Plan Orchestrator

Plan Orchestrator Skill

Produces concrete planning artifacts from specs and feature requests. Where project-planner advises on strategy ("how should we estimate?"), this skill emits files and structured backlogs ready to drive implementation: ORCHESTRATE-<topic>.md, feature task trees, sprint commitments, and visual roadmaps.

When to Use

Activate when the user wants a deliverable (file, backlog, milestone list), not just guidance:

• "Generate ORCHESTRATE from docs/specs/SPEC-auth-*.md"
• "Break this feature into tasks with estimates"
• "Plan next sprint, 80 hours capacity"
• "Generate a 6-month roadmap as mermaid"
• "Turn this brainstorm into a planning doc"

Do not use for: routing a single task to a command (use task-analyzer), saving session state (use session-state), or general "how do I plan?" advice (use project-planner).

Differentiation Map

Four Planning Modes

This skill consolidates four commands. The user's phrasing selects the mode:

1. Spec → ORCHESTRATE (orchestrate:plan)

Trigger: spec file mentioned, or "implementation plan from SPEC".

Flow:

1. Discover specs in docs/specs/SPEC-*.md (and BRAINSTORM-*.md lacking specs).
2. Parse selected spec: title, phases/increments, acceptance criteria, key files, cross-repo references.
3. Detect cross-repo work (paths under ~/projects/<other>/); enforce same branch name across repos.
4. Confirm plan with user (ORCHESTRATE+worktree / ORCHESTRATE-only / modify / cancel).
5. Generate ORCHESTRATE-<topic>.md with sections: Objective, Phase Overview, Phase 1..N, Friction Prevention, Acceptance Criteria, Commit Strategy, Verification, Session Instructions.
6. Create worktree at ~/.git-worktrees/<project>/feature-<topic> from dev if selected.
7. Update .STATUS and .gitignore.
8. Branch on output:
   • orchestrate-worktree (default) or orchestrate-only → instruct the user to cd into the worktree and start a new session — do NOT begin implementation (STOP-new-session mode).
   • orchestrate-dispatch → skip the STOP instruction; instead run the confirm-before-dispatch gate (below), then dispatch a background Agent call from this same live session. See "orchestrate-dispatch mode" below for the full flow.

Key constraint: ORCHESTRATE file lives in the worktree root, never the main repo.

orchestrate-dispatch mode

A third output value. Same self-containment guarantee as STOP-new-session mode, but execution happens via a background Agent call dispatched from the live planning session instead of a fresh human session opening a new terminal. Use this when the token/attention cost of a cold-started human session outweighs the extra safety mechanisms below (see "When to use which" in the pipeline-orchestrate guide).

GRILL-file precondition (warn-only, not a hard block): if no GRILL-*.md exists for the spec's topic, print an advisory warning and proceed anyway — grilling resolves judgment calls before dispatch, but a well-scoped, low-ambiguity spec doesn't strictly require it. The backstop below (ungrilled-ambiguity handling) covers the case where it turns out an unresolved judgment call was needed after all.

Confirm-before-dispatch gate: after the ORCHESTRATE file and worktree are generated (steps 5-7), before calling Agent, run an AskUserQuestion gate — not a suppressible prompt. Show the generated ORCHESTRATE summary (Phase Overview table) and the worktree path. Options:

• dispatch-now — call Agent immediately with the self-containment prompt (below).
• review-first — pause; let the user inspect/edit the ORCHESTRATE file before dispatching.
• cancel — stop; ORCHESTRATE file and worktree remain, no dispatch.

This gate fires unconditionally. --yes auto-accepts other prompts in this skill but does not suppress this one — it is a design requirement, not a prompt-refiner echo, because dispatch hands off execution to an unsupervised agent and deserves an explicit human checkpoint every time.

Self-containment prompt shape: the dispatched Agent's entire prompt is exactly:

"Working directory: <absolute worktree path>. Read ORCHESTRATE-<topic>.md in full from that directory, then execute it."

The path is a structural requirement, not context — the Agent tool has no cwd parameter (confirmed: ~/projects/dev-tools/CLAUDE.md — "There is no cwd parameter on the Agent tool"), so without it the dispatched agent has no way to locate the worktree or the ORCHESTRATE file itself. Beyond the path, no other context is passed — no summary, no paraphrase, no conversation history, and no confinement instruction (a legitimate cross-repo dispatch, per Cross-Repo Detection above, is expected to touch a second repo's paired worktree — don't contradict that here). This is the same durable artifact a fresh human session would read under STOP-new-session mode; self-containment becomes structural (inherited from the ORCHESTRATE file's own required completeness) rather than a discipline checklist.

Ungrilled-ambiguity backstop: if the dispatched agent hits genuine unresolved ambiguity (whether or not a GRILL file existed), it must leave that phase's checkbox unchecked, add a one-line blocker note directly in the ORCHESTRATE file, and stop — never guess.

Concurrency cap (scoped to orchestrate-dispatch only): soft cap of 2 concurrent orchestrate-dispatch dispatches per session. A 3rd+ concurrent dispatch requires an explicit AskUserQuestion confirmation before calling Agent. Scoping rule (the part most likely to be silently gotten wrong): the counter tracks ONLY background Agent calls made via this orchestrate-dispatch flow — it does NOT count unrelated background Agent calls the session may also have running for other purposes (research agents, doc agents, etc.). Maintain the count as an explicit running tally of dispatches labeled orchestrate-dispatch (e.g. tag each dispatch's description so it's identifiable), not a raw count of "all currently-running background agents." This is not a hard block — a deliberate larger fan-out is still possible, just confirmed.

Failure/hang detection: never trust the Agent tool's completion notification alone — cross- check it against the dispatched ORCHESTRATE file's own checkboxes and Phase Overview status column.

• Notification fires, but re-reading the ORCHESTRATE file shows no checkboxes moved since dispatch → flag as a suspected silent failure. Do not offer merge. Surface this to the user explicitly.
• No completion notification within the hang-detection window (below) → surface the crash/hang case explicitly. Do not wait silently past the window.

Hang-detection window formula: 2 × the dispatched ORCHESTRATE file's own stated phase-effort estimate — not a fixed wall-clock constant. The effort estimate comes from the Phase Overview table's Effort column, which Mode 1's existing template already requires for every phase. Example: a phase overview row stating Effort: Med for the currently-dispatched phase sets the flag-as-hung threshold at 2× whatever wall-clock duration this session maps Med effort to for this project — reuse the same effort→duration mapping the session already uses elsewhere, don't invent a new one here.

Confirmed-failure disposition: on a confirmed silent-failure or hang (per the detection rule above), never auto-delete the worktree or branch — leave both in place for inspection. Add a .STATUS note in the same HELD style already used for prior held work in this repo (factual: what was dispatched, what was observed, that it's paused pending manual inspection — not narrative guesswork about the cause).

Resumability: re-dispatching against the SAME ORCHESTRATE file must be idempotent. Before dispatch, read the file's Phase Overview status column and per-phase checkboxes (the same tracking this mode already requires agents to maintain). If any phases are already checked, the self-containment prompt is identical between the initial and a resumed dispatch (same worktree, so the same path) — resumption is driven by the file's own content (an agent reading a file with Phase 1 fully checked off naturally continues from the first unchecked phase), not by a modified or parameterized prompt. Never restart from Phase 1 on a resume; never invent a second tracking mechanism alongside the checkboxes.

.STATUS auto-write scoping: at dispatch time, auto-write only the factual fields to the Active Worktrees entry — branch name, worktree path, and PR link once one is opened. All of that data already exists at dispatch time; no new logic is needed to produce it. The narrative/purpose prose column stays manual — do not attempt to auto-generate it (auto-generated prose reads worse than hand-written, per this skill's own .STATUS history).

Model selection for the dispatch (distinct from workflow-engine's per-stage routing): default to omitting model — the dispatched agent inherits the session's default tier, correct for a fully-specified, mechanical ORCHESTRATE file. This is a whole-dispatch choice, not the same knob as skills/orchestration/workflow-engine/SKILL.md's "Model Routing (Haiku for cheap stages)" section, which governs individual reasoning stages inside a multi-stage run and says to omit model for architecture/cross-file-reasoning stages specifically — that rule is about per-stage cost tuning within an already-running pipeline, not about this one-shot dispatch prompt. Pin model: "opus" here only when the ORCHESTRATE file's own Phase notes flag an unresolved design judgment call for the dispatch as a whole (e.g. a Phase scoped as "pick the lower-friction option, document the choice") — not as a blanket default for orchestrate-dispatch.

2. Feature Plan (plan:feature)

Trigger: "plan feature", "break down", "scope this feature".

Output: user stories → task tree grouped by layer (backend/frontend/infra) → time estimates → dependencies → risks → acceptance criteria → total estimate.

Options: --scope mvp|full|enterprise, --format markdown|jira|github, --include-tests, --output <file>.

3. Sprint Plan (plan:sprint)

Trigger: "plan sprint", "next sprint", capacity/duration mentions.

Flow: gather backlog → prioritize → fit to capacity → name sprint goal → commit + stretch items → highlight dependencies and risks.

Options: --duration <days> (default 14), --capacity <hours>, --goal <text>, --from github|jira.

4. Roadmap (plan:roadmap)

Trigger: "roadmap", "quarterly plan", "milestone view".

Output: phased milestones with progress bars, dependency callouts, timeline gantt-style strip.

Options: --horizon <months> (default 6), --format mermaid|markdown|html, --output <file>, --update to refresh existing roadmap.

ORCHESTRATE Template (Canonical)

# <Topic> — Orchestration Plan

> **Branch:** `feature/<topic>`
> **Base:** `dev`
> **Worktree:** `~/.git-worktrees/<project>/feature-<topic>`
> **Spec:** `<spec-path>`

## Objective
<1-2 sentences from spec>

## Phase Overview
| Phase | Increment | Priority | Effort | Status |

## Phase 1..N: <Name>
**Scope:** ...
- [ ] 1.1 <task>
**Key files:** `<file>` (NEW or update)

## Friction Prevention
- Context first, verify CWD, no autonomous starts, test per phase.

## Acceptance Criteria
- [ ] <from spec>

## Commit Strategy
Conventional commits per phase.

## Verification
<auto-detected test command>

## Session Instructions
cd <worktree-path> && claude
> "Read ORCHESTRATE-<topic>.md and start Phase 1."

(Desktop app, no persistent shell: `EnterWorktree({ path: "<worktree-path>" })` switches the
session's cwd directly instead of opening a new terminal — the worktree already appears in
`git worktree list` since this skill created it.)

Auto-Detection

See ../references/verify-gate-detection.md for the full detection table and how to use it.

Rebase Strategy (Spec → ORCHESTRATE mode)

Before finalizing the worktree, check whether the base branch (dev) has recent automated commits:

• If the last 5 commits on dev include chore: or bot commits → suggest a rebase before PR.
• If the history is clean → standard merge workflow, no rebase suggestion needed.

Cross-Repo Detection

Scan spec for paths under ~/projects/dev-tools/<other>/. If detected:

• Primary repo + secondary repo(s) listed.
• Same branch name (feature/<topic>) enforced across all worktrees.
• Generate paired worktrees on confirmation.

ADHD-Friendly Outputs

• Visual hierarchy: headers, tables, progress bars, checkboxes.
• Quick-wins vs long-term clearly labeled.
• Numbered next steps at end of every artifact.
• Total estimate always surfaced (no "TBD" without a placeholder).

Integration

Test-plan scaffolding (default-on)

When this skill emits an ORCHESTRATE artifact, it also emits a test-plan scaffold by default. Pass --no-tests to suppress the section.

Tier-inference rule

Infer tiers from the shape of the artifact/change being planned:

Unselected tiers print as N/A — <reason> (never empty stubs).

Emission rules

• Emit test stubs red-first (failing placeholder, not passing no-op).
• Each stub carries # TODO(author): delete if not contract-bearing until the author confirms the contract.
• Scaffold templates live in ../../workflow/brainstorm-insights/references/scaffold-templates.md; point to that file — do not duplicate templates inline.

--yes non-suppression

--yes auto-accepts prompts only; the test-plan and Documentation sections are CONTENT and are still emitted under --yes. Only --no-tests/--no-docs remove them.

Opt-out

--no-tests suppresses the entire test-plan section. Default is on.

Documentation scaffolding (default-on)

When this skill emits an ORCHESTRATE artifact, it also emits a Documentation section by default. Pass --no-docs to suppress the section.

Which docs to emit

Derive which documentation artifacts are needed by running the existing doc-scorer rubric from skills/orchestration/references/doc-impact-rubric.md (tiered thresholds — see that file for the current type list, per-type weights, and threshold rule). Do not invent a new rubric, and do not hardcode a parallel type list here — always defer to that file as the single source of truth for which types exist and what their thresholds are.

For each doc type the scorer evaluates, pre-check ([x]) boxes that meet that type's threshold, and mark the rest N/A — score <N>. The template for the Documentation section — including the Site Consistency checklist — lives in ../../workflow/brainstorm-insights/references/scaffold-templates.md.

Lifecycle split

Count-cascade exclusion

Auto-docs emission touches only semantic docs — CHANGELOG [Unreleased] ×2 mirror, guide/refcard/tutorial prose. It never touches version or count lines. Version/count updates stay in bump-version.sh.

Opt-out

--no-docs suppresses the entire Documentation section. Default is on.

Common Pitfalls

• Writing ORCHESTRATE to the main repo — always worktree root.
• Auto-starting implementation after generating ORCHESTRATE — STOP; new session required.
• Generating spec from brainstorm silently — offer /brainstorm s instead; let user capture spec interactively.
• Overwriting existing ORCHESTRATE without confirmation — detect and ask before regenerating.
• Cross-repo branch name drift — enforce identical feature/<topic> across paired worktrees.

Example Invocations

"Generate ORCHESTRATE from docs/specs/SPEC-auth-2026-02-15.md"
"Break down 'avatar upload' as MVP-scope feature plan"
"Plan a 7-day sprint, 40 hours capacity, goal: ship OAuth"
"Generate quarterly roadmap as mermaid, save to ROADMAP.md"