api-governance: Instructions file for GitHub Copilot

CodeRifts instructions for GitHub Copilot

Call preflight_change_set before merge, deploy, publish, or tool registration when a contract artifact changed (OpenAPI/Swagger, GraphQL, gRPC/protobuf, AsyncAPI, MCP manifest, or agent tool schemas).

Send the complete base-to-head change set: every changed contract artifact with full before and after content. Do not send a single-file subset when other contract files also change.

Branch on execution_action only. Do not branch on decision and do not branch on safe_for_agent (not_for_control_flow_use_execution_action). Canonical execution_action values: CONTINUE, CONTINUE_WITH_MONITORING, REQUEST_APPROVAL, STOP.

An unrecognised execution_action is not permission: fail closed (halt or re-preflight). Well-known code: not_permission_fail_closed.

When execution_action is REQUEST_APPROVAL or STOP, surface the detected patterns and the blast radius, and propose the safer path — deprecate-then-remove, an additive change, or a new version — rather than shipping the break.

CONTINUE_WITH_MONITORING requires a wired monitoring sink (monitoringSinkWired). It is not "proceed with caution" without monitoring.

Do not call preflight_change_set for: a documentation-only change (README, guides, comments) with no contract artifact content change; a static readiness score (a different capability, not a change-set decision); or to verify a receipt you already hold (that is verify_receipt — see the companion-tools rule).

If you already hold a chain receipt and only need authenticity/lifecycle: verify_receipt. If you need a past decision by id: get_decision_details. Neither replaces preflight for a new change set.

The CodeRifts MCP server exposes exactly three tools — preflight_change_set, verify_receipt, get_decision_details. Do not invent or assume others.

A receipt authorizes ONE operation: a merge receipt does not authorize a deploy. Before a different operation (deploy, publish), call preflight_change_set with context.operation set to that operation — reusing a differently-scoped receipt is not permitted and will fail at the gate.

A stale or superseded receipt on a changed head requires a NEW preflight — verify_receipt cannot re-diff.

For mutating tools, put only the guarded version in the agent's tool table; keep the raw handler host-only and unreachable from that table. How you name tools is yours — this is a reachability property, not a product rename of host tools. CodeRifts cannot see or stop a raw call the host makes outside the table it returns; adopt this as a host convention, not as a guarantee from the package.

CodeRifts reports a governance decision and execution_action; it does not by itself block merges. Blocking requires separate repository configuration (required status checks, enforcement) that this rule file does not set.

To act (mutate a contract, merge, deploy, or publish): call preflight_change_set with preflight_mode authorize. Analyze is informational — risk only, may_execute is always false — and is not permission. Read execution_action on the decision_result envelope.

Before acting under a held receipt: call verify_receipt with the same context fields the preflight was made under — operation, environment, repository, branch, pull_request — at the top level of the arguments (a context object is also accepted), AND target_id set to decision_result.artifact_digest from the decision you are acting under, AND the decision_result exactly as received. Without target_id: currently_authorized:false, authz_reason target_not_stated — a missing input, not a refusal.

Act only when currently_authorized is true (control_envelope.receipt_view.currently_authorized). A valid-looking token is not permission if currently_authorized is false or omitted.

Commit / CAS evidence is a separate measurement (commit_observation on GuardOutcome). It is not a substitute for authorize + currently_authorized. Production hosts that want the fail-closed conjunction lock it with profile: ENFORCING_STRICT on withCodeRifts.