Skip to content

Agents

An agent is a Markdown contract: runtime instructions plus tool allow-lists interpreted by an executor plugin. Coro swaps agents at phase boundaries so each step has a narrowly scoped mandate.

Below is the full roster shipped in base intelligence—not every agent joins every workflow, but recognition helps when reading logs.

Agent fileRole in one sentence
spec-writerTurns tracker payloads into reviewer-ready specs the Planner can execute.
plannerReads the repo after spec hand-off, detects language and risk, emits work items, and may reroute FAST vs STANDARD vs DEEP—or convert to campaigns.
analyzerDEEP only — Produces authoritative design notes before decomposition when public surface or cross-cutting risk demands it.
coderImplements planned work locally, invokes convention skills on demand, and opens PRs.
code-reviewerRuns as a constrained sub-phase inside coding to sanity-check conventions, tests, and plan alignment—fast feedback before humans see the PR.
pr-reviewerMerge gatekeeper: human SCM coordination, policy checks, iterative fix loops with the coder—then merge.
evaluatorPost-merge verifier: rebuilds/tests against merged commits, validates acceptance criteria, captures memory deltas and improvement proposals.
qaDEEP lane enrichment after merge—extra testing rigour layered on evaluator outputs when the lane demands it.
campaign-architectCampaign prelude — Captures shared architecture across many child repos/issues before decomposition.
campaign-plannerFans out coordinated child jobs with explicit dependency and contract scaffolding.
campaign-integratorReconciles cross-child contracts and verifies integration checkpoints near the epic close.
campaign-evaluatorAggregates campaign-level outcomes and surfaces systemic follow-ups.
memory-curatorThe only subtractive memory editor—merges and compacts append-only memory safely under human review.

Text “diagram” of a typical STANDARD job

Tracker / CLI ──► spec-writer ──► planner ──► coder ⇄ code-reviewer
▼ (PR open)
pr-reviewer ⇄ humans
merge / complete WI loop
evaluator ──► memory + proposals (optional paths)

Campaign flows insert architect → planner loop → per-child STANDARD/FAST/deep forks → integrator → campaign-evaluator. DEEP inserts analyzer before planning and activates QA after merge verification.

Exact YAML phase lists live beside each workflow file in the overlay you ship; when in doubt consult the workflow reference.

Agents reference tables