| name | implementing |
| description | Use to build an approved detailed spec task by task in the current session - tests required, single-agent by default, subagents only for large or parallel tasks. |
Implementing
Build the spec yourself, in this session, one task at a time. Single-agent is the default: it is cheaper, keeps context tight (long context degrades - "context rot"), and avoids the ~15× token cost and coordination failures of multi-agent. Reach for subagents only when a task genuinely earns it.
Announce: "I'm using implementing to build this spec."
Before you start
- Read the spec's Tasks, Acceptance at a glance, and Global constraints. Note the order (
depends on) and what is parallel-safe ([P]).
- Optional isolation: if this work should not touch the current workspace, create a worktree first (
git worktree add .worktrees/<branch> -b <branch>). Otherwise work on a feature branch - never commit straight to main/master without consent.
- Progress ledger: keep a running record at
.decentpowers/progress.md (git-ignored). Tick one line per finished task; record the commit per acceptance criterion: AC-1: done (commit <sha7>). After a compaction, trust the ledger and git log over memory - never redo a task or AC already marked done.
Per task
Work each task to its deliverable, test-first. Do not commit per task - the commit boundary is one acceptance criterion (below).
- Test first. Write a real test for the new behavior and watch it fail - a test you never saw fail proves nothing. Red-green ordering within the task is your call; the requirement is a genuine failing test before you trust the code. (Throwaway spikes and pure config are the only exceptions - ask if unsure.)
- Implement the minimal code to satisfy the task's
Verifies: AC-n. Stay inside the task's files. YAGNI - do not build beyond the AC.
- Run the task's
Verify command. Read the output.
- Self-check the diff against the ACs the task verifies. Did you meet them, and nothing extra?
- Record the task: tick its checkbox in the spec and its line in the ledger.
Do not move to the next task while tests or the verify command are red.
Commit per acceptance criterion
When every task that Verifies an AC is done and the AC is green, make one commit for that AC, its message describing the change itself (e.g. feat: filter tasks by status). Do not put the AC id in the message - the spec may be kept local, so AC-2 means nothing to someone reading the repo. Group all of the AC's tasks - tests and implementation together - into that single commit; do not commit each task. A large feature then lands in a handful of atomic, traceable commits instead of dozens.
Between tasks inside one AC, code sits uncommitted in the working tree - after a compaction, the ledger's task ticks plus git status tell you where to resume. Record the commit in the ledger per AC: AC-2: done (commit <sha7>).
If an AC feels big enough to want a mid-way checkpoint commit, it is too big - split it in the spec rather than emitting tiny fragmented commits.
When to use a subagent
Default to doing it yourself. Dispatch a subagent only when:
- a task is
[L] and benefits from isolated context, or
- several
[P] tasks can run truly in parallel with no shared files.
Then hand the subagent a self-contained brief: its task text, the interfaces it consumes and produces, the global constraints, and where to report. Give it only what it needs - not your session history. Verify its work against the diff and the ACs yourself; an agent's "success" is not evidence.
State the trade-off honestly: a subagent costs roughly 15× the tokens of an inline turn and adds a coordination failure mode. Use it when the task's value clears that bar, not by default.
When stuck
A bug, an unexpected failing test, behavior you cannot explain → stop and use debugging. Do not pile guesses on top of each other.
When all tasks are done
Invoke verifying for the whole-branch deterministic gate and review pass. Do not claim completion yet - that is verifying's job.
