fix(agent): close orphaned assistant(tool_calls) in repair_message_sequence (#56980)

[hashbender]

What does this PR do?

repair_message_sequence handles stray tool messages, consecutive-assistant merges, and consecutive-user merges, but had no pass for the inverse case: an assistant message with tool_calls whose ids are not answered by the tool messages that follow. This happens after an interruption (/stop, process kill, resume mid tool-loop) when some or all of a turn's tool calls never got a result appended.

Strict OpenAI-compatible providers (DeepSeek v4, Moonshot/Kimi) reject that message shape outright:

An assistant message with 'tool_calls' must be followed by tool messages
responding to each 'tool_call_id'. (insufficient tool messages following
tool_calls message)

Scope, honestly

This does not fix an actively-reproducing HTTP 400 in the main request pipeline. sanitize_api_messages (called right after repair_message_sequence on every call, in conversation_loop.py and chat_completion_helpers.py — the single path CLI, gateway, and webui all funnel through) already injects a stub result for orphaned tool_call_ids before the request reaches the provider. Verified this directly against the real pipeline for every orphan shape in NousResearch#56980; none reproduce the 400 on current main.

What this PR actually fixes: repair_message_sequence operates on the canonical, persisted messages list, while sanitize_api_messages only patches the ephemeral per-call api_messages copy. Any caller that runs repair_message_sequence without a follow-up sanitize_api_messages call — sub-agents, MoA reference-model calls, plugins, or the existing run_agent.py forwarder used only by tests — stays exposed to the exact symptom in NousResearch#56980. This closes that gap at the source instead of depending on a second, separately-invoked function to catch it downstream every single call.

Fix

Added a pass to repair_message_sequence: for each assistant(tool_calls), scan the tool messages immediately following it, and for any tool_call_id still unanswered, insert a synthetic error result:

{"role": "tool", "tool_call_id": "<id>", "name": "<tool name>", "content": "Tool execution was interrupted before a result was returned."}

Stubs the missing result instead of stripping tool_calls from the assistant message, to avoid leaving {"role": "assistant", "content": None} when a turn's calls are all orphaned, and to keep the model's stated intent visible in the transcript.

No-op when every tool_call_id already has a matching result — the existing "ongoing dialog" pattern (a complete assistant(tool_calls)+tool pair followed by a user redirect) is left untouched, as documented in the function's docstring.

Follow-up fix (NousResearch#57036): the id lookup in this pass initially read tool_calls[].id directly. Switched it to AIAgent._get_tool_call_id_static / _get_tool_call_name_static — the call_id || id extractor sanitize_api_messages and the context compressor already use — since Codex/Responses-API-style tool_calls key the id as call_id, not id. Without this, the pass silently had no effect on that shape.

Related

Fixes NousResearch#56980 (design gap this closes), fixes NousResearch#57036 (extractor follow-up).

Testing

• tests/run_agent/test_message_sequence_repair.py: 6 new tests — fully unanswered trailing tool_calls, partially answered parallel tool_calls, orphaned tool_calls followed by an injected user/wakeup message, idempotency, call_id-keyed (Codex-shaped) orphans, and a non-regression check that a complete pair before a user redirect is left untouched.
• Full file: python -m pytest tests/run_agent/test_message_sequence_repair.py -v — 31/31 passing.
• Regression: tests/agent/test_context_compressor.py, tests/agent/test_replay_cleanup.py, tests/agent/test_close_interrupted_tool_sequence.py — 159/159 passing.
• Full test_run_agent.py suite — 414/414 passing.

Type of Change

• Bug fix / hardening (non-breaking)

Checklist

• No new dependencies
• No changes to message serialization format
• No breaking changes to OpenAI/Anthropic provider compatibility
• Existing "ongoing dialog" pattern (complete pair + user redirect) covered by a regression test
• Uses the same call_id||id extractor as the rest of the file (post repair_message_sequence orphan-close pass doesn't use canonical call_id||id extractor NousResearch/hermes-agent#57036)

---

Mirror-of: NousResearch#57013
NousResearch#57013