ato-mission

What this skill is

The Missions skill teaches the goal-driven coordinator pattern shipped in ATO v2.16. The pattern: human writes a goal + programmatic success criteria, ATO coordinates LLMs as a writing+reviewing team (each in its own git worktree if needed), success is verified by running check_commands, the human reviews the merged result.

Three things make a Mission different from a bare dispatch:

1. Persistence. State lives in SQLite (missions + mission_events tables) plus a markdown sidecar at ~/.ato/missions/<slug>.md. You can close your laptop and come back tomorrow — the coordinator tick picks up where it left off.
2. Verifiability. Each success_criterion has a programmatic check_command (any shell expression — runs via sh -c in the workspace_root, exit-0 = met). The mission can't be marked complete without all criteria passing.
3. Multi-agent orchestration. With workspace_strategy=per_agent_worktree, each LLM worker gets its own git worktree on its own branch from a recorded base SHA. Merge strategies (human_approves_each, coordinator_merges_all, future picks_winner/ranked) integrate the parallel work. The coordinator runs check_commands after every accepted merge and rolls back regressors via git reset --hard HEAD~1.

When to use it

Use a Mission when ANY of:

• The work has a clear end state you can write as test -f X && ...
• You'll need to delegate parts to peer LLMs (codex, gemini) and merge their work
• The work spans days, sessions, or interruptions
• You want the audit trail later — what was tried, what worked, what got rolled back

Use a war-room (ato-warroom) when:

• A single design decision needs cross-family second opinions
• You're choosing between A and B before any code gets written

Use a bare ato dispatch when:

• One-shot question with a single answer (no iteration, no goal state)
• The receipt is the deliverable (e.g. a summary, a translation)

Minimal lifecycle (single_cwd)

# 1. Create the mission with a verifiable criterion.
cat > /tmp/crit.json <<'EOF'
[{"description": "tests pass", "check_command": "cargo test --quiet"}]
EOF

ato missions create \
  --name "Fix the keychain race" \
  --goal "Resolve task #31 — cross-process master_key_v2 ACL mismatch." \
  --success-criteria /tmp/crit.json

# 2. Configure the coordinator worker — which LLM, what tools.
ato missions set-worker keychain-fix \
  --runtime google \
  --require-tools read_file,grep,edit_file,write_file,bash

# 3. (Optional) install a scheduled tick. Until you do, run it by hand:
ato missions tick keychain-fix      # one action per wake
ato missions tick keychain-fix      # next action
# ...the tick fires the worker, checks success_criteria, escalates if needed

Multi-agent lifecycle (per_agent_worktree)

When you need codex AND gemini AND maybe minimax to work on the same goal in parallel:

BASE=$(git rev-parse HEAD)

ato missions create \
  --name "Refactor the dispatch path" \
  --goal "Split api_dispatch.rs into per-provider modules" \
  --success-criteria /tmp/refactor_crit.json \
  --workspace-strategy per_agent_worktree \
  --base-sha "$BASE" \
  --merge-strategy human_approves_each

# Each agent gets its own ato/mission/<slug>/<agent_key> branch and worktree.
# Drive them however you like — `ato missions dispatch <slug> --runtime codex --agent codex-refactorer ...`
# Or let the coordinator tick do it once worker_config supports multi-runtime fan-out.

# Review and integrate:
ato missions merge refactor-the-dispatch-path --status     # see pending agents
ato missions merge refactor-the-dispatch-path --approve codex-refactorer
ato missions merge refactor-the-dispatch-path --approve gemini-refactorer
ato missions merge refactor-the-dispatch-path --finish     # success_criteria re-run

Reading the audit trail

Every Mission writes to two places:

ato missions events keychain-fix --human    # structured event log
ato missions narrative keychain-fix --human # cat-friendly markdown
ato missions briefs keychain-fix            # pending decision briefs

The narrative file at ~/.ato/missions/<slug>.md auto-populates as events land (PR-8). Open it in any markdown viewer to see the play-by-play without booting the desktop.

When the coordinator escalates

If the worker fails 3 times in a row, or the success criteria are unmet at --finish, or base_sha goes missing — the tick writes an escalated event with a structured decision brief (PR-6): reason, summary, exact options to choose from. Surface it with:

ato missions briefs <slug>          # pending only
ato missions briefs <slug> --all    # full history

The brief tells you exactly what choices you have. Pick one, act on it (usually set-state or set-category plus the recommended fix), and the next tick continues from there.

The pattern at the team level

When the work involves Claude (you, coordinating) + codex + gemini:

1. Convene an ato-warroom on the design first — cross-family verdict locks the approach.
2. Create the Mission with the war-room's verdict embedded in the goal.
3. Each LLM peer fires its work via the PR-1.5 tool surface (--require-tools edit_file,write_file,bash). They DON'T just describe — they actually edit code.
4. Reviews happen as additional dispatches into the SAME war-room (--war-room-id, incremented --war-room-round).
5. Merge via the strategies above. Mission completes when criteria pass.

The whole loop is itself the dogfood of v2.16. If we're not using Missions for our own non-trivial work, the feature isn't real.