Orchestrator Lanes

Orchestrator lanes — solo operator

Inside a Lane Pilot chat (LANE_PILOT_AGENT_TYPE set / tools lane_pilot_* present) do not follow this file: it is the terminal harness. Dispatch with lane_pilot_dispatch_writer (the whole plan at once, order through depends_on), poll with lane_pilot_wait_writer; Lane Pilot runs plan critique, acceptance and the merge into main itself, so run-init, run-controller, run-supervisor, lane-ctl, wt-create and wt-merge-main are not yours to call (the guard refuses them). Only references/decomposition.md still applies there.

Info

If $ARGUMENTS is exactly info, print references/info.md verbatim (Russian), then stop. Do not score. Do not run-init.

Load: karpathy-guidelines, lane-contract, project-life, resume-project, project-design, ui-ux-pro-max. Before filling tasks/*.yaml: Read this skill's references/task-yaml.md.

Docs: FILE-CONTRACT.md, ROUTING.md, SOLO-ORCHESTRATION.md, PLATFORM-CAPABILITIES.md (Claude Code + Codex features we use), docs/decisions/ADR-codex-effort.md under the lane-stack / ~/.agents/docs/.

You are the only person who merges to main. Human never merges.

Planning vs run

If the user is in a planning session (project-life: «планируем / не запускай / спланируй / обсудим / пока план») or has not said «делай / реализуй / в работу / запускай ран»:

• Do not announce score, run-init, or dispatch writers.
• Follow project-life → Planning session. Write .agents/plans/ only.
• New app or service talk («архитектор / новое приложение / новый сервис»): load app-architect. Same draft plan, living artifacts/. Still no run.
• Claude Code Plan mode is not this. Do not write ~/.claude/plans/. If the host is in Plan mode, tell the user to switch to default.

Phase 0 starts only after «делай / реализуй / в работу».

UI / visual initiative: root docs/DESIGN.md plus the full apps/<name>/docs/DESIGN.md for every UI app you touch. If any of those are missing, spawn design-lead before run-init. UI task read_first lists both root and that app's DESIGN.md. Do not put them in writer owns_paths unless the outcome is tokens/docs.

Phase 0 — Score (announce once)

+2 multi-problem · +2 UI/state/auth/pay · +2 backend/API · +2 multi-surface · +2 needs verify · +3 prod/billing/security

Writer source of truth = adoc / agents-doctor → .agents/routing.profile.yaml (lanes.main_write, writer.model, writer.reasoning_effort, workspace.mode).

• Every task YAML must set lane: <main_write> exactly (e.g. lane: codex).
• Never hardcode lane: kimi unless adoc says so. run-validate rejects mismatch.
• run-controller defaults --provider / model / effort from the same profile.
• Selectable writers: kimi / qwen / grok / agy / codex. Codex Sol remains recovery + night review / onboard / docs.
• Workspace (adoc tab Work): in_place | worktree | auto — see Phase 2.

Task decomposition

Rules (one outcome per task, disjoint owns_paths, real depends_on only, size budgets) live in references/decomposition.md. Read it before run-init and before filling YAML; skipping it is how multi-task runs stall.

Phase 1 — Files

Not .agents/plans/ for coding execution. Strategy stays in .agents/plans/; promote to a run when implementing.

run-init "$(pwd)" <slug> --score <score>
# Fill PLAN.md, SPEC.md (required content when score≥7 or ≥2 tasks), tasks/*.yaml
# YAML is technical: objective=outcome, read_first=existing files,
# context_selectors=line windows, interfaces=signatures or [].
# Do not copy writer recovery (git checkout, edit tool, CONTINUATION) into the task.
plan-critique --run-dir "$(pwd)/.agents/runs/<slug>"   # stages.plan_critique (adoc)
# Every error/warn id → artifacts/critique-reply.json (take | skip+note)
run-validate --run-dir "$(pwd)/.agents/runs/<slug>" --phase pre-dispatch
run-board "$(pwd)"

Plan critique (configure in adoc → Stages):

1. Structural coverage + independent LLM review (agy uses LanePlanCritique JSON schema).
2. Writes artifacts/critique.json + critique.md. Findings with error/warn are inbox.
3. PM must reply before run-controller / writers (advisory and gate):

{"schema_version":1,"items":[{"id":"structural:verify_heavy:003","verdict":"skip","note":"L1 is enough"}]}

Bulk skip: plan-critique --ack --note 'reason'.
run-validate --phase pre-dispatch and run-controller start refuse unreplied ids.
wiki/ TODO/ .agents/plans/ (and legacy docs/plans/) owns_gap is info — no reply.

PLAN.md

DAG table, goals, out-of-scope, verification plan (L1 vs L2), risk notes.

SPEC.md (professional, not a stub)

Required when score ≥ 7 or ≥ 2 tasks. Must include, in English:

1. Goal — one paragraph
2. Interfaces — stable exports/routes/types the writer must honor
3. Invariants — what must not break
4. Out of scope — explicit non-goals
5. Definition of done — observable, testable

Reject template-only text (“Record interfaces, invariants…”). run-validate enforces this.

Task YAML

Before writing tasks/*.yaml, Read references/task-yaml.md in this skill. That file is the field map + good/bad examples. lane-contract is the schema/gate sibling; do not skip the reference and invent a novel in interfaces.

Where the TZ lives:

Immutable after first start. verification[] = L1 focused only (see below). run-validate --phase pre-dispatch rejects the bad column.

Repeated correction rule: if you retype the same path/command fix twice in a project, persist it as a rule first (lane-memory lesson "<rule>" --for pm|writer|both, kept on the hub), then regenerate the plan.

Phase 2 — Isolation (workspace from adoc)

This skill is Claude Code dev-orchestrator only. Codex / Grok / Kimi / Qwen / AGY must not load it and must not wt-create / git worktree add.

Read .agents/routing.profile.yaml → workspace.mode (default auto if missing). in_place wins over score, risk, and multi-write. Never override adoc In-place.

Also: high-risk write → prefer worktree even under auto; avoid parallel writers on overlapping blast radius.

Phase 3 — Dispatch (durable, bounded)

run-controller start → one run-supervisor watches
provider slots (default 5) release ready DAG tasks
complete → owns → L1 verify → accept (progressive)
retry once only if `recovery.retry_ok`; eligible 2nd failure → Codex Sol
trusted STATUS:partial / `provider_partial` → block + replace task (not retry, not Sol)
task blocked → siblings continue; dependents of blocked upstream cascade-blocked

Never one Claude subagent per writer. Writers = durable processes (kimi/…).
Never PM run-controller start|watch|status — only Agent(run-supervisor).
Never PM nohup/async ad-hoc monitors.

Claude agent names are roles, not brands. Daytime coder = process from adoc. Roster: agents/claude/README.md.

Verification tiers (L0 / L1 / L2)

L1 rules (PM when authoring YAML)

• Prefer: npm run test:unit -- path/to/spec --silent, single-package typecheck, path-scoped vitest/jest.
• Do not put bare monorepo npm run build / root npm test on every task when the run has ≥2 tasks — that is L2.
• Multi-task + full-package build in L1 → run-validate warns or rejects (score≥7).
• Acceptance = behavior, not “entire monorepo green”.

L1 paths under worktree (temples-admin class bugs)

verification[].cwd is almost always project_cwd (the worktree). Relative script args resolve there, not in the main checkout.

Canonical patterns:

1. Product test under owns (best): tests/test_foo.py +
   python3 -m unittest discover -s tests -p test_foo.py with cwd: project_cwd.
2. Worktree-local pre-authored check:
   worktree/.agents/runs/<slug>/artifacts/<id>/check.py exists on disk before
   pre-dispatch validate; command uses that relative path.
   PM may Write only basename check.py under
   .agents/runs/<slug>/[artifacts/<id>/]check.py (or under .worktrees/... same shape) — not helper.py, not state.json / reports.
3. in_place (adoc Work → In-place): one tree — main .agents/runs/... paths work.

run-validate --phase pre-dispatch rejects missing script files under verification cwd. Fix paths before first lane start (YAML is sha-pinned).

L2

After all accepted: run-validate --phase pre-merge, then one build/test pass, then merge.

Phase 4 — Accept (progressive)

When A verifies while B runs: accept A now. Done only with acceptance.json.
No daytime LLM review; medium/high → nightly tier.