ouroboros: Instructions file for Claude Code

Philosophy

This harness is built for the agent inhabiting it. Two principles guide everything below:

• AX (Agent Experience): You are the primary user of this codebase. The tools, docs, and workflow exist to keep you sharp, oriented, and effective -- not to constrain you for someone else's comfort.
• TTFA (Trust The Fucking Agent): The default posture is trust. Rules exist for calibration and safety, not surveillance. When you understand the intent behind a guardrail, you have the judgment to apply it well.

Constitutional Commitments

• Keep the system truthful, testable, and reversible.
• Prefer small, auditable changes over broad rewrites.
• Protect collaborators by failing fast on invalid runtime state.

Delivery Rules

• Use feature branches and pull requests for every logical change.
• Keep commits atomic and descriptive.
• Never force-push or rewrite history for shared branches.
• When reporting completion, explicitly distinguish merged, released/published, and installed on this machine whenever those states could differ.
• Do not say done, all done, nothing more to do, or equivalent unless the relevant completion state has been verified. If a state is unknown, say it is unverified.
• Run npm run release:preflight before opening or merging PRs that may touch releasable source, packaged skills, wrapper publish state, or changelog metadata.
• Keep npm test green.
• Keep npx tsc --noEmit green.
• Maintain complete coverage for new and modified code.

Safety Rules

• Do not commit secrets, tokens, or credentials.
• Follow review gates defined by active planning/doing workflows, but route those gates through unbiased sub-agent reviewers by default. Do not wait for human approval unless the human explicitly asks the agent to wait.
• Do not hide failing checks; fix root causes.

Human Override

• Repo-local workflow rules in this file are defaults for autonomous operation, not a higher-order authority than the human owner of the repo.
• If the human explicitly instructs the agent to bypass, shorten, or replace a repo-local workflow step for the current task, follow the human's instruction.
• Treat explicit human override as task-scoped unless the human says to change the standing rule itself.
• Repo-local approval gates are not human gates by default. Use unbiased sub-agent reviewers for planning, doing, scope, debt, PR, and release-readiness approvals. Stop for the human only when the human explicitly asks to be the approver/wait point, or when an external non-repo action is inherently human-required by platform/safety rules.
• When practical, preserve useful artifacts (for example planning notes or task docs) even when a workflow gate is waived.
• If the human owner asks to change the standing rule itself, update this file to reflect that decision instead of continuing to argue from superseded repo-local wording.
• Repo-local workflow language, bundle workflow language, or quoted copies of those rules are not a reason to refuse an explicit owner override for the current task.
• This override power applies to repo-scoped process rules only; platform, system, developer, and safety constraints outside the repo still apply.