collab-evals

name	collab-evals
description	Run collab/multi-agent eval scenarios (symbolic RLM, large-context, pause/resume, multi-hour checkpoints) and capture manifest-backed evidence.

Collab Evals

Use this skill to run repeatable collab evaluation scenarios and record evidence. Keep scope to evals; do not implement unrelated fixes.

Quick start

0. Confirm feature readiness:

• Run codex features list and verify multi_agent is enabled.
• In this skill, "collab" refers to the same multi-agent tooling path; collab naming remains in legacy keys like RLM_SYMBOLIC_COLLAB and manifest.collab_tool_calls.

1. Pick the scenario(s):

• Large-context symbolic RLM with collab subcalls.
• Multi-hour refactor with checkpoints.
• 24h pause/resume context-rot regression.
• Multi-day initiative (48–72h) with multiple resumes.
• Additive config simulation: verify updates merge into existing user config without destructive overwrite.
• RLM default-capability simulation: verify built-ins-first behavior (default/explorer/worker/awaiter) before custom overlays.
• Docs relevance simulation: delegated doc-audit stream checks stale/irrelevant guidance before hard-gating.

2. Ensure task context:

• export MCP_RUNNER_TASK_ID=<task-id>

3. Run the scenario using codex-orchestrator start <pipeline> --format json and record the manifest path.

Evidence checklist

• Manifest path under .runs/<task-id>/cli/<run-id>/manifest.json.
• Log path under .runs/<task-id>/cli/<run-id>/runner.ndjson.
• Findings recorded in docs/findings/<date>-<topic>.md.
• Task mirror update in docs/TASKS.md and task spec.

Guardrails

• Collab is additive; keep MCP as the control plane for approvals and audit trails.
• Cap collab event capture with CODEX_ORCHESTRATOR_COLLAB_MAX_EVENTS when needed.
• If pause/resume is required, use control endpoints or codex-orchestrator resume with manifest evidence.
• Keep top-level defaults on latest baseline model and model_reasoning_effort >= high; avoid role sprawl in eval lanes.
• Treat fallback profiles as exception paths in scenarios and record why baseline settings were insufficient.
• For non-trivial eval runs, include at least one delegated research/review stream and summarize evidence in parent output.

Skill split policy

• Keep scenario/mock/simulation coverage in this skill, while scope stays focused on collab/multi-agent behavior.
• Propose a dedicated simulation skill only when repeated non-collab simulation workflows make this skill too broad.

Related skills

• collab-subagents-first: for production stream ownership patterns mirrored in eval scenarios.
• delegation-usage: for delegation MCP setup and lifecycle semantics.
• long-poll-wait: for multi-hour/multi-day eval monitoring with durable checkpoints.