Paths: File paths (references/, ../ln-*) are relative to this skill directory.

Type: L3 Worker Category: 4XX Execution

Task Reviewer

MANDATORY after every task execution. Reviews a single task in To Review and decides Done vs To Rework with immediate fixes or clear rework notes.

This skill is NOT optional. Every executed task MUST be reviewed immediately. No exceptions, no batching, no skipping.

Purpose & Scope

• Resolve task ID (per Input Resolution Chain); load full task and parent Story independently via the configured tracker provider (getTask, getStory).
• Check architecture, correctness, configuration hygiene, docs, and tests.
• For test tasks, verify risk-based limits and priority (≤15) per planner template.
• Update only this task: accept (Done) or send back (To Rework) with explicit reasons and fix suggestions tied to best practices.

Inputs

Resolution: Task Resolution Chain. Status filter: To Review

Phase 0: Tools Config

MANDATORY READ: Load references/environment_state_contract.md, references/storage_mode_detection.md, and references/input_resolution_pattern.md

Extract: task_provider = Task Management -> Provider (linear | github | file). Operations stay provider-agnostic; load the selected provider transport reference only when performing tracker I/O.

Task Storage Mode

Tracker operations used by this skill: getTask, getStory, updateStatus (Done | To Rework), addComment, createTask (for [BUG] side-effect tasks).

Mode Detection

Detect operating mode at startup:

Plan Mode Active:

• Steps 1-3: Resolve task and load context (read-only, OK in plan mode)
• Generate REVIEW PLAN (files, checks) → write to plan file
• Call ExitPlanMode → STOP. Do NOT execute review.
• Steps 4-9: After approval → execute full review

Normal Mode:

• Steps 1-9: Standard workflow without stopping

Plan Mode Support

Conditional read: load references/plan_mode_pattern.md Workflow A only when Plan Mode is active. Tool policy: follow host AGENTS.md MCP preferences; load references/mcp_tool_preferences.md and references/mcp_integration_patterns.md only when host policy is absent or MCP behavior is unclear.

CRITICAL: In Plan Mode, plan file = REVIEW PLAN (what will be checked). NEVER write review findings or verdicts to plan file.

Review Plan format:

REVIEW PLAN for Task {ID}: {Title}

| Field | Value |
|-------|-------|
| Task | {ID}: {Title} |
| Status | {To Review} |
| Type | {impl/test/refactor} |
| Story | {Parent ID}: {Parent Title} |

Files to review:
- {file1} (deliverable)
- {file2} (affected component)

| # | Check | Will Verify |
|---|-------|-------------|
| 1 | Approach | Technical Approach alignment |
| 2 | Clean Code | No dead code, no backward compat shims |
| 3 | Config | No hardcoded creds/URLs |
| 4 | Errors | try/catch on external calls |
| 5 | Logging | ERROR/INFO/DEBUG levels |
| 6 | Comments | WHY not WHAT, docstrings |
| 7 | Naming | Project conventions |
| 8 | Docs | API/env/README updates |
| 9 | Tests | Updated/risk-based limits |
| 10 | AC | 4 criteria validation |
| 11 | Side-effects | Pre-existing bugs in touched files |
| 12 | Destructive ops | Safety guards from destructive_operation_safety.md (loaded in step 4) |
| 13 | Algorithm correctness | Loop invariants, collection keys, unbounded ops, shared state leaks |
| 14 | Event channels | Channel name consistency in diff |
| 15 | CI Checks | lint/typecheck pass per ci_tool_detection.md |

Expected output: Verdict (`Done | To Rework`) + Issues + Fix actions

Progress Tracking with TodoWrite

When operating in any mode, skill MUST create detailed todo checklist tracking ALL steps.

Rules:

1. Create todos IMMEDIATELY before Step 1
2. Each workflow step = separate todo item; multi-check steps get sub-items
3. Mark in_progress before starting step, completed after finishing

Todo Template (~11 items):

Step 1: Resolve taskId
  - Resolve via args / Story context / kanban / AskUserQuestion (To Review filter)

Step 2: Load Task
  - Load task by ID, detect type

Step 3: Read Context
  - Load full task + parent Story + affected components

Step 3b: Goal Articulation Gate
  - State what specific quality question this review must answer (<=25 tokens each)

Step 4: Review Checks
  - Verify approach alignment with Story Technical Approach
  - Check clean code: no dead code, no backward compat shims
  - Cross-file DRY: Grep src/ for new function/class names (count mode). 3+ similar → CONCERN
  - Check config hygiene, error handling, logging
  - Check comments, naming, docs updates
  - Verify tests updated/run (risk-based limits for test tasks)

Step 5: AC Validation
  - Validate implementation against 4 AC criteria

Step 6: Side-Effect Bug Detection
  - Scan for bugs outside task scope, create [BUG] tasks

Step 7: Decision
  - Apply minor fixes or set To Rework with guidance

Step 8: Mechanical Verification
  - Run lint/typecheck per ci_tool_detection.md (only if verdict=Done)

Step 9: Update & Commit
  - Set task status, update kanban, post review comment
  - If Done: leave branch changes uncommitted for downstream branch ownership rules

Workflow (concise)

Use hex-graph first when semantic diff, clone groups, references, or review blast radius matter. Use hex-line first for local code/config/script/test reads when available. If MCP is unavailable, unsupported, or not indexed, continue with built-in Read/Grep/Glob/Bash and record the fallback in the review instead of blocking.

1. Resolve taskId: Run Task Resolution Chain per guide (status filter: [To Review]).
2. Load task: Load full task and parent Story independently. Detect type (label "tests" -> test task, else implementation/refactor).
3. Read context: Full task + parent Story; load affected components/docs; review diffs if available. Hex MCP acceleration: Prefer analyze_changes(path=project_root, base_ref="HEAD~1") for semantic risk snapshot when graph is indexed; use changes(path="src/", compare_against="HEAD~1") for AST-level diff review of structural changes. 3b) Goal gate: Before reviewing, state: (1) REAL GOAL: what quality question must this review answer? (2) DONE: what evidence proves quality is sufficient? (3) NOT THE GOAL: what would a surface-level rubber-stamp look like? (4) INVARIANTS: what non-obvious constraint exists? Load references/goal_articulation_gate.md only when this gate is ambiguous or disputed.
4. Review checks:

   Spec-first gate: Quick AC pre-check: scan task AC against implementation. If any AC is clearly unmet (BLOCKER-level) → immediate To Rework, skip remaining quality checks. Full AC validation still runs in Step 5. MANDATORY READ: Load references/clean_code_checklist.md, references/destructive_operation_safety.md

   • Goal validation (Recovery Paradox): If executor articulated a REAL GOAL (visible in task comments or implementation), validate it matches the Story's target deliverable. If executor framed the goal around a secondary subject (e.g., "implement the endpoint" instead of "enable user data export") → CONCERN: GOAL-MISFRAME: executor goal targets secondary subject, may miss hidden constraints.
   • Blueprint completion (advisory): If executor runtime data available (.hex-skills/runtime-artifacts/runs/ for this task), load PHASE_3 blueprint and PHASE_6 blueprint_status from executor checkpoints. Flag as CONCERN if: completion_pct < 100 without justifications for skipped items, or added files exceed 50% of planned without justification. If runtime data unavailable, check metadata.blueprint_status from executor summary. Not a BLOCKER.
   • Approach: diff aligned with Technical Approach in Story. If different → rationale doc