start-task

The loop for one task. Do the steps in order and do not write code before step 9. The routing table and plan rules live in reference.md; step 6 says when to read it (never Read plugin files directly).

Config: .claude/nauraiz-dev.json, with .claude/nauraiz-dev.local.json over it. Keys: branch.base, branch.pattern, plan.dir, plan.template, models.*, sharedPaths, gate. No config: base main, plans in .work/tasks, models small sonnet, normal and large opus, explore haiku.

Rules for every step:

• The owner decides. Never write that the owner approved, agreed or confirmed anything they did not say in this session.
• "Stop" and "ask" mean: end your turn with the question and wait. Do not continue on a guess.
• Never push, open a PR or merge unless the owner asked for it in this session.

Steps

1. Identify. Issue number from the request, else from the branch name (git branch --show-current, branch.pattern; feature/12-x is issue 12). Then make ONE call for steps 1 to 3: node "${CLAUDE_PLUGIN_ROOT}/scripts/task/preflight.mjs" <n> (add --branch <name> when the request names a branch, and --branch <current branch> when the issue number came from the current branch name). It prints JSON: issue (title, body, labels, state, comments; null if gh issue view failed), collision, base. Steps 2 and 3 read it; make no other gh or git call for them. No issue number: the request is the spec; run it without one and use only base. A number with issue null: say so and work from the request.
2. Collision check. With an issue number, even if gh issue view failed, read collision.verdict. If preflight cannot run, run node "${CLAUDE_PLUGIN_ROOT}/scripts/task/collision.mjs" <n> with the same flags instead; if that prints no verdict either, treat it as unknown.
   • taken: stop. Tell the owner who and where: the reason, otherAssignees, each openPRs number, author and why (linked, mention, url, branch or head), and each remoteBranches and remoteBranch name and authors. Nothing more until the owner says go on.
   • unknown: ask the owner whether to go on, quoting the reason. Wait.
   • ours or free: say the verdict in one line and go on.
3. Base check. Read base. behind null (fetched false, or the count failed): say its reason and go on. behind above 0: say "<n> commits behind origin/<base>" and, if ahead is 0 (no commits of its own), add "recreate it on origin/<base>?". Go on; recreate only on a yes.
4. Triage. Size the task and say the size and the model models.<size> in one line; if this session runs another model, say so. Small: one or two files, an obvious change (typo, copy, rename, one-line bug with a known cause). Normal: one area, about 3 to 10 files, a clear design. Large: across areas, a new schema, API, shared seam or dependency, more than ~10 files, or an open design. Torn: pick the larger. Small: make the change, verify it (the relevant test or gate command), report, done. Skip steps 5 to 8; never read reference.md.
5. Ask when it matters. Run the grill skill. It may end the turn to wait for answers, then hands back here.
6. Route the process. Run node "${CLAUDE_PLUGIN_ROOT}/scripts/show.mjs" skills/start-task/reference.md once (routing table and plan rules, for steps 6 and 8). Pick one process per phase from the routing table and say the picks in one line, e.g. "Process: grill → superpowers:writing-plans → superpowers:test-driven-development → gate". Invoke each with the Skill tool when its phase starts; never follow one from memory. Large: the Design phase runs the skill picked here. When superpowers:brainstorming offers to start superpowers:writing-plans, that is the Plan phase: use the Plan pick instead.
7. Explore cheaply. If understanding the task needs more than ~5 files, or any file over ~800 lines, dispatch the Explore subagent with its model set to models.explore (default haiku). Ask it for conclusions with file paths and line ranges, then read only those ranges yourself.
8. Plan. Write the plan by the plan rules in reference.md (read at step 6): branch: frontmatter, ### Allowed listing every path you will edit, ## Decisions (the owner's answers) and "Decisions to confirm". Any *_approval field is pending owner confirmation unless the owner said yes in this session; then quote the date and the owner's words. If any ### Allowed entry matches sharedPaths and was not asked about, run grill again before building.
9. Build, gate, report. Build with the routed skill. Then run the gate skill, or node "${CLAUDE_PLUGIN_ROOT}/scripts/gate/run.mjs" --plan <plan path> (exit 0 all pass, 1 a failure, 2 no gate configured). Run it once, from any folder in the project; it prints every command's exit code itself, so never append echo $?, pipes or cd. Re-run only after you changed code, or once with --plan <path> when a row says no plan for this branch. Read the table; never call anything green that failed or did not run. Report: what was built, the gate table, decisions to confirm, risks, anything unverified. Offer ship only if the owner authorised a PR this session.