kenoxa/run-insights

Cross-tool session analysis. Python scripts compress ~256MB raw data into ~100KB normalized analytics. Subagents analyze patterns, produce recommendations in 7 categories.

Phases

Every subagent prompt MUST be self-contained — include all prior-phase context.

Reference paths (backticked): dispatch to subagent — do NOT Read into mainthread.

Session ID: Generate per SPINE.md Sessions convention. Reuse across all phases.

Phase Trace: Log row at collect, analyze, synthesize, present. Include dispatch count.

| Phase | Agent type | Reference | |-------|-----------|-----------| | Collect | @miner | references/collect-miner.md | | Analyze | @miner (source-expert) | references/analyze-source-expert.md | | Synthesize | @synthesizer | references/synthesize-synthesizer.md | | Present | @visualizer | references/present-visualizer.md |

1. Collect

Dispatch @miner (collector) → references/collect-miner.md.

Output: .scratch/<session>/insights-collect-collector.md

2. Analyze

For each provider in summary.provider_breakdown with sessions >= 1, dispatch one @miner (source-expert) in parallel → references/analyze-source-expert.md. Cap: 6 per batch (per SPINE.md).

Null-telemetry short-circuit: if all of a provider's sessions have empty tool_calls AND empty skills_used, do not dispatch. Write a 1-paragraph "no usable telemetry — N sessions detected, no per-call attribution available" note to .scratch/<session>/insights-analyze-{provider}-expert.md and proceed.

| Role | Input | Output | |------|-------|--------| | {provider}-expert | provider-specific sections + per_project entries where providers includes {provider} + cross-cutting sections relevant to that provider (friction, subagent, operational_health, cross_tool — apply per focus block in ref) | .scratch/<session>/insights-analyze-{provider}-expert.md |

Include relevant analytics data inline in each dispatch prompt.

3. Synthesize

Dispatch 1 @synthesizer → references/synthesize-synthesizer.md.

Input: all insights-analyze-*.md files + cross_tool + sample_prompts sections from analytics.json.

7 recommendation categories: Skills, Hooks, MCP Servers, Plugins, Agents, CLAUDE.md rules, Anti-patterns.

Output: .scratch/<session>/insights-synthesize-synthesizer.md

4. Present

Terminal summary (always): read synthesizer output. Activity stats table (sessions by provider, avg duration, top projects), then recommendations by category sorted by priority. Lead with highest-impact. Include only non-empty categories.

HTML dashboard: Guard: .scratch/<session>/insights-synthesize-synthesizer.md exists and non-empty. Inform user before dispatch.

Dispatch @visualizer → references/present-visualizer.md.

Inputs: .scratch/<session>/insights-synthesize-synthesizer.md, .scratch/<session>/analytics.json

Output: .scratch/<session>/history-insights.html

Anti-Patterns

• Recommendations without session evidence ("seen in N sessions" or "e.g., session X on date Y")
• Conflating project-level and global patterns
• Creating new skill when existing skill just needs better description triggers
• Ignoring cross-tool patterns (highest-value findings)
• Auto-applying recommendations instead of presenting as proposals