2026-10-13-unreal-agent-v0-3-1-fixes-sigterm-handling-but-retains-false-success-responses-provider-dependent-dupli

v0.3.1 fixes SIGTERM handling but retains false-success responses, provider-dependent duplicate outputs and unbounded Bash capture. Three outside reports against v0.1.x and v0.2.0 code, filed on 2026-09-22 and 2026-09-23, are still open. In #13, a provider returns HTTP 200 with a response.failed event (the reproduction uses insufficient_quota), and the runner records it and exits 0. The maintainer said persisting failed responses is intended “to capture usage in session logs” and that he would look at exiting nonzero; no fix commit exists, and the v0.3.1 coordinator still ignores Response.Failure. In #11, parallel Bash fan-out across turns produced two function_call_output items for one call, and a provider 400 that repeats on every resume, which poisons the session. The maintainer called it “a known limitation for some provider-model pairs”, citing the launch post’s footnote that two results for one call (running, then final) are “underspecified in the Responses API documentation” and were rejected by “some models on some inference providers (not OpenAI)”. The Anthropic adapter avoids the pattern (item 6); the Responses-API providers do not. #3 remains open on its other half: no Bash timeout and no on-disk output cap (item 4). Fixed in the v0.3.0 Go tag and v0.3.1 binary release: SIGTERM now cancels through the runner’s context (#16), and skill frontmatter is parsed as YAML (#15, 0c9baa7).

Channel: tagged-release. Half: defect. Date: 2026-09-22..2026-10-06.

Operator consequence: In CI, do not trust the exit status of unreal-agent-runner: scan the JSONL for a failed response before you count a run as a success. For a provider-model pair that rejects repeated native tool outputs, replaying the same poisoned history can repeat the 400. Start a fresh session or constrain fan-out on the affected pair; this is not a claim that every OpenRouter, Fireworks or Ollama model rejects the pattern. Upgrade to v0.3.x if you stop runners with SIGTERM, as container runtimes and CI timeouts do. All of these reports came in within about 36 hours of launch. The issue tracker is the only public defect list: there is no SECURITY.md and no published repository advisory was found. Dependency advisories still apply to the benchmark environment (items 11 through 13).

Receipt

• https://github.com/unreallabsai/unreal-agent/issues/13#issuecomment-5822471686
• https://github.com/unreallabsai/unreal-agent/issues/11#issuecomment-5800680734
• https://github.com/unreallabsai/unreal-agent/issues/16#issuecomment-6021820979
• https://github.com/unreallabsai/unreal-agent/blob/v0.3.1/harness/coordinator/loop.go#L231-L238
• https://github.com/unreallabsai/unreal-agent/commit/4be534be93e8f686b0be256062267af77c8e66df
• https://github.com/unreallabsai/unreal-agent/commit/0c9baa7310fabd37d045c106cac8a5ca43e8b1cd