v-collect

You are running /v:collect — the collect → scope-gate → review tail of the Compound V pipeline, on an already-dispatched run. It does not re-dispatch workers (use /v:resume for that). It re-normalizes results, re-runs the file-scope gate, and re-runs the Review Gate — the deterministic, idempotent back half of a run.

The run-id is {{args}}.

The run-dir layout and per-job status semantics are in skills/compound-v/state-machine.md; the manifest each job is checked against is execution-manifest.md; the canonical job_result shape is schemas/job_result.schema.json.

Steps

1. Locate the run. If {{args}} is empty, list the subdirectories of docs/superpowers/execution/ and ask which run to collect. The run dir is docs/superpowers/execution/<run-id>/. If it does not exist, stop and say so. If no jobs have been dispatched yet (phase earlier than DISPATCHED), tell the user there is nothing to collect and stop.

   Branch on the manifest kind. If manifest.yaml carries a fast_path block (a v2.9 pre-eval-backed fast-path run — its phase is FASTPATH_DISPATCHED), do not run the ordinary three-pass review tail (steps 3–4). Run the fast-path authoritative sequence below instead (CR5-2). Step 2 (collect/normalize) still applies to the single implementer job. A legacy (non-fast_path) manifest uses steps 2–5 unchanged.

2. Collect — normalize each job's output. The collector is per-job (required --job-id + --run-dir; no positional run-dir). Loop over every dispatched <job-id>, normalizing its heterogeneous worker output into a canonical results/<id>.json:

   # for each dispatched <job-id>:
   python3 scripts/compound-v-collect-results.py \
     --job-id <job-id> \
     --run-dir docs/superpowers/execution/<run-id>/ \
     --scope <scope-verdict.json> \
     --worker-output <last-message> \
     --schema schemas/job_result.schema.json
   

   Output lands in <run-dir>/results/<job-id>.json. The collector validates each result against schemas/job_result.schema.json. Its files_changed / violations / blocked fields are git-derived (folded in from --scope), never model-self-reported; --worker-output feeds only the human summary. It prints no fabricated cost metrics.

3. Scope gate — git-diff authority, every job. Re-run the scope gate on each job against its write_allowed from the manifest. The script takes a mutually-exclusive mode flag (--worktree OR --repo) plus the allowed globs (--allow-file) — no positional path. Use the mode matching how the job was dispatched (mirrors phase-3-parallel-opus-dispatch.md):

   # worktree jobs:
   python3 scripts/compound-v-scope-check.py --worktree <wt-dir> --allow-file <allow-globs-file>
   # direct (in-repo) jobs:
   python3 scripts/compound-v-scope-check.py --repo <repo-dir> --baseline <pre-dispatch-commit> --allow-file <allow-globs-file>
   

   Exit codes: 0 = pass, 1 = blocked (violations present), 2 = usage/git error. It derives changed files with git -C <worktree-or-repo> diff --name-only HEAD ∪ git ... ls-files --others --exclude-standard. Any path outside write_allowed ⇒ the job is BLOCKED: mark it blocked in state.json, retain its worktree, do not merge it, and HALT the run — surface the offending paths to the user. A BLOCKED job is corrected and re-dispatched via /v:resume, not silently merged.

4. Review Gate — three passes (Opus), AC-gated. If every job passes the scope gate, run spec-reviewer's three passes:

   • SPEC — each job satisfies its own acceptance in the manifest.
   • QUALITY — code quality, no regressions, no fabricated metrics.
   • INTEGRATION — cross-job seams build, and the composite change satisfies the feature-level acceptance_criteria. DONE is gated on all three. Unresolvable reviewer ISSUES ⇒ HALT (do not merge).