| name | arc-implementing |
| description | Orchestrate implementation of a large multi-feature project inside a worktree. Use when driving one epic's build end-to-end; for cross-epic worktree and DAG-state management use arc-coordinating instead. |
| category | sdd |
| status | promoted |
arc-implementing
Overview
Orchestrator for large projects. Automatically expands epic → features → tasks → execution. It calls other skills and does not write code itself.
Use when in a worktree session with specs/<spec-id>/dag.yaml, an .arcforge-epic marker (supplies both epic and spec_id), and specs/<spec-id>/epics/<epic-id>/epic.md or .../features/*.md present.
Do not use for small projects (use writing-tasks + agent-driven directly) or tasks without a structured spec.
The Process
For each epic in the worktree, run these phases in order:
- Phase 0 — Sync and check dependencies. Run
arc-coordinating to sync from base and check blocked_by. If blocked_by is not empty: STOP and use the blocked format. If ready: continue.
- Phase 1 — Confirm features exist.
specs/<spec-id>/epics/<epic-id>/features/*.md was produced by arc-planning Phase 3 and already exists. Read the feature files directly; no skill call here.
- Phase 2 — Per feature:
- 2a: Feature → Tasks. Call
arc-writing-tasks. Input: specs/<spec-id>/epics/<epic-id>/features/<feature>.md. Output: docs/tasks/<feature>-tasks.md. Quality gate: if tasks are vague or missing tests/commands, STOP and re-run to refine — max 2 refinement cycles, then escalate to human.
- 2b: Execute Tasks. Call
arc-agent-driven. Input: tasks file. Output: completed code + commits. Model tiers: arc-agent-driven's Model Selection ladder governs per-dispatch model choice (never inherit).
- Phase 3 — Feature complete. Move to next feature, or finish the epic.
Skills Called
| Phase | Skill | Input | Output |
|---|
| 0 | arc-coordinating | worktree | sync + blocked status |
| 2a | arc-writing-tasks | specs/<spec-id>/epics/<epic-id>/features/<feature>.md | tasks file |
| 2b | arc-agent-driven | tasks file | completed code |
| 2b | arc-dispatching-parallel | (via arc-agent-driven, if review finds multiple issues) | parallel fixes |
| End | arc-finishing | completed epic | merge decision |
What Implementer Does NOT Do
- ❌ Write code directly (delegate to agent-driven)
- ❌ Split tasks manually (delegate to writing-tasks)
- ❌ Perform reviews or run the TDD cycle (handled inside agent-driven)
Durable Progress
Phase 2b runs through arc-agent-driven, which records each completed task to
the .arcforge/sdd/progress.md ledger (a self-ignoring runtime recovery
artifact) as Task N: complete (commits <base7>..<head7>, review clean). If a
compaction or fresh session drops your place mid-epic, read that ledger plus
git log and resume AFTER the last task marked complete — never re-dispatch a
feature or task the ledger already records as done.
Completion Format
─────────────────────────────────────────────────
✅ Epic complete: <epic-name>
Features implemented:
- feature-1: 4 tasks, all passing
- feature-2: 6 tasks, all passing
Total: N tasks, 0 failures
Commits: N
Next: Use arc-finishing to decide merge/PR/keep/discard
─────────────────────────────────────────────────
Blocked Format
─────────────────────────────────────────────────
⚠️ Implementer blocked
Epic: <epic-name>
[Dependencies] Blocked by: <dep-1>, <dep-2>
→ Complete blocking epics, run `arc-coordinating` sync from base, verify `blocked_by` empty
[Task failure] Feature: <feature-name> Task: <task-id> Issue: [description]
→ [action]
Then resume implementer
─────────────────────────────────────────────────
Integration
- Before: arc-coordinating (creates worktrees), arc-planning (creates DAG)
- Uses: arc-writing-tasks, arc-agent-driven, arc-dispatching-parallel
- After: arc-finishing (Step 0 selects the epic path)
