speq-plan git:20261011.5b73638A

Immutable. This exact content never changes and is served at /api/v1/blob/e31cacb77bbfe76b.

---
name: speq-plan
description: Plan a feature or change through a clarifying interview, producing spec deltas, plan.md, and decision-log.md via planner-agent with adversarial review. Use when the user asks to plan, spec, design, or scope a new feature, a change or removal of existing behavior, a refactor, or a fix — before any implementation.
model: sonnet
---

# Spec Planner (Orchestrator)

This skill is a thin orchestrator. It runs the clarifying interview, collects context, and delegates the planning work (spec delta authoring, test mapping, task decomposition) to `planner-agent`.

## Required Skills (for the orchestrator)

Invoke before starting:
- `/speq-cli`: spec discovery

`planner-agent` and `plan-reviewer` invoke their own required skills.

## Workflow

### 0. Load Project Hook (orchestrator)

Check for `.speq/plan-hook.md` in the repo root.
- **Present:** read it. Announce "Loaded project hook: .speq/plan-hook.md". Its content is authoritative: it can add to, change, or override any part of this workflow. If the hook conflicts with this workflow, the hook wins.
- **Absent:** continue without mention.

### 1. Discovery (orchestrator)

Run the speq CLI:

```bash
speq domain list
speq feature list
grep -rhE -A3 '^## ADR:|^\*\*(ID|Status):\*\*' specs/_decision/ 2>/dev/null || echo none   # ADR digest, per /speq-cli
```

Name the features in the list that relate to the request, per `/speq-cli`'s absence check.

Note whether `specs/architecture.md` exists. Pass its path, do not read it.

Collect only enough context for good interview questions, not a full exploration.

### 2. Clarifying Interview (orchestrator)

Apply the Socratic Method via `AskUserQuestion`. Never assume. Decompose the problem space with MECE partitioning:

- **Probe**: surface hidden assumptions with open-ended questions
- **Partition**: present alternative solutions as MECE options
- **Challenge**: test design tradeoffs with counterexamples

Keep each exchange as one line: `- Q: <question> → A: <answer>`, a few words each. This is the only copy of the interview: step 3 writes it into `decision-log.md` and the briefs point to it.

### 3. Plan Name (orchestrator)

**Split check.** If the work spans more than one independent phase, or more than about 4 features, propose one plan per phase in the interview. Plan the first phase now. Each later plan names the earlier plan in its Existing Context.


Pattern: `<verb>-<feature-scope>[-<qualifier>]`

Once the name is set, write `specs/_plans/<plan-name>/decision-log.md` with the H1 `# Decision Log: <plan-name>` and an `## Interview` section holding the step 2 lines. Do not paste the interview into any brief.

| Verb | When |
|------|------|
| `add` | New feature |
| `change` | Modify existing |
| `remove` | Deprecate/delete |
| `refactor` | Restructure, same behavior |
| `fix` | Bug or spec mismatch |

### 4. Delegate to planner-agent

Spawn the planner sub-agent with everything it needs:

```
Delegate to planner-agent — Plan <plan-name>

## Plan Name
<plan-name>

## User Intent
<1-3 sentence summary of what the user wants>

## Clarifying Interview Results
Read `## Interview` in `specs/_plans/<plan-name>/decision-log.md`. Keep it as is. Do not copy it into other files.

## Existing Context
<the exact `speq domain list` / `speq feature list` / `speq feature get` calls you ran, each followed by its output (the ADR digest, not the full `speq decision-log show`; name features you want read, do not paste `speq feature get` output); name `specs/architecture.md` by path when it exists, do not paste it>

## External Research
<any research already conducted, or "none — agent to research as needed">

## Project Hook
<if active: note ".speq/plan-hook.md — read it and apply it" — otherwise omit this section>

## Your Task
Produce spec deltas and plan.md per the `planner-agent` workflow. Tag tasks requiring deep reasoning with [expert] so the implementer orchestrator can route them to implementer-expert-agent.

Return the list of files created and the validation result.
```

### 5. Review planner-agent output (orchestrator)

When the sub-agent returns:

1. Confirm `speq plan validate <plan-name>` passed. Re-run if uncertain.
2. List all created files.
3. If the sub-agent escalated a question, resolve it with the user and respawn with the clarification.

### 6. Adversarial Plan Review (orchestrator)

Spawn `plan-reviewer`, a diabolus advocatus, to challenge the plan before handoff. Maximum 2 rounds total.

**Round 1:**

```
Delegate to plan-reviewer — Review <plan-name> (round 1)

## Plan Name
<plan-name>

## User Intent
<verbatim original request>

## Clarifying Interview Results
Read `## Interview` in `specs/_plans/<plan-name>/decision-log.md`.

## Plan Artifacts
plan.md, decision-log.md, every specs/_plans/<plan-name>/**/spec.md delta, and specs/_plans/<plan-name>/architecture.md when present

## Project Hook
<if active: note ".speq/plan-hook.md — read it and apply it" — otherwise omit this section>
```

It writes its findings to `specs/_plans/<plan-name>/review/round-1.md` and returns only `PLAN REVIEW round 1: BLOCKERS: <n>, ADVISORY: <n>, INTENT: <n>, HUMAN: <n> — <path>`. `INTENT` counts the BLOCKERs on the Intent Fidelity axis alone; `HUMAN` counts every BLOCKER tagged `Escalation: HUMAN` per `/speq-plan-review`'s Escalation Class section — the rest are `MECHANICAL`.

**If `INTENT > 0`:** the reviewer holds that the plan solves a different problem than the one asked. Read the Intent-Fidelity findings from the round file. Present them via `AskUserQuestion` before any revision. The user accepts the plan as-is, or gives guidance and you respawn `planner-agent` manually.

**If `INTENT == 0` and BLOCKER findings exist:**

1. Respawn `planner-agent` with the path to `review/round-1.md`. Instruct it to read the BLOCKER findings from that file, execute each `Fix:` line, log a `[plan-review]`-prefixed `## Review Findings` entry in `decision-log.md` only for a blocker whose fix changed a design decision or settled a HUMAN finding (no entry for MECHANICAL fixes), re-run `speq plan validate`, and return its per-finding `Resolved:`/`Could not resolve:` report per `/speq-planning`'s Revision Mode.
2. **Round 2 runs only if round 1's `HUMAN` count was greater than 0.** A `HUMAN: 0` round 1 means every BLOCKER was `MECHANICAL` — the reviewer already judged none needed adversarial re-checking, only a fix. Skip to the ship decision below without respawning `plan-reviewer`.
3. **`HUMAN > 0`:** compute Plan Size (the plan is `small` when all three hold: the plan-name's verb, per the verb table, is `fix`; `decision-log.md`'s `## Design Decisions` section is empty; the plan directory has no `architecture.md`; otherwise `full`), then respawn `plan-reviewer` for round 2 with the path to `review/round-1.md` plus `Plan Size: small | full`, so it confirms each round-1 BLOCKER is resolved before checking for new ones (or, on `small`, confirms and stops there). Do not run a third adversarial round, even if round 2 raises new BLOCKERs.

**BLOCKERs remaining after round 2, split by `Escalation`:**

- **`MECHANICAL` remainder:** respawn `planner-agent` once more with the still-open `MECHANICAL` findings from `review/round-2.md`; same Revision Mode format. Don't interrupt the user for these unless one comes back `Could not resolve:` or `speq plan validate` fails — then it joins the `HUMAN` remainder below, with the `Could not resolve:` reason as the question.
- **`HUMAN` remainder** (plus any `MECHANICAL` finding a fix pass could not close): read those findings and use `AskUserQuestion`, one line of context each plus a pointer to the round file for detail. The user accepts the risk and proceeds, or gives guidance and you respawn `planner-agent` manually.

**ADVISORY findings** never loop and are never persisted. Read them from the last round file that actually ran and carry them into step 7's report — round 1's, if round 2 was skipped entirely (`HUMAN: 0`) or ran confirm-only (`Plan Size: small`); round 2's otherwise.

### 7. Explain next steps (orchestrator)

Print the plan's summary to the terminal per `/speq-plan-pr`'s `references/pr-body-template.md` and its terminal-print rule, composed from `plan.md` and `decision-log.md`. This step never touches git or GitHub.

Then:
- Print `ADR candidates: none`, or list the title of each `Promotes to ADR: yes` entry in `decision-log.md`. Each one is a proposed ADR, and `/speq-record` accepts it. To drop one, set it to `Promotes to ADR: no` and run this skill again
- Report ADVISORY findings from step 6, read from the round file
- If step 6's round 1 was all-`MECHANICAL` (round 2 skipped) or ran a round-2 `MECHANICAL` follow-up, name it in one line ("N mechanical findings fixed") — do not restate each one
- Tell the user to run `/speq-implement <plan-name>` to continue
- Tell the user to run `/clear` to implement with a fresh context window
- If Claude Code is in "plan mode", call `ExitPlanMode` and ask to proceed with cleared context

## Spec Hierarchy (reference)

```
specs/
├── <domain>/<feature>/spec.md     # Permanent
├── _plans/<plan-name>/            # Active
└── _recorded/<plan-name>/         # Archived
```

## Work Split (reference)

| Step | Performed by | Why |
|------|--------------|-----|
| Discovery, interview, coordination | This skill (pins Sonnet) | Conversational, tool-call heavy |
| Spec delta authoring, plan context and decision log, task decomposition | `planner-agent` sub-agent | Reasoning-heavy; a defect here compounds through implementation |
| Adversarial review, revision loop | `plan-reviewer` sub-agent | Catches intent drift, infeasibility, and ambiguity before implementation |

Each sub-agent pins its own model and effort in its frontmatter, so planning quality does not depend on the parent session's configuration.

## Anti-Patterns

| Pattern | Why Wrong |
|---------|-----------|
| Skipping the clarifying interview | Content comes from user answers, never assumptions |
| Asking the user about a `MECHANICAL` finding | Round count is not the escalation test — `Escalation: HUMAN` is; fix mechanical findings directly, no interruption |
| Running round 2 when round 1's `HUMAN` count is 0 | A round with nothing judgment-worthy left doesn't need a second adversarial pass — fix and ship |
| Persisting ADVISORY findings or looping on them | Report-only; they never gate |