davepoon/check

Run CIAgent checks on this repo's agent

The repo has agentci_spec.yaml (if it does not, use the onboard skill instead). Your job: run the right check for the change that was just made, read the result correctly, and never paper over a failure.

Which command

| Situation | Command | |---|---| | Spec or wiring changed, or no API keys | ciagent test --mock | | Agent code / prompt / retrieval changed | ciagent test --yes --format json | | Result differs from last run, or flakiness suspected | ciagent test --runs 3 --yes | | Knowledge base changed | ciagent generate-checks --dry-run, review, then apply | | The LLM judge's verdicts look wrong | ciagent judge-audit |

Live runs (test without --mock, judge-audit, generate-checks) call model APIs on the user's keys. Mock mode is free. If the user has not already approved live runs in this session, prefer --mock or ask.

Reading results

Exit codes: 0 pass (including flaky-but-passing), 1 correctness failure (with --runs N: failed in every run), 2 infra or config error — fix the setup, not the agent.

With --format json: per-query entries carry layer results (correctness / path / cost) and the answer text; with --runs N a top-level stability block lists flipped queries with flip_source.

Flip sources route the work:

• agent-variance — the agent's answer changed between runs → fix the agent (prompt, retrieval, temperature).
• judge-flake — same answer, the LLM judge changed its verdict → fix the eval (tighten the rubric or replace with a deterministic check).
• infra-error — a judge API call failed → retry; fix nothing.
• mixed — ambiguous; look at the answers yourself.

Rules

• A correctness failure means the agent lost a fact it used to state. Fix the agent, or — only if the check itself is factually wrong — fix the check. Never weaken or delete a correct check or baseline to make a run green; report the failure to the user instead.
• After intentionally changing agent behavior, re-record the affected golden: delete its baseline file and rerun ciagent bootstrap --runner <runner> --queries <file> --yes for that query, or update the spec's expectations — with the user's confirmation.
• Report results in one or two sentences: score, what failed and in which layer, flip sources if any, and the command you ran.