Skip to main content

co-agent

Second opinions, decisions, ADRs, and implementation pipelines with peer review. The current host chairs the work.

Six skill modes​

ModeResultExecution requirement
reviewEvidence-checked findings and disagreementsAvailable peers; solo allowed with notice
decideOption comparison and the host's recommendationAvailable peers; solo allowed with notice
adrAlternatives, tradeoffs, and an ADR draftAvailable peers; solo allowed with notice
sync-contextShared AGENTS.md and Kiro steering bridgeLocal context validation
consensusHost implementation with plan and final review gatesREADY peer required
harnessWorktree implementation by an eligible peer or explicitly selected host modeFresh READY raw-CLI reviewer required; host owns verification and commits

When no eligible implementer is READY, harness requires an explicit, authorized host-mode choice through --allow-host-implementation. Configuration and readiness errors must be repaired first. Both modes require a fresh, gate-eligible reviewer; see the harness contract.

/co-agent:setup is a separate command that probes installed/authenticated peer CLIs and records readiness before the gated workflows run.

Claude chairs Claude Code sessions; Codex chairs Codex sessions. The host does not call itself as an external peer. External responses are advisory: the host checks evidence, attributes findings, and reports meaningful disagreement.

The active CLI candidates are Kiro plus the opposite host CLI: Codex when Claude hosts, or Claude when Codex hosts. Antigravity (agy) and the legacy Gemini CLI are retired. Refresh setup after migrating old provider settings. The v2.0.0 migration guide describes the supported peers and implementation choices.

Components​

The Claude package declares co-agent, gate-chair, harness-analyst, pr-autofix-planner, and pr-autofix-implementer; its skills are co-agent, pr-autofix, and decision-reconcile. The planner and implementer are internal PR workers that require prepared inputs. Codex exposes these workflows through generated overlays, including command wrappers.

Configuration and gates​

Use /co-agent:configure to inspect merged configuration, its source, and host-specific peer availability. The defaults file defines profiles, model lists, effort, timeouts, context limits, call budgets, and retry bounds. A deep profile can use multiple models per peer; a default profile uses the single model setting. Do not infer a provider's supported model catalog from a configured string.

Canonical defaults

pr_gate.enabled and push_gate.enabled default off. These optional local hooks are separate from mandatory CI and branch protection. Missing, failed, or stale required review coverage is not a successful review.

Context synchronization​

/co-agent:sync-context distills CLAUDE.md into a marked AGENTS.md and creates .kiro/steering/project-context.md as a bridge. The source hash detects drift; unmarked handwritten files are protected. Use /co-agent:configure set autosync on to opt into automatic synchronization prompts.

PR autofix applies review feedback. Decision reconcile finds contradictions across ADRs and current code. Ordinary code review or an unqualified “decide” request does not by itself request a multi-AI panel; explicitly invoke co-agent when peer input is wanted.