Skill: implementation-planner

When to Use

Use this skill before writing any code when:

• The change touches more than 2 files
• A new endpoint, agent capability, or routing behaviour is being added
• The approach is not obvious from the existing code structure
• You are unsure whether a change will break existing tests

Instructions

Step 1 — Understand the current state

1. Read the relevant module CLAUDE.md if one exists.
2. Run pytest -x to confirm the baseline is green.
3. Read the key files you expect to touch.

Step 2 — Write the plan

Produce a plan in this structure:

## Goal
One sentence: what does this change accomplish?

## Approach
2-3 sentences: the technical approach chosen and why.

## Files to change
- `path/to/file.py` — what changes and why
- `tests/test_something.py` — new or updated tests

## Files to read first
- `path/to/related.py` — to understand the interface

## Risks
- Any behaviour that might regress
- Any risky module being touched (triggers risky-module-review skill)

## Acceptance checks
- [ ] pytest -x passes
- [ ] changelog entry added
- [ ] no hardcoded secrets
- [ ] new tests cover the new behaviour

Step 3 — Get implicit approval before coding

If the plan involves a risky module (auth, key store, agent tools), invoke the
risky-module-review skill before proceeding.

Step 4 — Implement

Follow the plan step by step. Update the plan if reality diverges from the
written plan (annotate with # REVISED: comments).

Step 5 — Verify

After implementing:

1. Run pytest -x.
2. If tests fail, fix and re-run before reporting done.
3. Update docs/changelog.md.
4. Update .claude/state/agent-state.json with completed steps.

Acceptance Checks

• Plan was written before implementation
• All listed files were changed
• All tests pass
• Changelog updated
• No risky module changed without risky-module-review

Failure / Retry Behaviour

If a step fails (syntax error, test failure, unexpected interface):

1. Do NOT silently swallow the failure.
2. Annotate the plan with the failure.
3. Diagnose before retrying — read the error carefully.
4. If blocked, write the blocker to .claude/state/NEXT_ACTION.md and stop cleanly.