bmady (Bmady) — dogfood bilan

Index + template: README.md. Newest first.

2026-09-17 — profile 2: the five personas and five gates walked on a scratch CLI, the commit swept the skills mirror again (run 01a0afa6-aaf4)

• Status: validated end to end, one known sweep — analyst → elicit_brief → pm → review_prd (approve) → architect → approve_arch (approved) → select_stories (S1–S3, now, WIP 3) → dev → qa (no blocker, confidence high) → final_review (ship) → commit_changes → done. Neither new exhaustion exit was reached (both loops approved on their first turn); their witness is the dry run below.
• Versions: bot bmady 0.1.1 (dsl: 2 + the two C145 exits, wave 3b of #1344) · iterion 5fa790d82 for the bots; the engine the branch binary v3.157.0+c31559517 (built on the wave-3a branch from main at v3.157.0; main is at v3.159.0 as this is written, one PR of engine work ahead), served through the host's Anthropic-compatible facade (z.ai) as the runs' provenance records.
• Method: launched FROM a scratch greeter repository, --var brief="Add a --lang option (en or fr) to the greet CLI: fr prints 'Bonjour, NAME'. Keep the standard-library-only rule of ADR-0001 and the existing tests green.", --store-dir the operator's workspace store, --sandbox none, --merge-into none, caps --max-cost-usd 1.5 --max-duration 10m, every LLM node on claude_code/claude-opus-5. The five gates answered with iterion resume --answer …; the operator pauses count against the wall, so the 90 % duration guard fired as dev started and the run was resumed with --max-duration 40m --max-cost-usd 4.
• Result: $1.83, 23 min 18 s from launch to done including the five pauses. The analyst asked the right clarifications (language set, casing, exit code), the PM wrote acceptance criteria with the exit-code map, the architect chose a ValueError subclass carried through the existing except-ladder and named the ordering risk, the dev implemented S1–S3 with a GREETINGS table as the single source of truth plus tests asserting exit 3 and the exact stderr line, QA found no blocker; commit_changes produced d5a39fd on the run worktree — 43 files, of which 16 under .claude/skills/ are iterion's own mirror, swept in by the deterministic git add -A (the bot's comment calls it safe on the assumption that the repository ignores the mirror; this scratch does not). Storage branch iterion/run/01a0afa6-aaf4-7549-960f-70206b791fd0 → d5a39fd.
• Value: the fifteen prompts' paragraph breaks reach the models — proven in the run's events.jsonl for every system prompt and all five gate instruction blocks, whose # titles and paragraphs now arrive as the authors laid them out — and the two exits hold in validate --exec --strict: the false pass (every condition false, so both loops exhaust) now ends finished through derive_prd -> architect and approve_arch -> select_stories where main's copy ended failed_resumable (LOOP_EXHAUSTED).
• Findings / misses: the mirror sweep is the commit-side twin of #1364 (the gate-side refusal of ?? .claude/) — same cause, same class rule to apply (iterion's scaffold is never run output); recorded on #1364, not fixed in this wave.
• Engine hardening: none in this wave.
• Lessons for next run: budget the wall for the operator's own latency at five gates (--max-duration 30m); pass --merge-into none again on resume.

2026-06-14 — full BMAD flow validated end-to-end via API-driven gates (run 019ec66b)

• Status: validated. All five BMAD personas + all five human gates ran to a shipped, tested, committed feature. I drove the gates programmatically (resume API), not Playwright.
• Method: dedicated worktree studio :4899, worktree: auto (no sandbox), all nodes claude_code/opus (no claw/gpt → no forfait shape flakiness). --var brief="Add a --short flag to iterion version that prints only the semantic version, for shell scripts", merge_into=none.
• Flow: analyst (Mary) → elicit_brief → pm (John) → review_prd → architect (Winston) → approve_arch → select_stories → dev (James) → qa → final_review → commit → done. Each gate paused as paused_waiting_human; answered via POST /api/runs/{id}/resume with the node's output-schema fields (clarifications / action / approved / selected_story_ids+priority+wip_limit / action:ship).
• Quality of each persona (high):
  • Analyst surfaced 5 genuine open questions (build-date misconception, dev un-stamped handling, leading-v, --json composition, flag spelling) — real ambiguities, no solutioning.
  • PM produced precise acceptance criteria incl. "no regression to RawVersion / desktop auto-updater" and "--json wins deterministically".
  • Architect correctly judged this "exposure work, not new machinery" and reused the existing FullVersion/Version seam (3-layer minimal design).
  • Dev implemented exactly to spec AND wrote tests unprompted (cmd/iterion/version_test.go + pkg/cli/version_test.go).
  • qa judge: blockers=[], confidence high.
• Result: final_commit f452cccf on storage branch iterion/run/dawn-thrash-midnightkazoo-9da0 (not merged). The feature is correct: ShortVersion() reads only appinfo.Version (commit suffix structurally impossible), strips one leading v; --short prints it; --json takes precedence; version_test.go covers both; docs in docs/cli-reference.md + CLAUDE.md.
• Finding (confirms Doki): the commit also swept in .claude/skills/bmady-*.md (6 files) — the runtime skill mirror. The clone carries main's .gitignore (which lacks the .claude/skills/ entry added on the c082-board-emit branch, commit a9b6d671), so Bmady's commit_changes included the mirror as if it were source. This is exactly the data-loss/noise Doki flagged and the gitignore fix prevents — a code bot DID sweep the mirror into its commit. Re-running with the gitignore fix in place would produce a clean feature-only commit.
• Lessons for next run: Bmady is reliable + high-value for a scoped feature when the gates are driven attentively; the all-claude_code/opus pipeline avoided the forfait shape flakiness that bit Seki. Land the .claude/skills/ gitignore fix (already on this branch) before using Bmady on a repo where the commit matters.