codex-loop-orchestra AGENTS.md

LOOP Discipline

1. Decomposition discipline (single-pass Plan-and-Solve)

• Planning is a SINGLE PASS: read the task, then emit the complete decomposition in one planning event — packets/*.json (exactly 4 fields: goal, authorized_paths, acceptance, constraints, plus packet_id) and dag.json.
• No incremental re-planning mid-wave. Plans change only at an adjudication event (dead-letter, merge conflict, L3 escalation, wave finale).
• Each packet must be self-contained: an Executor with only the 4 fields and its worktree can finish it. If a packet needs "context from the main thread", the decomposition is wrong — split or rewrite it.
• Intra-wave packets must have non-intersecting authorized_paths; the DAG assertion enforces this before any dispatch.
• No 5th required packet field. No idempotency_key, no embedded runtime_resources.

2. Anti-polling negative discipline (what Sol rounds are NOT for)

Sol is invoked on exactly two event types: planning and adjudication. Sol rounds are NEVER used for:

• Waiting — native wait-all blocks without producing rounds.
• Polling — status lives in data/events.ndjson; scripts consume it.
• Tallying — report counting is the missing-item check script's job.
• Retry decisions — config/retry_classes.yaml + retry script decide.
• State recap — the state machine holds state; never ask Sol "where were we".

Any transition decidable by if/else, exit code, or a state machine does not go through an LLM. Sol's ideal round count per task approaches 2 + anomaly count (one planning round, one finale round, plus one per genuine anomaly). If a run uses more, find which of the five forbidden uses leaked back in.