Review Knowledge

Goal

Maintain knowledge base health and surface entries for review, helping the user internalize accumulated knowledge through periodic reflection.

When to Use

• User explicitly requests a knowledge review (e.g., "review knowledge", "check knowledge base")
• At the end of a significant work phase or project milestone
• When the user wants to understand the current state of knowledge on a topic
• Periodically (e.g., weekly/monthly) to keep the knowledge base healthy

Execution

Delegate the review work to a Sonnet subagent to minimize main context consumption.

1. Determine the review mode from user input:

• No argument or "health" → health
• topic:<keyword> → topic with the specified keyword
• fix → fix

1. Determine the knowledge base path (default: .claude/knowledge/entries/)
2. Precompute deterministic graph facts from the MAIN agent's Bash (subagents run in a

sandbox that blocks code execution, so these must run here):

python3 "{plugin_root}/scripts/kb_graph.py" --root .claude/knowledge/entries stats
python3 "{plugin_root}/scripts/kb_graph.py" --root .claude/knowledge/entries lint --json

Notes: pure stdlib, ~1s, no index needed. lint exiting 1 just means it found findings — capture stdout either way. If python3 is unavailable or the commands fail, set both outputs to (unavailable) and continue; the subagent then derives everything by reading files as before.

1. Spawn a subagent with the following configuration:

Agent(
  subagent_type: "general-purpose",
  model: "sonnet",
  description: "review knowledge base",
  prompt: |
    You are a knowledge base review agent. Read the procedure file and follow it precisely.

    ## Input
    ### mode
    {mode}

    ### topic_keyword (only for topic mode)
    {topic_keyword}

    ### current_date
    {current_date}

    ### graph_stats (deterministic, precomputed by kb_graph.py)
    {graph_stats}

    ### graph_lint (deterministic, precomputed by kb_graph.py)
    {graph_lint}

    ## Instructions
    1. Read the procedure file at: {plugin_root}/skills/review-knowledge/procedure.md
    2. Follow the procedure step by step for the specified mode
    3. The knowledge base is at: {project_root}/.claude/knowledge/
    4. Return the full review report in the output format specified in the procedure
)

Replace the placeholders:

• {mode} — review mode: health, topic, or fix
• {topic_keyword} — keyword or tag for topic mode (empty for other modes)
• {current_date} — today's date in YYYY-MM-DD format (needed for stale entry detection)
• {graph_stats} / {graph_lint} — stdout captured in step 3 ((unavailable) on failure)
• {plugin_root} — the plugin's installation path (shown in the skill loading message as "Base directory for this skill")
• {project_root} — the project working directory

1. Report the subagent's result to the user

IMPORTANT: The procedure file path uses the plugin's base directory, NOT the project directory. Read the "Base directory for this skill" line from the skill loading message to determine the correct path.