Review Work Item

!${CLAUDE_PLUGIN_ROOT}/bin/accelerator config context --skill review-work-item --fail-safe !${CLAUDE_PLUGIN_ROOT}/bin/accelerator config agents --fail-safe

If no "Agent Names" section appears above, use these defaults: accelerator:reviewer, accelerator:codebase-locator, accelerator:codebase-analyser, accelerator:codebase-pattern-finder, accelerator:documents-locator, accelerator:documents-analyser, accelerator:web-search-researcher.

!${CLAUDE_PLUGIN_ROOT}/bin/accelerator config review work-item --fail-safe

Work items directory: !${CLAUDE_PLUGIN_ROOT}/bin/accelerator config path work --fail-safe Work item reviews directory: !${CLAUDE_PLUGIN_ROOT}/bin/accelerator config path review_work --fail-safe

Work Item Review Template

The template below defines the frontmatter and body structure that every work item review must carry. Read it now — use it to guide what information you record in Steps 3-4 and what shape you persist in Step 4.8.

!${CLAUDE_PLUGIN_ROOT}/bin/accelerator config template work-item-review --fail-safe

You are tasked with reviewing a work item through quality lenses and then collaboratively iterating the work item based on findings.

Initial Response

When this command is invoked:

1. Check if a work item path or ID was provided: invoke the resolver:

   ${CLAUDE_PLUGIN_ROOT}/bin/accelerator work resolve <argument>
   

   The resolver respects work.id_pattern and accepts paths, full IDs (PROJ-0042), and bare numbers.

   • Exit 0: stdout is the absolute path. Read the work item immediately and FULLY, then begin the review process.
   • Exit 1: unrecognised input. Print the resolver's error and offer to run /list-work-items.
   • Exit 2: ambiguous match. The resolver lists candidates with source-category tags. Ask the user to disambiguate by re-running with a full ID or path.
   • Exit 3: no match. Print the resolver's error and offer to run /list-work-items.
   • If optional focus arguments were provided (e.g., "focus on testability"), note them for lens selection.

2. If no work item path or number provided, respond with:

I'll help you review a work item. Please provide:
1. The path to the work item file (e.g., `{work_dir}/0042-my-work-item.md`)
2. (Optional) A work item number shorthand (e.g., `/review-work-item 42`)
3. (Optional) Focus areas to emphasise (e.g., "focus on testability")

Tip: Use `/list-work-items` to find the work item you want to review.

Then wait for the user's input.

Available Review Lenses

Lens	Lens Skill	Focus
Clarity	clarity-lens	Unambiguous referents, internal consistency, jargon handling
Completeness	completeness-lens	Section presence, content density, kind-appropriate content
Dependency	dependency-lens	Implied couplings not captured — blockers, consumers, external systems
Scope	scope-lens	Right-sized, single coherent unit of work; decomposition; orthogonality
Testability	testability-lens	Measurable criteria, verifiable outcomes, verification framing

Note: completeness flags an absent Dependencies section; dependency flags an empty or underspecified section whose contents fail to name every coupling the work item implies.

Process Steps

Step 1: Read and Understand the Work Item

1. Read the work item file FULLY — never use limit/offset

2. Parse the frontmatter to note kind (bug, story, spike, epic, etc.) and status

3. Read any documents referenced in the References section — these provide context the lenses may need; do not read source code

4. Check for existing reviews: Glob for review documents matching {work_reviews_dir}/{work-item-stem}-review-*.md. If any are found:

   • Read the most recent review document (highest review number)
   • Note the previous verdict, review pass count, and key findings
   • Inform the user: "I found {N} previous review(s) of this work item. The most recent (review {N}, verdict: {verdict}) will be used as context."
   • The agents do NOT receive the previous review — they review the work item fresh. But the aggregation step (Step 4) should reference the previous review when composing cross-cutting themes and the assessment.
   • If the prior review file exists but cannot be parsed (malformed frontmatter), warn the user and proceed as if no prior review exists.

   The new review creates a new file with the next review number (e.g., -review-2.md). Previous review files are never modified.

Step 2: Select Review Lenses

By default, run every lens registered in BUILTIN_WORK_ITEM_LENSES unless the user has provided focus arguments or config restricts the selection. The five work item lenses cover orthogonal concerns, so there is no relevance-based auto-selection.

If the user provided focus arguments:

• Map the focus areas to the corresponding lenses
• Include any additional lenses that are clearly relevant
• Briefly explain which lenses you're running

If no focus arguments were provided:

Run all built-in work item lenses unless:

• A lens is listed in disabled_lenses — remove it from the active set
• The user's configured core_lenses has filtered this to a subset (see below)

When core_lenses is set in config, apply it as the minimum required set; add any remaining non-disabled lenses up to max_lenses. This means users who previously pinned core_lenses to the Phase 4 work item lenses (completeness, testability, clarity) will also receive scope and dependency on upgrade, unless they add those names to disabled_lenses or set max_lenses to their subset size.

Present the selection briefly — enumerate the chosen lenses with a one-line focus each — then use the AskUserQuestion tool with two options:

1. Yes, use the proposed lenses — run the review with the selected lenses
2. No, specify which lenses to use — adjust the selection before running

The confirmation gate is preserved even though the default always selects every lens; the gate is useful when focus args or config have narrowed the set.

Example (default path, no focus args, no core_lenses restriction):

I'll review this work item through all work item lenses (clarity, completeness,
dependency, scope, testability).

Wait for the user's answer before spawning reviewers. If they choose option 2, ask which lenses they want using a plain-text question only — do NOT use AskUserQuestion for this follow-up (the lens list is too large for the 4-option limit). If any lens name is unrecognised, seek clarification. Once confirmed, update the selection and re-present it using the same AskUserQuestion proceed/adjust pattern. This loop is user-controlled with no hard termination limit.

Step 3: Spawn Review Agents

For each selected lens, spawn the {reviewer agent} agent with a prompt that includes the paths to the lens skill and output format files. Do NOT read these files yourself — the agent reads them in its own context.

Compose each agent's prompt following this template:

You are reviewing a work item through the [lens name] lens.

## Context

The work item is at [path]. Read it fully.
Also read any source documents listed in the work item's References section.

## Analysis Strategy

1. Read your lens skill and output format files (see paths below)
2. Read the work item file fully
3. Read referenced documents from the work item's References section if present
4. Evaluate the work item through your lens, applying each key question
5. Reference specific work item sections in your findings using the `location`
   field (e.g., "Acceptance Criteria", "Requirements", "Frontmatter: kind")

IMPORTANT: Do not evaluate the codebase — work item content (and any documents
it explicitly references) is the sole artefact under review. Do not run
codebase exploration agents or read source files unless the work item's
References section explicitly links to them.

## Lens

Read the lens skill at the path listed in the Lens Catalogue table in the
review configuration above. If no review configuration is present, use:
${CLAUDE_PLUGIN_ROOT}/skills/review/lenses/[lens]-lens/SKILL.md

## Output Format

Read the output format at:
${CLAUDE_PLUGIN_ROOT}/skills/review/output-formats/work-item-review-output-format/SKILL.md

IMPORTANT: Return your analysis as a single JSON code block. Do not include
prose outside the JSON block.

Spawn all selected agents in parallel using the Task tool with subagent_type: "!${CLAUDE_PLUGIN_ROOT}/bin/accelerator config agent reviewer --fail-safe".

IMPORTANT: Wait for ALL review agents to complete before proceeding.

Handling malformed agent output:

If an agent's response is not a clean JSON block, apply this extraction strategy:

1. Look for a JSON code block fenced with triple backticks (optionally with a json language tag)

2. If found, extract and parse the content within the fences

3. If the extracted JSON is valid, use it normally

4. If no JSON code block is found, or the JSON within it is invalid, apply the fallback: treat the agent's entire output as a single finding with "suggestion" severity (marked synthetic: true), attributed to that agent's lens

   Note: "suggestion" severity is used here (not "major" as in review-plan) so a single flaky agent cannot deterministically force a REVISE verdict when work_item_revise_severity is major or higher.

When falling back, warn the user that the agent's output could not be parsed and present the raw agent output so the user can see what the agent found. Include remediation guidance: "Try re-running with a narrower lens selection, or file a bug with the raw output above."

Step 4: Aggregate and Curate Findings

Once all reviews are complete:

1. Parse agent outputs: Extract the JSON block from each agent's response (see the extraction strategy in Step 3). Collect the summary, strengths, and findings arrays from each.

2. Aggregate across agents:

   • Combine all findings arrays into a single list
   • Combine all strengths arrays into a single list
   • Collect all summary strings

3. Deduplicate findings: Where multiple agents flag the same section with similar concerns, consider merging — but only when the findings address the same underlying concern from different lens perspectives. When in doubt, keep findings separate.

   When merging:

   • Combine the bodies, attributing each part to its lens
   • Use the highest severity among the merged findings
   • Use the highest confidence among the merged findings

4. Prioritise findings:

   • Sort by severity: critical > major > minor > suggestion
   • Within the same severity, sort by confidence: high > medium > low

5. Determine suggested verdict:

   If review configuration provides verdict overrides above, apply those thresholds instead of the defaults below:

   • If work_item_revise_severity is none, skip the severity-based REVISE rule (major count rule still applies independently)
   • If any findings at or above the work item revise severity ({work item revise severity}) exist → suggest REVISE
   • If {work item revise major count} or more "major" findings exist → suggest REVISE
   • If fewer major findings than the threshold, or only minor/suggestion → suggest COMMENT
   • If no findings at all (only strengths) → suggest APPROVE

   Verdict meanings:

   • APPROVE — work item is ready for implementation
   • REVISE — work item needs changes before implementation
   • COMMENT — observations only, work item is acceptable as-is

   When presenting a COMMENT verdict with major findings, note: "Work item is acceptable but could be improved — see major findings below."

6. Identify cross-cutting themes: Look for findings that appear across multiple lenses — issues flagged by 2+ agents reinforce each other and should be highlighted in the summary.

7. Compose the review summary:

   ## Work Item Review: [Work item Title]
   
   **Verdict:** [APPROVE | REVISE | COMMENT]
   
   [Combined assessment: synthesise each agent's summary into 2-3 sentences
   covering the overall quality of the work item across all lenses]
   
   ### Cross-Cutting Themes
   [Issues that multiple lenses identified — these deserve the most attention]
   - **[Theme]** (flagged by: [lenses]) — [description]
   
   ### Findings
   
   #### Critical
   - 🔴 **[Lens]**: [title]
     **Location**: [work item section]
     [First 1-2 sentences of body as summary]
   
   #### Major
   - 🟡 **[Lens]**: [title]
     **Location**: [work item section]
     [First 1-2 sentences of body as summary]
   
   #### Minor
   - 🔵 **[Lens]**: [title]
     **Location**: [work item section]
     [First 1-2 sentences of body as summary]
   
   #### Suggestions
   - 🔵 **[Lens]**: [title]
     **Location**: [work item section]
     [First 1-2 sentences of body as summary]
   
   ### Strengths
   - ✅ [Aggregated and deduplicated strengths from all agents]
   
   ### Recommended Changes
   [Ordered list of specific, actionable changes to the work item, prioritised by
   impact. Each should reference the finding(s) it addresses.]
   
   1. **[Change description]** (addresses: [finding titles])
      [Specific guidance on what to modify in the work item]
   
   ---
   *Review generated by /review-work-item*
   

8. Write the review artifact to {work_reviews_dir}/:

   Derive the review filename using the work item stem and the next available review number. The work item stem is the basename of the work item path without the .md extension. For example, if the work item is {work_dir}/0042-improve-search.md and no prior reviews exist, the review filename is {work_reviews_dir}/0042-improve-search-review-1.md.

   To determine the next review number:

   mkdir -p {work_reviews_dir}
   # Glob for existing reviews of this work item
   ls {work_reviews_dir}/{work-item-stem}-review-*.md 2>/dev/null
   # Extract the highest number, increment by 1. If none exist, use 1.
   

   Extract the work item's stable 4-digit identifier from its filename using ${CLAUDE_PLUGIN_ROOT}/bin/accelerator work show {path} --field id (or parse the 4-digit prefix from the filename directly).

Populate frontmatter

Before writing the work item review file, capture metadata and substitute the unified base fields and per-type extras into the template's frontmatter block:

1. Invoke ${CLAUDE_PLUGIN_ROOT}/bin/accelerator corpus metadata derive to obtain Current Date/Time (UTC):.
2. Substitute every field below with the indicated value:
   • type: ← work-item-review
   • id: ← the review filename stem (without .md), always quoted as a YAML string
   • title: ← Work Item Review: {work item title}
   • date: ← the Current Date/Time (UTC): value
   • author: ← the author value resolved per create-work-item/SKILL.md:578-580
   • producer: ← review-work-item
   • status: ← complete
   • last_updated: ← the same Current Date/Time (UTC): value
   • last_updated_by: ← the same value resolved for author
   • schema_version: ← 1 (bare integer, not quoted)
   • parent: ← typed-linkage ref to the parent work item ("work-item:NNNN"). Fill when the review names a parent; otherwise omit the key.
   • target: ← "work-item:<4-digit-id>" (e.g. "work-item:0042"), the typed-linkage ref to the work item under review, per ADR-0034. Always fill — every review has a target.
   • relates_to: ← list of typed-linkage refs to related reviews or artifacts (["work-item-review:NNNN", ...]). Fill when prior reviews are explicit; otherwise omit the key.
   • work_item_id: ← the same 4-digit identifier as the target payload's id portion (transitional alias — see Migration Notes; the visualiser's read_ref_keys consumes this scalar today)
   • reviewer: ← the reviewer value resolved per create-work-item/SKILL.md:578-580
   • verdict: ← the verdict from Step 4.5 (APPROVE | REVISE | COMMENT)
   • lenses: ← the list of work-item lens names used
   • review_number: ← N (the next available review number from the glob above)
   • review_pass: ← 1 (initial-write pass count; re-reviews bump per the Step 7 flow)
3. Write the file with the substituted frontmatter block, followed by the review summary composed in Step 4.7 and the per-lens results as a final section:

{The full review summary from Step 4.7}

## Per-Lens Results

### {Lens 1 Name}

**Summary**: {agent summary}

**Strengths**:
{agent strengths}

**Findings**:
{agent findings — each with severity, confidence, location, and body}

### {Lens 2 Name}

...

The target: field stores the work item's stable 4-digit identifier as a typed-linkage key (e.g. "work-item:0042") per ADR-0034, providing resilience against work item renames. A work_item_id: field is also emitted as a transitional alias carrying the same 4-digit identifier — the visualiser's read_ref_keys consumes it as the primary work-item cross-reference key today. Both fields encode the same edge; the duplication is bounded by the visualiser consumer update.

Validate the frontmatter: after writing the review file, run

${CLAUDE_PLUGIN_ROOT}/bin/accelerator corpus frontmatter validate --file <the review file path>

If it exits non-zero, the document violates the canonical frontmatter standard; report the emitted violation and fix the frontmatter before completing.

Step 5: Present the Review

Present the composed review summary from Step 4.7 to the user.

After presenting, offer the user control before proceeding to iteration:

The review is complete. Verdict: [verdict]

Use the AskUserQuestion tool with four options:

1. Address findings — edit the work item to resolve issues
2. Change the verdict — currently: [verdict]
3. Discuss specific findings — explore any finding in more detail
4. Re-run specific lenses — adjust focus and re-review

Step 6: Collaborative Work Item Iteration

After presenting the review:

1. Discuss findings with the user:

   • Ask which recommendations they want to address
   • Clarify any findings that need more context

2. Edit the work item based on agreed changes:

   • Use the Edit tool to modify the work item file directly
   • Make targeted edits to the relevant work item sections (Summary, Context, Requirements, Acceptance Criteria, etc.)
   • Do NOT modify the status field — that is a separate workflow decision
   • Preserve the work item's existing frontmatter and section structure

3. Summarise changes made:

   I've made the following changes to the work item:
   - [Change 1] — addressing [finding]
   - [Change 2] — addressing [finding]
   - [Skipped] — [finding discussed and decided not to address, with reason]
   

Step 7: Offer Re-Review

After edits are complete, use the AskUserQuestion tool with two options:

1. Yes, run another review pass — re-run the relevant lenses to verify the changes address the findings
2. No, done — exit without re-reviewing

If the user chooses option 1:

• Re-run only the lenses that had findings in the previous pass
• Use the same spawn pattern and JSON extraction strategy from Steps 3-4
• Compare previous findings against new findings to determine resolution status
• Present a shorter, delta-focused review:

  ## Re-Review: [Work item Title]
  
  **Verdict:** [APPROVE | REVISE | COMMENT]
  
  ### Previously Identified Issues
  - [emoji] **[Lens]**: [title] — Resolved / Partially resolved / Still present
  
  ### New Issues Introduced
  - [emoji] **[Lens]**: [title] — [brief description]
  
  ### Assessment
  [Whether the work item is now ready for implementation or needs further iteration]
  

After composing the re-review summary, update the review artifact as a single write operation:

1. Read the full content of the existing review document at {work_reviews_dir}/{work-item-stem}-review-{N}.md

2. If the existing review file's frontmatter cannot be parsed (malformed YAML or missing --- delimiters), warn the user and write a fresh -review-{N+1}.md file instead of appending in place

3. In memory, update exactly four frontmatter fields — verdict, review_pass, last_updated, and last_updated_by — preserving all other fields and body content verbatim. The date field retains the original-review timestamp; only last_updated advances on re-review. (last_updated_by may match reviewer if the re-reviewer is the same person, but is computed independently.)

   Pre-0066-artifact handling: when the re-reviewed artifact lacks last_updated: and/or last_updated_by: (it was written pre-0066), insert those fields rather than treating their absence as malformed-frontmatter. Only an unparseable YAML block or missing --- delimiters triggers the fresh--review-{N+1}.md fallback.

4. Append the re-review section at the end of the content

5. Write the complete modified content back to the same file in one operation

The document reads chronologically: initial review, per-lens results, then re-review sections in order. The frontmatter always reflects the latest verdict and pass count:

## Re-Review (Pass {N}) — {date}

**Verdict:** {verdict}

### Previously Identified Issues
- {emoji} **{Lens}**: {title} — {Resolved | Partially resolved | Still present}

### New Issues Introduced
- {emoji} **{Lens}**: {title} — {brief description}

### Assessment
{Whether the work item is now ready for implementation or needs further iteration}

If the user declines or the re-review shows all clear, the review is complete.

Important Guidelines

1. Read the work item fully before doing anything else

2. Spawn agents in parallel — the work item lenses are independent and should run concurrently for efficiency

3. Synthesise, don't concatenate — your value is in compiling a balanced view across lenses, identifying themes, and prioritising actionable recommendations

4. Do not modify the work item's status field — a REVISE verdict does not automatically change the work item's status; that transition belongs to a separate workflow decision by the team

5. Do not run codebase exploration agents — the reviewer agents stay inside the work item and any documents it explicitly references; source code is out of scope for work item review

6. Be balanced — highlight strengths alongside concerns

7. Prioritise by impact — structural issues that would block implementation matter more than surface-level polish

8. Handle malformed agent output gracefully — use the suggestion severity fallback (not major) so a single flaky agent does not force a REVISE verdict

9. Use emoji severity prefixes consistently — 🔴 critical, 🟡 major, 🔵 minor/suggestion, ✅ strengths. IMPORTANT: Use the actual Unicode emoji characters (🔴 🟡 🔵 ✅), NOT text shortcodes.

What NOT to Do