--- name: rdc:co-develop description: "Usage `rdc:co-develop ` — peer-aware Claude/Codex co-development over clauth, with ask --wait as the default delegation pattern." --- > **⚠️ OUTPUT CONTRACT (READ FIRST):** `guides/output-contract.md` > Checklist-only output. No tool-call narration. No raw MCP/JSON/log dumps. > One checklist upfront, updated in place, shown again at end with a 1-line verdict. > **Sandbox contract:** This skill honors `RDC_TEST=1` per `guides/agent-bootstrap.md` § RDC_TEST Sandbox Contract. Under the flag, do not create/send/reply to live co-development turns; report the intended clauth command and mark the external relay action as skipped. # rdc:co-develop — Peer-Aware Co-Development Use this skill when an RDC workflow needs Claude and Codex to work together in one governed session. This skill uses the clauth `codevelop_*` route only. Do not use `chitchat_*` for this workflow. ## Arguments Use the high-level delegation commands by default: - `rdc:co-develop ask ""` — delegate to partner and wait for the reply - `rdc:co-develop reply ""` — reply to a partner request with evidence - `rdc:co-develop inbox` — drain your inbox when you are explicitly acting as the worker - `rdc:co-develop start ` — create a session and return startup details - `rdc:co-develop resume --peer ` — join as a peer - `rdc:co-develop send --session --from --to ` — send a turn - `rdc:co-develop poll --session --peer ` — drain messages for a peer - `rdc:co-develop status [session-id]` — show session status and queues - `rdc:co-develop stop ` — stop the session ## Default Delegation Pattern When you need partner help, do not hand-write `/codevelop/send` JSON unless the CLI fails. Call the partner like a subagent: ```powershell clauth codevelop ask --to codex --wait --task "Review this plan and identify blockers." clauth codevelop ask --to claude --wait --task "Review this implementation and advise." ``` `ask --wait` sends the request, captures the `turn_id`, polls the sender inbox, and returns the matching partner reply to the current live agent. This preserves the sender's active conversation context. When you receive a partner request, complete the bounded task and reply with: ```powershell clauth codevelop reply --turn turn-0001 --verdict pass --summary "Reviewed plan." --evidence "Checked files X and Y" --files-changed "" ``` For manual worker intake: ```powershell clauth codevelop inbox --peer codex clauth codevelop inbox --peer claude ``` ## Required Transport HTTP routes: ```text POST /codevelop/start POST /codevelop/join POST /codevelop/send POST /codevelop/poll GET /codevelop///stream GET /codevelop//status GET /codevelop POST /codevelop/stop ``` MCP tools: ```text codevelop_start codevelop_join codevelop_send codevelop_poll codevelop_stream codevelop_status codevelop_stop ``` ## Session Roles Default peers: | Peer | Role | Responsibility | |---|---|---| | `claude` | `supervisor` | RDC loop, work items, validation, integration | | `codex` | `implementation_partner` | implementation, focused audit, repair, evidence | Other valid turn roles: `planner`, `builder`, `reviewer`, `validator`, `advisor`, `release_guard`, `infra_operator`. ## Message Envelope Always send structured JSON. The receiving peer must treat the JSON object as the source of truth. ```json { "session_id": "", "turn_id": "turn-0001", "from": "claude", "to": "codex", "type": "audit_request", "role": "reviewer", "skill": "rdc:review", "task": "Audit the current diff for regressions before merge.", "context": { "repo": "{PROJECT_ROOT}", "work_item_id": "", "branch": "develop", "owned_files": [] }, "expect": { "response_format": "CO_DEVELOP_REPLY", "evidence_required": true, "commit_allowed": false } } ``` Required reply shape: ```json { "session_id": "", "turn_id": "turn-0001", "from": "codex", "to": "claude", "type": "reply", "verdict": "pass", "summary": "One paragraph.", "evidence": [], "files_changed": [], "commits": [], "blockers": [], "next": [] } ``` ## Escalation Rules Invoke co-development when: - Dave asks for partner review, audit, second opinion, or co-development - a task crosses multiple apps, packages, repos, APIs, database surfaces, or deployment surfaces - `rdc:plan` creates 5 or more work packages, or 3 or more work packages with integration risk - `rdc:build` has file overlap or integration risk - `rdc:review` finds failures that need judgment - the same implementation, test, build, deploy, or migration path fails twice - architecture, auth, deployment, schema, source-of-truth, or bridge-mode interpretation is uncertain - a production-facing release path is about to proceed Do not invoke by default for simple read-only explanations, one-line command output, one-file typo fixes, or already decomposed low-risk work. ## In-Session Escalation — Truth-Gain Protocol Once a co-development session is running, bound the debate with the shared escalation protocol: `guides/escalation-protocol.md`. In short: after each peer round, both agents rate marginal truth-gain (Δ). When two consecutive rounds are sub-5% for both agents, stop debating and decide — build a scored rubric (options × weighted criteria, both scoresheets recorded). Non-critical decisions adopt the rubric winner; critical decisions escalate HITL to the human decision table (interim: surface to Dave + record as work-item note / CodeFlow memory until `human_items` ships). See the guide for the full criticality list and acceptance criteria. ## Workflow 1. Start or resume a session with `codevelop_start` / `codevelop_join`. 2. Join the local peer with a stable `peer_id`. 3. Delegate with `clauth codevelop ask --to --wait --task ""`. 4. Treat the returned JSON as the partner result and continue your active work. 5. If acting as the worker, drain inbox with `clauth codevelop inbox --peer `. 6. Reply with `clauth codevelop reply --turn ...`. 7. Record evidence in the calling RDC skill report or work-item implementation report. 8. Stop the session with `codevelop_stop` when the work is done. ## Completion Evidence Every co-development turn must report: - session ID and turn ID - peer IDs - files changed - commands/tests/routes checked - commits, if any - blockers or decisions needed - next action