Working in chunks

A chunk is a roadmap deliverable with acceptance criteria. Its implementation has an objective, tasks, actions, checks, and resulting artifacts.

Before starting, connect your agent, review the roadmap, and confirm the project’s check and delivery instructions. Save or identify unrelated changes so the agent can preserve them.

Complete one chunk in your conversation

The skill resumes interrupted roadmap work first. Otherwise it selects one ready chunk. It reads requirements and code, records a task plan, implements the change, runs relevant checks, and commits the completed result.

Required checks and acceptance criteria must pass before completion. If verification fails, work remains unfinished. The agent should diagnose the failure, correct it, and rerun the affected checks.

A direct small request can have a work plan without a roadmap chunk. Use the direct request instead of selecting unrelated roadmap work.

Start a separate session from a terminal

From the project root:

magistr next --agent codex

Use --agent claude for Claude Code. This command starts an agent session; it is different from invoking a skill inside an existing chat. These terminal commands support coding agents Claude, Copilot, and Codex, subject to installation and authentication.

Check the effective choice without starting work:

magistr next --show-config

Configuration explains saved preferences and overrides.

Bound a longer run

From the project terminal, this runs up to three iterations with a fresh Codex session for each:

magistr next-loop --agent codex --max-iterations 3

To focus on a workstream, add --track "Account access", replacing the quoted example with an actual roadmap workstream name or an unambiguous fragment. The default limit is ten iterations; specify your intended limit explicitly.

The loop stops when its limit is reached, no eligible work is ready, or it cannot continue safely. A failed check or unresolved agent attempt may require attention before another iteration can start.

Inside a chat, $next-loop or /next-loop completes one chunk and writes a handoff. Continue that bound only in a fresh agent session. The terminal command manages the sequence of new sessions automatically.

Resume after interruption

Reopen the same project and inspect the saved work before editing. The roadmap workflow gives interrupted work precedence over new pending work. For a loop started in a terminal, rerun the same command and settings.

An unfinished loop retains its saved agent settings. Conflicting options are refused instead of silently changing its agent or model. An unresolved agent attempt must be inspected; it must not be blindly retried. See interrupted attempts.

Recognize completion

The final report should identify the delivered behavior, current passing checks, commit, and remaining limitations. A completed task is only one part of the chunk; a successful command does not by itself prove every acceptance criterion.

Pushes and deployments follow your explicit project instructions. Keep those instructions available to every session, including longer runs.