[blafourcade]

🎯 What & why

Lean aidd-dev:01-plan so it runs on Mistral medium under opencode as well as on Claude models, and writes the plan folder every time. Instructions drop from 968 to 478 words (-51%); everything a run reads, scripts excluded, from 1750 to 1322 (-24%).

🛠️ How it works

• Three actions behind one mermaid router, SKILL.md: capture (read the source silently), explore (loop on the code until nothing blocks), conclude (phases, files, validation, check, approval). Same vocabulary as brainstorm.
• Medium ends its turn on any text-only message, so the router carries what must hold on every turn: "Speak only to ask or to give the final reply" and "Plan only". The mermaid cites no file path: the router guard needs a table row per action.
• The plan folder is checked by validate-plan.cjs, derived from the two templates (sections, order, headers, front matter, no comments or placeholders, phase rows match phase files, listed files to modify exist, acceptance rows map to tasks). A changed template changes the verdict with no edit to the script. Tests: validate-plan.test.js.
• One read-only subagent checks the plan against plan-checks (Requirements, Reading, Evidence), once, and returns only its ✗ lines. No score.
• The Phases table gains Depends on (earlier phase numbers, or none), validated as strictly lower numbers.
• A phase that builds a screen layout draws a numbered wireframe with a legend, else omits the section. The validator rejects none, N/A, or a frame without region (1) and a legend.

🧪 How to verify

• pnpm exec lefthook run pre-commit
• node scripts/check-tests-leave-git-alone.js -- node --test scripts/__tests__/validate-plan.test.js (28 tests, mutation proofs for the wireframe and Depends-on rules)
• Nine scripted scenarios (empty request, ticket, UI, regression, trivial, correction, explore more, file source, auto mode):
  • Claude haiku passes all nine with the judge called once. sonnet and opus write the plan on S2, S3, S5; plans checked against the code.
  • mistral-medium-2604: plan written in 17 of 18 runs of the final router, no source file modified in the last 7 S2/S9 runs.
  • Original skill, 4 scenarios on haiku: 9 bench failures (documentation impact missing on all four, comments left, a "Confidence Score").

⚠️ Heads-up

• Costs more than the original: haiku 0.21 $ against 0.07 $ over S2 to S5; mistral-medium about 0.4 $ against 0.2 $ per run. The validator and the check add calls; the lean text does not pay for them.
• Medium still stops without writing the plan about once in 18 runs (a text-only turn). No wording fixes it; a retry loop outside the skill would.
• Medium plan content is weaker than Claude's: wrong version source (cli/package.json instead of the marketplace), invented external URLs, tests of the touched file not always cited. Running the check on a stronger model through the opencode config is the lever left.
• Depends on follows the sub-agents work; how 02-implement consumes it is not touched here.
• Single runs per cell: only large gaps are signal.

🔗 Linked issue

Closes #1075

✅ 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