improve-skills

name: improve-skills description: Report or fix a behavior gap in Asta skills. Use when an agent does the wrong thing, doesn't do what was asked, or lacks a capability that's wanted. metadata: internal: true allowed-tools: Bash(git clone *) Bash(git fetch *) Bash(git worktree *) Bash(git checkout *) Bash(git add *) Bash(git commit *) Bash(git tag *) Bash(gh issue view *) Bash(jq *) Bash(eval_env *) Bash(asta auth print-token *) Read Write Edit

Improve Asta skills

1. Capture the issue (bug or feature idea)

Write the gap to <project-root>/.asta/improve-skills/<slug>.md (absolute path, dated slug): the skill(s), prompt, current vs. desired behavior, and relevant setup state. Seed it from an existing issue/doc/thread if one exists (gh issue view <num> --repo allenai/asta-plugins --json body --jq .body > <project-root>/.asta/improve-skills/<slug>.md).

If stopping here (report): file it as a new issue (skip if it already is one): gh issue create --repo allenai/asta-plugins --title "<summary>" --body-file <project-root>/.asta/improve-skills/<slug>.md.

2. Assemble the case set

Set a dev root for the repos, then add or reuse cases in the asta_skills suite (under $DEV_ROOT/asta-bench-private):

export DEV_ROOT=<your dev dir>
[ -d "$DEV_ROOT/asta-bench-private" ] || git clone git@github.com:allenai/asta-bench-private "$DEV_ROOT/asta-bench-private"

Two groups:

• Cases for the behavior (skip if regression-checking only). Reuse existing cases that fit; otherwise add new entries on a branch in asta-bench-private (see the asta_skills eval README for case format and metric handlers).
• Regression guards. Existing cases that could be affected. Start by searching the suite's data.json for cases referencing the changed skill(s), but use judgment — cross-skill effects matter (a description change can shift routing on cases targeting other skills; see asta-plugins#67 for a routing-competition example).

3. Run baseline

Set up a clean origin/main worktree for the baseline:

[ -d "$DEV_ROOT/agent-baselines" ] || git clone git@github.com:allenai/agent-baselines "$DEV_ROOT/agent-baselines"
[ -d "$DEV_ROOT/asta-plugins" ]    || git clone git@github.com:allenai/asta-plugins "$DEV_ROOT/asta-plugins"
export ASTA_TOKEN="$(asta auth print-token --raw --refresh)"
git -C "$DEV_ROOT/asta-plugins" fetch -q origin
git -C "$DEV_ROOT/asta-plugins" worktree add --detach "$DEV_ROOT/asta-plugins-main" origin/main 2>/dev/null \
    || git -C "$DEV_ROOT/asta-plugins-main" checkout -q --detach origin/main
export ASTA_MAIN="$DEV_ROOT/asta-plugins-main"

From $DEV_ROOT/agent-baselines, define eval_env to run each inspect command in the solvers/inspect-swe project:

cd "$DEV_ROOT/agent-baselines"
eval_env() { uv run --project solvers/inspect-swe --no-group astabench \
    --with "$DEV_ROOT/asta-bench-private" --frozen --reinstall-package astabench-private -- "$@"; }

Run the full case set (behavior + guards) in one eval:

eval_env inspect eval astabench/asta_skills \
    --sample-id <case_id_1>,<case_id_2>,... \
    --solver agent_baselines/solvers/inspect_swe/agent.py@inspect_swe_solver \
    --sandbox docker:solvers/inspect-swe/sandbox_compose.yaml \
    --model anthropic/claude-sonnet-4-6 \
    --epochs 3 --working-limit 600 \
    -S agent=claude_code \
    -S strict_reproducibility=true \
    -S skills=$ASTA_MAIN/plugins/asta-preview/skills \
    --log-dir logs/baseline
# model & agent above are the asta-plugins low-cost convention
# -S version and the ASTA_IMAGE env var are left unset for baseline -> auto / :latest

Read per-case-per-metric scores: eval_env inspect log dump --header-only "$(ls -t logs/baseline/*.eval | head -1)" | jq '.reductions[0].samples'. For every arm you run (here and in steps 5–6), also read the transcript (eval_env inspect view --log-dir logs, or the sample messages), not just the score — confirm the metric reflects the intended behavior, not a coincidence (the gap reproduces, any lift is real, a held guard isn't masking change). For a new case, its gap-capturing metric should score below ceiling; if at ceiling, the gap isn't reproducing — investigate.

If stopping here (reproducible test + baseline, TDD red): file an issue per step 1 with a ## Reproducible test section — the baseline numbers, the transcript read (so the fixer inherits the why, not just the number), and the case(s): link the asta-bench-private PR if you added any (push the branch, open the PR), or name the existing case(s) you reused.

4. Make the skill change(s)

From the baseline transcript, hypothesize why the skill(s) produce the gap and what will close it.

# Worktree for the PR branch:
git -C "$DEV_ROOT/asta-plugins" worktree add "$DEV_ROOT/asta-plugins-<pr-branch>" -b <pr-branch>

Then, in $DEV_ROOT/asta-plugins-<pr-branch>, edit existing skill(s) or add a new one under plugins/asta-preview/skills/<name>/ (the canonical tree — edit it directly). Fixes may span multiple skills. Commit and push:

cd "$DEV_ROOT/asta-plugins-<pr-branch>"
git add -A && git commit -m "<summary>" && git push -u origin <pr-branch>

5. Run the PR arm (same case set)

Back in $DEV_ROOT/agent-baselines, pin to what the baseline resolved:

eval "$(eval_env inspect log dump "$(ls -t logs/baseline/*.eval | head -1)" | jq -er '.samples[0].metadata
    | "ASTA_IMAGE=\(.asta_image)\nAGENT_VERSION=\(.agent_version)"')"
export ASTA_IMAGE   # env var the sandbox compose substitutes into `image: ${ASTA_IMAGE}` (no flag sets the image); AGENT_VERSION stays a shell var for -S version below

Rerun the step-3 eval with the exported ASTA_IMAGE and:

• -S version="$AGENT_VERSION"
• -S skills=$DEV_ROOT/asta-plugins-<pr-branch>/plugins/asta-preview/skills
• --log-dir logs/pr-arm

Confirm new-case metrics fire and the regression guards hold at baseline. If a guard drops, your change regressed that case — find the cause and fix it (or explain it if the drop is benign) before opening the PR.

6. Ablation

For a multi-part fix, hypothesize which parts carry the result and which may be unnecessary or harmful, then ablate to test (e.g. asta-plugins#67).

Create the variant — a worktree off the PR branch with one part undone:

git -C "$DEV_ROOT/asta-plugins-<pr-branch>" worktree add "$DEV_ROOT/asta-plugins-ablate" -b ablate-<short-desc>
cd "$DEV_ROOT/asta-plugins-ablate"
git checkout main -- plugins/asta-preview/skills/<skill>/SKILL.md   # undo one part: a whole file, or hand-edit to undo part of one
git add -A && git commit -m "ablate: <description>"

Rerun the step-5 eval on the variant (keeping its -S version pin and exported ASTA_IMAGE) with:

• -S skills=$DEV_ROOT/asta-plugins-ablate/plugins/asta-preview/skills
• --log-dir logs/ablation

A drop from the PR arm means that part is load-bearing — keep it. No drop means it isn't pulling weight — drop it from the PR (or note why it stays). If undoing a part recovers a regressed guard, it was causing that regression — drop or rework it.

7. Open the PR(s)

Draft a body per repo you touched, in <project-root>/.asta/improve-skills/ as <slug>-<repo>-pr.md — concise, linking the companion rather than restating it:

• asta-plugins (always): what it fixes — Resolves #<issue> if one exists, else the gap and the behavior it addresses — plus Validation: pinned setup (agent version, model, image tag + @sha256:), an arms table (baseline, PR, and ablation if you ran one) across cases and metrics, and the transcript read.
• asta-bench-private (only if you changed cases): the case(s) and what they measure; defer results to the companion.

Create each with gh pr create --repo allenai/<repo> --body-file <project-root>/.asta/improve-skills/<slug>-<repo>-pr.md, then cross-link their numbers with gh pr edit.

If you ran an ablation, tag its variant and link it from the asta-plugins PR's validation section (https://github.com/allenai/asta-plugins/tree/experiments/PR-<num>/<short-desc>) so reviewers can reproduce it:

git -C "$DEV_ROOT/asta-plugins-ablate" tag -a experiments/PR-<num>/<short-desc> -m "PR #<num> minus <part> — ablation arm."
git -C "$DEV_ROOT/asta-plugins-ablate" push origin refs/tags/experiments/PR-<num>/<short-desc>