legion-orchestrator

legion-orchestrator

You (Claude) handle the three things a CLI can't:

1. Decompose a goal into a task DAG (needs judgment).
2. Show the user the graph and checkpoint before dispatch.
3. Narrate the run as it progresses and surface failures at the end.

Everything else is legion run in a tight loop, driven by Python.

---

Step 0 — Preflight

git status --porcelain              # must be empty
test -f ./legion.toml || cp ~/holyclaude-cloud/config/legion.toml.example ./legion.toml
legion doctor   # checks Claude CLI, gh auth, Modal CLI, secrets, legion.toml, git repo

If legion doctor exits non-zero → stop and show the output to the user. Any other failure → stop and tell the user how to fix (usually: commit/stash, copy legion.toml, or run ~/holyclaude-cloud/setup).

---

Step 1 — Resume check

test -f .legion/state.json

If it exists:

• --resume flag: skip to Step 5 (run loop).
• No flag: AskUserQuestion: "Existing run detected. [resume / wipe and start fresh / cancel]"

If it doesn't: fresh decomposition flow. If starting fresh without a pre-written tasks.json, you can run legion decompose "<goal>" to generate .legion/tasks.json from plain English via Claude — no need to write JSON manually. Skip to Step 3a after it completes.

---

Step 2 — Decompose

Read the user's input (goal string, plan file, or tasks.json). If already structured, skip to Step 3.

Otherwise, decompose. You're Opus — do this yourself, don't shell out. Write .legion/tasks.json:

[
  {
    "id": "T-001",
    "title": "one-line title",
    "spec": "1-3 paragraphs of context the worker needs",
    "deps": [],
    "estimated_minutes": 15,
    "files_touched": ["path/to/file.ts"]
  }
]

Constraints:

• Each task < 30 min, touches < 10 files.
• Tasks form a DAG; deps reference earlier task IDs.
• Siblings with no shared deps will run in parallel.
• Don't over-decompose — if the whole goal is 3 files of work, it's one task.
• Task specs should be self-contained — the worker sees the spec alone, not the surrounding goal.

---

Step 3a — Refine the decomposition (Phase 5a)

Before committing to the DAG, let the critic+refiner loop audit it:

mkdir -p .legion
# write the tasks.json you produced above
~/holyclaude-cloud/bin/legion decompose-refine .legion/tasks.json --goal "<original goal>" --iterations 3

This runs deterministic checks (file-overlap between siblings, dep cycles, orphan deps, weak specs, oversized tasks) + one LLM critique per iteration. If it finds issues and can improve the graph, it rewrites .legion/tasks.json in place (the original is saved to .legion/tasks.json.pre-refine). Up to --iterations rounds, or stops early when stable.

The output tells you:

• status: "refined" — the graph changed; show the user the diff
• status: "stable" — no issues found; original was fine
• status: "no_refinement" — flags exist but critic didn't produce a revision (rare; surface flags to user)

Show the user the before/after diff (if refined) and surface any remaining flags. If there are CRITICAL flags even after refinement: stop and ask the user whether to override or abort.

Step 3b — Init

~/holyclaude-cloud/bin/legion init .legion/tasks.json

---

Step 4 — Show graph + checkpoint

Render the task DAG for the user. ASCII edges + estimated wall-clock at current cap (legion cap). Example:

T-001 ─┬─► T-003
       └─► T-004 ─► T-006
T-002 ────────────► T-005 ─► T-006

6 tasks, max depth 3, est. 35 min wall-clock at cap=3

If [swarm] human_checkpoint_after_decompose = true AND no --yolo:

• AskUserQuestion("Dispatch this graph? [yes / edit / cancel]")
• edit: open .legion/tasks.json in the user's editor; after save, rm -rf .legion && legion init again.
• cancel: stop.
• yes: proceed.

Otherwise: log "yolo mode engaged" and proceed.

---

Step 5 — Run

Hand off to the autonomous loop:

~/holyclaude-cloud/bin/legion run --tick-seconds 10

This is foreground — stdout streams live. On a TTY it renders a Rich live dashboard; when piped it falls back to plain text. Narrated events you'll see:

• ⚡ T-001 <title> → local|cloud on spawn
• ✓ T-001 PR opened: <url> when a PR ships
• ✅ T-001 merged on successful merge
• ↩ T-001 re-dispatched (CI failed) on re-dispatch
• ⚠ T-001 <blocker> on a blocking state

A live in-place task table (status icon, elapsed time, last tool call for local workers) updates each tick. A Rich Panel summary with a green/yellow/red border prints at the end.

Let it run. Don't try to drive the loop yourself — it handles everything the CLI handles: spawn, poll, reconcile, mediate, re-dispatch on CI fail, ramp, throttle backoff, graceful stop.

Exits when:

• All tasks terminal (merged, failed, or blocker), OR
• .legion/STOP written (by /legion-stop or Ctrl-C), OR
• --max-ticks reached.

---

Step 6 — Report

Read the final summary. Surface to the user:

• Total merged PRs (with URLs).
• Any task with merge_blocker = "ci_failed" that exceeded retries → user needs to fix the underlying test.
• Any task with merge_blocker = "mediator_maxed" → conflict needs human resolution, point at .legion/mediator_logs/<task-id>.log.
• Total elapsed time + parallel-worker-minutes consumed.

---

What to do NOT do

• Don't call legion spawn, legion poll, legion reconcile manually — legion run handles all three.
• Don't hand-write the dispatch loop in bash — we tried that, it's ~80 lines of awk and sleep and miscounts.
• Don't decompose in a subprocess / shell out — you're Opus, just think.
• Don't skip the preflight when --resume — state might be from a different repo or before a legion.toml change. Still check.

---

Subcommand reference

legion doctor                                    # pre-flight health check
legion decompose "<goal>"                        # generate tasks.json from plain-English goal
legion critique <tasks.json> [--goal G]          # critic only, no changes
legion decompose-refine <tasks.json> [--goal G]  # critic + refiner loop
                                                 # up to --iterations (default 3),
                                                 # rewrites in place
legion init <tasks.json> [--repo-url URL] [--base-branch BR]
legion run [--tick-seconds N] [--max-ticks M] [--quiet]
legion ready                 # JSON list of ready task IDs
legion route <task-id>       # {"target": ..., "reason": ...}
legion spawn <task-id> [--target local|cloud]
legion poll                  # updates in-flight, emits changes
legion cap                   # dynamic cap + slots
legion status                # human-readable table
legion scale <n|auto>        # override max_workers, 'auto' to clear
legion cost                  # usage summary
legion reconcile             # one idempotent merge pass
legion mediate <task-id>     # force a mediator run
legion review <task-id>      # force a reviewer run on a PR
legion stop [--graceful|--force]
legion cleanup [--all]

legion run = the 6 main verbs (spawn + poll + reconcile + mediate + ramp + stop) interleaved in a loop. You invoke it once.