odylith-discipline

Odylith Discipline

Governance-Learning Default

Before acting on a durable error, escaped defect, failed mechanism, failed simulation, bad generated artifact, semantic drift, quality-gate miss, latency breach, architecture decision, validation result, or release-risk learning, search Casebook and related governance truth first. Read prior failed mechanisms, failed fix attempts, rejected approaches, guardrails, and validation history; do not repeat a fix path that already failed. Capture new mechanism-level learning in Casebook or Compass, and update Radar or plans, Registry, and Atlas when the learning changes planned work, component contracts, or flows.

Use this skill when Odylith Discipline behavior, credit-safe hard laws, learning-spine signals, or benchmark-proved pressure behavior need validation or implementation.

Default Flow

• Run ./.odylith/bin/odylith discipline status --repo-root . for the human-readable local readiness and credit posture.
• Put a proposed move in a file and run ./.odylith/bin/odylith discipline check --repo-root . --intent-file <path> for a human-readable local admissibility summary before acting on ambiguous, risky, proof-sensitive, delegation-sensitive, or public-claim-sensitive work.
• Run ./.odylith/bin/odylith validate discipline --repo-root . for deterministic proof. Add --case-id <id> for a focused case and --json for machine-readable proof.
• For benchmark proof, use the existing benchmark lane: ./.odylith/bin/odylith benchmark --profile quick --family discipline --no-write-report --json.
• Keep Guidance Behavior as the first deterministic pressure-family lane under this layer: ./.odylith/bin/odylith validate guidance-behavior --repo-root . and ./.odylith/bin/odylith benchmark --profile quick --family guidance_behavior.

Rules

• Discipline checks are local deterministic computation. They must not call Codex, Claude, provider APIs, subagents, full validation, benchmark execution, broad scans, or projection expansion on the hot path.
• Codex and Claude are first-class host families for the shared contract across dev, dev-maintainer/source-local, pinned dogfood, and consumer lanes.
• Hard laws stay deterministic: CLI-first governed truth, fresh proof for completion claims, visible-intervention proof, queue non-adoption, bounded delegation, benchmark proof for public claims, consumer-lane mutation guard, and explicit model-credit authorization.
• Passing checks stay quiet. Visible nudges, Observations, Proposals, and fallbacks need concrete recovery value and visible proof.
• Odylith Discipline may emit evidence, proof needs, recovery affordances, and owner surfaces, but it does not script final voice. Intervention and Chatter keep the rendered words live, evidence-shaped, and non-mechanical.
• Durable learning requires deterministic validation, benchmark evidence, or Tribunal/governance promotion. Store compact practice signals, not raw transcripts, secrets, broad context, ephemeral temp paths, shell descriptors, or full corpus payloads.
• Do not expose internal acronyms or planning shorthand in user-facing output. Default chat summaries should contain only the decision, practical reason, nearest safe move, proof need, and whether the hot path stayed local.
• Keep raw labels such as hard_law_results, pressure_features, retention_class, raw_transcript_retained, and counter names out of user-facing output unless the user explicitly asks for JSON verification details.