retro

Files (1)

• SKILL.md

SKILL.md (first lines)

# retro

The reflector step. Evidence first, opinions last. Keep it under one page.

1. Numbers:
   - `hbs report --repo <repo>` — per-phase retries, review rounds, loop triggers, time, tokens/cost; weak spots.
   - `hbs optimize` — whether the system as a whole is getting cheaper, faster, more accurate, and hallucinating less; which experiments this run was part of (`exp`/`arm` in the ledger).
   - `hermes kanban --board <board> list` — anything still blocked or archived mid-run.
   - Budget: spent vs `router.yaml` budgets.
2. Lesson votes (this is how the playbook learns what actually helps):
   - Grep card bodies for injected lessons: `grep -h "^- L[0-9]" build/runs/*.prompt.md 2>/dev/null | sort -u`, and check the kanban card bodies of this run.
   - For each lesson that was in play: if the node it targeted went clean (no retry, no changes requested), `hbs lessons vote <id> helpful --source <repo>`; if that node failed in the way the lesson should have prevented, vote `harmful`. No evidence either way → no vote.
3. Write `build/12-retro.md`:
   - **Outcome**: shipped / capped / parked, and the verify result.
   - **Numbers**: the report table (trimmed) and total cost.
   - **What cost the most**: top 3 by retries + review rounds + loops, each with the concrete cause (quote the verify failure or review finding).
   - **Human overrides**: gates rejected, reruns, parked assumptions.
   - **Lesson votes cast**: id → helpful/harmful, one reason each.
   - **Scout needs**: 1–3 capability gaps phrased as search queries (e.g. "playwright e2e test generation skill"). Add them under `needs:` in `~/.build-framework/scout/sources.yaml` only if the human agrees.
4. Append the build-log line. Complete the card (or update STATUS in chat mode).

Do not propose skill edits here — that is `distill`.

From Kshintak/new-build-framework, MIT.

More from Kshintak/new-build-framework