The orchestration policy and CLI that selects a Codex stage, skill, model role, and reasoning effort. Auto owns routing; selected skills own methodology.
A stable capability class (luna, terra, or sol) resolved to a configurable Codex model identifier at process launch.
One fresh Codex execution with a single responsibility, explicit access level, selected skill, and structured output contract.
A durable, validated stage output and source record used by ContextAssembler to derive target-specific Stage Context.
The structured output of the plan stage. It informs approval but is not executable authority.
The immutable version-1 proposed-plan snapshot pinned to a run at approval. It is the durable authority for downstream execution and cannot be rewritten by a stage.
The approved-plan version and its sha256: canonical-JSON digest. Downstream context and results use it to prove that they refer to the pinned plan.
A compact, typed, target-specific projection of the approved plan and relevant current state. It is neither a transcript nor a source of truth.
A structured review issue with a P0-P3 priority, concrete evidence, and required change. P0/P1/P2 block completion; P3 is advisory.
A P0/P1/P2 review finding present in the latest completed review. A finding omitted by a later clean review is resolved; attempt counts never control the review-and-revision loop.
The durable review-to-implementation transition created by an unresolved finding. Attempt history stays in SQLite while the next stage receives only its bounded Stage Context.
The durable SQLite record of a task, its classification, selected stages, events, outputs, and current state.
A structured request emitted when a stage cannot continue without a different role or a human decision. Evidence, not additional effort alone, determines the route.
The explainable record of which policy rule selected a stage, skill, model role, and reasoning effort.
A durable paired Auto-versus-Normal experiment rooted at one immutable commit.
A trial is one isolated candidate worktree. Its variant is either the full Auto route or one ordinary direct Codex run.
Participant metrics measure the candidate run. Evaluator metrics measure the separate blind review and never count toward either candidate.
A trial with no valid quality observation because its participant or evaluator failed, or because its same-repetition peer is invalid or missing. Existing participant evidence remains durable even when paired evaluation is skipped.