orchestrate

orchestrate

Before the first delegation of a run, read references/model-map.md once for tiers, pins, fallbacks and verification routing.

Shapes

Pick the shape the task needs, compose freely, chain shapes across turns and read each result before choosing the next. No shape is mandatory.

• Map — parallel cheap readers over the surfaces involved, returning one structured map (file → responsibilities → key symbols → line ranges → coupling). Runs before the seat reads anything large.
• Build — the seat cuts the work at file seams and specifies each piece; workers build, in parallel where the pieces touch different files.
• Check — on one diff, at the same time: cheap single-dimension checkers (one acceptance criterion or one risk each), the test/lint gate, and a cross-family reviewer when verification is owed. Findings reach the seat as candidates; fix-ups fold into the next piece instead of blocking it.
• Sweep — one bulk transform pipelined over a discovered list: rename sites, test updates against a changed API, rule-driven deletions.
• Second look — one cross-family reader on a finished result.

Scout inline, fan out the reading. The seat discovers the work-list with cheap listing (ls, grep, diff stat); reading files to understand them is Map work.

Scale to the ask. "Understand this package" → three to five readers. "Find every site" → finders until two rounds return nothing new. A one-line fix → no fan-out. A worker earns its slot with a named distinct slice; when the next slice has no name, the fan-out is done. A workhorse stage that must read more than about ten files is a Map stage in disguise: fan the reading out to cheap readers and hand the workhorse their extracts.

The instrument follows the lane. Codex stages are seat-dispatched: the seat writes the prompt and schema files, starts the helper in the background with a labeled call, keeps working, and harvests the run dir when the harness reports the exit (references/codex-exec.md § Dispatch patterns). One call per stage, up to four at a time, no relay agent in between. Claude stages: one short stage → a plain Agent dispatch with model pinned, effort inheriting the session; fan-out or multiple stages → a Workflow of agent() calls, every lane a labeled row in one tree. A Workflow carries a Codex stage only when a per-item pipeline must mix lanes; then the foreground codex-worker adapter makes one call over files the seat wrote. A one-row tree is an effort adapter, never a shape: only deep verification of another lane's work under a seat running below high earns it, with effort pinned on the agent() call. Invoking /orchestrate is the Workflow opt-in. pipeline() by default; a barrier only where a stage needs every prior result. Too small or too ambiguous to delegate well → state the sequential fallback and do it in the main loop.