---

name: harness-pr-reviewer description: Checks a submitted PR against the environment's baseline assets (CLAUDE.md, memory, naming, asset classification) and attaches a review comment with a merge recommendation. 5 steps — diff read, 8-area consistency check, self-catch, comment, merge recommendation. user-invocable: true allowed-tools: ["Bash", "Read", "Grep", "Glob"] model: sonnet complexity_routing: base: sonnet high: opus escalate_when: - adversarial - cross_project - high_stakes

Note: The original developer is the forge-harness original developer (development source + meta-monitoring home). In external user install environments, the install environment user themselves is the baseline integrity gate operator (following path B generalization baseline / SKILL_detail.md §External User Environment Adaptation Path §).

harness-pr-reviewer — Hub Gate Operation Rule Automation

When a PR is submitted, checks consistency against the user environment's baseline assets (CLAUDE.md · memory · naming · asset classification) and attaches a review comment. 5-step: diff read → 8-matrix check → self-catch → comment attachment → merge recommendation.

Activation Triggers

Renamed from hub-cc-pr-reviewer (2026-09-12, operator decision). The old name said where it was born (the hub's own cc reviewing its own PRs). What it actually is — measured on qasp/pmh PRs in 2026-09 — is the standalone doorway through which a field harness verifies its own submitted PR with FH's review capability. Old-name utterances ("hub-cc-pr-reviewer", "hub review", "hub cc review") still route here; downstream forks that carry the old directory name (PMH) keep working until they sync. No redirect stub directory is shipped (same policy as phantom-quench).

1. PR #N input: "Review PR #N" / "Check PR #N" / "hub review" (old-name alias) / "harness PR review" / "baseline consistency check"
2. Action leader cc → hub sync point: Large decision area PR catch (following Option C Hybrid policy — memory creation / CLAUDE.md change / CATALOG round / skill v0.x evolution / policy change / asset synergy branch judgment)
3. Hub cc session entry: Layer A auto-read recent external commit catch (auto-discover new PRs)

Natural Language Triggers (General user phrasing — activates without internal vocabulary)

Example phrasing	Intent
"Is it okay to submit this PR?"	PR review request
"This change seems inconsistent with existing rules"	Baseline consistency check
"Please review before merging"	PR review gate
"Does this change affect other parts?"	Consistency check

Activation criteria: "Review PR #N" / "Add review comment" / "Baseline check" → Run this skill directly
(pr-review-watcher deprecated as of v0.2.0 — recommend using gh pr view --json reviews directly)

Exceptions (this skill does NOT apply):

• Small patches (typo / 1-line cross-ref addition / sync / minor adjustments) → Follow Option C Hybrid policy (direct push allowed area / review skip)
• Original developer simple correction commands ("This is wrong, redo it") → Immediate correction (no review / direct handling)

Processing Steps (5-step)

Step 1. PR Diff Read

Read the PR diff + metadata. If this cc authored the change, the diff read can be skipped (directly state changed areas in PR body).

Detail: See SKILL_detail.md §Step 1 Diff Read — gh pr diff + gh pr view commands — read when executing the diff read.

Step 2. Baseline Consistency Check — 8-Matrix Auto-Generation

#	Area	Check path
1	CLAUDE.md (hub identity + asset ownership + sync policy)	Grep PR diff vs CLAUDE.md baseline areas
2	Memory accumulation (accumulated naming/decision baseline + asset synergy branch judgment + active onboarding + bidirectional self-validation, etc.)	Grep PR diff vs memory feedback_*.md key areas (External environment: skip this item if memory files absent → SKILL_detail.md §External User Environment Adaptation Path §)
3	Naming baseline (accumulated naming baseline + new naming candidate area)	Catch new naming candidates from PR diff / check adherence to existing naming
4	Asset synergy branch judgment (meta/hub seed vs action leader persistent location)	Check PR changed asset location consistency
5	Simplification guard (P15 asymmetry catch + R7 over-engineering)	New asset creation vs existing asset reinforcement judgment / body length check
6	Dimension separation baseline (## Plugins / ## Skills / ## Agents)	Check dimension separation consistency on CATALOG changes
7	Branch criteria (large decision PR mandatory vs small patch direct push)	Check if PR is a large decision area (Option C Hybrid)
8	Hub gate operation consistency	Check if PR itself is a hub gate operation proof path

Matrix result = Consistent ✅ / Partially Consistent ⚠️ / Inconsistent ❌.

Step 3. Layer 5 Self-Catch Matrix

Self-precision catch areas after first cc review (following previous PR self-catch patterns):

• Check adherence to frontmatter description plain text only baseline (project baseline)
• Check honest documentation of generalization effect weakening areas
• Check explicit statement of gap between accumulated history (original developer environment) vs external user starting point (0 instances)
• Check explicit statement that audience-specific guides are limited to original developer environment
• Check explicit statement of organization-specific areas

Self-catch areas 0 items = skip this entire catch matrix — do not pad with token-filling to make the section look populated.

🟥 Self-catch is a CUE, not a verdict. Its output is the input to Step 3.5 below. A self-catch that finds nothing does not clear Axis 2 or Axis 3 — measured 2026-09-08/11: a reviewer session with no external cue ran the load-bearing gate 0/51, and a PMH session reviewing qasp PR #13 dispatched the cross-family panel only after the operator pushed (pmh-dev #77). The self-check is the reviewer; it cannot be its own decorrelation.

Step 3.5. Axis 2 · Axis 3 — mandatory dispatch lane (mechanical trigger, typed degrade)

Axis 1 above is wired at mandatory strength (--pr mode, typed verdict). Until 2026-09-12 Axis 2 (adversarial panel) and Axis 3 (phantom-quench) had no dispatch block in this skill at all — only the self-catch matrix — so they fired only when someone said so. This step closes that with a trigger that is decided from the PR, never from the reviewer's feel.

Trigger — mandatory if ANY of these holds (compute all three; record which fired):

(i)   Step 2 matrix has ≥1 ❌  (Inconsistent)
(ii)  the PR touches a load-bearing FH asset: SKILL.md · SKILL_detail.md · .claude/rules/*.md ·
      knowledge/shared/rules/*.md · templates/** · scripts/**/*.sh · scripts/**/*.py ·
      plugins/*/agents/**   (same list as CLAUDE.md §FH Improvement 4-Axis Auto-Gate)
      → gh pr diff "$PR" --name-only | grep -E '(^|/)SKILL(_detail)?\.md$|^\.claude/rules/|^knowledge/shared/rules/|^templates/|^scripts/.*\.(sh|py)$|^plugins/[^/]+/agents/'
(iii) a merge recommendation (Step 5) will be issued for this PR

None of the three → record axis2: not-triggered(reason) · axis3: not-triggered(reason) in the comment capsule and continue. This is the only non-run path, and it is stated, never silent.

Axis 2 — cross-family adversarial panel. Dispatch auto-decorrelation on the PR diff (reuse, not a new engine — its Step 1 load-bearing predicate, Step 4.5 residency screen and Step 6 degrade ladder all apply unchanged). What this step owns is only the decision to dispatch. The result lands as the typed crossfamily: value: panel(<families>) with the source-grounded findings, or one of declined · DEGRADED_SINGLE_FAMILY · DEGRADED_PANEL_UNUSED · UNKNOWN with grounds — e.g. consent OFF, no different-family sidecar reachable, quota exhausted. A degrade value is allowed; an absent value is not. 🟥 A judgment-type question put to the panel (by-design vs fail-open, is this a regression) needs reps ≥ 3; a 1-of-3 split is recorded as unresolved, never as CONCUR (field_verdict_crossfamily_gate.md §4-2).

Axis 3 — phantom-quench. Run /phantom-quench over the PR's changed documentation surfaces (any *.md in the diff, plus every path/citation the PR body asserts). Mechanical N/A only when gh pr diff --name-only has zero *.md files and the PR body cites no path — record axis3: N/A(no doc surface). A phantom finding is a ❌ in the comment capsule, and it flips the Step 5 recommendation the same way an Axis 1 M-tier block does.

Degrade direction. This is a review surface (reversible): tooling-down does not block the comment, but the merge recommendation must carry the degrade value verbatim and read NOT-CONVERGED while Axis 2 is a DEGRADED_*/UNKNOWN value on a load-bearing PR. Silent same-family pass is the one outcome this step forbids.

Step 4. Review Comment Attachment

Mandatory before any gh pr comment: run /public-surface-audit over the composed comment text. A PR comment is a paste on a public surface, and the repo's mechanical privacy floor does not reach it — the pre-commit confidentiality guard scans staged tracked content and has no view of PR-body text (.claude/rules/fh_4axis_gate.md §Reviewer-visible evidence says so explicitly). This step's own inputs make that acute: Step 2 matrix #2 greps the operator's local memory files, so an unfiltered paste can carry absolute home paths and private memory prose onto a public PR.

verdict CLEAN                     → attach (automatic, within this skill's authority)
verdict REVIEW / LEAK             → do NOT attach. Redact the flagged spans, re-scan, then attach
verdict NOT_CONFIGURED, or the
  skill is unavailable            → do NOT attach automatically. This is an irreversible surface
                                    (a posted comment is public the instant it lands and may be
                                    mirrored before deletion) → **fail-closed**: hand the composed
                                    text to the operator, or take an explicit logged override

Never paste raw Step 2 grep output. Write a sanitized capsule — what was checked, what it returned, what was found — never the matched lines themselves. Same rule as the marker: the file is a local artifact, the capsule is what crosses the boundary.

Noise control — two rules that decide what reaches the comment (absorbed 2026-09-04 from the AI-Native SDLC playbook's REVIEW.md; FH already tiers findings M/S/R, but had no cap and no exclusion list, so a long PR produced a long comment that buried the one finding that mattered):

• Nit cap = 5. A nit is any finding that would not change behavior, leak data, or breach a baseline rule (style · naming · wording · ordering). Report at most five nits, then one line: +N more nits (not listed). M-tier and S-tier findings are never capped.
• Do-not-report list. Skip anything a mechanical gate already enforces on this repo (the pre-commit 4-axis gate, validate CI, regression_guard, public-surface scan) and generated or vendored files. Reporting what a gate already blocks is double coverage that reads as noise; if a gate should have caught it and did not, that is an S-tier finding about the gate, not a nit.

Then attach the review comment (8-matrix results + self-catch + refinement suggestions + merge recommendation) via gh pr comment.

Detail: See SKILL_detail.md §Step 4 Comment Template — gh pr comment heredoc template — read when attaching the comment.

Step 5. Admin Override Merge Recommendation

User decision delegation (this skill = review/recording automation / no merge authority):

• Read the branch-protection state at run time — never from this line. This repo moved to enforce_admins: true + required_approving_review_count: 0 on 2026-07-20, and an earlier version of this bullet still claimed false: a gate skill was recommending an override on a field that had already flipped. Protection is also two independent layers (legacy + rulesets, strictest wins), so one object is never the effective answer — check both: gh api repos/{owner}/{repo}/branches/main/protection and gh api repos/{owner}/{repo}/rules/branches/main
• Self-approve is impossible when this cc authored the PR → after a completed review, --admin is the normal route, not a shortcut
• Self-approve blocked (GHE policy) → admin override path adherence
• When this cc authored the change, admin override path is mandatory
• N+1th operation proof = baseline stabilization acceleration path

Detail: See SKILL_detail.md §Step 5 Merge Command — gh pr merge command (executed after user decision, not by this skill) — read when the user authorizes merge.

User Approval Gate

Stage	Approval
Step 1~3 check auto-activation	Automatic (editable afterward)
Step 4 review comment attachment	Automatic only after /public-surface-audit on the comment text returns CLEAN. REVIEW/LEAK → redact and re-scan; NOT_CONFIGURED or audit unavailable → fail-closed, hand to the operator (a posted comment is public on landing)
Step 5 admin override merge execution	User decision (this skill = recommendation only / no merge authority)

Constraints

• This skill = review/recording automation / no merge authority — user admin override or other reviewer merge decision
• No single-person decision application — following fact-checker rule (narrow 1 / broad N+1 / this cc self-catch joins fact-checker count)
• Simplification guard consistency — when creating/modifying this skill, update SKILL.md only. No new auxiliary files
• Markdown editing discipline mandatory — Edit first. No Write
• Frontmatter description plain text only baseline — avoid markdown bold

The three rules above were previously each attributed to a memory feedback_*.md file. Those files do not exist (verified 2026-08-11 against the operator's memory root, with a known-positive control in the same run). The rules stand on their own; only the pointers were dead, and citing a non-resolving file as the authority is the phantom-reference class this skill is supposed to catch. Do not re-attach a memory citation here unless ls resolves it in the same run that cites it.

Detail: See SKILL_detail.md §Sister Asset Utilization Path, §External User Environment Adaptation Path, §Disable Path, §Persona Synergy Catch — cross-ecosystem utilization, external-environment fallback, own-PRS disable resolution, and deep-insight simultaneous-activation handling — read when operating in an external user environment, resolving own-PRS conflict, or coordinating with deep-insight.

Done When

This is a gate/routing skill — its output is a merge verdict — so every judged condition below names its adversarial pairing. No judge-only path.

All 5 Steps completed
  — mandatory-pass: each step produced its output or is marked N/A with a reason

+ Baseline consistency check 8-matrix results output (OK/WARN/BLOCK each item)
  — measured: count items REPORTED vs items ACTUALLY CHECKED; the two must
    match. Matrix #2 (memory baseline) reports SKIPPED when no memory file
    resolves — it is never folded into the pass count (see §References)

+ Axis 1 run in --pr mode with a typed verdict read from
  REGRESSION_GUARD_RESULT_FILE
  — mandatory-pass: result is `pass` or `block`. `skip` and exit 3 are NOT
    passes; they mean Axis 1 did not examine this PR and the recommendation
    may not cite it as green

+ Step 3.5 trigger computed from the PR (matrix ❌ count · load-bearing path
  grep · merge-recommendation requested) and recorded with which clause fired
  — measured: the three booleans appear in the comment capsule. Absent = the
    step did not run, which is the pre-2026-09-12 defect (pmh-dev #77)

+ Axis 2 dispatched via auto-decorrelation when triggered, landing a typed
  `crossfamily:` value (panel(...) or a DEGRADED_*/UNKNOWN value WITH grounds)
  — mandatory-pass: the value exists and is one of the closed enum. A
    load-bearing PR whose value is DEGRADED_*/UNKNOWN leaves the merge
    recommendation NOT-CONVERGED; judgment-type panel questions carry reps ≥ 3
    or are recorded unresolved

+ Axis 3 phantom-quench run over the PR's doc surfaces when triggered, or
  N/A(no doc surface) decided by `--name-only` grep, never by feel
  — mandatory-pass: a phantom finding is a ❌ that flips the recommendation

+ /public-surface-audit run over the composed comment text BEFORE any
  gh pr comment
  — mandatory-pass, fail-closed: CLEAN attaches; REVIEW/LEAK redact-and-rescan;
    NOT_CONFIGURED or audit unavailable hands to the operator. A posted comment
    is irreversible, so tooling-down is a block, never a free skip

+ Review comment attached via gh pr comment command
  — mandatory-pass: the comment URL is returned by the command

+ Admin override merge recommendation output (merge execution is user's
  decision)
  — judged; adversarial pairing: the branch-protection state is re-read at run
    time from BOTH layers (`.../branches/main/protection` and
    `.../rules/branches/main`, strictest wins) in the same run that recommends.
    A recommendation citing this file's prose instead of a live read is
    unfounded — that exact defect already shipped once here, on a field that
    had flipped

+ External verification path: an isolated Critic agent can reach the same
  verdict from the artifacts alone
  — judged; adversarial pairing: the reviewer-visible evidence must be
    reproducible WITHOUT the author's local files. Any verdict resting on a
    gitignored local artifact ships labelled LOCAL-ONLY ATTESTATION -
    UNVERIFIED, which leaves the condition UNMET rather than met
    (`.claude/rules/fh_4axis_gate.md` §Reviewer-visible evidence)

→ Mandatory when PR contains SKILL.md / rules / templates changes: run Axis 1 (backward check) in --pr mode, against the PR's head branch — before the merge recommendation is issued. If regression_guard reports an M-tier block, the merge recommendation must change to ❌ regardless of other checks.

# Precondition: the PR head branch must exist locally. This skill reads the PR via `gh pr diff`
# without checking anything out, so fetch the head ref first or --pr has nothing to resolve.
PR_BRANCH="$(gh pr view "$PR" --json headRefName -q .headRefName)"
git fetch origin "$PR_BRANCH":"refs/remotes/origin/$PR_BRANCH"   # skip if already present
bash templates/regression_guard.sh --pr "origin/$PR_BRANCH"

⚠️ Do not run it with no arguments. Bare bash templates/regression_guard.sh diffs the working tree, and this skill's own workflow leaves the reviewer standing on a clean main — so the bare form returns REGRESSION_GUARD_RESULT=skip with exit 0 100% of the time, and the mandatory Axis-1 gate never examines the PR at all. Measured 2026-08-11 on a clean checkout: bare form → rc=0 / result=skip; --pr <branch> on the same commit → rc=0 / result=pass having actually read the changed SKILL.md. Canonical form is --pr {BRANCH} (.claude/rules/fh_4axis_gate.md).

Read the verdict from the typed channel, not the exit code — exit 0 means pass or skip (not-checked). Run with REGRESSION_GUARD_RESULT_FILE=/tmp/rg.$ and read result= from that file: skip means Axis 1 did not examine this PR (no matching file, or the wrong invocation form) — record it as "Axis 1 N/A", never as a green check. exit 3 means the invocation itself failed (unresolvable branch) — also not a pass; fetch the ref and re-run. A merge recommendation that cites an unexamined Axis 1 as PASS is the 2026-07-22 fail-open class.

References

⚠️ **The memory filenames below were audited 2026-08-11 and none of them exist — 8/8 absent in the operator's own memory root, i.e. they were never reachable, not merely absent externally. They are kept, struck, as the record of a phantom-reference class: a gate skill citing rule bodies that resolve nowhere, while matrix #2 silently "skips" and the run still reports an 8-matrix pass. Matrix #2 is therefore a 7-matrix in practice — report it as matrix 2: SKIPPED (no resolvable memory baseline) rather than folding it into the pass count (not found ≠ 0). Re-populate this list only with paths verified by ls in the same run that cites them.

Arithmetic reconciled 2026-08-11: the "8/8" above is now true of the list below — all 8 cited filenames are struck (re-verified in one run: 8 cited / 8 absent, with a known-positive and a known-negative control). Previously only 7 were struck while the 8th (feedback_autonomous_commit_proposal) was still cited live, and three more were cited as live authority up in §Constraints and §Step 3 — where an executor actually reads, since References is not on the execution path. Those live citations are removed; the rules they carried are stated directly. A struck entry in References is not a fix if the same name is still load-bearing above.

• Rule body: memory feedback_command_tower_gate.md + memory feedback_field_to_hub_sync_protocol.md — absent (verified 2026-08-11)
• Consistency rules: feedback_simplification_evidence · feedback_markdown_edit_discipline · feedback_skill_frontmatter_description_plain_text · feedback_bidirectional_self_validation · feedback_reference_own_hub_assets_first — absent (verified 2026-08-11)
• Sister skills: cross-ecosystem-synergy-detection (sister asset cluster baseline) · verify-bidirectional (bidirectional self-validation automation / self-catch auxiliary axis) · harvest-loop (weekly audit automation / operation proof accumulation cross-link)
• Autonomous commit proposal §2.19 baseline: memory feedback_autonomous_commit_proposal.md — absent (verified 2026-08-11). The rule it stood for is live and lives in CLAUDE.md §AI Contribution Model: development-source automation is allowed, PR submission requires explicit human approval. Cite that, not this filename.