fix(moa): advisory references end on a user turn + get a reference-role system prompt

[teknium1]

Summary

Two related fixes to MoA reference-model calls:

1. Reference calls no longer 400 on Anthropic models that reject assistant prefill. The advisory reference view now always ends with a user turn instead of a trailing assistant turn.
2. Reference models now know their role. Each reference call is prefixed with an advisory system prompt so the model acts as an analyst advising the aggregator, instead of assuming it is the acting agent and refusing ("I can't access repositories/URLs from here") or trying to call tools it doesn't have.

Root cause — prefill 400

_reference_messages() in agent/moa_loop.py builds an advisory transcript by keeping user/assistant text turns and dropping the system prompt and tool messages. Mid-tool-loop, the last assistant turn carries reasoning text plus tool calls — its text survives the trim while the following tool result is dropped, leaving a trailing assistant turn. Anthropic (and OpenRouter→Anthropic) treat that as an assistant prefill to continue, and no-prefill models (e.g. Claude Opus 4.8) reject it with 400 ... must end with a user message. GPT-5.5 tolerates it, so only the Anthropic reference failed.

Root cause — role confusion

References got the bare trimmed conversation with no framing, so they assumed they were the acting agent being asked to do the task.

Changes

• agent/moa_loop.py:
  • _reference_messages() strips trailing assistant turns so the advisory view ends on a user turn; intervening assistant context is preserved.
  • New _REFERENCE_SYSTEM_PROMPT prepended in _run_reference() — frames the model as an advisory analyst (no execution, no tools, reason about state, hand guidance to the aggregator).
• tests/run_agent/test_moa_loop_mode.py: updated the test that encoded the buggy trailing-assistant shape; added test_reference_messages_ends_with_user_not_assistant_prefill and test_run_reference_prepends_advisory_system_prompt; updated the facade test for the leading system message.

Validation

	Before	After
Reference view ending	trailing assistant (prefill)	trailing user
Anthropic no-prefill reference (Opus 4.8)	400 must end with a user message	succeeds
Reference role framing	none (acts as the agent)	advisory analyst system prompt
MoA test suite	—	15/15 green

Infographic

[github-actions]

🔎 Lint report: hermes/hermes-36fe5816 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: 11611 on HEAD, 11611 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 6094 pre-existing issues carried over.

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

[OutThisLife]

Reviewed the full diff against current main. Both fixes are correct, the root-cause analysis holds against the actual code, and the diff is minimal and well-scoped — it fixes the whole bug class (both _reference_messages consumers) rather than one call site.

Fix 1 — trailing assistant strip (the 400)

Traced it in agent/moa_loop.py: _reference_messages() keeps assistant text even when it accompanied tool_calls, then drops the following tool result, so mid-tool-loop the advisory view ends on an assistant turn → Anthropic reads it as a prefill and Opus 4.8 rejects. GPT-5.5 tolerating it explains why only the Anthropic reference failed. Accurate.

Two things I specifically checked and they hold:

• Ordering with the fallback — the while strip runs before the if not trimmed: degenerate fallback, so an assistant-only view that strips to empty still falls back to the latest user turn.
• Intervening assistants preserved — only trailing assistants pop, so prior-turn Q&A context survives.

Bonus not mentioned in the description: this also stabilizes the per-turn reference cache. Before, each tool-loop iteration appended an assistant turn → new _ref_cache_key signature → references re-ran every iteration (and re-hit the 400). After, the advisory view is constant across a turn, so references run once. Net cost reduction on top of the correctness fix.

Fix 2 — advisory system prompt

Correctly added in _run_reference() at dispatch time, after _reference_messages() builds the cached/signed view — so the constant prompt doesn't perturb the cache key, and it's the only system message the reference sees. New list per call, dicts only read across threads → no aliasing in the fan-out. No prompt-cache concern (auxiliary call_llm calls; a constant prefix is cache-friendly).

Nits (non-blocking)

• The updated facade test asserts "reference advisor" in ref_msgs[0]["content"].lower() — mild change-detector coupling to prompt wording. The sibling test_run_reference_prepends_advisory_system_prompt already asserts against the _REFERENCE_SYSTEM_PROMPT constant, which is the robust form; the facade test could do the same.
• On the P3 label: defensible since _run_reference catches the 400 and returns [failed: ...], so it degrades gracefully rather than crashing. But "silently loses a reference's advice on every multi-iteration MoA turn with an Anthropic reference" is arguably closer to P2.