> Part of [AGENTS.md](../AGENTS.md) — project guidance for AI coding agents.
A recursive self-improvement toolchain for plan2code contributors. Collects run metrics, aggregates by prompt generation, diagnoses weak steps via AI, and proposes surgical prompt edits.
## Metrics Data Flow
```
Collect → Aggregate → Analyze → Improve → Apply
```
1.**Collector** reads project artifacts (`specs/<feature>/`) → writes `RunMetrics` JSON per run
2.**Aggregator** groups runs by prompt SHA fingerprint (cohorts) → `aggregated.json`
3.**Analyzer** invokes AI with aggregated metrics + prompt contents → diagnosis markdown
4.**Improver** invokes AI with diagnosis → validated `PromptEdit[]` proposals (char limit + verbatim checks)
Community submissions arrive as GitHub issues labeled `community-feedback` on `jparkerweb/plan2code`, created by the finalize prompt's post-Step-6 submission flow; the "Fetch community submissions" option requires an authenticated `gh` CLI to list/close them.
The collector parses an optional `## User Feedback` table from `overview.md`:
```markdown
## User Feedback
| Field | Value |
|-------|-------|
| Rating | 8 |
| Reason | Smooth workflow |
| Went Well | Planning was thorough |
| Went Poorly | Some tasks unclear |
```
Feedback is collected during finalize (Step 5) or retroactively via the CLI. Pipe characters in values are escaped as `\|`. The aggregator computes `avg_user_rating` and `feedback_count` per cohort.
On start the CLI asks for the plan2code repo path (default: saved `plan2codeRepoPath` in `~/.plan2code-metrics.json`, else auto-detected up to 5 parent dirs) and the project path (default: cwd). Data is stored in `<project>/.plan2code-metrics/` (runs/, aggregated.json, proposals/).
## Prompt fingerprints and the one-time regrouping
`collectPromptVersions` reads every `src/plan2code-*.md` file from the folder (so a new skill is covered with no code change) and keys each hash by name: the eight prompts recorded earlier keep their keys (`plan`, `revise_plan`, `document`, `implement`, `finalize`, `init`, `init_update`, `quick_task`), any other file becomes its name without `plan2code-` and `.md`, dashes as underscores (`3_implement_review`, `git`).
The cohort key hashes every fingerprint that is not the `sha256:missing` sentinel, so the first run recorded after this change has more entries than any earlier one and starts a new cohort. Runs recorded before stay as they are and group together as the previous prompt generation; nothing is rewritten or dropped. This happens once, at the change, and again only when a prompt file is edited or added.
## The scrape pattern (one source)
What the collector reads from a plan draft (the `METRICS_JSON` comment, the confidence labels and the four breakdown labels) lives in `src/web-console/metrics-pattern.json`. The console's bait scanner (`src/web-console/patches.mjs`) reads it, `plan2code-metrics/src/metrics-pattern.json` is its copy (bundled by tsup), and `SCRAPE` in `src/skill-scripts/common.mjs` mirrors it for `specs.mjs check`. `scripts/test-web-console-hardening.mjs` fails when any copy differs. After editing the pattern, copy the file to `plan2code-metrics/` and update `SCRAPE`. A breakdown label then a number counts in prose, and in a table row only when the number ends the row, so an ordinary risk-register row is not refused.