radar-analyze

Radar Analyze

Analyze recent Claude Code and Cowork sessions to generate actionable workflow insights.

Uses the bundled workflow-analyzer for session parsing and enrichment. Claude does the reasoning.

Arguments

• $ARGUMENTS — Optional:
  • --days N — How many days of history to analyze (default: 7)

Parse the days value from $ARGUMENTS if provided. Default to 7.

Workflow

Step 1: Parse & Enrich Session Data

Run the workflow-analyzer CLI to parse and enrich sessions from all configured sources (Claude Code + Cowork):

# Bundled binary (preferred)
node "${CLAUDE_PLUGIN_ROOT}/bin/workflow-analyzer/dist/cli.js" parse --since ${DAYS} --output /tmp/workflow-analyzer-parsed.json

# Fallback if bin/ not available:
# npx @flippyhead/workflow-analyzer@latest parse --since ${DAYS} --output /tmp/workflow-analyzer-parsed.json

If the npx command fails, surface the error output directly to the user. Do not swallow the error or exit silently.

Read the output file. It contains { sessions: [...], sessionGroups: [...] }.

If sessions is empty, report "No activity in the last N days" and stop.

Otherwise, note the summary: how many sessions, which sources (claude-code, cowork), how many session groups.

Step 2: Check Previous Insights

Use the get_insights MCP tool to check for existing insights:

1. Call get_insights with status: "new" — unresolved insights
2. Call get_insights with status: "noted" — acknowledged but not acted on
3. Call get_insights with status: "dismissed" — things the user doesn't want to see again

Note which deduplication keys and categories to avoid repeating. If the MCP tool is unavailable, skip this step.

Step 3: Fetch User Goals

Call get_lists with pinned: true to load the user's current priorities and goals. These are used by the Decision Support analysis below.

If unavailable, skip goal-based analysis.

Step 4: Analyze

With all data gathered, analyze the parsed sessions and produce insights across four modules. For each insight, you MUST provide a concrete action — never just describe a problem without telling the user what to do about it.

Module A — Root Cause Diagnosis:

Look at tool failures in the session data. For each tool with a notable failure rate:

• Read the actual error messages (not just counts)
• Diagnose the root cause: is it an auth issue, config problem, API bug, user error, or transient?
• If fixable: provide the specific command or config change to fix it (action type: run or action type: install)
• If not fixable: acknowledge it so the user stops worrying (action type: acknowledge)
• Skip tools with very low failure counts (<3 calls) or very low failure rates (<20%)

Module B — Direct Automation:

Look for patterns that could be automated:

• Permission confirmations: Count short messages like "yes", "y", "a", "ok", "sure" that appear to be tool permission confirmations. If a specific tool triggers many confirmations, suggest an allowedTools config entry (action type: install with settings.json patch)
• Repeated prompts: Look at user messages across sessions for frequently typed prompts. If a prompt appears 5+ times, suggest creating a skill or alias (action type: install with skill content)
• Repeated tool sequences: Look for the same sequence of 3+ tool calls appearing across sessions. Suggest automation if a pattern repeats 5+ times.

Module C — Decision Support:

Using the session groups and the user's pinned goals from Step 3:

• Calculate time allocation by project/topic (sessions and minutes)
• Compare against stated goals — flag misalignments
• Note cross-platform patterns (e.g., "researched X in Cowork but never implemented in Claude Code")
• If a project is consuming disproportionate time without being in the goals, surface a decision (action type: decide)

Module D — Knowledge Nudges:

Look for repeated topics across sessions:

• Find user prompts that ask about the same thing in multiple sessions (sign the answer should be saved)
• Detect when the user provides the same context/background repeatedly at session start (should be in CLAUDE.md or memory)
• For repeated topics, generate a consolidated summary of what to save (action type: save)

Insight Format

For each insight, produce:

{
  "module": "root-cause | direct-automation | decision-support | knowledge-nudges",
  "severity": "alert | action | suggestion | info",
  "title": "One-line summary",
  "observation": "What the data shows (with specific numbers)",
  "diagnosis": "Why this is happening (optional)",
  "action": {
    "type": "install | run | save | review | decide | acknowledge"
  },
  "evidence": [{"metric": "specific numbers"}],
  "effort": "low | medium | high",
  "impact": "low | medium | high",
  "confidence": 0.0-1.0,
  "deduplicationKey": "unique-key-for-this-insight"
}

Action types:

• install: { type: "install", artifact: "filename", content: "file content to install" }
• run: { type: "run", command: "command to run", explanation: "why" }
• save: { type: "save", content: "what to save", destination: "AI Brain / CLAUDE.md / etc" }
• review: { type: "review", summary: "what to look at", links: [] }
• decide: { type: "decide", question: "decision to make", options: ["option 1", "option 2"] }
• acknowledge: { type: "acknowledge", message: "FYI only, no action needed" }

Aim for 5-10 total insights. Prioritize high-impact/low-effort actions. Skip insights that match dismissed deduplication keys from Step 2.

Step 5: Publish

Write insights to a temp file and use the CLI to publish:

# Bundled binary (preferred)
node "${CLAUDE_PLUGIN_ROOT}/bin/workflow-analyzer/dist/cli.js" publish --insights /tmp/workflow-analyzer-insights.json

# Fallback if bin/ not available:
# npx @flippyhead/workflow-analyzer@latest publish --insights /tmp/workflow-analyzer-insights.json

The insights JSON file should contain:

{
  "insights": [...],
  "metadata": {
    "period": { "since": "ISO date", "until": "ISO date" },
    "sessionCount": N,
    "sources": ["claude-code", "cowork"],
    "modulesRun": ["root-cause", "direct-automation", "decision-support", "knowledge-nudges"]
  }
}

Write this file before running the publish command. If publish fails, fall back to saving insights via create_report or capture_thought MCP tools.

If no brain MCP tools are available, skip publishing entirely — the terminal output from Step 6 is the primary output in terminal-only mode.

Step 6: Summary

Output a brief summary:

• How many insights were generated, by module
• The report period and session count (including Cowork if any)
• Top 2-3 "quick win" recommendations (highest impact, lowest effort)
• If brain is connected: "Insights also saved to your brain. Review at /insights in the AI Brain web UI."
• If terminal-only: no brain reference in the output