Skip to content

Relay

Relay takes a single, well-specified task and runs it through the same phased pipeline as everyday Orchestration — plan → architect → implement → review — but with one change: every phase is executed by a CLI vendor lane instead of a sub-agent. The Conductor stays in charge, hands the task from one vendor to the next, and writes each phase’s output to a .ptah/specs/TASK_* folder so the whole run is auditable and resumable.

Where Forge and Race give every vendor the same prompt and pick a winner, Relay gives each vendor a different prompt — one per phase — and runs them as a sequential pipeline. The diversity shows up in one place: the review phase is always handled by a different vendor family than the one that wrote the code, so you get genuine cross-vendor review baked into delivery.

  • You want orchestration’s structured delivery — a real plan, an architecture doc, an implementation, and a review — but you want the work done by external CLI vendors rather than in-process sub-agents.
  • Cross-vendor review on your own changes — the implementer and the reviewer are deliberately different vendor families.
  • An auditable, resumable run — every phase is persisted to .ptah/specs, so you can review each artifact or resume a run that timed out.

The Conductor creates a .ptah/specs/TASK_* folder, restates the task with explicit acceptance criteria, and hands the planning and architecture phases to a reasoning-strong vendor lane. Each phase writes its deliverable (task-description.md, then implementation-plan.md) to disk. You review and approve these documents before any code is written — the same checkpoints as a normal Orchestration run.

The approved plan is handed to a coding-strong vendor lane, which implements the change in place on your working branch and logs its work to tasks.md. The artifact from each phase becomes an input to the next — that hand-off is the “relay baton”.

A different vendor family reviews the implementation against the acceptance criteria and writes code-logic-review.md. Because the reviewer never wrote the code, this is a true peer review, not a self-check. If the review surfaces a real issue, the Conductor relays a fix phase before declaring the task done.

The Conductor runs the project’s tests/build/lint on the change, then produces a summary that cites which vendor produced each artifact. You see the diff and approve before anything is committed.

  • Runs in place on your active branch — no worktrees by default, since it’s one task rather than N competing attempts.
  • Never commits without your sight and never auto-merges to main.
  • Vendors never commit — each lane does its phase and hands back; the Conductor owns verification and the final commit.
  • Resumable — the .ptah/specs/TASK_* folder lets a timed-out run pick up where it left off instead of starting over.

Natural language triggers:

  • “Relay this task across the panel”
  • “Run this as a vendor pipeline — plan, build, then a different vendor reviews”
  • “Orchestrate this with CLI vendors instead of sub-agents”

Explicit harness: select Tribunal Conductor from the harness picker, then describe the task and its acceptance criteria.

MovePrompt per vendorShapeProduces
CouncilSame questionParallel panelA cited verdict (no code)
ForgeSame coding taskParallel + reviewBest of N implementations, merged
RaceSame coding taskParallel + verifyOne verified winner
RelayDifferent prompt per phaseSequential pipelineDelivered change + spec artifacts
  • Needs a clear task and acceptance criteria — a vague brief produces a vague plan, and the whole pipeline inherits the fuzz.
  • Cost — roughly one vendor call per phase (more if a high-stakes phase is fanned out to multiple vendors). Ptah announces the lanes and cost before spending.
  • Heterogeneous, not a panel — Relay deliberately bends Tribunal’s “diversity is the signal” thesis; the diversity here is the cross-vendor review phase, not N answers to one question.