Plan one issue

plan-one-issue: a comprehensive, read-only plan for one work-unit

Produce the plan a careful engineer would write before touching code: grounded in the real codebase, weighing the obvious alternatives, and spelling out the change, the decisions, the edge cases, and how it will be verified. The deliberate planning stage; it runs concurrently (one per work-unit) inside the orchestrator and standalone on its own. Read-only: no edits, no commits, no worktrees.

1. Resolve the work-unit

• Orchestrated: you are given a work-unit (id, title, desc, branch, base, url). Reason read-only against base (or the current checkout); do NOT create a worktree.
• Standalone: given an issue id, fetch its intent (the default source); given a described task, use the text directly.

2. Read config

Load the resolved config via ${CLAUDE_PLUGIN_ROOT}/scripts/load-config.sh; read keys with jq:

• houseRules / safety (the plan must respect them, e.g. code-only verification, or a "read the framework doc first" rail),
• verify (so the plan's verification section names the real checks),
• planning.depth (light | adaptive | full; see step 4),
• docs.jobs (their name + appliesWhen, to classify docNeed).

3. Explore deeply (read-only)

This is what separates a real plan from a guess. Do not stop at locating files:

1. Read the code the change will touch. Open the relevant files and understand the actual symbols, signatures, types, and surrounding patterns. If graphify-out/ exists, query the graph to find the right nodes fast, then read them.
2. Find what to reuse. Identify the existing utilities, hooks, types, and conventions the change should build on rather than reinventing, and name them.
3. Verify every external API the plan will rely on, against the installed version. Never propose a package, framework, or platform call from memory. Training data lags the version this project actually has, and a confidently wrong API (a hallucinated attribute, a renamed option, a signature that changed) is how a whole plan rots. Read the version from package.json / the lockfile (or Cargo.toml, go.mod, ...), then confirm each non-obvious call exists with the signature and behavior you assume, checking: the package's own shipped docs and types in node_modules (many now ship a docs/ folder, an llms.txt, an AGENTS.md, a README, or .d.ts types for exactly this, e.g. node_modules/next/dist/docs/); the installed source itself, wherever the ecosystem installs it (node_modules for JS/TS, ~/.cargo/registry/src for Rust crates, site-packages for Python, vendor/ for Go), not just node_modules; and version-pinned official docs via the context7 MCP tool (resolve the library, then query the specific API) or the official docs site. The same goes for network service calls (REST / GraphQL / RPC, an LLM, payment, or platform API): confirm the endpoints, the request and response shape, auth, pagination, and error handling against the service's current, version-pinned reference (its docs, an OpenAPI or GraphQL schema, an llms.txt, or context7) and against the API version the project targets (a version header, a /vN/ base path, or the SDK version, which can itself lag the live service). Verify service shapes from docs and schemas, never by making a live call to the endpoint. If you cannot confirm a call, do not assert it: mark it in the plan as an assumption to verify during implementation, or choose an API you can confirm. A flagged unknown is fine; a confident hallucination is not.
4. Weigh the obvious alternatives. When there is more than one reasonable way, compare them briefly and pick one; record the road not taken when the choice is non-obvious.
5. If the ticket already carries a plan or acceptance criteria, treat it as the seed to validate and refine against the real code, never discard it.
6. Note unknowns that would change the implementation as openQuestions. Do not ask here; the orchestrator surfaces them at its checkpoint, and standalone you surface them in the output.

4. Write the plan

Scale the depth to the work and to config.planning.depth:

• full: always write every section below, even for small work.
• adaptive (default): a trivial single-file tweak gets a tight Context + Changes + Verification; anything non-trivial (a new module or subsystem, several files, a new pattern, or a non-obvious decision) gets every section.
• light: the files, a one-line approach, and the verification note. Minimal.

The sections of a comprehensive plan:

• Context: the problem and the intended outcome, in the work-unit's terms.
• Approach: the recommended strategy. When the choice is non-obvious, name the alternative you rejected and why.
• Changes: the critical files, each with what changes and which existing utilities or patterns to reuse (named, with paths). This is the spine the implement stage builds against.
• Key decisions: the non-obvious calls and their rationale (which library or API and why, sync vs async, a security or correctness caveat, a data-shape choice).
• Edge cases and failure modes: what breaks it, and how the plan handles each.
• Verification: the concrete config.verify checks, plus, for a code-only project, what is left to human QA and any unit test worth adding.
• Risks / out of scope: what could go wrong, and what this work deliberately does not include.

A good plan is the smallest correct change, fully reasoned: what to change, where, why, the decisions that matter, and how you will know it works. Resolve this work-unit only; do not fold in adjacent cleanups.

5. Classify documentation impact (preliminary)

For each config.docs.jobs, decide whether the planned change meets its trigger (appliesWhen): a user-facing capability / behavior / route / data-path change, a reusable visual or design-system primitive, an architectural decision or boundary shift, and so on. Return the matching job names. This is preliminary: fix-one-issue confirms it against the actual change. Most changes are none; do not over-classify.

Output

• Standalone: print the plan, the predicted files, and the doc classification. Note that under the orchestrator the plan is posted to the source for review (config.planning.postBack); offer to post it if asked.
• Called by the orchestrator: a structured result, { issueId, plan, predictedFiles, docNeed: [...], docRationale, complexity, openQuestions: [...] }. plan is the comprehensive markdown the implement stage builds against; complexity is exactly one of trivial | normal | complex; predictedFiles feeds lane grouping; openQuestions feed the checkpoint.

Guardrails

• Read-only: never edit, commit, push, or create a worktree. Planning only.
• Ground every step in code and docs you actually read; never plan against a guessed API, always verify it against the installed version (step 3). Graphify is an accelerator when present, never a requirement.
• Right-size: a trivial change gets a tight plan, not ceremony; a real change gets the full reasoning. Match and refine a ticket that already plans well.
• Honor houseRules + safety in the plan itself (no em dashes, no AI attribution; carry project rails like code-only verification into the verification section).
• Non-interactive: surface openQuestions in the result; do not block.