Fix relay worker final-summary capture for Codex PTY runs

[miyaontherelay]

Problem

OpenKaren relay execution is much more stable now, but the final Telegram-facing summary is still unreliable for Codex PTY workers.

The remaining failure mode is no longer primarily broker/session reuse. The narrow issue is final-answer capture.

In live relay runs, worker roles often finish with waitStatus: "idle" or waitStatus: "exited", but OpenKaren still cannot consistently recover a clean final answer. The resulting Telegram summary often degrades into one of these:

• raw Codex PTY chrome
• verifier finished without relay output.
• noisy partial fragments

What is already fixed

This issue is not about redoing the earlier relay reuse work from scratch. The following work has already materially improved the runtime:

• stale in-memory RelaySession reuse was removed
• new turns rebuild session state from persisted broker state
• repeated-turn hangs became much less frequent
• relay worker model normalization now avoids unsupported ChatGPT-account models for Codex workers by falling back to gpt-5.4
• focused validation has been passing:
  • npm run test:agent-runner
  • npm run typecheck

Current diagnosis

The remaining problem is that Codex PTY workers do not provide OpenKaren a dependable structured final-answer surface.

Right now OpenKaren tries to infer the final answer from:

• PTY worker log tails
• worker_stream output chunks
• idle / exited worker lifecycle states

That inference is still weak.

Concrete live symptoms observed

• workers can finish as idle or exited
• verifier sometimes leaves no clean natural-language summary block
• PTY logs are dominated by terminal chrome, MCP boot lines, watchdog noise, and other Codex UI fragments
• useful text can exist, but is often buried in noisy PTY output

Concrete evidence already seen in live runs

• unsupported relay worker model path previously showed up as:
  • The 'opencode/gpt-5-nano' model is not supported when using Codex with a ChatGPT account.
• after relay model normalization, live workers began using gpt-5.4, which improved execution viability but did not fully solve final summary extraction
• some live artifacts still ended with:
  • verifier finished without relay output.
• some other artifacts captured noisy but clearly useful text inside a larger PTY blob, for example summary-like content such as:
  • normalize unsupported Codex account models to gpt-5.4, and filter Codex UI noise so downstream roles keep real worker output. Fresh proof: npm run test:agent-runner passed ...

Partial mitigation already added

Verifier prompting has already been tightened to ask for an explicit final marker:

• OPENKAREN_FINAL: <final reply>

The extractor also now prefers that marker when present.

However, live behavior still suggests that some workers exit before leaving a dependable recoverable final line.

What needs to happen for a real fix

1. Treat clean exited as a first-class completion candidate

Do not assume exited means unusable output.

If a worker exits cleanly and leaves a valid final summary marker or a valid final natural-language completion block, OpenKaren should treat that as a successful completion path.

2. Capture the final answer from a deterministic surface, not just noisy PTY logs

Preferred order should be something like:

1. explicit OPENKAREN_FINAL: marker
2. last meaningful natural-language block near completion/exit
3. fallback to shaped worker stream/log delta only if it looks like a real answer

3. Tighten prompt contracts for orchestrated roles, especially verifier

The verifier should be required to emit exactly one final-answer marker at the end.

If needed, planner/reviewer/implementer can also be asked to avoid UI-heavy chatter and keep handoff blocks minimal.

4. Avoid treating Codex UI noise as useful completion text

The extractor should continue excluding things like:

• Starting MCP servers
• Booting MCP server
• terminal chrome
• status spinners
• watchdog idle logs
• /model to change
• https://chatgpt.com/codex...
• handoff prompt scaffolding that was echoed back

5. Add tests for the real end-state, not just plumbing

We need focused tests that prove:

• idle + explicit OPENKAREN_FINAL: returns the final summary
• exited + explicit OPENKAREN_FINAL: also returns the final summary
• noisy PTY blobs with one buried useful summary extract the useful summary
• noisy PTY blobs with no useful summary fall back to the current safe degraded text
• handoff prompts / echoed context are not mistaken for final answers

Suggested acceptance criteria

• A live relay run with Codex workers returns a clean Telegram-ready summary instead of PTY chrome
• A live verifier that exits after emitting OPENKAREN_FINAL: is treated as successful and the marker text is returned
• verifier finished without relay output. is only used when there is truly no recoverable final answer
• PTY UI noise is not surfaced as the final Telegram reply
• Focused tests cover idle and exited completion with explicit summary markers and noisy-log fallbacks

Nice-to-have

If the SDK exposes a better structured final-message/event surface in the future, OpenKaren should switch to that instead of relying so heavily on PTY logs.