BP mode fails with gpt-5.4-mini: model doesn't follow babysitter orchestration instructions

[tmuskal]

Summary

ALL agents fail BP/Predefined and BP/Create Interactive with gpt-5.4-mini on Ubuntu. The model produces output and writes the file, but doesn't go through the babysitter SDK run/iterate lifecycle.

Verification Report (codex + gpt-5.4-mini BP/Predefined)

• model-response: PASS (558K chars)
• file-creation: PASS (3487 bytes)
• stop-hooks: PASS
• hooks-adapter-session: PASS
• babysitter-run-completion: FAIL — no .a5c/runs/ directory
• babysitter-completion-proof: FAIL — no .a5c/runs/ directory

Root Cause

gpt-5.4-mini ignores the $babysitter:yolo command prefix and does the task directly instead of invoking the babysitter plugin. The same agents pass with gpt-5.5 and gemini-3.1-pro-preview, confirming this is a model capability issue.

Affected Cells

12 cells: BP/Predefined Interactive (6 agents), BP/Predefined BH (5 agents), BP/Create Interactive (2 agents) — all Ubuntu, all gpt-5.4-mini.

Possible Fixes

1. Stronger babysitter invocation prompt for smaller models
2. Force babysitter invocation via a pre-hook that wraps the prompt
3. Mark gpt-5.4-mini as unsupported for BP mode

[tmuskal]

Also affects gemini-cli with gpt-5.5 in BP/Predefined — model responded (18K chars) but didn't write file or create babysitter run. Same root cause: the babysitter plugin invocation prefix wasn't followed by the model.

Updated affected count: 14 cells total (12 gpt-5.4-mini all agents + 1 gemini-cli gpt-5.5 + 1 hermes Windows gpt-5.5).

[tmuskal]

Addressed via option 1 (stronger invocation prompt) in 60b8125e7.

The BP-mode prompts in the live-stack runner (packages/adapters/cli/tests/live-stack/primary-live-runner.ts, buildPrompt) sent a bare /babysitter:yolo / $babysitter:yolo / /yolo <task> with no instruction that orchestration is mandatory — so weaker models (gpt-5.4-mini) treated it as preamble and did the task directly.

Now every BP-mode prompt (predefined / create / resume) appends an unmissable directive:

=== ORCHESTRATION REQUIRED === You MUST execute this through the babysitter run lifecycle (create/resolve a run, iterate, post effects, return the completion proof). Do NOT produce the output directly: writing the result file without an active .a5c/runs/<runId>/ run is a FAILURE. Invoke the babysitter command/skill first and drive it to completion.

Key constraint respected: the directive is appended, never prepended — the slash/$-command must remain the first token or native command recognition breaks for claude-code/codex. Guard tests assert (a) the command token stays first in all BP modes, (b) the directive is present in predefined/create/resume, and (c) vanilla (non-BP) prompts are untouched. Full primary-live-runner suite 173/173 from repo root.

Caveat / honesty: this is a best-effort prompt-hardening. I can't verify it actually flips gpt-5.4-mini without a live run against that model — but it cannot regress models that already orchestrate. If gpt-5.4-mini still ignores it on the next live run, the durable resolution is option 3 (mark gpt-5.4-mini unsupported for BP mode), since this is a confirmed model-capability limit. Leaving this open until a live run on gpt-5.4-mini confirms the outcome.