Divide and conquer

The project documents and their rules are in project-documents.md; its terms (authorized, obligation, lazy, pointer) apply throughout.

Identify the authorized behavior and originating agreement for the batch: specifications with tasks, or a user-approved batch in working state. Read the applicable requirements. For remote tracker work, read the task tracker configuration that AGENTS.md points to for the configured workflow; if AGENTS.md points to none, ask the user how the remote tasks are read and updated, and keep remote updates pending until they answer. Local-only work runs without tracker setup.

The goal is the entire authorized batch implemented on a single integration branch, each task built by the least expensive agent that can do it reliably, with each task's implementation and validation accounted for. The run has three phases: divide routes every task, conquer builds them in parallel, combine merges and reviews once.

The tasks form a task graph with blocking relationships, not a list of implementation steps. Its frontier contains tasks whose prerequisites have been satisfied.

Communicate primarily through context pointers to the originating agreement, tasks, research notes, and previous commits. Do not duplicate information already available through those pointers.

Run subagents in the background where possible for maximum concurrency. Without a tool that dispatches subagents, routing keeps every task here: build the frontier tasks one at a time in this session, following the kept-task path in step 7, and do the merges and final fixes here too.

Divide

1. Read the originating agreement and tasks to understand the task graph. For a batch in working state, derive its work boundaries and dependencies from there. Pin the starting commit and scope for final review.

2. Optionally use an exploration subagent for relevant codebase or external-documentation research. Save its Markdown notes outside the repository where future implementer subagents can read them.