advance-plan

advance-plan

Role: the campaign driver. Takes a named chunk of an approved plan, has a planner decompose it into a forward-only DAG of phases, secures one approval, then walks the DAG dispatching one change-runner per phase, unattended.

Prerequisites

An approved feature plan from plan-system (a planning/tasks/plan-*.md file), or a named chunk the maintainer points at. ~/.agentic-sdk/<project>/project.edn exists with :lanes and :spine set.

Procedure

Read skills/shared/references/orchestration.md before driving. Its laws bind this skill: context is the budget, the work flows forward-only as a DAG, the stack stays linear on the tip, and agents decide and record rather than stall.

1. Check for a resume. If ~/.agentic-sdk/<project>/runs/<slug>/checkpoint.edn exists for this chunk, resume: read it and the plan at ~/.agentic-sdk/<project>/runs/<slug>/plan.edn, skip landed phases, re-dispatch any in-flight phase, continue the open ones. Resume from the checkpoint and plan, not from history.
2. Plan up front, in a sub-agent. Dispatch one planner agent running plan-work. Its prompt is self-contained: the named chunk (or "next runnable work") and a pointer to the feature plan. The planner reads what is already landed (the commit log, the modules and tests that exist), picks the chunk in dependency order, and decomposes it into a forward-only DAG of phases and tasks. For each task it names the planned commit and the specialist that will execute it (a writer with the relevant write- recipe or write-tests, a reviewer, and so on). It tags each phase with a risk tier that sets the model and effort: HIGH-RADIUS (a native edge between languages, a query or codegen path, a security seam, a cross-module or wire-format change) runs at the full tier; MECHANICAL (a test-only phase, a scaffold, a single-module pure refactor) runs at a cheaper model and lower effort. Each phase carries its own tests, its check-security pass where untrusted input is involved, and its verification lanes in its definition of done. The planner writes the full plan to the gitignored ~/.agentic-sdk/<project>/runs/<slug>/plan.edn and returns a compact summary: the phases, their dependency edges, any ADR conflict it found. Hold the summary, never the full plan.
3. Approve once. Surface the plan summary to the maintainer: the phases, their order, the dependency rationale, any ADR conflict. This is the only up-front gate. On approval, write the initial checkpoint and a campaign-context note (~/.agentic-sdk/<project>/runs/<slug>/context.md, gitignored) holding the invariants every phase shares: the base-on-tip rule and the base commit, the env contract, the commit and review policy. Each dispatch points to this note instead of retyping invariants, paying their token cost once. Then proceed unattended.
4. Auto-dispatch the DAG. Walk the DAG in topological order. For each phase whose dependencies have all landed, dispatch one change-runner via the Agent tool, at the model and effort the phase's risk tier names. Its prompt is self-contained but lean: the phase name, a pointer to its entry in plan.edn (cited, not pasted), and a pointer to the campaign-context note for the shared invariants. The change-runner reads its phase from the plan, runs implement-change for it end to end in its own context, and returns exactly one line and nothing else: