9thlevelsoftware/legion/review-panel

Input: Phase CONTEXT.md content, SUMMARY.md content, files_modified list
Output: Ordered list of 2-4 reviewer agents with assigned rubrics
 
Step 1: Extract review signals from phase artifacts
  Read the phase CONTEXT.md and all SUMMARY.md files. Extract:
  - Primary domains touched (engineering, design, marketing, testing, product, infrastructure)
  - File types produced (.md, .ts, .js, .css, .py, config files, etc.)
  - Keywords from phase goal and task descriptions
  - Combine into a composite task description for the recommendation algorithm
 
Step 1.0: Preconditions Verification (MANDATORY before composition)
 
  Before running the scoring and filtering steps below, verify all preconditions. Missing
  preconditions trigger documented degraded behavior, not silent success.
 
  a) **agent-registry.md Section 3 structure** — read `skills/agent-registry/SKILL.md`.
     Verify Sections 1 through 6 exist as named headings. If Section 6 (Memory Boost) is
     absent: **skip the memory boost step below** and log WARN:
     `"agent-registry Section 6 not found. Skipping memory-based agent score boost."`
     Do not fail the review — proceed without the boost.
 
  b) **OUTCOMES.md schema** — if `.planning/memory/OUTCOMES.md` exists, verify it conforms
     to the memory-manager schema (required front-matter keys: `task_type`, `outcome`,
     `agent_id`). If malformed: skip memory boost with WARN; do not fail.
 
  c) **Agent frontmatter — `division` field** — every candidate agent .md file must have a
     `division` key in its frontmatter. If absent on any candidate: log WARN and exclude
     that candidate from the panel. Do not guess the division from the filename prefix.
 
  d) **Review-capability declaration** — an agent is eligible for the panel iff its
     frontmatter has a non-empty `review_strengths` array. This is the canonical marker;
     do NOT infer eligibility from specialty substring matches. See Step 3 below.
 
Step 2: Score agents using agent-registry Section 3
  Pass the composite task description to the agent-registry recommendation algorithm:
  - Step 1: Parse extracted keywords as task terms
  - Step 2: Match agents — exact (3 pts), partial (1 pt), division (2 pts)
  - Step 3: Rank by score, break ties by specificity
  - Step 6 (optional, per precondition 1.0.a): Apply memory boost if OUTCOMES.md exists
    AND agent-registry Section 6 is present AND OUTCOMES.md is schema-conformant.
 
Step 3: Filter to review-capable agents
  From the ranked list, keep only agents whose frontmatter declares a non-empty
  `review_strengths` array (the declarative eligibility marker established in Step 1.0.d).
 
  **Reference roster — current review-capable agents** (verify against agents/ directory
  at runtime; this list is informational, not authoritative):
 
  Testing division (6 agents — verified against CLAUDE.md):
  - testing-qa-verification-specialist, testing-api-tester
  - testing-workflow-optimizer, testing-performance-benchmarker
  - testing-test-results-analyzer, testing-tool-evaluator
 
  Design division (review-capable):
  - design-brand-guardian, design-ux-architect, design-ux-researcher
 
  Engineering division (review-capable):
  - engineering-senior-developer, engineering-backend-architect
  - engineering-frontend-developer, engineering-infrastructure-devops
 
  Product division (review-capable):
  - product-sprint-prioritizer, product-feedback-synthesizer
 
  Project Management (review-capable):
  - project-manager-senior, project-management-project-shepherd
 
  **Note:** `testing-evidence-collector`, `engineering-devops-automator`, and
  `marketing-content-creator` do NOT exist in the Legion 48-agent roster. They MUST NOT
  appear in this list. If found in a prior version of this skill file, treat as a doc bug
  and verify removal. Canonical agent IDs are in CLAUDE.md Division table.
 
  If an agent from a non-review-capable role (e.g., a marketing-platform specialist without
  review_strengths, or an XR-immersive role) scores highly, skip it and take the next
  eligible agent.
 
Step 4: Cap panel size and enforce diversity
  - 2 reviewers: single-domain phase (only one division touched)
  - 3 reviewers: standard phase (2 divisions involved)
  - 4 reviewers: cross-domain phase (3+ divisions involved)
 
  Diversity rule: no more than 2 reviewers from the same division.
  If 3+ agents from Testing score highest, keep the top 2 and pull the
  next-highest from a different division.
 
  Mandatory: at least one Testing division agent on every panel.
  This inherits from agent-registry Step 5 (Mandatory Roles) and from
  the existing review-loop principle that testing-qa-verification-specialist is
  always included.
 
Step 5: Assign rubrics
  For each selected reviewer, look up their domain rubric from Section 2.
  The rubric is keyed by agent ID. If no specific rubric exists for an agent,
  use the division default rubric.
 
Step 6: Present panel to user for confirmation
  Display the composed panel via AskUserQuestion:
 
  ## Phase {N}: {phase_name} — Review Panel
 
  **Panel Size**: {count} reviewers (based on {domain_count} domain(s) detected)
  **Domains Detected**: {comma-separated domains}
 
  | # | Agent | Division | Rubric Focus | Score |
  |---|-------|----------|--------------|-------|
  | 1 | {agent-id} | {division} | {rubric_name} | {score} pts |
  | 2 | {agent-id} | {division} | {rubric_name} | {score} pts |
 
  **Why this panel**: {1-sentence rationale linking detected domains to selected agents}
 
  Options:
  - "Use this panel" (Recommended)
  - "Add a reviewer" — add one more agent from the ranked list (up to max 4)
  - "Replace a reviewer" — swap one reviewer for an alternative
  - "Other" — enter custom agent IDs
 
  If user selects "Add a reviewer": show next-ranked eligible agent, confirm
  If user selects "Replace a reviewer": show which reviewer to replace and alternatives
  If user selects "Other": accept custom agent IDs, validate each exists, assign default rubric

panel:
  reviewers:          # ordered list, length 2-4 per Step 4
    - agent_id: {string — must match filename in agents/}
      division: {string}
      rubric_name: {string — matches Section 2 rubric key}
      score: {integer}
      rationale: {string}
  domains_detected: [{string}, ...]
  mode: "full" | "security-only" | "design-only" | "marketing-only"
  confirmed_by_user: true