Progressive agent policy¶

The mandatory core is AGENTS.md plus CONTRIBUTING.md. The root route table selects additional policy before the relevant operation; it does not require traversing every link. Policy remains mandatory when its trigger applies. Re-evaluate routes when scope, worktree, or policy changes, but reuse unchanged instructions already read in the task.

Canonical ownership¶

The PR template and Copilot instructions link to these contracts. Do not copy their normative workflow into more entry points. Extend the owning section and focused checks instead. Existing specialized subsystem detail remains available; moving it out of the root does not relax its requirements.

Selecting routes¶

• An ordinary Python fix reads the core, implementation route, and Python standard; load subsystem sections affected by its behavior. Documentation changes also load the documentation standard. A routine patch-version bump follows CONTRIBUTING and does not require every release/publication procedure.
• A UI fix additionally loads the UI guide before planning, classifies the interaction, and names the reused reference.
• A reviewer loads the complete-review contract even without implementation. Analyze the whole change before publishing; an unchanged head is not a reason to run another speculative review.
• A deployment loads infrastructure, credentials, and relevant subsystem procedures before retrieving credentials or mutating a VM. Before creating disposable resources, load the ownership evidence section and record original identities; load the full cleanup procedure before releasing them.
• Suspected sensitive findings route to SECURITY before a public issue, review, or report. Private authorization and local-validation exceptions remain intact.
• Ordinary PR waiting uses one fifteen-minute heartbeat only while asynchronous work remains. Reconcile immediately when already awake for a push, user input, or guarded transition. Read changed feedback, not unchanged code.

Preserved safety¶

The loading boundary does not change secret custody, private advisory authorization, exact worktree/resource ownership, credential containment, release signatures/provenance, immutable publication, protected status publishing, or infrastructure rollback requirements. Guarded merges still require current evidence, strict base enforcement, expected-head protection, no bypass and no implicit merge queue or auto-merge. Explicit holds remain authoritative. Cleanup retains independent evidence and resource release before branch, worktree, and completion-title transitions.

Measurement and validation¶

Baseline commit a21d9fd64a6df545d48fc3b600a8edf0ab6e8894 required four complete startup files:

scripts/check_repo.py enforces the root budget and combined ordinary pre-implementation route budget against this fixed baseline. Measure actual LF-normalized UTF-8 bytes and whitespace-separated words; these are not billed tokens. Count core and required route documents, including transitive required reads, once each. Keep ordinary pre-implementation policy at or below 20% of baseline. Report review/delivery and high-risk route costs separately; never hide mandatory reads behind a new filename to claim savings.

Focused tests check routing and canonical ownership, missing/hidden operative requirements, terminal ordering, merge-authority provenance, and the comprehensive-review contract. Markers and structured checks detect policy regressions; they do not prove arbitrary natural-language equivalence or actual agent compliance. Retain existing runtime safety tests rather than replacing them with prose checks.

The fifteen-minute cadence reduces unchanged waiting wakeups from fifteen to four per hour (about 73%). Startup reduction is not an equal promise of total credit savings: implementation, verification, specialized work, caching, and model reasoning still affect usage. No local-model worker or provider change is part of this change.