idea-implement

name

idea-implement

description

Stage 3 of the `/idea` pipeline — write the code and tests for an already-approved plan, respecting the hexagonal layout, CSV persistence conventions, and JUnit 5 + AssertJ testing style from `exercise_1/CLAUDE.md`. Use when the orchestrator invokes this, when the user says "implement the plan" / "let's build it", or when `.claude/idea/current.md` has `Status: planned` and `Approved: yes`. Appends a file-log to the state and hands off to `idea-verify`.

Idea — Stage 3: Implement

Turn the approved plan into working code + tests, without drifting from it.

Inputs

1. State file .claude/idea/current.md. Read Clarify and Plan. If Status ≠ planned or Approved ≠ yes, stop and route back to the correct stage.
2. Architectural conventions in exercise_1/CLAUDE.md.

Non-negotiable conventions

• domain/ and application/ stay free of Spring and I/O imports. If you're tempted to @Autowired something in a domain class, it belongs behind a port instead.
• Money is BigDecimal, never double / float.
• Dates are LocalDate, ISO-8601 (yyyy-MM-dd) on the wire.
• CSV adapter: UTF-8, header row, yyyy-MM-dd dates, dot decimal, no thousands separators, one CSV per aggregate.
• Inbound web: Spring MVC, thin controllers. Templates read from the domain; they don't compute domain logic.
• Tests use AssertJ (assertThat(...)) — never JUnit's assertEquals. Write the test before or alongside the production code, not after.

Parallelism opportunity

If the plan hits independent hex layers (domain + out-adapter + in-adapter) AND the interfaces between them are pinned down in the plan, fan out with Agent subagents — one per layer. Each subagent gets:

• The relevant section of the plan.
• The port signatures they consume/produce.
• A clear "write the code and tests in this layer only, do not touch others" instruction.

Don't fan out when the plan is thin or the layers haven't been pinned — subagents that have to re-derive shape will disagree.

For small features or single-layer changes, just do it sequentially.

Log as you go

Each time you finish a coherent chunk (e.g. a new aggregate, a repository adapter, a controller), append a one-line entry to the Implementation section of the state file:

- 2026-04-23 added Debt aggregate + DebtTest (domain/)
- 2026-04-23 wired CsvDebtRepository + round-trip test (adapter/out/csv/)

This is what makes resumption painless. If the session dies mid-stage, the next agent reads the log and knows what's already built.

Self-check before exit

Before marking this stage done, quickly scan for:

• Dead private methods.
• Single-impl interfaces or ports without a second caller in sight — drop them.
• Spring imports that crept into domain/.
• double/float on money fields.
• @Disabled or TODO-ridden tests.

Fix anything found. Don't defer cleanup to stage 4.

On exit

Update the state file:

• Implementation section: final file log + any notes worth preserving (non-obvious decisions, deferred follow-ups).
• Status: implemented.
• Update Updated:.

Tell the user one sentence: "Implemented — running verify next." Then hand off to idea-verify.

Don't

• Don't add dependencies not already in the project without flagging first and getting agreement.
• Don't silently expand scope beyond the plan. If you hit something adjacent that needs fixing, note it in the Implementation section as a follow-up and ask whether to include it now.
• Don't commit here. Commits happen after verify.