| name | ami-orchestrator |
| description | Dispatch the orchestrator subagent for any task in the dependence-forecastability (ami) project. Use for every code change, bug fix, test run, review, planning, or release — even one-liners. Invoked as /ami-orchestrator <task>. Pre-loads pytest -n auto (parallel, never piped), active release plan, and all stage-gate constraints so the orchestrator gets full project context from the first message. |
AMI Orchestrator Dispatcher
Routes tasks through the orchestrator subagent with all project constraints
pre-loaded. No boilerplate. No forgotten flags.
Step 1 — Capture the task
The task is everything after /ami-orchestrator, or the user's most recent
message if invoked without args.
Step 2 — Gather context (run in parallel)
Before dispatching, collect:
git branch --show-currentgit log --oneline -5
Step 3 — Dispatch
Call Agent with subagent_type="orchestrator" using this prompt template
(fill in the angle-bracket placeholders):
Task:
Project: dependence-forecastability (src/forecastability/)
Branch:
Active plan: docs/plan/v0_5_0_review_hardening_ultimate_plan.md
Recent commits:
<GIT LOG ONELINE -5>
Test constraints — propagate to every specialist without exception
| Constraint | Why it matters |
|---|
uv run pytest -q -ra -n auto | Parallel workers cut wall-clock time; -ra surfaces all non-passing reasons |
Never pipe pytest output — no | tail, | grep, | tee, no 2>&1 redirect | Piping truncates output and makes the same tests appear to run multiple times (worker output interleaving fools naive grep) |
| Read pytest stdout and exit-code directly | The only reliable signal of pass/fail |
Lint gate: uv run ruff check . | Must be green before any commit |
Type gate: uv run ty check | Must be green before any commit |
Files touched (if known): <extract from user message, else "derive from task context">
Risks pre-screened by caller: <any risks the user mentioned, else "none">
Please:
- Decompose this task into specialist assignments
- State routing order and stage gates between each step
- Flag any concerns or blockers before work begins
Behaviour notes
- If the task is ambiguous, ask one clarifying question before dispatching.
- "Run tests" with no other context → dispatch
tester subagent with
uv run pytest -q -ra -n auto and no piping.
- Update the active-plan path when switching to a different release branch
(check
docs/plan/ for the current plan file).
- Never dispatch orchestrator for read-only investigation that produces no
edits (that's just regular Claude conversation).
