/we:develop

Implement the chunk, run the local gates, commit, report, stop. The Lead's brief outranks every
default below; it never outranks a stop. Your final message is the report the Lead receives.
Contract, finish sequence and report fields:
${CLAUDE_PLUGIN_ROOT}/references/worker-dispatch.md § Dev-only worker contract, § Finish sequence,
§ Report.

1. Locate the plan

• A path argument is the plan. A key means docs/plans/<KEY>-story.md. No argument: ask the Lead
  or, invoked by a human, the human.
• --phases N,M limits the chunk to those ### Phase blocks; absent means all phases.
• Invoked without a brief, run ${CLAUDE_PLUGIN_ROOT}/skills/story/references/plan-format.md § The refined scan
  and stop on a failed item. A briefed worker skips it; the Lead ran it.
• An unanswered ## Open Fork stops you. When the brief or the plan already carries the answer,
  build on it and name it in the report.
• With a ticket key, read the ticket with its comments (${CLAUDE_PLUGIN_ROOT}/references/ticketing.md).
  A clarifying comment: build the newest statement and name the conflict. A comment that changes
  scope after the plan was approved: stop and hand it back as a question.

2. Worktree and branch

• Dispatched with isolation: "worktree": you are already in your own worktree. Rename its
  worktree-… branch to the brief's branch (git branch -m <type>/<KEY>-<slug>) before the first
  commit. Dispatched with cwd=<path>: work there on the branch it has.
• A brief naming a base branch: git merge <base> first; your chunk builds on it.
• Run the bootstrap the brief names (repo source: .weside/orchestrate.md) and check its result
  before the first gate; a missing venv or database surfaces later as an unrelated error.
• Invoked by a human outside a worktree: EnterWorktree(name="<type>/<KEY>-<slug>"), then the same
  rename and bootstrap.

3. Implement

Read the plan completely, then implement its phases in order, inline. A file listed under both
your phases and a phase outside your --phases scope is a shared seam: name it in the report. Per
phase:

1. Tests per the brief's test discipline; brief silent → test_discipline in .weside/config.json
   (tdd: failing test first at each seam · tests-after, the default: tests after the code, same
   change · off: no new tests unless the plan asks). No implementation-coupled or tautological
   tests; mock only at system boundaries.
2. A new field flows end to end, not only into the model.
3. Regenerate and commit the generated artifacts the brief lists. Leave gate baselines to the Lead.
4. Commit the phase by path: <KEY>: phase <N> — <what> with the Co-Authored-By: trailer of the
   model running you.

A refactor chunk's last step deletes the characterization tests that a test on the new seam now
asserts; the report's retired: field names each one and the test that replaces it.

4. Local gates

Run the commands yourself; they are deterministic. Scope and exceptions are in the contract:
affected tests only, service-bound tests skipped and listed, the named integration suite for a
critical chunk. A gate failure: fix, commit, re-run. The same gate red three times: stop.

5. Finish, when the brief makes you the last writer

Run the finish sequence (the deterministic gates over the branch, code-review at high through the Skill tool only for
money, auth, tenant isolation or a migration, the ordered verification, the deviations from the plan recorded in it, the
affected gates again) and commit after each step. The brief says "not yours": skip it and say so. Invoked by a human: run it.

6. Report

A critical chunk gets its code-review at high (worker-dispatch.md § Dev-only worker contract) unless
step 5 just ran it; any other chunk gets no LLM review. Push only when the brief says Push: yes (git push -u origin <branch> && git ls-remote --heads origin <branch>; an empty answer is a blocker). Then end with the § Report fields as your final
message, one line per AC with its evidence. Stopping early is a report too: what you
completed, why you stopped, blockers: <reason>.