/run_agent/test_run_agent.py:81-94 enforces that every dictionary return from run_conversation() carries a non-None final_response.

Because AIAgent.chat()'s direct result["final_response"] now receives that guaranteed key on the reported error paths, this PR's behavior is implemented on main.