research-hub Multi-AI Router

This skill is the router in a router/leaves architecture. The leaves are codex-delegate and gemini-delegate. The router's only job is to write a coordination plan; the leaves read their assigned task brief from that plan and execute.

When to Invoke

Invoke this skill only when a single round of work will need two or more delegates. Examples:

• Codex writes the test scaffolding and Gemini drafts the zh-TW companion doc — same round, both delegates.
• Three parallel Codex runs against independent subtrees of the same repo, with a final Claude reconciliation step.
• A sequence of mixed handoffs: Codex generates data, Gemini summarises in zh-TW, Codex writes a plot script.

If only one delegate is needed this round, use the leaf skill directly. Do not produce a router plan for a one-task delegation — it adds ceremony with no benefit.

When Not to Invoke

• Single delegate this round (just Codex, just Gemini): use the leaf skill.
• Single-LLM research-hub auto runs that pass --llm-cli codex or --llm-cli gemini: the flag is a research-hub feature, not a router decision. See "Single-LLM research-hub routing" below for reference.
• Pure-Claude planning, review, or judgment work: keep it in Claude.

Prerequisite check

The router emits commands that go through the research-hub CLI in many cases. Before producing a plan, verify the CLI exists:

research-hub doctor

If that command is not found (vs. emitting a health report), the host has loaded the skill instructions but the Python CLI is missing. Stop and tell them:

This skill orchestrates research-hub CLI commands across multiple AIs, but the CLI itself isn't installed. Please run:

pip install research-hub-pipeline
research-hub setup --persona researcher   # or analyst | humanities | internal

Then re-run your request.

Do not produce a plan that calls research-hub ... if the CLI is missing — the user can't execute it. Plans that only call codex-delegate / gemini-delegate (no research-hub CLI) are fine without it.

Output artifact: .coord/multi_ai_plan.md

The router's only output is .coord/multi_ai_plan.md at the workspace root. Every plan has YAML frontmatter and a per-task brief block.

Schema:

---
plan_id: <short-slug>             # e.g. paper-corpus-2026-05
created_utc: 2026-05-09T12:34:56Z
goal: <one paragraph>
success_criteria:
  - <observable check 1>
  - <observable check 2>
tasks:
  - id: t1
    agent: codex                  # codex | gemini | claude
    brief_path: .ai/codex_task_t1.md
    depends_on: []
    success_criteria:
      - <verification command or assertion>
  - id: t2
    agent: gemini
    brief_path: .ai/gemini_task_t2.md
    depends_on: [t1]
    success_criteria:
      - <verification>
risks:
  - <known risk>
---

A ready-to-paste template lives in references/multi_ai_plan_template.md.

The router writes the plan and writes each per-task brief at the path declared in brief_path. Leaves read the brief and execute.

Hand-off contract

The plan is the source of truth. After the router writes it:

1. Each leaf is invoked with the brief path as input. Leaves do not read the plan — they read their own brief.
2. Claude (or another reconciler) reads result.json from each leaf, compares against success_criteria in the plan, and decides accept / re-run / fall back.
3. Plans are append-mostly: if a re-run is needed, append a new task with a fresh id rather than mutating an existing one.

Single-LLM research-hub routing (reference only)

The research-hub auto command takes a --llm-cli flag to pick which CLI handles long mechanical work like crystal generation. This is a research-hub feature, not router logic. Use it directly without invoking this skill:

research-hub auto "TOPIC" --with-crystals --llm-cli codex
research-hub auto "TOPIC" --with-crystals --llm-cli gemini

Pick codex for token-heavy mechanical generation; pick gemini for long-context reading or CJK output. If you also need a second delegate in the same round (e.g. translate the resulting brief to zh-TW with Gemini), then this becomes a multi-AI plan and the router applies.

Check available CLIs:

python -c "from research_hub.auto import detect_llm_cli; print(detect_llm_cli())"

Guardrails

• Do not produce a plan with a single task — that is a misuse; use the leaf skill directly.
• Do not fabricate citations or metadata in either the plan or the per-task briefs.
• Do not assume Gemini is Chinese-only; route by task character (long context, CJK prose, second-opinion review), not by language alone.
• Do not overwrite .coord/multi_ai_plan.md of a different plan_id — append a new file .coord/multi_ai_plan_<plan_id>.md if the workspace already has an active plan.
• Do not overwrite vault notes or delete clusters without explicit user approval.