plan-change

wb <args>    # wb not on PATH: python <this skill's directory>/../../lib/wb.py <args>

Steps

1. Load the ticket. sdd get is not it — read .workflow/<KEY>/triage.json. No triage yet? Run triage-task first, or work from what the user stated and record that in questions.

2. Get the bar. repo profile. It prints the preset, the repo's own conventions, and the gates that apply. Follow the conventions it reports; do not introduce a second test runner or package manager. LOW confidence means the preset is a guess nobody reviewed: say so, let the user settle it. Once the file list is known, repo gates <paths> resolves the bar those paths land in — in a monorepo that is not the repo's, and the audit fails a plan that declares a preset below it.

3. Pick a persona from references/personas.md. It shapes what counts as done, not the format.

   A change already on disk (a light-path ticket finish moved here) stays: plan around it, never revert it to plan from scratch.

4. Confirm everything in the code. Assume nothing. Every claim about this codebase needs a file:line and the text of that line, because the audit in step 7 reopens each one. A claim you cannot cite is a question, not a fact — put it in questions. Search with the strongest tool available and quote only lines you actually read: ../../shared/code-search.md.

5. Find the raised bar. repo zones <path>... with every file you plan to touch. Billing, auth, user data, migrations and secrets are held to a higher standard whatever the preset says.

6. Write .workflow/<KEY>/sdd.json. Shape and required fields: references/sdd.md.

7. Audit. sdd audit <KEY>. It must pass. On failure, fix the plan — never the citation to match a wrong claim, and never proceed to implementation. Then sdd render <KEY> for the human-readable copy.

   The audit reports a tier. Give every files[] entry a lines estimate (added plus removed); an entry without one makes the plan standard. An estimated total within the repo's bound (100 by default), no critical zone and no bug/support ticket qualifies as light, which waives steps and product — nothing else. impl check measures the real diff against the bound, so an estimate written low only moves the failure later. Do not aim for a tier: write the plan the change needs and let the audit compute it. Citations, the file list, verify and rollback are required at every tier.

   The first audit is strict and reads the working tree, so get the line numbers right while the plan is still cheap to change. It records the commit it ran against; every later audit is anchored there, so a plan can be corrected while it is being implemented. Drifted citations are reported, not failed. A quote that was never in the code fails at every stage.

Rules

• Smallest change that achieves the objective. Fixes and improvements to existing behaviour touch as little as possible. No refactor riding along, no abstraction for a second case that does not exist yet, no dependency that replaces code already in the repo.
• List every file before touching any. implement-change stops on a file that is not in the list, so an incomplete list is a plan that fails to run.
• Verification is commands, not intentions. verify holds the exact lines that will be executed, using this repo's runner. Shell is refused, so a command needing a variable declares it in verify_env ({"PYTHONPATH": "lib"}) rather than inlining VAR=x. Variables that change how a process loads code are refused there.
• An open question is output. A plan that guesses reads the same as a plan that knows; say which one it is.

Next: implement-change reads sdd get <KEY> --section files|steps.