UAPI Multi-Agent Coordination Directive
Active coordination directive for UAPI multi-agent work: Codex coordinates, operator approves gates, agents avoid unnecessary operator interruptions and secret scans.
Promotion Metadata
- Source feedback id:
90 - Source feedback ids:
90 - Promotion request id:
24 - Feedback category:
beneficial - Feedback source:
ai-lane - Feedback created at:
2026-06-22T16:43:04.350934+00:00 - Target slug:
[REDACTED:high-entropy] - Review state:
review - Reviewer note: AI lane auto-promoted redacted feedback into internal catalog context.
Source Feedback Body
Feedback 90
- Product:
uapi - Category:
beneficial - Related slug:
[REDACTED:high-entropy] - Source:
ai-lane - Created at:
2026-06-22T16:43:04.350934+00:00
title: UAPI Multi-Agent Coordination Directive project: uapi kind: runbook state: active tags: [docs, runbook, uapi, coordination, multi-agent]
UAPI Multi-Agent Coordination Directive
Authority split: - Human operator: final approval for access, secrets, credentials, destructive actions, live infrastructure mutation, budget, purchases, and policy exceptions. - Codex: day-to-day coordinator for the UAPI multi-agent wave. Codex owns schema names, integration gates, conflict resolution, and final synthesis. - Grok: hardware, autorouter, 140T/NPU, SN850X, Colab/cloud GPU, and friendly route evidence. - GLM 5.2 or Claude CLI: Interpreter/Conductor first-wedge builder behavior, mock-vs-real adapter review, gates, and demo acceptance.
Agent question rule: do not bounce ordinary architecture or task-splitting questions to the operator. Use the Explorer runbooks first, then ask Codex through the coordination channel or publish a blocked diary/feedback item. Interrupt the human operator only for secrets, login/access, paid/external actions, destructive changes, live lab mutation, or genuinely ambiguous product decisions.
Required first answer for any CLI asking who is running coordination: Codex coordinates the project; the human remains final approving operator. Continue from Explorer source of truth, keep a diary, and report blockers with audit ids.
Safety rule: agents must not search broad home directories for API keys or secrets. Use environment variables supplied by the operator or the specific approved session context only. If a needed credential is missing, report missing access instead of scanning.
Immediate source contexts to fetch: master program control, assigned agent runbook, Codex core contracts, auth identity contracts, job routing contracts, embedding storage contracts, and the integrated UAPI operating plan.