git:20260703.939a14f to git:20260713.d210701

6 added, 18 removed. Audit A to A.

---

name: plugin-feedback

- description: "Record or apply development-skills plugin feedback. Mode via argument: produce (factual chronicle of plugin interactions this conversation) | ingest <report-path> (apply FIX verdicts, challenged against the Iron Rules). Runs on /plugin-feedback."

- disable-model-invocation: true

+ description: "Record or apply development-skills plugin feedback; runs on /plugin-feedback. Mode via argument: produce (factual report of plugin interactions this conversation) | ingest <report-path> (apply only evidence-backed simplifications)."

+ user-invocable: true

argument-hint: "produce | ingest <report-path>"

---

- # Plugin Feedback

- `$ARGUMENTS` selects the mode.

- ## produce

- Factual chronicle of development-skills interactions in this conversation — pure record, no judgment. Write to `docs/reports/development-skills-feedback-YYYY-MM-DD.md`, self-contained:

- - **Context** — language, size, test count, framework; the task; why the plugin was involved.

- - **Chain of thought** — exhaustive chronological dump of every plugin interaction (skill trigger, phase read, gate, routing decision, agent spawn, tool call, verification, deviation). Per step: plugin instruction quoted · agent action · reasoning · outcome.

- - **Friction points** — the steps that wasted effort or misled: file + instruction, what happened.

+ # Plugin feedback

- ## ingest <report-path>

+ `produce` writes `docs/reports/development-skills-feedback-YYYY-MM-DD.md` with the task context, plugin/skill actions, observed outcomes, friction, and reproducible eval ideas. Record events and evidence, not private reasoning.

- The report is input, not truth — most friction is model behavior or edge cases; a change must EARN its place against Iron Rules 3 (simplicity) and 5 (signal). SKIP if any holds: model ignored a clear instruction · fix adds a rule/exception for a one-time event · marginal gain for more words · already covered · shortening would fix it better (the only good FIX direction). FIX only when an instruction is wrong/misleading/repeatedly wasteful AND the fix is net simplification.

+ `ingest <report-path>` treats the report as a hypothesis. Change the plugin only when an instruction is demonstrably wrong or repeatedly wasteful and the fix is simpler than the current text. Prefer deletion or merging; do not add an exception for one model mistake.

- - Apply FIX verdicts only, surgically; adding words requires removing at least as many elsewhere (net ≤ 0).

- - Write verdicts + summary (FIX/SKIP counts, files changed) to `docs/reports/ingest-YYYY-MM-DD.md`; run pre-commit if a `.py` was touched.

- - Expect most friction to SKIP. If >30% FIX, re-examine your rigor.

+ Add an eval only when it can check an observable outcome through the Pydantic fresh-context schema. Report fixes, rejected suggestions, changed files, and verification. Expect most suggestions to be rejected.