legion-dispatcher

name: legion-dispatcher description: Decide whether to run a single legion task as a local Claude Code subagent or as a cloud Modal worker. Used by the orchestrator skill per task. Inputs are the task spec + current swarm load + heuristics from legion.toml. Output is "local" or "cloud" with a one-line reason.

legion-dispatcher

Per-task router. Called once per task by the orchestrator's dispatch loop.

Inputs

• task: {id, title, spec, deps, estimated_minutes, files_touched}
• load: {in_flight_local, in_flight_cloud, max_workers, throttle_active}
• heuristics from legion.toml [dispatch]:
  • local_file_threshold (default 5)
  • cloud_minutes_threshold (default 5)
  • always_cloud_patterns (regex list)

Decision rules

Apply in order. First match wins.

1. always_cloud_patterns matches the title → cloud. Reason: "matched pattern ".
2. Throttle active AND task is read-only / quick → local. Reason: "throttle backoff".
3. estimated_minutes ≤ 2 AND len(files_touched) ≤ 2 → local. Reason: "trivial task".
4. Task spec mentions "browser", "scrape", "playwright" → cloud. Reason: "browser work; cloud has dedicated chromium".
5. Task spec mentions "benchmark", "measure", "profile", "long-running" → cloud. Reason: "throughput-bound".
6. len(files_touched) < local_file_threshold AND estimated_minutes < cloud_minutes_threshold → local. Reason: "small enough for local".
7. Default → cloud. Reason: "default route".

Output

Return JSON: {"target": "local" | "cloud", "reason": "..."}.

If cloud:

• The orchestrator will shell out to modal run ~/holyclaude-cloud/modal/worker.py::run_task with the task fields.

If local:

• The orchestrator will spawn a Claude Code Agent subagent (general-purpose) with the framed task prompt and the same constraints (work on a worktree branch, push when done).

Why an LLM call is overkill here

These rules are the prior. For Phase 1, don't call Opus — just apply the rules. Adding an LLM tiebreaker is a Phase 2 polish that will marginally improve routing at meaningful cost (one Opus call per task, on top of the per-worker cost). The rules cover ~95% of cases.

If you want to eyeball the routing before dispatch, the orchestrator already shows the task graph with assignments at the human checkpoint.