fix(engine): render local-history reruns as Responses-API input items (analysis-smoke EngineAgentExhaustedError)

[inf-quantavius]

What

The nightly [Inference] LLM Ops Analysis Smoke (dev) failed on 2026-07-28 (run 30326543437): both fixed-analysis runs (iar_55bgjz7omQ8Oq5BHy4toe task scope, iar_v2ydWpePBNlNqHerYtFXP agent scope) ended status=error stage=halo with:

HALO run failed: EngineAgentExhaustedError: agent root-... exhausted after 10 consecutive failures

Root cause

Dev ClickHouse (default.inferences, 03:40–04:00 UTC window) shows the underlying LLM failures: 29 gpt-5.5 requests to v1/responses rejected with

litellm.BadRequestError: OpenAIException - Missing required parameter: 'input[2].content'

Dissecting a failing request: input[2] is {"role":"assistant","tool_calls":[...]} with no content and input[3] is {"role":"tool","content":...} — chat-completions shapes inside a Responses-API input array.

Mechanism:

• OpenAiAgentRunner.run rebuilds every attempt's input from local history: [m.model_dump(exclude_none=True) for m in agent_context.to_messages_array()] — chat-completions-shaped AgentMessages.
• The SDK runs on the Responses API (OpenAIProvider default use_responses=True; all requests land on v1/responses). A healthy first attempt sends only [system, user] (valid), and mid-run turns use the SDK's own Responses items (function_call_output — visible in the 88 successful multi-turn requests in the same window).
• But every rerun-from-local-history path — mid-stream failure recovery (INF-3504/INF-3308), refusal "Continue." retries, and the final_answer reprompt — re-renders the stored chat shape. Once history contains a tool turn, the assistant tool_calls message has no content → deterministic 400.
• is_retriable_llm_error deliberately retries non-terminal 400s ("a clean rerun from local history fixes stale server-side state") — but here the local history itself renders invalid, so the rerun can never succeed. 10 identical failures → EngineAgentExhaustedError.

In the failing window all 29 failed requests carried chat-style tool_calls (and no function_call_output); all 88 multi-turn successes carried function_call_output (and no tool_calls). The failure is deterministic once any resend path triggers after a tool turn — which is probabilistic per run, hence the smoke passing 07-26/07-27 and failing 07-28.

Not a monorepo regression — none of the 110 monorepo commits since the last green run touch this path. Prod is affected too: the same error signature appears in prod default.inferences daily since at least 2026-07-17 (11–61 requests/day), i.e. real HALO/analysis runs intermittently dying the same way.

Fix

Add a render boundary, engine/agents/responses_input.py::to_responses_input, used by OpenAiAgentRunner.run in place of the raw model_dump:

• system/user/assistant text messages → {"role", "content"} message items (None content → "", since message items require content)
• assistant messages with tool_calls → optional text message item + one {"type": "function_call", "call_id", "name", "arguments"} item per call
• role="tool" results → {"type": "function_call_output", "call_id", "output"} items

Order is preserved, so the existing context invariant from trim_incomplete_tool_turn (every call answered by its result) carries over to call/output pairing.

Tests

• New tests/unit/agents/test_responses_input.py: tool-turn pair conversion, text+tool_calls assistant, None-content edges, compacted-summary passthrough.
• Updated test_runner_retries_refusal_after_tool_result_without_replaying_tool_output to the Responses shapes.
• Full unit suite: 490 passed; ruff check / ruff format clean.

Provenance

• Requested by: CI Notifications (Slack U0B5KMSTCKX)
• Source: Slack thread
• Quant turn: 95a7ee0d