[blafourcade]

🎯 What & why

Lean aidd-refine:01-brainstorm from 13 files and 1506 words to 8 files and about 940, so it runs on Mistral small and medium under opencode as well as on Claude models, at a lower token cost.

🛠️ How it works

• The loop is one action. probe and integrate become explore: take stock, decide whether to conclude, pick one point, ask, wait, repeat. A skill that alternates between two files forces a re-read on every turn.
• Three actions, one per phase: capture (once), explore (loop), conclude (once). The router in SKILL.md is one mermaid whose nodes carry the file paths. The action table is gone: it repeated each file's goal.
• References answer one question each: discovery-map (what to explore, points in order, no labels so the model has nothing to quote), probing (how to ask, 7 techniques), persistence (where the brief lives). Single-use references moved into their action.
• Rules that must hold on later turns live in SKILL.md, not in an action read once: a stop or a summary request produces the brief; never talk about the skill's steps or files.
• The brief is Clear / Assumed / Open. A summary can no longer present a guess as a decision. A persisted brief carries status: approved.
• The router guard test now also mutates citations inside fenced blocks, so a mermaid router is checked like a table one. This removes the 10-batch exception.

🧪 How to verify

• pnpm exec lefthook run pre-commit
• node scripts/check-tests-leave-git-alone.js -- node --test 'scripts/__tests__/**/*.test.js' (477 tests)
• Measured against the original skill on the same scripted scenarios, one run per scenario: Claude haiku / sonnet / opus cost -29% / -47% / -19%; on "stop" the new skill produces the brief with headings on 3 of 3 models, the original on 0 of 3. mistral-medium-2604 holds the loop over 22 turns with no exit from the skill and no empty reply.
• Two real backlog issues (feat(framework): show installed versions and notify available updates #980 as S, feat(aidd-dev): carry documentation impact through feature delivery #927 as M) replayed with a simulated author: issue decisions recovered on par with the original (15 against 15).

⚠️ Heads-up

• mistral-small-2603 asks a question on the first turn in about 60% of runs, whatever the wording (20 runs per variant), and does not always produce the brief on "stop".
• mistral-medium-2604 asks more questions per message than the original (5.4 ? against 2.1). Four fixes were tested over 30 messages each, none helped, none is kept.
• Invented points in the brief are slightly higher than the original (22 against 16 across the replayed dialogues), partly because the S issue is ambiguous without repo context.
• Single run per cell for the Claude comparison: only the large gaps are signal.
• The auto mode of aidd-orchestrator:01-sdlc is untouched: brainstorm still waits for an answer.

🔗 Linked issue

Closes #1071

✅ I certify

• I DO CERTIFY I READ EACH LINE OF THE PULL REQUEST BECAUSE I AM A SOFTWARE ENGINEER, NOT A AI PUPPY.

🤖 Generated with Claude Code

[alexsoyes]

J'ai rien à dire 😗

[alexsoyes]

https://www.loom.com/share/fd1d935553e048b998ab1419a0232734