Eval regression
Use deterministic repository tests first. Stop when the target has no behavioral diff.
Resolve the plugin from the argument or cwd. Its catalog is evals/evals.json; the shared runner is ../ai-agent-bench/scripts/run_evals.py. Require the user to choose agent, model, and effort. Never select a costly model or high effort silently.
Normal check
- Select cases by changed paths with
--changed-from <base>, or name them with --case. Do not select the full catalog implicitly.
- Run the command without
--run. The runner prints the cases, modes, repeat count, session count, per-session timeout, and maximum duration without starting an agent.
- Present that plan and stop for explicit cost approval.
- After approval, repeat the same command with
--run. The default is candidate-only, one run per case, 180 seconds per session, and at most four sessions.
- Report every failed assertion, timeout, non-zero exit, duration, and token count. Missing evidence is inconclusive.
Escalation
- Compare base and candidate only when the user asks, or when a failed candidate check needs to distinguish a regression from an existing failure. Extract the base with
git archive; run the same selected cases, agent, model, effort, repeat, and timeout on both; compare reports with --compare.
- Repeat only a failed or observably unstable case. Three repeats are a stability benchmark, not a default.
--all, --repeat > 1, a larger --max-sessions, or a longer timeout needs a new run plan and explicit approval.- Routing cases stop at the first Skill selection. A tool assertion is appropriate there because routing is the contract; they must not execute the selected skill.
Remove temporary base copies after the comparison. Do not edit or commit the target.
Deterministic assertions are tool, tool_not, clean_worktree, changed_files_exact, file_contains, file_not_contains, transcript_contains, and tool_sequence. Use a semantic judge only when no filesystem, command, tool, ordering, or assistant-output observation can express the contract.
1---2name: eval-regression3description: Use when the user asks to check a plugin or skill with behavioral evals, compare it with HEAD, or investigate a behavioral regression.4---56# Eval regression78Use deterministic repository tests first. Stop when the target has no behavioral diff.910Resolve the plugin from the argument or cwd. Its catalog is `evals/evals.json`; the shared runner is `../ai-agent-bench/scripts/run_evals.py`. Require the user to choose agent, model, and effort. Never select a costly model or high effort silently.1112## Normal check13141. Select cases by changed paths with `--changed-from <base>`, or name them with `--case`. Do not select the full catalog implicitly.152. Run the command without `--run`. The runner prints the cases, modes, repeat count, session count, per-session timeout, and maximum duration without starting an agent.163. Present that plan and stop for explicit cost approval.174. After approval, repeat the same command with `--run`. The default is candidate-only, one run per case, 180 seconds per session, and at most four sessions.185. Report every failed assertion, timeout, non-zero exit, duration, and token count. Missing evidence is inconclusive.1920## Escalation2122- Compare base and candidate only when the user asks, or when a failed candidate check needs to distinguish a regression from an existing failure. Extract the base with `git archive`; run the same selected cases, agent, model, effort, repeat, and timeout on both; compare reports with `--compare`.23- Repeat only a failed or observably unstable case. Three repeats are a stability benchmark, not a default.24- `--all`, `--repeat > 1`, a larger `--max-sessions`, or a longer timeout needs a new run plan and explicit approval.25- Routing cases stop at the first Skill selection. A tool assertion is appropriate there because routing is the contract; they must not execute the selected skill.2627Remove temporary base copies after the comparison. Do not edit or commit the target.2829Deterministic assertions are `tool`, `tool_not`, `clean_worktree`, `changed_files_exact`, `file_contains`, `file_not_contains`, `transcript_contains`, and `tool_sequence`. Use a semantic judge only when no filesystem, command, tool, ordering, or assistant-output observation can express the contract.
reidemeister94