dimitritholen/1337-claude/implement

Work the open tasks of the plan one at a time, in dependency order: tasqx_list_tasks under the project (tasqx_list_projects) if the plan lives there, else the plans/ file.

$ARGUMENTS picks the scope. With --one (or the bare word one or single), run single-ticket mode: take the first open, unblocked task exactly as the loop would, run the full cycle below for it, retry and diagnose rules included, then stop without starting the next task. End the reply with exactly one line:

Ticket <id> done. Next: <id> <title> — run /clear, then /1337:implement --one.

When no open task remains, say the plan is done instead. This keeps the orchestrator's context from growing across tickets, and the next run resumes from tasqx or the plan file, so nothing needs to survive the /clear. Anything else in $ARGUMENTS, or nothing, runs the loop over every open task.

For each task:

1. tasqx_start_timer and annotate the approach.
2. Pick the tier and dispatch:

• Tiered mode: run skills/tier/route.py once for the step, exactly as

hooks/tiered.md shows: ``bash python3 "${CLAUDE_PLUGIN_ROOT}/skills/tier/route.py" <<'EOF' {"task": "<the task in one or two lines>", "steps": [{"id": 1, "title": "<step>", "brief": "<files, the change>"}]} EOF ` then dispatch 1337:builder` at the tier it printed.

• Otherwise, pick the tier yourself per hooks/orchestrator.md's ladder

(haiku/sonnet/opus). The builder brief asks for /1337:tdd red-green-refactor. When the task comes from a spec with an Expectations block, the brief includes the task's expectation rows verbatim. Each non-judgement row becomes a test, red first, at the row's seam: metric rows assert measured values (computed style, contrast ratio), visual rows use the project's screenshot comparison if it has one, else become metric rows.

1. Verify with 1337:checker, which runs those tests.
2. QA: not per task. Once per pull request, after the PR's last task has

checker PASS and before the PR is opened or merged, only when the PR changes something user-facing (UI, interaction, CLI/TUI output, user-visible copy). Dispatch 1337:qa (opus) in mode smoke with the specs or acceptance criteria of the PR's tasks and the branch's diff against the base. Never pass the builder's report or reasoning; qa stays independent. The judgement rows of those tasks' expectations go into qa's brief (every other row is already a test). Internal-only PR: skip it and say so in the annotation.

1. git diff, then /1337:review in spec mode against the task's spec.
2. Annotate what was delivered.
3. Commit the task.

On checker FAIL: retry once, one tier up from the routed/chosen tier, never twice, never for a step already at Opus; if it fails again, stop and hand off to /1337:diagnose.

1337:qa reports PASS, FAIL or SKIP on its first line. It needs the project contract docs/agents/qa.md (format in skills/exploratory-test/references/project-contract.md); without it qa returns SKIP, which the task annotation records as a skip, never as a pass. On FAIL, send the confirmed blocker and major findings, verbatim from qa's report, to 1337:builder under the usual tier rules, re-run 1337:checker, then 1337:qa again in smoke mode on the fixed surfaces. At most two qa rounds; still failing, stop and hand the findings to the user. Only a confirmed blocker or major blocks the merge. Minor and polish findings become tasks in the tracker (or are fixed when the project's rules say so). Product questions go to the user, not the builder. A smoke run takes roughly 10-15 minutes, hence once per PR, never per task or builder step. The deeper full run is a nightly job, not part of this flow: see docs/qa-nightly.md.

A task is not done while one of its expectation rows has no passing test.

Never tasqx_complete_task without the user's acceptance.

One routing call licenses exactly as many builder dispatches as it routed steps — route each task on its own rather than batching several under one call. Review the diff before moving to the next task; nothing after a dispatch skips straight to another dispatch or a commit.

Outside orchestrator mode, run this same loop yourself in the main session instead of dispatching. Tiered and orchestrator mode are separate switches (hooks/lib/mode.sh), so the route.py step in 2 still applies whenever tiered mode is on; without it, pick the tier yourself.