Plan

Turn the caller's intent into one bounded, testable behavior in the place that already owns the work. Prefer the caller's tracker, if any. When no durable tracker or issue reference is available, use the caller's conversation or supplied text; the runtime snapshots those resolved intent bytes so later contexts can read and hash the same source. Do not make the model restate those facts in a packet.

Workflow

1. Resolve the intent source and choose one active behavior. When that source is not already durable, have the runtime pass its exact bytes to the validate skill's scripts/validate.py snapshot-intent --source -, resolved relative to wherever that skill package is installed (a repo checkout: skills/validate/scripts/validate.py; an installed skill package: .agents/skills/validate/scripts/validate.py), and use the returned intent_ref for later phases.
2. Route the work by type (see Ground-truth routing) and name its ground truth first. Then inspect only enough real context to make paths, interfaces, and evidence concrete: hydrate only the context sources this decision needs and carry their citations forward. Existing research and specialist skills are advisory inputs, never a merged context store.
3. Ensure the source contains acceptance examples, important non-goals, and the allowed write scope. Use lightweight prose or Given/When/Then only where it removes ambiguity; do not require both normal and edge ceremony for every change.
4. Name the first useful acceptance check.
5. If authorized and the source is writable, update that bead or issue in place. Otherwise return a concise proposed amendment to the caller.

Planning produces no AgentOps packet. A durable caller-owned source stays in place; the runtime carries its reference and the digest of its exact resolved bytes to detect later acceptance drift. Only when no durable source exists does the runtime store those bytes under their digest as a content-addressed snapshot. That fallback is derived automatically and is not another model-authored planning artifact.

Bound the work around the caller-visible outcome, not individual files, gates, or reviewer comments. Decomposition is useful only when it reduces reasoning cost; it must not multiply invocations or proof artifacts.

Scope admission

In a repository with generated projections, write scope names generator-owned outputs as a class — the hand-edited sources plus all outputs of the owning regen commands — never as a hand-enumerated path list. Hand enumeration is falsified the first time a regen command rewrites a companion the author did not list: the 2026-07-15 heal-skill fold burned two implement lanes and three intent revisions (.agents/ao/intents/sha256/d1db59d4...2b81 superseded by f5fd7c3c...af75 superseded by 26a4f2be...eb48) before scope was restated as a class.

Before freezing acceptance, run a complexity admission: enumerate the generated companions, parity twins (for example a skills-codex/ mirror), and test files that assert on the paths being changed. Anything this pass finds that the scope does not admit will surface later as an out-of-scope diff or a broken gate.

Ground-truth routing

Every plan needs a ground truth outside the planner's own reasoning. Before freezing acceptance, classify the work and name its ground truth, its control experiment, and its deviation ledger from the row below.

The Extend row is already the repo's default discipline: behavior-first acceptance, RED -> GREEN, the smallest real change. The Integrate row is the one that is cheap to skip and expensive to have skipped — run the stock control experiment before you design, or you will re-plumb what the substrate already documents and inherit bugs you built yourself.

Trigger: the Integrate-row mechanics — the stock-quickstart control run and the deviation ledger from the documented flow — apply only to integration-class work (adopting or wiring in an external substrate, runtime, tracker, or service). Routine feature work on this project uses the Extend row and does not incur them.

A plan is done only when it passes the fresh-context test: a cold context, given the intent source alone, could execute it without the author's conversation. If execution needs facts that live only in the planning conversation, move them into the source before freezing.