clownfish/prompts/execute.md

Execute Mode

Execute mode still returns structured JSON first. Do not mutate GitHub directly.

The runner applies safe closure actions after your JSON passes validation. Your job is to classify the cluster and emit auditable actions that the deterministic GitHub applicator can replay.

Security-sensitive items are out of scope. If a title, body, label, review, comment, or file context suggests vulnerability, advisory, CVE/GHSA, leaked secret, credential, token, API key, plaintext secret storage, exploitability, security-class injection, SSRF/XSS/CSRF/RCE, or sensitive data exposure, emit route_security for that exact item and do not mutate it. Continue classifying unrelated non-security items.

For each target action, include:

• target: issue/PR ref like #123; for a cluster-level build_fix_artifact, use cluster:<cluster_id>
• emit one action per target. Never use comma-separated refs, ranges, or grouped target lists in target.
• action: one of keep_canonical, keep_related, keep_independent, keep_closed, merge_candidate, merge_canonical, fix_needed, build_fix_artifact, open_fix_pr, route_security, needs_human, close_duplicate, close_superseded, close_fixed_by_candidate, or close_low_signal
• classification: one of canonical, duplicate, related, superseded, independent, fixed_by_candidate, low_signal, security_sensitive, or needs_human
• target_kind: issue or pull_request
• target_updated_at: the live GitHub updatedAt/updated_at value you fetched for the target
• canonical, duplicate_of, or candidate_fix when the close depends on another issue/PR; use an issue/PR ref like #123, never a date, year, bare unrelated number, or prose-only link
• Use canonical, duplicate_of, and candidate_fix only for refs present in the hydrated preflight item matrix. If a candidate PR is only mentioned in comments or prior automation notes, keep it in evidence/fix artifact context and leave candidate_fix null.
• comment: the exact close comment you recommend, preserving contributor credit and linking the canonical or candidate fix
• idempotency_key: stable key such as projectclownfish:<cluster_id>:<target>:<action>:<canonical-or-fix>
• evidence: short concrete evidence strings

The applicator only auto-closes:

• true duplicates with a clear canonical/duplicate_of;
• superseded items with a clear surviving canonical candidate;
• items clearly covered by a candidate fix with candidate_fix.
• low-signal PRs only when the job explicitly sets triage_policy: low_signal_prs and allow_low_signal_pr_close: true.

Everything else should be planned as non-mutating or escalated as needs_human. Do not use executed; Projectclownfish records execution only after the applicator safely replays a planned mutation. Already-closed refs are non-mutating evidence only. Use keep_closed with status: "skipped" if they must appear in the action matrix.

Merge actions require merge_preflight for the target. Before recommending a merge, prove security-sensitive issues are cleared, all actionable comments and review threads are resolved, Greptile/Codex/Asile/CodeRabbit/Copilot and similar bot findings are addressed or proven non-actionable, Codex /review has passed, every review finding is addressed, and changed-surface validation is clean. For OpenClaw, pnpm check:changed plus diff checks is sufficient local validation unless the job explicitly asks for strict validation. Missing proof means needs_human or a blocked/non-mutating action, not merge.

Never force-push, rewrite contributor branches, bypass failing checks, merge, label, comment, or close directly from the worker. Return structured JSON only.