| name | arc:implement |
| description | Orchestrator — full implementation cycle from requirements to PR. Runs skills in the correct order with feedback loops. Handles single tickets, all tickets from a PRD, or bug fixes. Use when implementing a feature end-to-end, running the full development cycle for a ticket, or fixing a bug. Triggers on "arc:implement", "implement task", "implement ticket", "implement feature", "full cycle", "implement issue", "work on ticket", "fix bug", "fix issue".
|
| user-invocable | true |
| argument-hint | {ticket-id} | prd:{feature-name} [--fix] [--yes] [--stop=phase] [--skip=phase] [--test-first] [--max-retries=N] |
| license | MIT |
| metadata | {"author":"supa-magic","version":"1.0.0","category":"workflow","tags":["orchestrator","implementation","full-cycle","automation"]} |
/arc:implement $ARGUMENTS
Orchestrator — full implementation cycle from requirements to PR.
Usage
/arc:implement JRA-123 Pair mode — confirm at each gate
/arc:implement JRA-123 --yes Recommended path, no pauses
/arc:implement JRA-123 --yes --skip=test-mutate Skip mutation testing
/arc:implement JRA-123 --yes --stop=code-review-request Autopilot until PR, then pair mode
/arc:implement JRA-123 --skip=test-mutate --skip=ac-enrich Pair mode, skip mutations and enrich
/arc:implement JRA-123 --test-first TDD: test → code instead of code → test
/arc:implement JRA-123 --fix Bug fix with existing ticket
/arc:implement --fix Bug fix — interactive, ask for details
/arc:implement JRA-123 --fix --yes Bug fix, autopilot
/arc:implement prd:checkout --yes All tickets from PRD, autopilot
| Argument | Format | Default | Effect |
|---|
ticket-id | Positional | — | Single ticket to implement |
prd:{feature-name} | Positional | — | Implement all tickets from a PRD |
--fix | Flag | false | Bug fix — short cycle (no plan, enrich, or mutate). With ticket-id or interactive |
--yes, -y | Flag | false | Follow recommended path without pausing |
--stop={phase} | Flag | — | Autopilot UNTIL this phase, then switch to pair mode |
--skip={phase} | Flag (repeatable) | — | Skip phase entirely |
--test-first | Flag | false | TDD/RDT order: test → code instead of code → test |
--max-retries={N} | Flag | 3 | Max iterations in feedback loops |
Instructions
Step 1: Parse Arguments
- Parse flags:
--fix, --yes, --stop, --skip, --test-first, --max-retries
- First non-flag token →
target:
- Starts with
prd: → multi-ticket mode, extract feature-name
- Looks like a ticket ID (e.g.,
JRA-123, #42) → single ticket mode
--fix without target → interactive bug fix mode (gather details from developer)
- Otherwise → single ticket mode, treat as
ticket-id
- If no target and no
--fix → error: "Target required. Usage: /arc:implement {ticket-id} or /arc:implement prd:{feature-name}" and stop
Step 2: Route by Mode
| Mode | Target | File |
|---|
| Bug fix | ticket-id + --fix | fix.md — short cycle for bug fixes |
| Single ticket | ticket-id | single.md — full cycle for one ticket |
| Multi-ticket (PRD) | prd:{feature-name} | multi.md — build dependency graph, run waves |
The --yes Flag
--yes does NOT disable analysis. Every skill still thinks, classifies, and forms a recommendation. --yes only removes the pause — the skill follows the recommended path without waiting for developer input.
Pair mode (without --yes):
1. Skill analyzes → forms recommendation
2. Shows recommendation to developer
3. Waits: "agree" / "no, do it differently"
--yes mode:
1. Skill analyzes → forms recommendation (same logic)
2. Executes recommended path without waiting
Quality does not drop in --yes mode. Analysis always happens — --yes is about speed, not quality.
Feedback Loops
Full cycle (single.md):
/arc:ac validate → fail → /arc:ac update → /arc:ac validate
/arc:test review → needs work → /arc:test write → /arc:test review
/arc:test validate → red → /arc:code or /arc:test write → /arc:test validate
/arc:test mutate → survivors → /arc:ac update → /arc:test write → /arc:test mutate
/arc:ac verify → partial → /arc:code → /arc:ac verify
/arc:code review-resolve → comments → fix/respond → push → /arc:code review-resolve
Bug fix (fix.md) — shorter loops:
/arc:ac validate → fail → /arc:ac update → /arc:ac validate
/arc:test validate → red → /arc:code or /arc:test write → /arc:test validate
/arc:ac verify → partial → /arc:code → /arc:ac verify
/arc:code review-resolve → comments → fix/respond → push → /arc:code review-resolve
Each loop: maximum --max-retries iterations (default 3). After that → stop and escalate to the developer.
Troubleshooting
Error: No project context
Cause: .arcana/project-context.md doesn't exist.
Solution: Run /arc:project init first.
Error: Task too large
Cause: /arc:plan determined the task exceeds one context window.
Solution: Use /arc:prd update to split the task into smaller tickets, then /arc:implement each one.
Stuck in a feedback loop
Cause: A skill keeps failing after max retries (e.g., tests stay red, AC keeps failing validation).
Solution: Orchestrator stops and shows the developer what's failing. Manual intervention needed — the issue may be in the AC, the code, or the test setup.