Review Clawhub Profile Proposal

Review ClawHub Profile Proposal

Overview

Use this owner-side workflow for a public PR that proposes a candidate ClawHub profile at proposals/<GHSA-ID>/clawscan.yml after a private vulnerability report. The goal is to validate the candidate, keep sensitive details private, and, if accepted, promote the public profile behavior into the bundled internal/profiles/clawhub/clawscan.yml.

Safety Boundary

• Treat the GitHub private vulnerability report as the source for sensitive malicious skill details.
• Do not paste live exploit details, private report text, private artifacts, or suspicious skill payloads into public PR comments, public docs, commit messages, or committed baseline summaries.
• The public proposal PR should start with only proposals/<GHSA-ID>/clawscan.yml. Do not trust it until reviewed.
• Do not run the suspicious skill. ClawScan scans skill files as data; it should not execute the skill's behavior.
• Do not promote a candidate unless the private report, proposal diff, and benchmark proof all line up.

Review Workflow

1. Read the current repo and PR state.

   • Check git status --short --branch before editing.
   • Fetch the PR branch and inspect the changed files.
   • Confirm the public PR initially contains only proposals/<GHSA-ID>/clawscan.yml plus any baseline summary commit produced by the maintainer gate.
   • Confirm the proposal file defines a clawhub profile.
   • Confirm the proposal does not edit official bundled profile files yet: internal/profiles/clawhub/clawscan.yml, internal/profiles/clawhub/prompt.md, or internal/profiles/clawhub/output.schema.json.

2. Read the private report privately.

   Use the private vulnerability report only to understand:

   • what malicious behavior must be caught
   • why the current built-in clawhub profile misses or under-detects it
   • what evidence should be preserved privately
   • whether any proposal text or config comments leak sensitive details

   Keep public comments high-level, such as "validated against the private report" or "needs private-case follow-up"; do not quote private details.

3. Validate the candidate profile.

   Prefer the manual GitHub Actions workflow when available:

   SkillTrustBench Profile Gate
   

   Dispatch it with the PR number and proposal path. It should run:

   clawscan benchmark SkillTrustBench \
     --ids https://huggingface.co/datasets/cuhk-zhuque/SkillTrustBench-results/resolve/main/data/evaluation_subset_10pct.jsonl \
     --config proposals/<GHSA-ID>/clawscan.yml \
     --profile clawhub \
     --output ./artifacts/skilltrustbench-candidate.json
   

   If running locally, use the same command. The --ids source is the public SkillTrustBench leaderboard subset and is mutually exclusive with --limit and --offset.

4. Review the benchmark artifact.

   Check:

   • the artifact is full JSON and preserved as a workflow artifact or private maintainer artifact
   • benchmark.id is cuhk-zhuque/SkillTrustBench
   • benchmark.split is benchmark
   • benchmark.idsCount is 556
   • benchmark.idsSha256 matches the planned subset hash 903a036e4b7b16ee28e22d5d9db57a00b3764cfe41e43144acad67921e5196c2
   • scanner and judge statuses are acceptable
   • evaluation metrics are not an unacceptable regression
   • the candidate catches the private reported behavior when private proof is available

5. Update the accepted baseline summary.

   Use the repo script so the compact baseline is generated from the full candidate artifact:

   go run ./scripts/update-skilltrustbench-baseline \
     --artifact ./artifacts/skilltrustbench-candidate.json \
     --output benchmarks/skilltrustbench-leaderboard-10pct/<YYYY-MM-DD>.json \
     --profile clawhub \
     --profile-source proposals/<GHSA-ID>/clawscan.yml \
     --subset-case-ids-sha256 903a036e4b7b16ee28e22d5d9db57a00b3764cfe41e43144acad67921e5196c2 \
     --workflow-url <workflow-url>
   

   If the PR merges, the newest dated JSON file in benchmarks/skilltrustbench-leaderboard-10pct/ is the latest accepted baseline for the bundled clawhub profile. The script fails if the candidate artifact's selected-ID hash does not match the planned subset. No post-merge rerun is required.

6. Decide.

   If rejected:

   • leave a public PR comment with non-sensitive reasons
   • keep details that identify the malicious payload in the private report
   • do not edit official bundled profile files

   If accepted:

   • promote the accepted public clawhub profile behavior into internal/profiles/clawhub/clawscan.yml
   • add a dated baseline under benchmarks/skilltrustbench-leaderboard-10pct/ from the candidate artifact in the same PR
   • preserve or remove proposals/<GHSA-ID>/clawscan.yml according to the issue/PR instruction; default to preserving it as public proposal trail unless the maintainer explicitly chooses to remove it
   • keep prompt/schema changes maintainer-owned; edit internal/profiles/clawhub/prompt.md or internal/profiles/clawhub/output.schema.json only when that is the accepted change

7. Verify the promoted built-in profile.

   Run at least:

   go test -count=1 ./...
   go vet ./...
   go run ./cmd/clawscan profiles -v
   go run ./cmd/clawscan --help
   

   For benchmark proof after promotion, run:

   clawscan benchmark SkillTrustBench \
     --ids https://huggingface.co/datasets/cuhk-zhuque/SkillTrustBench-results/resolve/main/data/evaluation_subset_10pct.jsonl \
     --profile clawhub \
     --output ./artifacts/skilltrustbench-clawhub.json
   

   Use a smaller proof only when explicitly accepted; the official gate uses the subset ID source above.

8. Update the PR.

   • Push the promotion commit to the PR branch if that is the chosen review path.
   • Keep the PR body/comments free of private exploit details.
   • Add proof with commands, artifact links, and residual risk.

Promotion Patch Shape

The maintainer promotion commit usually touches:

internal/profiles/clawhub/clawscan.yml
benchmarks/skilltrustbench-leaderboard-10pct/<YYYY-MM-DD>.json

It may also touch:

internal/profiles/clawhub/prompt.md
internal/profiles/clawhub/output.schema.json
docs/
tests

Do not include private artifacts, malicious payload details, or generated dist/ output in ordinary promotion commits unless the issue explicitly asks for them.

Handoff Shape

End with:

• verdict: accepted, rejected, or blocked
• proposal path and PR/ref reviewed
• private report checked, without sensitive details
• benchmark command and artifact location
• baseline summary update status
• bundled profile files changed
• exact verification commands and results
• commit SHA or reason no commit was created
• residual risk and next owner action