Quick Plan

Quick Plan

You produce a single, unsliced, consistently-phased implementation plan for a small, well-understood task and write it directly to .rpiv/artifacts/plans/ with status: ready. One non-interactive pass — no slice decomposition, no skeleton-then-fill, no in-skill review. The workflow's grade and validate stages (or your own review, standalone) are the gates; this skill only emits the artifact those gates parse.

The expected shape is one phase (the small-task default). Add a phase only when the task genuinely splits into independently-verifiable units that share no file. This is the trimmed mimic of blueprint's artifact shape minus every step that exists to produce confidence in the artifact.

Input

$ARGUMENTS — one of four shapes (the skill is sound however the caller wires it):

• Flags (workflow dispatch) — --research <path> --goal <path> [--acceptance <path>]. Read ALL files FULLY (no limit/offset). The research doc is the grounding; the goal file is the verbatim brief — every ask it names is either implemented by a phase or deferred under ## Out of Scope. The acceptance file, when given, is the goal-derived inventory (items: frontmatter, ids a1…): the plan MUST record a per-item acceptance: disposition — implemented naming the phase, or deferred with a reason — every item, no omissions (the completeness gate anchors on the inventory, and validate executes each item's evidence command against the finished tree).
• Research artifact path — a .md path under .rpiv/artifacts/research/. Read it FULLY (no limit/offset); it is the grounding for the plan.
• Injected research — when dispatched by a workflow reads: ["research"] stage, the research doc is already in context; treat $ARGUMENTS as the task description / goal.
• Free-text (standalone small task) — $ARGUMENTS is the task description; research is done or the shape is obvious.

If $ARGUMENTS is empty AND no research doc is in context, print an error and stop — there is nothing to plan from.

Metadata

node "${SKILL_DIR}/../_shared/now.mjs"
echo
node "${SKILL_DIR}/../_shared/git-context.mjs"

Copy values verbatim. <iso> is the first tab-separated field (use as date:); <slug> is the second.

Flow

1. Input → 2. Optional single targeted research → 3. Write plan (status: ready)

Steps

Step 1: Read the input

• If a research artifact path was given, read it FULLY. Extract the task, the Code References (file:line), the Integration Points, and any Open Questions already resolved.
• If a --goal file was given, read it FULLY and enumerate every ask it names; each is either implemented by a phase or deferred under ## Out of Scope with a reason (when the research narrowed the brief, its rationale is the deferral reason).
• If an --acceptance file was given, read it FULLY and treat its items: as the enumeration of record — dispose of every item id (Step 3's acceptance: mapping). An item whose evidence command the plan's work would not make pass is either genuinely covered by a phase or deferred with a reason; never quietly reinterpret an item's statement.
• Read the key source files the research points at — the anchors the plan will cite — so the phase's code blocks and file:line references are real and current.
• Determine the goal (the --goal brief, the task description, or the research's stated goal).

Step 2: Optional single targeted research (at most ONE dispatch)

You may dispatch at most one codebase-pattern-finder agent to find the implementation pattern to model the code after — only if the research did not already surface one. Never parallel, never run_in_background: a background completion cannot re-drive this session, so the skill would end its turn before writing the plan and the stage would fail with no artifact.

Spawn the agent using the Agent tool:

• subagent_type: codebase-pattern-finder
• Prompt: "Find the implementation pattern I should model after for {task}. Single pass."

If the research already gave you the pattern (Code References + Integration Points), skip this step entirely — most small, well-understood tasks need no dispatch.

Step 3: Write the plan (status: ready)

Write the plan in one pass to .rpiv/artifacts/plans/<slug>_<description>.md (<slug> from the Metadata block; <description> is a brief kebab-case summary). Set frontmatter status: ready directly — there is no in-progress/in-review transition. Do not author a risks: frontmatter array (see Important Notes).

The artifact MUST be consistently phased: frontmatter phases: array length == number of ## Phase N: headings == scalar phase_count. Any drift and planPhaseRecords halts the workflow at plan time. Single phase is the default.

Use this template (single-phase shown; replicate the ## Phase N: block per phase only when genuinely multi-phase):

---
date: {<iso> from Metadata block}
author: {`author:` from Metadata block}
commit: {Current commit hash}
branch: {Current branch name}
repository: {`repo:` from Metadata block}
topic: "{task name}"
tags: [plan]
status: ready
phase_count: 1
phases:
  - { n: 1, title: {Phase 1 title}, files: [{every repo-root-relative path Phase 1 creates/edits}], depends_on: [] }
acceptance:
  - { id: a1, disposition: implemented, phase: 1 }
  - { id: a2, disposition: deferred, reason: "{one-line reason}" }
last_updated: {same <iso> as date:}
---

# {Task Name} Implementation Plan

## Overview

{1-3 sentences: what this plan implements and why.}

## Out of Scope

{One line per goal ask this plan does not implement — "- {ask} — deferred: {reason}". Omit the section when every named ask is covered.}

## Phase 1: {Descriptive Title}

### Overview
{One sentence: what this phase delivers.}

### Changes Required:

#### 1. {repo-root-relative/path/to/file.ext}
**File**: `repo-root-relative/path/to/file.ext`
**Changes**: {NEW | MODIFY} — {one-line summary}

```{language}
// Copy-pasteable code — the full function/section to add, or the exact edit,
// grounded in the current tree. implement applies this without guessing.

Success Criteria:

Automated Verification:

• {One self-contained command that exits 0 when the criterion holds, write-scoped to THIS phase's files: set — e.g. npm run check:files -- <this phase's paths>}
• {Read-only whole-tree check — e.g. npx tsc --noEmit -p tsconfig.base.json}
• {Phase-scoped test — e.g. npx vitest run <this phase's test path>}

Manual Verification:

• {Human check — UI, real-conditions behavior, edge case}

Populate each `phases[].files:` from that phase's `#### N.` / `**File**:` paths — every repo-root-relative path the phase creates or edits — or the plan-time coverage floor (`planCitationCheck`) flags a gap. Make every `#### Automated Verification:` command write-scoped to its own phase's `files:` set; phases may run concurrently under implement, so an unscoped command that rewrites the wider tree corrupts a sibling phase's edit.

Populate `## Out of Scope` from the Step 1 goal-ask enumeration: one one-line deferral with a reason per goal ask no phase implements.

Populate `acceptance:` ONLY when an `--acceptance` file was given (omit the array entirely otherwise): one entry per inventory item, in item order — `disposition: implemented` with the covering `phase:` number, or `disposition: deferred` with a one-line `reason:` (and a matching `## Out of Scope` line). Every item id from the inventory appears exactly once; never invent an id the inventory doesn't carry.

Then print the path and a one-line summary: `quick-plan written: {N} phase(s), {M} files`.

## Important Notes

- **Consistently phased, always.** `phases:` length == `## Phase N:` heading count == scalar `phase_count`. A half-phased plan throws at plan time. Single phase is the small-task default — only split when units are independently verifiable and share no file.
- **No `risks:` frontmatter — this is the correctness lightening.** The `grade` skill's `correctness` risk-flag adjudication (the heaviest correctness sub-check, with its mechanics-evidence and verify-at-implement duties) fires **only** when the artifact carries `risks:`. With none declared, `risk_rulings` is omitted and adjudication auto-skips — the grade runs the cheap spot-check path. Do NOT invent a `"correctness-simplified"` dimension (grade does not recognize it and would deadlock the gate); the lightening is realized by the plan's shape, not by a new dimension.
- **One Write, `status: ready` directly.** No skeleton-then-fill, no progressive `Edit`, no 3-state status machine. The artifact is gated externally (the workflow's `grade` + `validate`, or your own review standalone).
- **At most ONE `codebase-pattern-finder` dispatch.** Never parallel, never `run_in_background` — a background completion cannot re-drive this session and the stage fails with no artifact. Skip the dispatch when research already surfaced the pattern.
- **Non-interactive.** No `ask_user_question` checkpoint. A checkpoint without blueprint's dimension sweep is pure latency on a fast-path preset whose output is immediately grade-gated. Resolve ambiguity from the research and the real code; if a genuine fork can't be settled, make the most defensible call and let the grade panel catch it.
- **Defer explicitly, never silently.** Every goal ask no phase implements gets a one-line `## Out of Scope` deferral with a reason — the completeness gate blocks on a named ask that is neither addressed nor deferred.
- **Dispose of every acceptance item.** With `--acceptance` given, the `acceptance:` mapping is the machine-readable form of that same rule: every inventory id, exactly once, `implemented` (with phase) or `deferred` (with reason). The completeness gate reads the inventory directly, so a missing or mislabeled disposition is a blocking gap — and validate will run each item's evidence command regardless of what the plan claims.
- **Ground every citation.** Every `file:line` in prose or code comments uses a repo-root-relative path and is verifiable at the current revision — the plan passes the deterministic `plan-cite-check` floor.
- **NEVER edit source files.** This skill produces a plan document, not implementation. Source editing is `implement`'s job.
- **Drops blueprint's ceremony by design:** multi-slice decomposition, skeleton-then-fill, the per-slice `slice-verifier` loop, the dual post-finalization review (`artifact-code-reviewer` + `artifact-coverage-reviewer`), the 6-dimension sweep, and parallel research are all absent here — the workflow's own gates own that confidence.