Retry ModelBehaviorError so a truncated stream no longer kills the run

[inf-quantavius]

Problem

The openai-agents SDK raises ModelBehaviorError: Model did not produce a final response! (agents/run_internal/run_loop.py) when a model stream truncates without a terminal event. On the root agent turn this escaped OpenAiAgentRunner's retry loop — is_retriable_llm_error only knew openai/httpx error classes, so the except at openai_agent_runner.py re-raised it, _drive() in engine/main.py called output_bus.fail(exc), and the whole run failed within seconds on a single transient stream truncation.

Failing examples (dev smokes):

• 2026-06-16 — run 988eaa3d, conversation 520914bc
• 2026-07-14 — run a91ba571-4fd3-4960-a2a2-de2fbb4dbac8

Subagents only survived this via guarded_invoke's blanket except Exception in subagent_tool_factory.py (#67), which converts the crash into a tool-failure result rather than retrying.

Fix

Classify ModelBehaviorError as retriable in engine/agents/llm_retry.py::is_retriable_llm_error. That's the whole behavioral change — it slots into the runner's existing mid-stream recovery machinery (INF-3504 / INF-3308):

• rerun from local conversation history, with a trailing incomplete tool turn trimmed (AgentContext.trim_incomplete_tool_turn) so the replayed message array stays valid
• full-jitter exponential backoff between attempts (retry_backoff_base / llm_retry_backoff_base_seconds, 0 in tests)
• bounded by the MAX_CONSECUTIVE_LLM_FAILURES = 10 circuit breaker → EngineAgentExhaustedError when exhausted

Retrying is safe here because HALO rebuilds every request from the local AgentContext (no server-side conversation state), which is the same reason non-terminal 400s are already retried. All ModelBehaviorError variants (no final response, malformed tool-call JSON, nonexistent tool) are nondeterministic sampled-output misbehavior, so a re-sampled turn can succeed. Deliberately not retried: MaxTurnsExceeded (must terminate the run), UserError, and guardrail tripwires.

Bonus: subagent turns now retry in-runner before falling back to guarded_invoke's tool-failure result.

Tests

• tests/unit/agents/test_llm_retry.py: ModelBehaviorError variants are retriable; MaxTurnsExceeded / UserError are not.
• tests/unit/agents/test_openai_agent_runner.py::test_runner_retries_model_behavior_error_when_stream_has_no_final_response: fake stream yields a partial assistant message then raises the exact production error; asserts the rerun replays local history, the run completes, and the failure counter resets.

Ran locally: pytest tests/unit → 462 passed; pytest tests/integration -m "not live" → 53 passed (11 Pyodide-sandbox failures are pre-existing on a clean tree on this host — see #67-era note); ruff check / ruff format clean; basedpyright → 0 errors.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes when agent runs retry vs fail on root turns; behavior is bounded by the existing circuit breaker and deliberately excludes deterministic SDK errors like max turns.

Overview
Treats the Agents SDK ModelBehaviorError as a retriable LLM failure so truncated streams (e.g. “Model did not produce a final response!”) no longer abort the root agent run.

is_retriable_llm_error now returns true for all ModelBehaviorError variants (missing final response, bad tool JSON, unknown tools), while MaxTurnsExceeded and UserError stay non-retriable. Those errors plug into the existing OpenAiAgentRunner mid-stream recovery: rerun from local AgentContext, trim incomplete tool turns when needed, backoff, and the consecutive-failure circuit breaker.

Tests cover classification and a runner case where a partial assistant message is preserved and the second attempt completes the run.

^{Reviewed by Cursor Bugbot for commit 2c77514. Bugbot is set up for automated code reviews on this repo. Configure here.}