kaizen-agentic/docs/TELEMETRY.md

50 lines
1.9 KiB
Markdown
Raw Permalink Normal View History

# Telemetry and Agent Effectiveness Tracking
WP-0001 T04 design — aligned with ADR-004 and WP-0004 ecosystem integration.
## Two layers (do not merge)
| Layer | Question | Mechanism |
|-------|----------|-----------|
| **Project** | How is agent *X* performing in *this repo*? | `kaizen-agentic metrics record``.kaizen/metrics/` |
| **Fleet** | How are coding sessions performing *across repos*? | agentic-resources Helix Forge |
kaizen-agentic **does not** ship a parallel session transcript ingestion pipeline.
## Project telemetry (implemented)
Memory-enabled agents record per-session outcomes at close:
```bash
kaizen-agentic metrics record <agent> --success --time <s> --quality <0-1>
kaizen-agentic metrics record <agent> --success --time <s> --quality <0-1> --emit-event
kaizen-agentic metrics optimize [agent]
kaizen-agentic memory brief <agent> # includes Performance Summary
```
Optional fleet correlation via `HELIX_SESSION_UID` (see
[integrations/helix-forge-correlation.md](integrations/helix-forge-correlation.md)).
## Fleet telemetry (agentic-resources)
Helix Forge owns session capture, digest storage, baselines, and weekly retro.
kaizen-agentic consumes correlation fields only.
## CLI install / usage analytics (future)
Potential v1.1 additions (not yet implemented):
- Opt-in anonymous counters on `install` / `memory init` (no PII, no project paths)
- Aggregate effectiveness reports via `metrics list` across a monorepo checkout
## tele-mcp evaluation (deferred)
2026-08-20 09:43:49 +02:00
[tele-mcp](https://forgejo.coulomb.social/coulomb/tele-mcp) is a candidate MCP adapter
for IDE-level telemetry (WP-0001 note). Assess before depending on it. Project and
fleet layers above satisfy INTENT's "measurable agents" requirement without tele-mcp.
## Feedback loop
2026-08-20 09:43:49 +02:00
User experience feedback uses [FEEDBACK.md](FEEDBACK.md) and Forgejo issue templates —
separate from execution metrics.