Team

/team — two-agent task execution

Read ~/clawd/agent-comms/team/PROTOCOL.md FIRST and follow it exactly, then ~/clawd/agent-comms/team/LESSONS.md. This skill is the Claude Code entry point; the protocol file is the source of truth for classifier, transport, ledger, tiers, and failure rules.

Paths here are home-anchored (~/…) on purpose. agent-comms/team/... is relative and does not resolve from the usual cwd, and ~/agent-comms is a DIFFERENT repo with no team/ dir. A relative path silently skips the protocol and the lessons file.

Procedure

0. Recon before you design. Before proposing any adapter, wrapper, bot, or browser automation, enumerate what the target already exposes — list its API routes, CLI commands, and exported functions FROM SOURCE. Never accept "there is no X" from a README, a report, or the other agent; a negative capability claim rules out whole designs, so verify it directly. Two agents agreeing on an inherited premise is one check, not two. (See LESSONS.md 2026-08-10.)

1. Classify the task with the PROTOCOL.md Classifier. Announce the lead in one sentence.

2. Create the task (you are the receiving agent): node ~/clawd/agent-comms/team/bin/team-ledger.mjs create --slug <slug> --instruction "<task>" --origin claude-code --lead <claude|openclaw> then update --task <id> --status in_progress.

3. If YOU lead:

   • dashclaw_session_start for the task.
   • Decompose. For each piece that belongs to OpenClaw's domain, send a DELEGATION envelope (PROTOCOL.md Transport) via Bash: openclaw agent --agent main --message "<envelope>" --timeout 300 Send the delegation, then log a delegation event (send first, then log). Log the returned answer as received context; MoltFire logs its own reply event itself.
   • Do your own share of the work. Apply the Tier rules from PROTOCOL.md (guard/record/approval) to every consequential action.
   • Integrate, verify, write artifacts to agent-comms/team/tasks/<id>/artifacts/.
   • Log a result event, report to the user in this session, log done, update --status done, dashclaw_session_end.

4. If OPENCLAW leads: send the LEAD_HANDOFF envelope (PROTOCOL.md Transport) via openclaw agent --agent main --message ... --timeout 1800 (the whole task runs inside that call — LEAD_HANDOFF window, not the 300s delegation window), log a lead_assigned event, then act as specialist: answer any DELEGATION envelopes MoltFire sends back during the task, and relay MoltFire's final report to the user verbatim when it arrives in the command output.

5. Failure handling: PROTOCOL.md Failure rules verbatim. Delegation timeout (300s) → log error, retry once, degrade to solo with a note. LEAD_HANDOFF timeout (1800s) → check the task's events-openclaw.jsonl first; fresh events mean the lead is alive — poll the ledger, never spawn a second lead. Cap 10 exchanges.

Hard rules

• You log events ONLY with --from claude.
• Specialists (either direction) answer once and stop. No sub-delegation.
• Never approve anything on Wes's behalf; Tier 2/3 waits for his approval via DashClaw. Fail closed if DashClaw is unreachable.
• No secrets in envelopes, events, or artifacts.