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; otherwise the caller's conversation or supplied text, which the runtime snapshots so later contexts read and hash the same bytes.

Prompt

Plan bead ag-1234: "ao gate check lists the probe-coverage row". Shape it in
the bead itself: one active behavior, acceptance examples, non-goals, write
scope as a class (cli/internal/gates/** plus regen outputs), first check
`cd cli && go test ./internal/gates/...`. Update the bead in place.

It's working if

• The bead or issue text itself gains acceptance, non-goals, and write scope; no plan file appears under .agents/ in the diff.
• Write scope names a regen class (skills/** plus every output of scripts/regen-all.sh), not a hand-enumerated path list.
• The plan names one first check as a runnable command, such as bash scripts/check-x.sh, and a fresh context given only the source can start Implement.

Workflow

1. Resolve the intent source and choose one active behavior. When the source is not durable, have the runtime pass its exact bytes to the validate skill's scripts/validate.py snapshot-intent --source - (under skills/validate/ in a checkout, .agents/skills/validate/ when installed) and carry the returned intent_ref into later phases.
2. Route the work by type (Integrate, Extend, or Greenfield) and name its ground truth, control experiment, and deviation ledger first from references/ground-truth-routing.md. Then inspect only enough real context to make paths, interfaces, and evidence concrete, carrying citations forward; research and specialist skills are advisory inputs.
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. Write-scope checks (folded from the retired scope skill):
   • patterns are normalized repository-relative paths;
   • includes cover the behavior without granting unrelated directories;
   • excludes do not contradict required changes;
   • generated companions that must move with the sources are explicit;
   • no ownership, scheduling, Git, hook, retry, release, or delivery state.
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: the runtime carries the source's reference and digest to detect acceptance drift. Bound the work around the caller-visible outcome, not files, gates, or reviewer comments; decompose only when it reduces reasoning cost.

Scope admission

At scope, read boundaries.md in the rpi skill's references directory for what Plan does not own. 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), because a hand-enumerated list is falsified the first time a regen command rewrites an unlisted companion. Before freezing acceptance, enumerate the generated companions, parity twins such as skills-codex/, and tests asserting on the changed paths; anything unadmitted here surfaces later as an out-of-scope diff or a broken gate.

A plan is done only when it passes the fresh-context test: a cold context, given the intent source alone, could execute it. Move any fact that lives only in the planning conversation into the source before freezing.