[blafourcade]

🎯 What & why

Lean aidd-dev:02-implement the way brainstorm and plan were, so it runs on Mistral medium under opencode as well as on Claude models, and give it a script that keeps a run on track. The plan skill gets three small additions the implement side relies on.

🛠️ How it works

• Router. One mermaid whose boxes carry the file paths, and four one-line rules (speak only to ask or answer, code only, push only when asked, do what a hook asks). The plan router now has the same shape: the architecture guard never required its action table.
• Three linear actions. prepare (5 steps), execute (8 numbered steps, the last one goes back to step 1), finalize (6). Flat numbered steps because Medium skipped the late steps of nested lists (commit, status, run record).
• A script that says what comes next. validate-implement.cjs checks plan and phase statuses and their dependencies, validates run.md, lists the files left uncommitted, and prints NEXT: take phase 2. Medium follows what a script prints far better than a rule it has to remember.
• --baseline. Run once in prepare, it records the files the user had already modified; a later commit holding one of them is refused with the command to undo it. It records only while the plan is pending, so a later call cannot take the run's own work for the past.
• run.md. One row per task (scenario, observed outcome), the head the run validated, and what was left. The script refuses a missing row, an unfilled template, or a head that is not a commit. This is what the review skill will read.
• Commits. commits.md: one commit per validated task by default; AI should auto commit: never produces none. stop.md: blocked and replan needed, each a one-line row; "stop at once" so no other task or phase is taken first.
• Plan side. A "Contributor checks" finding that becomes a task of the last phase, a "Runtime" check, validate-plan.cjs rewritten without loops (same output on the 73 plans of this repo), and the plan router without a table.
• Removed because it did not work. Delegates to sub-agents (never used spontaneously, 5x the cost when forced), a closed vocabulary for results, a feasibility step before coding, negative git rules, a plugin for opencode, and handling of byte-order marks and quoted statuses (0 occurrences in 322 plan files).

🧪 How to verify

• pnpm exec lefthook run pre-commit
• node scripts/check-tests-leave-git-alone.js -- node --test 'scripts/__tests__/**/*.test.js' (694 tests)
• Every rule of the two scripts has a mutation that turns its own test red: 36 for implement, 38 for plan. 14 tests of the first draft were dropped because no mutation needed them.
• Previous skill against this one, same scenarios, one run per cell (/ separates previous / new):

Scenario	Haiku	Medium (mistral-medium-2604)
Full plan, 3 phases	passes / passes	passes / 1 of 3 runs clean (8 of 8 on the 2-phase plan)
never in project memory	passes / passes	1 commit / 0 commits
A file the user modified before the start	passes / passes	the file is committed / left alone
Commit hook rejecting a message	passes / passes	not compared / passes, no bypass
Task that cannot hold in scope	passes / passes (4 of 4)	fails / fails

• Sonnet and Opus pass the full plan with the new skill. Words: router and actions 548 against 665; everything the skill holds 790 against 732.

⚠️ Heads-up

• Medium: the 8 runs on the 2-phase plan (4 dirty, 3 never, 1 rejected hook) all ended implemented with the script on OK. The 3 runs on the 3-phase plan, the longest, are the unstable ones: one clean, one implemented but ended on a script error (an invented head sha and a baseline re-run, both closed since without a new full run), one stopped mid-phase on a truncated reply. The previous skill was run once on that plan. So far the failures follow the length of the run, not the skill text.
• Medium writes pass: not applicable on a criterion that cannot hold and marks the plan implemented, with both versions. mistral-small-2603 never reads an action, with both versions.
• Medium commits once or per phase, not per task; the per-task commit is only reliable on Claude models.
• An absent commit policy means "commit each validated task", not "never". This is a decision, see the closing note on the commit-policy issue.
• Parallel phases are left to the orchestrator.
• On plan, Medium skipped the last action in one of 3 runs without the table and in 0 of 3 with it. Too few runs to tell.
• Not run here: Windows and Linux, which CI proves.

🔗 Linked issue

Closes #1076

Differs from its text on purpose: more words overall (790 against 732, though the part read on every run is 18% shorter), a commit per validated task by default, and no result on mistral-small-2603. Refs #927, #896, #928, #955.

✅ 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

https://claude.ai/code/session_01XeK3mFt8TGPAoKbzZAsaNf