Skill v1.0.0
currentAutomated scan100/100version: "1.0.0" name: cadence description: Use when changing production code or tests in projects with enabled docs/llm-orchestrator/cadence.json. Not for ordinary documentation or questions. license: MIT compatibility: Claude Code or Codex; bash 3.2+, git and python3 allowed-tools: Bash(${CLAUDE_SKILL_DIR}/scripts/orch-cadence-gate.sh ), Bash(${CLAUDE_SKILL_DIR}/scripts/orch-cadence-check.sh ), Bash(${CLAUDE_SKILL_DIR}/scripts/cadence-detect.sh *)
The cadence
Read the project's docs/llm-orchestrator/LAWS.md and configuration first. On Codex also read its CODEX.md for installed tools, evidence and trust. Absent/disabled cadence and ordinary documentation/questions are out of scope. User instructions and project amendments govern.
Check workflow: proportional uses the paths below. Any other value, or none, is a configuration error; fix it ("workflow": "proportional") before working, and never guess a process.
Proportional paths
Choose by impact, uncertainty, contracts and reversibility, not line count. Reassess if scope grows. Honour requested reviewers and models.
| Path | Use when | Required work | |
|---|---|---|---|
| Simple | Clear, local, reversible and low risk | Edit, inspect the full diff, run applicable existing checks, deliver. | |
| Standard | Bounded behavior with meaningful edge cases | State acceptance criteria, implement, check, run the Standard review (requesting-code-review), resolve findings. | |
| Full | Interacting architecture, difficult recovery, substantial uncertainty or consequential contracts | Review the spec, implement, run the Full review, resolve findings, independently verify. |
Simple needs no extra agent, spec, worktree, report or invented test; Standard adds only one independent review and fixes for its findings. Writers can inspect diffs and run checks, but cannot provide independent review or gates.
Read the Full path only when required. Simple and Standard need only this page.
Check once, with evidence
Use relevant checks and meaningful regression tests. Honour build/dependency restrictions. No default broad suites or mutation batteries for Simple/Standard. Change existing tests only when the task says so; report each change and why.
Run each check as one plain foreground command, so it appears in the transcript as itself. A Stop hook reads that transcript and warns if a reply says Verification: PASS when no check ran and passed in the turn. It only warns, and it reads only that label — never the prose around it.
Reuse checks across turns/commits preserving covered inputs. Rerun for changes, failures or unresolved concerns. Unrelated green cannot erase failure. Use the shared completion line: Verification: PASS — checks and scope, Verification: PENDING — remaining validation, or Verification: BLOCKED — unavailable prerequisite. For low-risk work with no meaningful automated check, use Verification: NOT APPLICABLE — reason; manual diff inspection performed. This cannot waive failures, uncertain writes or stale/unavailable required checks.
Finish and clean up
Keep maintained specifications, research conclusions, design decisions and runbooks. Reviews, temporary plans/logs and disposable copies belong outside Git.
Use this skill's scripts/orch-task-resources.py --help for task resources. Acquire consumer leases before use; finish after consumers stop and deliverables are preserved. Stop is not completion. Never force removal, adopt old paths or discard another session's work.
Protected policy
The lock and numbered-ruling mechanism is binding. Propose rule changes as a patch; only the person runs scripts/cadence-ruling.sh. Honour commit/push/merge authorization separately from implementation. Name dispatched models. Hooks require installation, enablement and trust; Git policy checks do not prove execution or review quality.