Context
A Codex-hosted ooo qa run reported "the Ouroboros QA tool call itself failed" (a retry succeeded), then ouroboros_session_status for the returned qa-f7e392ab id answered Session not found. Reproduced on main (d5c2f9f) with a raw JSON-RPC stdio probe against ouroboros mcp serve --runtime claude-cli --llm-backend claude_code:
- Server boots and registers 36 tools including
ouroboros_qa in ~6 s — "MCP does not come up" is not reproduced.
ouroboros_session_status(session_id="qa-…") → Session not found: No events found for session — always, because QAHandler.handle generates qa_session_id but never persists anything (src/ouroboros/mcp/tools/qa.py).ouroboros_qa with pass_threshold: "0.8" or iteration_history: "[]" (strings, a common serialization from GPT-family hosts) is rejected by inputSchema validation before the handler runs — the whole tool call fails.- A valid call takes ~18 s for a trivial artifact through the claude CLI backend; adversarial QA on a real app can exceed the 60 s default host tool timeout on Codex.
Handler-level defects reproduced with a faked LLM adapter:
- LLM verdict string overrides the score:
{"score":0.42,"verdict":"pass"} renders Score: 0.42 / 1.00 [PASS], Verdict: pass, while meta says passed: false, loop_action: continue — three surfaces disagree; qa-judge.md says score decides.
- Non-dict
iteration_history entries raise AttributeError out of handle() (no Result.err), since build_qa_subagent runs above the try:.
This issue is the umbrella for a series of single-subsystem PRs; each PR references it and fixes exactly one defect.
Context
A Codex-hosted
ooo qarun reported "the Ouroboros QA tool call itself failed" (a retry succeeded), thenouroboros_session_statusfor the returnedqa-f7e392abid answeredSession not found. Reproduced onmain(d5c2f9f) with a raw JSON-RPC stdio probe againstouroboros mcp serve --runtime claude-cli --llm-backend claude_code:ouroboros_qain ~6 s — "MCP does not come up" is not reproduced.ouroboros_session_status(session_id="qa-…")→Session not found: No events found for session— always, becauseQAHandler.handlegeneratesqa_session_idbut never persists anything (src/ouroboros/mcp/tools/qa.py).ouroboros_qawithpass_threshold: "0.8"oriteration_history: "[]"(strings, a common serialization from GPT-family hosts) is rejected by inputSchema validation before the handler runs — the whole tool call fails.Handler-level defects reproduced with a faked LLM adapter:
{"score":0.42,"verdict":"pass"}rendersScore: 0.42 / 1.00 [PASS],Verdict: pass, while meta sayspassed: false, loop_action: continue— three surfaces disagree;qa-judge.mdsays score decides.iteration_historyentries raiseAttributeErrorout ofhandle()(noResult.err), sincebuild_qa_subagentruns above thetry:.This issue is the umbrella for a series of single-subsystem PRs; each PR references it and fixes exactly one defect.