dispatching-chunks

Dispatching chunks

Assemble the assignment file for the next chunk and inject a drift preamble that names every unplanned change in completed upstream chunks. The executor sees what prior chunks actually did, not just what they declared.

When you're done

A file at work/<initiative>/.../chunks/N-<id>.md exists, containing: a ## Pinned locations table, the drift preamble (if any), and the chunk body verbatim. The slug comes from tracker.chunks[<N>].id; it is required — dispatchNext throws if it is missing rather than silently writing chunks/N-undefined.md. dispatchNext also injects the ## Pinned locations table at the top of the chunk file by resolving every declared and body-referenced symbol against graph.db (file:Lstart–Lend ranges), so executors should read the pinned ranges, not the whole files. Unresolved symbols (graph.db missing or stale) emit a warning row pointing at scripts/scan.js; they do not throw. The user has the path and is ready to hand it to an executing-chunks subagent.

Process

1. Pick the chunk. Either user-specified, or let the verb auto-pick the next pending chunk whose depends_on are all completed.

2. Invoke the verb:

   node greymatter/scripts/dev-pipeline.js dispatch <plan-path>
   

   The verb pulls the chunk body from the plan, synthesizes the drift preamble from upstream completed chunks, writes the assignment file under chunks/, and sets chunks.N.status: "in_progress" with a started_at timestamp. (Internal drift-propagation bookkeeping happens in the tracker; you don't need to touch it.)

3. Hand off. Tell the user the assignment file path and the prompts/executing-chunks/prompt.md path the subagent should use. Stop.

Anti-patterns

• Dispatching with unmet depends_on. Block and tell the user.
• Editing the chunk body in the assignment. Pass it verbatim. Edits belong in plan.md (which then re-flows on the next dispatch).
• Doing the chunk work yourself. This skill writes the assignment file and stops. The executor is a separate session/subagent.