80abf0f4d6
* Add analyze-project — Root Cause Analyst skill New forensic analysis skill for Antigravity. Reads brain/ artifacts (task.md, plans, walkthroughs + .resolved.N versions), classifies scope changes, rework shapes, root causes (SPEC_AMBIGUITY, REPO_FRAGILITY, etc.), clusters friction hotspots, and auto-updates project-health-state + prompt_improvement_tips.md. Full workflow and SKILL.md included. * Update SKILL.md Shorten frontmatter description to pass validation (<200 chars) * Update SKILL.md Fix YAML syntax + shorten description to pass validation (<200 chars); improve portability * Create sample_session_analysis_report.md Add examples/ with trimmed sample output report to demonstrate skill results * Update SKILL.md Fix non-portable hardcoded path in Step 11a (project-health-state) per bot feedback
4.4 KiB
Sample Output: session_analysis_report.md
Generated by /analyze-project skill on a ~3-week project with ~50 substantive sessions.
(Trimmed for demo; real reports include full per-conversation breakdown and more cohorts.)
📊 Session Analysis Report — Sample AI Video Studio
Generated: 2026-03-13
Conversations Analyzed: 54 substantive (with artifacts)
Date Range: Feb 18 – Mar 13, 2026
Executive Summary
| Metric | Value | Rating |
|---|---|---|
| First-Shot Success Rate | 52% | 🟡 |
| Completion Rate | 70% | 🟢 |
| Avg Scope Growth | +58% | 🟡 |
| Replan Rate | 30% | 🟢 |
| Median Duration | ~35 min | 🟢 |
| Avg Revision Intensity | 4.8 versions | 🟡 |
| Abandoned Rate | 22% | 🟡 |
Narrative: High velocity with strong completion on workflow-driven tasks. Main friction is post-success human scope expansion — users add "while we're here" features after initial work succeeds, turning narrow tasks into multi-phase epics. Not primarily prompt or agent issues — more workflow discipline.
Root Cause Breakdown (non-clean sessions only)
| Root Cause | % | Notes |
|---|---|---|
| Human Scope Change | 37% | New features/epics added mid-session after success |
| Legitimate Task Complexity | 26% | Multi-phase builds with expected iteration |
| Repo Fragility | 15% | Hidden coupling, pre-existing bugs |
| Verification Churn | 11% | Late test/build failures |
| Spec Ambiguity | 7% | Vague initial ask |
| Agent Architectural Error | 4% | Rare wrong approach |
Confidence: High for top two (direct evidence from version diffs).
Scope Change Analysis Highlights
Human-Added (most common): Starts narrow → grows after Phase 1 succeeds (e.g., T2E QA → A/B testing + demos + editor tools).
Necessary Discovered: Hidden deps, missing packages, env issues (e.g., auth bcrypt blocking E2E).
Agent-Introduced: Very rare (1 case of over-creating components).
Rework Shape Summary
- Clean execution: 52%
- Progressive expansion: 18% (dominant failure mode)
- Early replan → stable: 11%
- Late verification churn: 7%
- Exploratory/research: 7%
- Abandoned mid-flight: 4%
Pattern: Progressive expansion often follows successful implementation — user adds adjacent work in same session.
Friction Hotspots (top areas)
| Subsystem | Sessions | Avg Revisions | Main Cause |
|---|---|---|---|
| production.py + domain | 8 | 6.2 | Hidden coupling |
| fal.py (model adapter) | 7 | 5.0 | Legitimate complexity |
| billing.py + tests | 6 | 5.5 | Verification churn |
| frontend/ build | 5 | 7.0 | Missing deps/types |
| Auth/bcrypt | 3 | 4.7 | Blocks E2E testing |
Non-Obvious Findings (top 3)
- Post-Success Expansion Dominates — Most scope growth happens after initial completion succeeds, not from bad planning. (High confidence)
- File Targeting > Acceptance Criteria — Missing specific files correlates more with replanning (44% vs 12%) than missing criteria. Anchors agent research early. (High)
- Frontend Build is Silent Killer — Late TypeScript/import failures add 2–4 cycles repeatedly. No pre-flight check exists. (High)
Recommendations (top 4)
- Split Sessions After Phases — Start new conversation after successful completion to avoid context bloat and scope creep. Expected: +13% first-shot success. (High)
- Enforce File Targeting — Add pre-check in prompt optimizer to flag missing file/module refs. Expected: halve replan rate. (High)
- Add Frontend Preflight — Run
npm run buildearly in frontend-touching sessions. Eliminates common late blockers. (High) - Fix Auth Test Fixture — Seed test users with plain passwords or bypass bcrypt for local E2E. Unblocks browser testing. (High)
This sample shows the forensic style: evidence-backed, confidence-rated, focused on actionable patterns rather than raw counts.