ln-005-multi-agent-context-review（ln 005 多主体上下文审查）

Paths: File paths (shared/, references/, ../ln-*) are relative to skills repo root. If not found at CWD, locate this SKILL.md directory and go up one level for repo root.

Multi-Agent Context Review (Universal)

Runs parallel external agent reviews on arbitrary context, critically verifies suggestions, returns filtered improvements.

Purpose & Scope

• Standalone utility in 0XX category (like ln-003, ln-004)
• Delegate any context to codex-review + gemini-review as background tasks in parallel
• Context always passed via file references (never inline in prompt)
• Process results as they arrive (first-finished agent processed immediately)
• Critically verify each suggestion; debate with agent if Claude disagrees
• Return filtered, deduplicated, verified suggestions

When to Use

• Manual invocation by user for independent review of any artifact
• Called by any skill needing external second opinion on plans, decisions, documents
• NOT tied to Linear, NOT tied to any pipeline
• Works with any context that can be saved to a file

Plan Mode Support

Follows shared/references/plan_mode_pattern.md (Workflow B) and shared/references/agent_review_workflow.md Plan Mode Behavior. Step 7e (Compare & Correct): output findings to chat, apply edits only after user approval.

Parameters

Inputs

Context delivery rule: Context is ALWAYS passed via files.

• If context already exists as files (plans, docs, code) -> pass file paths directly
• If context is a statement/decision from chat -> caller creates a temporary file in .agent-review/context/ with the content, then passes the file path

Workflow

MANDATORY READ: Load shared/references/agent_review_workflow.md for Health Check, Ensure.agent-review/, Load Review Memory, Run Agents, Critical Verification + Debate, Aggregate + Return, Save Review Summary, Fallback Rules, Critical Rules, and Definition of Done. Load shared/references/agent_delegation_pattern.md for Reference Passing Pattern, Review Persistence Pattern, Agent Timeout Policy, and Debate Protocol.

Unique Steps (before shared workflow)

1. Health check: per shared workflow, filter by skill_group = 005.
2. Resolve identifier: If identifier not provided, generate review_YYYYMMDD_HHMMSS. Sanitize: lowercase, replace spaces with hyphens, ASCII only.
3. Ensure.agent-review/: per shared workflow. Additionally create .agent-review/context/ subdir if it doesn't exist (for materialized context files).
4. Materialize context (if needed): If context is from chat/conversation (not an existing file):

- Write content to .agent-review/context/{identifier}_context.md - Add this path to context_files list

1. Build prompt: Read template shared/agents/prompt_templates/context_review.md.

- Replace {review_title} with title or "Context Review" - Replace {context_refs} with bullet list: - {path} per context file - Replace {focus_areas} with filtered subset or "All default areas" if no focus specified - Save to .agent-review/{identifier}_contextreview_prompt.md (single shared file -- both agents read the same prompt)

Shared Workflow Steps

1. Launch Agents (background) — MANDATORY: before any foreground research:

Per shared workflow "Step: Run Agents". Prompt file is ready (step 5). Launch BOTH agents as background tasks NOW.

• {review_type} in challenge template = review_title or "Context Review"
• {story_ref} in challenge template = identifier

1. Foreground Research (while agents are running in background):

Agents are already thinking. Use this time for Review Memory + MCP Ref research.

Agents (background)                    Claude (foreground)
  codex-review ──┐                       7a) Load Review Memory
  gemini-review ─┤                       7b) Applicability Check
                 │                       7c) Stack Detection
                 │                       7d) Extract Topics
                 │                       7e) MCP Ref Research
                 │                       7f) Compare & Correct
                 ├── first completes ──→ 7g) Save Findings
                 └── second completes    8) Critical Verification (informed by memory + findings)

MANDATORY READ: Load shared/references/research_tool_fallback.md

7a) Load Review Memory

Per shared workflow "Step: Load Review Memory". Not passed to agents — used only in step 8 (Critical Verification).

7b) Applicability Check

Scan context_files for technology decision signals (skip 7c-7e if no signals found):

• No signals found → skip MCP Ref research, log "MCP Ref skipped: no technology decisions detected"
• Fewer than 3 topics with weight >= 3 → skip

7c) Stack Detection

Priority order:

1. tech_stack input parameter → use directly as query_prefix
2. docs/tools_config.md Research section → extract stack hints
3. Glob for indicator files:

1. Parse context_files for technology mentions (fallback heuristic)

Output: detected_stack = {query_prefix} or empty (generic queries)

7d) Extract Topics (3-5)

• Parse all context_files for technology decisions
• Score each by weight from 7b table
• Take top 3-5 with weight >= 3
• Format: {topic_name, plan_statement, file_path, line_ref}

7e) MCP Ref Research

Per research_tool_fallback.md chain: Ref -> Context7 -> WebSearch -> built-in knowledge.

For each topic:

• Query: "{query_prefix} {topic} RFC standard best practices {current_year}"
• Collect: {official_position, source, matches_plan: bool}
• Run queries in parallel where possible

7f) Compare & Correct

For each topic where matches_plan == false (high confidence):

• Apply surgical Edit to plan file (single-line or minimal multi-line change)
• Add inline rationale: "(per {RFC/standard}:...)"
• Record correction in findings

For each topic where finding is ambiguous:

• Record as "REVIEW NEEDED" (not auto-corrected)

IF Plan Mode -> corrections applied to plan-mode file directly.

Safety rules:

• Max 5 corrections per run
• Each correction must cite specific RFC/standard/doc
• Only correct when official docs directly contradict plan statement

7g) Save Findings

Write to .agent-review/context/{identifier}_mcp_ref_findings.md (per references/mcp_ref_findings_template.md).

IF Plan Mode -> output findings to chat, skip file write.

Display: "MCP Ref: {N} topics validated, {M} corrections, {K} confirmed"

1. Critical Verification + Debate: per shared workflow, with MCP Ref enhancement:

- If MCP Ref research completed (7b-7g), use findings to inform AGREE/DISAGREE decisions - Agent suggestion contradicts MCP Ref finding → DISAGREE with RFC/standard citation - Agent suggestion aligns with MCP Ref finding → AGREE with higher confidence - Agent suggests something MCP Ref didn't cover → standard verification (no enhancement)

1. Aggregate + Return: per shared workflow. Merge agent suggestions + MCP Ref corrections into unified output.
2. Save Review Summary: per shared workflow "Step: Save Review Summary". IF Plan Mode → output to chat, skip file save.

Output Format

verdict: CONTEXT_ACCEPTABLE | SUGGESTIONS | SKIPPED
mcp_ref_corrections:
  count: 2
  topics_validated: 5
  corrections:
    - topic: "OAuth 2.0"
      file: "plan.md"
      line: 42
      before: "Use implicit flow"
      after: "Use Authorization Code + PKCE (RFC 6749)"
      source: "ref_search_documentation"
  findings_file: ".agent-review/context/{identifier}_mcp_ref_findings.md"
suggestions:
  - area: "logic | feasibility | completeness | consistency | best_practices | risk"
    issue: "What is wrong or could be improved"
    suggestion: "Specific actionable change"
    confidence: 95
    impact_percent: 15
    source: "codex-review"
    resolution: "accepted | accepted_after_debate | accepted_after_followup | rejected"

• mcp_ref_corrections: present only when MCP Ref research ran. Omitted when skipped.
• Agent stats and debate log per shared workflow output schema.

Verdict Escalation

• No escalation. Suggestions are advisory only.
• Caller decides how to apply accepted suggestions.

Reference Files

• Shared workflow: shared/references/agent_review_workflow.md
• Agent delegation pattern: shared/references/agent_delegation_pattern.md
• Prompt template (review): shared/agents/prompt_templates/context_review.md
• Review schema: shared/agents/schemas/context_review_schema.json
• Research fallback: shared/references/research_tool_fallback.md
• MCP Ref findings template: references/mcp_ref_findings_template.md

Version: 1.1.0 Last Updated: 2026-03-06