| name | review |
| description | Run a periodic chunked code review against one shard of the repo. Use /review <shard> to start a review, /review next to rotate to the oldest-reviewed shard, /review act <findings-path> to triage findings and apply fixes, /review status to see what's pending. Spawns the reviewer/fixer/verifier subagents from .claude/agents/. |
| disable-model-invocation | true |
Chunked agentic-review skill
Orchestrates the periodic code-review cycle for envy-nx. Background and rationale: docs/decisions.md → "Process decisions". Framework artifacts: design/reviews/README.md.
This skill is never auto-invoked (disable-model-invocation: true). It runs only when the user explicitly types /review followed by a verb or shard name.
Argument dispatch
This skill is invoked as /review <args>. It is the only skill in the namespace; there is no separate /review-act or /review-status skill (Claude Code resolves slash commands to skill names, and a hyphenated form like /review-act would look up a skill named review-act which does not exist).
Dispatch rule on first arg:
act <findings-path> → triage and act (see the section below); spawns fixer + verifierfix <finding-id> → re-run the fixer on a single finding with revised guidance, after a verifier flag (auto-loads the active findings JSON from state.json)next → rotate to the oldest-reviewed shard and review itstatus → print the rotation table<shard-name> (any other value) → run a review on that shard- (no args) → print the rotation table (same as
status) plus a one-line usage hint
Reserved verbs (act, fix, next, status) take precedence over shard-name resolution. Do not add a shard named act, fix, next, or status to shards.json - there is no automated check for this collision (would need a dedicated hook), so the gate is this dispatch table. If the shard list grows past a single maintainer, add a check_shards.py hook that rejects reserved names.
Commands
/review <shard> - review one shard
-
Run python3 .claude/skills/review/scripts/review_prepare.py <shard>. The script validates the shard name, verifies the charter exists, computes today's date, git HEAD sha, charter sha256, and loads open followups. On non-zero exit, relay the stderr message and stop. On success, parse the JSON output - it contains all fields needed for the reviewer prompt.
-
Spawn the reviewer subagent (.claude/agents/reviewer.md) with this prompt:
Review shard: <shard>
Charter: <charter-path>
Charter sha: <charter-sha>
File globs: <comma-separated globs from shards.json>
Output path: .wolf/reviews/<date>-<shard>.json
Git sha: <head-sha>
Reviewed at: <iso-8601-utc>
Open followups: <list of {id, description, source_cycle} or empty>
Load the charter, the accepted-decisions ledger (design/reviews/accepted.md), and the followups list above. Walk the files matched by the globs. Emit findings as specified in the charter's severity rubric. For each open followup, decide whether to re-emit as a finding (prefix `[FU-NNN]` in the finding field) or skip it for this cycle (it stays open). Append today's date to each considered FU's `considered_in_cycles` in `.wolf/follow-ups/<shard>.json`. Write findings to the output path. Report back: total finding count, severity breakdown, output path, and which FU-NNNs were re-emitted vs skipped.
-
After the reviewer returns: update .wolf/reviews/state.json (create if missing) with the new last-run timestamp for this shard.
-
Print a one-line summary to the user: Reviewed <shard>: <N> findings (<critical>/<high>/<medium>/<low>). See <output-path>. Run /review act <output-path> to triage.
/review next - rotate to oldest shard
- Run
python3 .claude/skills/review/scripts/review_next.py. It prints the shard name with the oldest last_run (never-run first, ties broken by blast_radius desc).
- Run
/review <picked-shard>.
/review act <findings-path> - triage and act
-
Read the findings JSON at <findings-path>. Validate charter_sha against the current charter file's sha - if mismatched, warn the user the findings are stale and ask whether to proceed anyway or re-run review first.
-
For each finding, ask the user via AskUserQuestion:
F-XXX [<severity>/<category>] <file>:<line>
<finding>
Suggestion: <suggestion>
apply | defer | dispute?
- On defer: prompt for a one-line rationale and a
Re-evaluate when: trigger (optional). Append a new entry to design/reviews/accepted.md (next monotonic A-NNN ID).
- On dispute: prompt for a one-line rationale (why this isn't actually a problem). Append to
accepted.md as Decision: dispute (no Re-evaluate when:).
- On apply: keep in the apply-set for the fixer.
After each decision, run python3 .claude/skills/review/scripts/review_triage.py <findings-path> <finding-id> <decision> [rationale]. The script atomically writes triage_decision (and triage_rationale for defer/dispute) to the findings JSON. If the finding text starts with [FU-NNN], the script also closes the corresponding followup in .wolf/follow-ups/<shard>.json.
-
After triage:
- If apply-set is empty: report "no findings marked for fix; defers/disputes recorded in accepted.md" and stop.
- Otherwise: record
base_sha=$(git rev-parse HEAD) and base_branch (the currently checked-out branch). Then spawn the fixer subagent (.claude/agents/fixer.md) with the apply-set and findings path. The fixer writes fixer_outcome and fixer_commit per finding back to the JSON. Wait for it to return. Capture the fixer's worktree_path and fixer_branch from the agent result.
-
Persist the fixer's followups output. The fixer's report includes a structured followups array - things it noticed but didn't act on (per the minimum-scope rule). For each entry, append to .wolf/follow-ups/<suggested_home_shard>.json (creating the file if absent). Each new entry gets the next shard-local FU-NNN id, plus initial fields: status: open, source_cycle: <today>, source_shard: <current shard>, source_findings_path: <findings path>, considered_in_cycles: []. Multiple entries with different home shards land in different files.
-
Spawn the verifier subagent (.claude/agents/verifier.md) with the findings path and the diff range (git diff <base>...HEAD). The verifier writes verifier_verdict and verifier_note per finding back to the findings JSON, AND checks open followups for this shard against the diff (auto-closing any that the diff incidentally resolves). Report the verifier's per-finding verdicts and any followup auto-closures to the user.
-
Merge the fixer's branch and clean up the worktree. After the verifier returns:
- Verify fast-forward is possible: check that
base_sha still equals git rev-parse <base_branch> (i.e., the base branch has not moved since the fixer started).
- If fast-forward is possible: run
git merge --ff-only <fixer_branch> on <base_branch>, then git worktree remove <worktree_path> (retry with --force if locked), then git branch -d <fixer_branch>.
- If fast-forward is NOT possible (base branch moved during the cycle): do NOT merge. Warn the user:
"Merge conflict: <base_branch> has moved since the review started. The fixer's branch <fixer_branch> is intact in <worktree_path>. To merge manually: git merge <fixer_branch>. To clean up after: git worktree remove <worktree_path> && git branch -d <fixer_branch>." Stop here.
-
Report to the user. Print the number of commits ahead of origin and the verifier summary. If any verifier verdicts are non-confirmed, list the flagged findings with their verdicts and remind the user about /review fix F-NNN. The user may push when ready.
/review fix <finding-id> - re-run the fixer on one finding with revised guidance
For when the verifier flagged a fix as symptom-only, regression-risk, over-corrected, or not-applied. Targets a single finding by ID; auto-loads the active findings JSON from state.json so no path argument is needed.
- Locate the active findings JSON. Read
.wolf/reviews/state.json; pick the last_findings_path of the most recently active shard. If multiple shards have a finding with this ID (rare - IDs are file-local, not global), ask the user which shard.
- Validate the charter sha. If the findings JSON's
charter_sha doesn't match the current charter sha, warn the user the findings are stale and ask whether to proceed.
- Locate the finding by
<finding-id>. If not found in the active JSON, list available IDs and stop. If the finding's existing verifier_verdict is confirmed, warn the user that re-fixing isn't normally needed and ask whether to proceed.
- Display the finding's current state in the user's terminal (so they don't have to recall it):
- The original
finding and suggestion text (immutable from the reviewer)
- The most recent
verifier_verdict and verifier_note (if present)
- The
fixer_commit SHA (if present) so the user can git show <sha> to inspect the prior fix
- The current
retry_count (0 means this is the first retry)
- Prompt for revised guidance via
AskUserQuestion: free-text, one paragraph describing what the previous fix missed and what the new fix should address. Empty guidance means "try again with the same finding text"; usually you'll want concrete direction informed by the verifier's prior note.
- Record base state and spawn the fixer subagent. Record
base_sha=$(git rev-parse HEAD) and base_branch. Spawn the fixer with the finding's full payload, the verifier's prior verdict + note, and the user's revised guidance. The fixer follows its usual protocol: minimum-scope edit, gate on make lint && make test-unit, one commit. Commit message suffixed [F-NNN, retry-N] (where N is the new retry_count) so the original commit stays distinguishable in git log. Capture worktree_path and fixer_branch from the agent result.
- Fixer writes back the new
fixer_outcome, fixer_commit, and incremented retry_count to the findings JSON for this finding.
- Spawn the verifier subagent on the new diff range. The verifier overwrites the prior
verifier_verdict and verifier_note for this finding with the new ones (this is the one author-may-overwrite-own-field exception to the append-only invariant).
- Merge and clean up the worktree. Same protocol as
/review act step 6: fast-forward check on base_sha, git merge --ff-only, remove worktree, delete branch. If fast-forward fails, warn and leave intact.
- Report the new verdict to the user. Fixes are already on
<base_branch>. If still flagged, the user can /review fix F-NNN again (a fresh worktree will be created each time), push as-is and address the remainder manually, or demote the finding to accepted.md.
/review fix is targeted (single finding, no batch) and context-discovering (no path arg). It is NOT a redo of /review act - the original triage decisions stand; only the fixer phase re-runs for the one finding you named. To retry multiple findings, run /review fix F-NNN once per finding.
/review status - show review cadence
Run python3 .claude/skills/review/scripts/review_status.py and print its stdout verbatim - do NOT reformat, summarize, or duplicate the table. The script outputs a fixed-width aligned plaintext table. The only addition allowed is a one-line hint after the table (e.g., which shard /review next would pick).
Findings JSON schema
The reviewer writes initial findings; subsequent agents and commands ADD fields as the cycle progresses. The original reviewer fields are never modified after initial write - this preserves the forensic record of what the reviewer originally said.
{
"shard": "<shard-name>",
"charter_path": "design/reviews/charters/<shard>.md",
"charter_sha": "<sha256-of-charter-at-review-time>",
"reviewed_at": "<iso-8601-utc>",
"git_sha": "<git-rev-parse-HEAD-at-review-time>",
"findings": [
{
"id": "F-001",
"file": "<repo-relative-path>",
"line": 87,
"severity": "critical | high | medium | low",
"category": "correctness | security | maintainability | testability | docs",
"finding": "<1-3 sentence description of the issue>",
"suggestion": "<concrete fix direction, NOT a patch>",
"triage_decision": "apply | defer | dispute",
"triage_rationale": "<one-line - present for defer/dispute>",
"fixer_outcome": "applied | failed | skipped",
"fixer_commit": "<sha - present if outcome is applied>",
"fixer_failure_reason": "<captured output - present if outcome is failed>",
"retry_count": 0,
"verifier_verdict": "confirmed | symptom-only | regression-risk | scope-creep | not-applied | over-corrected",
"verifier_note": "<explanation - present for non-confirmed verdicts>"
}
]
}
Append-only invariant. The original reviewer fields (id, file, line, severity, category, finding, suggestion) are never modified after the reviewer writes them. Cycle agents only ADD new fields, or update fields they themselves previously added (e.g., the verifier overwrites its own verifier_verdict on a /review fix re-verify; the fixer increments its own retry_count). If the reviewer's findings turn out to be wrong (wrong file path, wrong severity), run /review <shard> again to produce a fresh findings JSON - never edit the immutable fields in place.
Practical consequence: the findings JSON is a self-contained record of one cycle. A reader of one file can reconstruct what the reviewer found, what the human decided, what the fixer did, and what the verifier concluded - useful for cross-session continuity, post-mortems, and /review fix UX.
The charter_sha field is load-bearing: when /review act or /review fix runs against a findings JSON whose charter has since changed, it warns the user before proceeding rather than acting on outdated criteria.
Triage state
/review act writes each finding's triage_decision (and triage_rationale for defer/dispute) to the findings JSON via jq per the schema above. The original reviewer fields stay immutable; only new fields accumulate. This means a partially-completed cycle can be picked up by a different session or even a different machine: anyone reading the findings JSON sees exactly which findings have been triaged and which haven't.
Followups JSON schema
The fixer often notices things during a cycle that don't fit the current finding's minimum-scope rule but warrant consideration eventually. The framework persists these as followups so they aren't lost between cycles.
.wolf/follow-ups/<home-shard>.json (per-machine ephemeral, gitignored - same lifetime as findings JSON):
{
"shard": "<home-shard-name>",
"entries": [
{
"id": "FU-001",
"source_cycle": "<iso-date - when the fixer noticed it>",
"source_shard": "<shard the fixer was working on when it noticed>",
"source_findings_path": ".wolf/reviews/<date>-<source-shard>.json",
"description": "<one sentence describing what to consider>",
"source_finding_ids": ["F-002", "F-007"],
"status": "open | closed",
"considered_in_cycles": ["<iso-date>", ...],
"closed_at": "<iso-date or null>",
"closed_via": "triage-applied | triage-deferred | triage-disputed | verifier-auto-resolved | manual",
"verifier_verdict": "auto-resolved-by-diff | partially-resolved-by-diff | null",
"verifier_note": "<sentence with commit SHA, or null>"
}
]
}
Lifecycle. A followup is created open by the fixer (via /review act step 4). It can transition to closed via four paths:
- Triage closure (most common): on the next
/review <home-shard> cycle, the reviewer re-emits the FU as a finding (with [FU-NNN] prefix). When the human triages that finding (apply / defer / dispute), /review act closes the FU with closed_via: triage-<decision>.
- Verifier auto-close: on any
/review act for this shard, the verifier checks open FUs against the fixer's diff. If a fix incidentally resolved a FU's concern, the verifier closes it with closed_via: verifier-auto-resolved. Independent bias control - the fixer focuses on its own finding and won't notice incidental followup resolution; the verifier reads the diff cold.
- Reviewer skip (no closure): if the reviewer chooses NOT to re-emit a FU (still applicable but not worth flagging right now), the FU stays open. Its
considered_in_cycles array grows so the human can spot stale FUs that have been considered many times without action - candidates for manual pruning.
- Manual closure: human edits the JSON to set
status: closed, closed_via: manual. Recovery path for verifier errors and stale FUs.
Cross-shard followups. The fixer may notice things outside its current shard (e.g., a WSL refactor noticed during a test-quality fix). The followup goes to the file named after the suggested home shard, so the next reviewer for that shard picks it up - not the source shard. Cross-shard observations would otherwise be orphaned.
Persistence trade-off. The followups file is gitignored (under .wolf/), so it lives only on the machine where reviews run. Load-bearing followups (those that absolutely must survive a machine wipe) should be promoted to durable artifacts at triage time: a PR, a GitHub issue, or a charter v2 entry. The followups file is the "remember next time" inbox, not the archive.
Rotation state
.wolf/reviews/state.json:
{
"shards": {
"<shard-name>": {
"last_run": "<iso-8601-utc>",
"last_findings_path": ".wolf/reviews/<date>-<shard>.json",
"last_finding_count": 12
}
}
}
Updated at the end of every /review <shard> run. Read by /review next and /review status. Per-clone - not synced via git.
Safety constraints (enforced by agent frontmatter, documented here for the reader)
- Reviewer cannot edit code.
.claude/agents/reviewer.md has tools: Read, Grep, Glob, Bash - no Edit/Write. This is intentional bias control: the reviewer cannot "pick easy issues" because it can't fix anything.
- Verifier cannot edit code. Same restriction. The verifier's job is to flag "fixed the symptom, not the cause" or "fix introduced an unrelated change". It escalates via report, never overrules.
- Fixer cannot ignore the human. The fixer only acts on findings the user marked
apply during /review act. If a fix turns out to be wrong (test fails, unintended scope), it marks the finding disputed (fix-broke-tests) with the failing output captured, rather than silently dropping it.
- Fixer hard DONE marker.
make lint && make test-unit must pass before the fixer reports completion. This is the deterministic verification per Anthropic's "give Claude a way to verify its work" guidance - not the agent's own judgment.
Worktree lifecycle
The fixer subagent is always spawned with isolation: "worktree" to prevent make lint && make test-unit from affecting the user's working tree. The lifecycle is fully managed by the skill - the user never interacts with worktree artifacts directly.
- Create. The skill spawns the fixer with worktree isolation. Claude Code creates a worktree under
.claude/worktrees/agent-<id> and the fixer creates a branch review/<shard>-<date>.
- Use. The fixer commits fixes on this branch. The verifier validates the diff and runs
make lint && make test-unit at the branch HEAD.
- Merge. After the verifier returns, the skill fast-forward merges the fixer's branch into the base branch. This is safe because the fixer branched from the base branch's HEAD at cycle start.
- Clean up. The skill removes the worktree (
git worktree remove) and deletes the local branch (git branch -d).
- Failure mode. If the base branch moved during the cycle (another commit landed), fast-forward is not possible. The skill leaves the worktree and branch intact and prints manual merge instructions.
/review fix follows the same lifecycle: fresh worktree per retry, merge and cleanup after the verifier returns.
Manual cleanup. If a session crashes mid-cycle, a dangling worktree may remain. Clean up with:
git worktree remove .claude/worktrees/<dir>
git branch -d review/<shard>-<date>
Out of scope (intentionally not implemented)
- Fan-out across all shards in one command. Reviewing all 8 in parallel would consume context and produce a noisy combined inbox. Manual rotation via
/review next is intentional.
- Auto-trigger on cron. WSL distros idle-shutdown after ~8s; cron is unreliable. Manual trigger also keeps the human in the loop on every cycle.
- Persistent committed findings archive. The fixer's commit history (on the base branch after merge) and
design/reviews/accepted.md are the historical record. Raw findings JSON in .wolf/reviews/ is forensic and ephemeral by design. The user may still open a PR for team review before pushing.
