| name | implement |
| description | Orchestrate workflow phases. Выполни план, приступай к работе, приступай к реализации, реализуй, начинай, делай, работай, execute plan, start working, begin implementation. |
Workflow Orchestrator
Execute the .claude/rules/workflow.md phases in order. Phase definitions, the sizing table
and the skip rules live there — do not duplicate them, refer to them.
1. Locate the plan
- The plan is the one the user named. Approved plans live in
plans/; a ## Progress
section, if present, is the resume point.
- No plan named → fall back to the newest (
ls -t plans/*.md | head -1) and say which one
you picked and that none was named — a guessed plan is not an approved one.
- Display:
━━━ PLAN LOADED ━━━ + filename + first heading.
- If no plan exists, ask the user to describe the task or enter plan mode first. HALT.
- Shape check — run it, do not eyeball it:
python3 .claude/scripts/plan-shape-gate.py <plan path>. Non-zero ⇒ print the output and
HALT: offer to rewrite the plan to shape (rewrite, never append), or to proceed once
the user says so. This is the gate's hard point — a plan is usually still uncommitted when
its implementation commits, so execution is the last moment where reshaping is free.
2. Phase 0 — Scope
Determine S/M/L per the sizing table (commits × entanglement — never file count). Run the
discovery gate: SYSTEMS.md → grep -rn "SYSTEM:" src/ → read the entry files. Output the
━━━ SCOPE ━━━ block. HALT for confirmation.
3. Phases 1–2 (L only)
- Phase 1: output the filled checklist from
.claude/rules-scoped/integration.md.
- Phase 2: failing tests before implementation, per the TDD cycle (
/tdd).
For S and M output "Phases 1–2: skipped — <S|M>, single commit" and proceed. S and M still
write tests in Phase 3; what L adds is the red run as a gate.
4. Phase 3 — Implementation
Output ━━━ PHASE 3: IMPLEMENTATION ━━━. Work the plan step by step:
- Tests after each step; breadth from the entanglement table, commands in
.claude/rules-scoped/testing-ops.md. Rebuild the container — the image copies source.
- Add
ARCH: / INVARIANT: / WHY: markers per documentation.md; regenerate SYSTEMS.md
if a SYSTEM: marker changed.
- Ambiguous step → rule on it and keep going, logging
Ruling: <what you decided> — <why> — <what it costs if wrong> in the run output (and in the
plan's ## Progress if steps remain). Halt only for the four stop conditions in
workflow.md Phase 3. Why: a wrong ruling costs rework the user can see and undo; a session
parked on a question costs a turn every time, and the answer is usually "yes, do it".
- After the final step: drive the affected flow live against
:8777 (streaming and
non-streaming as applicable, plus the refused principal for any access change) and record
the output verbatim — reality-facing acceptance per workflow.md Phase 4. A
tests/docs/config-only diff states the exemption instead.
5. Phase 4 — Review
M or L, or any review-gate trigger (hot path, new module, user said "done"/"push") → invoke
/review (mandatory, do not ask). An S with no trigger → output the skip reason. HALT
for approval. Do NOT gate on file count.
6. Phase 5 — Commit & Retro
Run .claude/scripts/pre-commit-gates.sh unpiped. Output ━━━ PHASE 5: COMPLETE ━━━.
HALT — wait for an explicit commit request. If steps remain, update the plan's
## Progress so the next session can resume from it.
