feat(evals): harness qualification bench#6
Merged
Merged
Conversation
A maintainer tool to benchmark a candidate agent CLI BEFORE wiring it into the
live loop — so the roster (codex now; pi agent, kilo cli, … later) gets qualified,
not just plugged in. Reuses Part A's agent_run adapter, so a candidate is invoked
exactly how the loop would invoke it (harness.kind=custom + a command template).
- evals/run.sh — per task: throwaway sandbox → agent_run (the configured harness)
→ the task's objective check → record JSONL. --harness/--command/--model/--repeat/
--label/--list. Runs on demand (needs the model endpoint + the CLI); NOT in CI.
- evals/score.py — aggregate results → per-task pass-rate (over --repeat), per-surface,
overall, latency; prints a table + writes a scorecard.json. aggregate() is unit-tested.
- evals/tasks/ — 4 starter tasks, objective binary scoring (no judge):
triage/do-issue → JSON verdict ∈ {do,try}
triage/skip-issue → verdict ∈ {skip,needs_info}
implement/failing-test → agent's edit makes a fixture's python test pass
review/planted-bug → review flags a hardcoded secret in a diff
- evals/checks_lib.py — shared, dependency-free check helpers (JSON extract, verdict
+ must_contain matching).
- evals/tests/test_checks.py — feeds canned outputs to the real checks + score,
validating the scoring logic deterministically (no agents). Wired into ci.yml
alongside shellcheck of run.sh.
Qualifying a new CLI: write a custom template → evals/run.sh --harness custom
--command "<tmpl>" → read the scorecard → decide.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Part B — a maintainer bench to qualify a candidate agent CLI before integrating it (codex now; pi agent, kilo cli, … later). Reuses Part A's
agent_runadapter, so a candidate is benchmarked exactly how the live loop invokes it. Scoring is objective binary — no judge.evals/run.sh— per task: throwaway sandbox →agent_run(configured harness) → the task's check → JSONL.--harness/--command/--model/--repeat/--label/--list. On-demand (needs the model endpoint + the CLI installed); not in CI.evals/score.py— per-task pass-rate (over--repeat), per-surface, overall, latency → table +scorecard.json.evals/tasks/— 4 starter tasks, all objectively scored:triage/do-issue→ JSONverdict∈ {do,try};triage/skip-issue→ {skip,needs_info}implement/failing-test→ the agent's edit makes a fixture python test passreview/planted-bug→ the review flags a hardcoded secret in a diffevals/tests/test_checks.py— feeds canned outputs to the real checks +aggregate(), validating the scoring logic deterministically, no agents. Wired into ci (+ shellcheck ofrun.sh).Qualify a new CLI: write a
customtemplate →evals/run.sh --harness custom --command "<tmpl>"→ read the scorecard → decide.Proof
self-guard + ci (shellcheck
run.sh+test_checks.py) on this PR. The live bench runs on demand.🤖 Generated with Claude Code