R-fix

Fix

Success

I := ∀ f ∈ actionable → applied ∨ deferred (issue ∃) ∨ skipped (user) ∧ PR comment posted V := gh pr view {N} --comments | grep "## Review Fixes Applied"

Two-pass pipeline: auto-apply high-C findings (C≥T, 2+ agents), then 1b1 for rest.

⚠ Continuous pipeline. ¬stop between phases. Stop only on: unrecoverable failure or Phase 8 completion.

/R-fix        → findings from conversation context
/R-fix #42    → gather findings from PR #42 comments

Pipeline

Pre-flight

Success: ∀ actionable → applied ∨ deferred ∧ PR comment posted Evidence: gh pr view {N} --comments | grep "## Review Fixes Applied" Steps: gather → triage → auto-apply → walkthrough → apply-1b1 → final-push → post-comment ¬clear → STOP + ask: "Do you have review findings to fix?"

Let: F := all findings | f ∈ F | C(f) ∈ [0,100] ∩ ℤ — confidence A(f) := {agents that flagged f} | cat(f) ∈ {issue, suggestion, todo, nitpick, thought, question, praise} src(f) := originating agent | actionable := {issue, suggestion, todo, nitpick} T := 80 — auto-apply threshold Q_auto := {f | cat(f) ∈ actionable ∧ C(f) ≥ T ∧ |A(f)| ≥ 2} Q_1b1 := {f | cat(f) ∈ actionable ∧ f ∉ Q_auto} O_push(N, scope, msg) { lint+test gate (max 3 retries) → stage specific files (¬git add -A) → commit fix(<scope>): <msg> → git push } D_subsumption := {d ∈ D | d.tag = "subsumption-violation"}

Diagnostics Bus

D := []   — ordered list of records; insertion-ordered; duplicates permitted iff (tag, file, line) tuple differs
d ∈ D := {tag: str, file: str, line: int, description: str, phase: str}

• Initial value: [] (empty)

• Append-only invariant: entries are never removed or mutated after insertion

• Lifecycle: written in Phase 1 (enforcement checks); rendered in Phase 8 when |D| > 0. D is per-invocation and ephemeral.

• Future invariant slots: agent_src trust (append here when landed) (when promoting, set phase to non-empty ^[a-z0-9-]+$; do not leave angle-bracket placeholders in shipped records)

Phase 0 — Load Taxonomy

Read ${CLAUDE_PLUGIN_ROOT}/skills/dev-review/review-classes.yml → extract classes[].class slugs → canonical_slugs. (Canonical YAML ships with dev-review; fix reads it cross-skill — single source, ¬duplicate copy that could drift.) File absent, unreadable, or parse error → HALT: [taxonomy-error] review-classes.yml {reason} at ${CLAUDE_PLUGIN_ROOT}/skills/dev-review/review-classes.yml — reinstall dev-core plugin. Used in Phase 1 steps 4–5 to validate class[] values against the live YAML (¬LLM memory).

Phase 1 — Gather Findings

1. PR# → gh pr view <#> --json comments,closingIssuesReferences; parse Conventional Comments from .comments[].body; capture SOURCE_ISSUE = .closingIssuesReferences[0].number (∅ if none — used in Phase 5 Defer to wire blocked-by). When SOURCE_ISSUE ≠ ∅, also resolve SOURCE_PARENT = gh api graphql -f query='query{repository(owner:"<O>",name:"<R>"){issue(number:<SOURCE_ISSUE>){parent{number}}}}' --jq '.data.repository.issue.parent.number // empty' — used in Phase 5 Defer to wire deferred issue as sibling under shared parent (see issue-triage:issue-triage "Deferred Follow-Ups — Sibling Rule"). 1a. Strip the /R-dev-review filtered block first — ∀ comment body: remove everything from <summary>Filtered by finding-verifier through the next </details> before parsing. Those f were dropped by the Phase 4 keep/drop filter (dev-review); re-ingesting them defeats the filter. dev-review posts them table-shaped (¬CC grammar) — strip explicitly anyway, ¬rely on shape alone.
2. ¬PR# → scan conversation for latest /R-dev-review output
3. F = ∅ → halt
4. ∀ f: parse → label, file:line, agent, root cause, class[], raw_callsites[], solutions, C(f)
   • class[] — 0–N canonical slugs from review-classes.yml + 0–1 candidate/<slug>; absent field → class[] = []
   • raw_callsites[] — [{file, line}] list; required when class[] ≠ []; absent when class[] = []
5. Malformed (missing mandatory fields ∨ C ∉ ℤ ∩ [0,100] ∨ free-text class label not in canonical list and not candidate/* ∨ candidate/<slug> violates ^candidate/[a-z][a-z0-9-]{1,48}$ ∨ class[] ≠ [] ∧ raw_callsites[] = []) → C(f) := 0 Step 5 fires first; step 5b applies only to findings that passed step 5 (C(f) ≠ 0 after step 5). 5b. [only if step 5 did not fire for f] Subsumption strip: ∃ bare-except ∧ missing-error-handling in same finding's class[] → strip missing-error-handling; D.append({tag: "subsumption-violation", file: f.file, line: f.line, description: "bare-except subsumes missing-error-handling, duplicate tag stripped", phase: "1"}); ¬set C(f) := 0

Phase 2 — Triage + Verify

Split into Q_auto, Q_1b1, skipped (praise).

Single-agent high-C verification: ∀ f where cat(f) ∈ actionable ∧ C(f) ≥ T ∧ |A(f)| = 1:

• Spawn fresh verifier (different domain from src(f))
• C_v ≥ T → f → Q_auto, |A(f)| := 2
• C_v < T → f → Q_1b1
• Batch ∥ (group by domain, 1 verifier/domain)

∀ f ∈ Q_auto: solution(f) := Solution 1 (recommended).

Proxy-fix ban (cls(f) ∈ {test-tautology, vacuous-guard, parallel-path-drift}): Solution 1 that widens a denylist, adds a grep, or copies an inventory/validate:full list is invalid — demote to Q_1b1 with note "oracle/SSoT required". Required fix: change the oracle / single SSoT (matcher, parser, one package.json script).

Display:

── Fix Plan ──
Auto-apply: |Q_auto| finding(s) (C≥80, 2+ agents)
1b1 review: |Q_1b1| finding(s)
Skipped:    |skipped| (praise)

Q_auto = ∅ ∧ Q_1b1 = ∅ → "No actionable findings", halt.

Phase 3 — Auto-Apply (High Confidence)

Q_auto = ∅ → skip to Phase 4.

∀ f ∈ Q_auto (sequential, inline — already verified by 2+ agents):

• Apply recommended solution
• succeeds → [applied]
• fails → stash restore, [failed], demote to Q_1b1

── Auto-Apply Results ──
  1. [applied] issue(blocking): SQL injection in users.service.ts:42 (92%)
  2. [failed → 1b1] nitpick: Unused import in dashboard.tsx:3 (85%) -- test failure
Applied: N | Failed → 1b1: M

Phase 4 — Push Auto-Applied

∃ applied → O_push(N, scope, "auto-apply N review findings" + list in body). Fail after 3 → halt. ¬∃ applied → skip.

Phase 5 — 1b1 Walkthrough

Q_1b1 = ∅ → skip to Phase 7.

∀ f ∈ Q_1b1 sequentially (excluding praise):

── Finding {i}/{|Q_1b1|}: {cat(f)} ──
{cat} — C(f)% — {src(f)}
  {file}:{line}

Root cause: {root cause}

Recommended: Solution 1 — {rationale}
Alternative: Solution 2 — {rationale}

Demoted from auto-apply → prepend: Auto-apply failed: {reason}

→ present choice(single per finding): Solution 1 | Solution 2 | Defer (→ create issue) | Skip

Defer → create linked follow-up issue via Skill(skill: "issue-triage:issue-triage") (¬raw gh issue create — global rule: issue mutations go through the skill so blocked-by + parent are wired atomically). Sibling rule: the deferred issue is a sibling of SOURCE_ISSUE under their shared parent (¬child of SOURCE_ISSUE) — see issue-triage:issue-triage SKILL "Deferred Follow-Ups — Sibling Rule":

Invoke Skill(skill: "issue-triage:issue-triage") in create mode (requires the issue-triage plugin installed). Pass:

• --title "{cat}: {summary}"
• --body "{details}"
• --blocked-by "#${SOURCE_ISSUE}" — omit when SOURCE_ISSUE = ∅
• --parent "#${SOURCE_PARENT}" — omit when SOURCE_PARENT = ∅

SOURCE_ISSUE (captured Phase 1, step 1) → blocked-by ensures the deferred finding is traceable back to the issue whose review surfaced it. SOURCE_PARENT (also captured Phase 1, step 1) → parent makes the deferred issue a sibling of SOURCE_ISSUE under the same epic, so gh issue view <epic> shows the full fan-out flat. ∄ SOURCE_ISSUE → create without --blocked-by or --parent; {details} MUST include **Origin:** PR #<N> so traceability is preserved. ∄ SOURCE_PARENT (SOURCE_ISSUE is top-level) → create without --parent (the deferred issue is also top-level).

{details} template:

**Origin:** PR #<N> review <comment-id> (deferred per `/R-fix` walkthrough).

{root-cause + agent finding text}

**Action:** {chosen path or open question}

issue-triage --blocked-by accepts issues only (¬PRs) — when the source review is on a PR with no closing-issue reference, fall back to no --blocked-by and rely on the Origin: PR #N body line.

── Walkthrough Complete ──
Accepted: N | Deferred (issues created): M | Skipped: K

acc := {f ∈ Q_1b1 | decision ∈ {solution1, solution2}}, each with chosen solution.

Phase 6 — Apply 1b1 Decisions

acc = ∅ → skip to Phase 7.

Group acc by class, then dispatch per class:

classes = { c | ∃ f ∈ acc: cls(f) = c }

∀ class ∈ classes:
  files_in_class = unique({ file(f) | f ∈ acc, cls(f) = class })

  |files_in_class| ≤ 3  →  single R-fixer agent for the class
  |files_in_class| > 3  →  shard by file: ⌈|files_in_class| / 3⌉ R-fixer agents
                            each R-fixer owns ≤3 files of the same class

unclassified = { f ∈ acc | cls(f) = ∅ }
¬∅ → single R-fixer agent for unclassified findings (same path as |files_in_class| ≤ 3)

R-fixer payload per agent:

• findings (with class + raw_callsites) for owned files
• chosen solution per finding
• diff context for owned files
• instructions: "re-read targets before editing; lint + test after each fix; sweep file for same-class anti-pattern — justify or fix any uncited hit."

R-fixer constraints: re-read targets before editing (Phase 3 may have changed them). CI fail → retry max 3; [failed] if stuck.

Proxy-fix ban (class ∈ {test-tautology, vacuous-guard, parallel-path-drift}): Forbidden: widen a denylist, add another grep, copy another inventory/validate:full list. Required: change the oracle / single SSoT (matcher, parser, one package.json script). A proxy "fix" → [failed], do not apply.

pattern-class findings (Lane B tag) → same class-shard dispatch as Lane A findings. (TODO: pattern-class tag and Lane B defined in Slice 3 — targeted recall. Until Slice 3 lands, this clause is a forward-reference only; pattern-class is not yet in review-classes.yml.)

Phase 6.5 — Falsification Gate

∀ class ∈ classes (from Phase 6):

Run falsification gate per ${CLAUDE_SKILL_DIR}/falsification.md. Gate emits boolean per class:

pass  →  fix accepted; continue
fail  →  fix tautological (RC-1); re-open each failed finding for that class
          (max 1 falsification-retry per finding — independent of CI retry budget in Phase 6)

New findings surfaced during falsification → parking lot: file as candidate finding for next PR cycle. ¬reopen current /R-fix loop. ¬increment 2-iter cap. Applies to same-class and cross-class anti-patterns alike.

Phase 7 — Final Push + Approve

1. ∃ Phase 6 changes → O_push(N, scope, "apply N review findings from 1b1" + list in body). Fail after 3 → halt.
2. ∃ PR → gh api repos/:owner/:repo/issues/<#>/labels -f "labels[]=reviewed"

Phase 8 — Post Follow-Up Comment

∄ PR → skip.

Tempfile per ${CLAUDE_PLUGIN_ROOT}/../shared/references/tempfile-convention.md:

[[ "$PR" =~ ^[0-9]+$ ]] || { echo "Invalid PR number: $PR" >&2; exit 1; }
TMPDIR=$(mktemp -d -t "dev-core-review-fixes-PR${PR}-XXXXXX")
trap 'rm -rf "$TMPDIR"' EXIT
BODY="$TMPDIR/body.md"

Write summary (below) to "$BODY" → gh pr comment "$PR" --body-file "$BODY"

## Review Fixes Applied

**Auto-applied (C≥80, 2+ agents):** N finding(s)
**Applied via 1b1:** M finding(s)
**Deferred (issues created):** J finding(s)
**Skipped:** K finding(s)
**Failed:** L finding(s)
**Enforcement diagnostics:** |D_subsumption| subsumption violation(s) (0 if none)

### Auto-Applied
- [applied] issue(blocking): SQL injection in users.service.ts:42 (92%)

### Applied (1b1)
- [applied] suggestion: Missing error boundary in dashboard.tsx:15

### Deferred
- nitpick: Variable naming in auth.service.ts:88 → #123

### Failed
- [failed] nitpick: Unused import in dashboard.tsx:3 -- test failure

### Parking Lot
_(omit section when parking_lot = ∅)_
- {class}: {file}:{line} — {description} (falsification-gate)

### Enforcement diagnostics
_(omit section when |D| = 0; group by tag when |distinct tags| > 1 using **[tag]** (N) sub-groupings)_
- `[subsumption-violation]` `auth.service.ts`:`42` — bare-except subsumes missing-error-handling, duplicate tag stripped

Edge Cases

Safety Rules

1. Human can git diff anytime — applied changes visible in working tree
2. ∃ PR → must post follow-up comment (Phase 8)
3. Fixer agents ¬have implementation context → spawn fresh
4. Stage specific files only — ¬git add -A (risk of .env, secrets)
5. Merge via the gate: label reviewed → auto-merge merges (merge commit) on green. ¬manual gh pr merge while any check is IN_PROGRESS/QUEUED

Chain Position

• Phase: Verify
• Predecessor: /R-dev-review (findings)
• Successor: /R-dev-review (re-review after fix) — LOOP
• Class: loop (bounded, max 2 iterations)

Task Integration

• /R-dev owns the dev-pipeline task lifecycle externally
• Sub-tasks created: none directly (findings are ephemeral — tracked in-skill via F, Q_auto, Q_1b1)
• Follow-up tasks: on success → TaskCreate new review task with metadata: { kind: "dev-pipeline", step: "review", follow_up: true, iteration: N+1, blockedBy: [this.id] }

Exit

• Success via /R-dev: fixes applied + committed + pushed + PR comment posted → TaskCreate follow-up review task → return silently. /R-dev picks up the new review task.
• Success standalone: print summary (Applied/Skipped/Deferred/Failed) + Next: /R-dev-review (re-verify). Stop.
• Failure (quality gate, ¬findings, unrecoverable): return error. /R-dev presents Retry | Skip | Abort.
• Loop cap: metadata.iteration ≥ 2 on entry → refuse another iteration; return with message "Max fix iterations reached — resolve remaining manually". /R-dev presents Abort.

$ARGUMENTS