attention-router

attention-router

Triage AI coding agents' questions. A 5-persona LLM council auto-resolves what's safe, surfaces only the highest-value decisions to you.

You run several AI coding agents in parallel. You miss half their questions because they're in terminals you tabbed away from. The other half are vague — "what should I do?" — and force you to rebuild context.

attention-router fixes both. Agents send structured asks to a local daemon. A council of 5 LLM personas votes. Naked questions are rejected with repair instructions. Safe decisions auto-resolve silently. Only the genuinely ambiguous, high-stakes forks reach you, ranked by urgency × score, never more than 3 at a time.

Table of contents

• Demo
• Install — Claude Code plugin (recommended)
• How it works
• The human side
• Other agents (raw HTTP)
• Configuration
• HTTP API reference
• Decision card format
• Architecture
• Development
• Status
• License

Demo

agent ── POST /asks/structure ──┐
                                ▼
                     ┌──────────────────────┐
                     │  attention-router    │
                     │  ┌────────────────┐  │
                     │  │ council (×5)   │  │  ← Claude personas vote
                     │  │ entropy / loss │  │     in parallel
                     │  │ reversibility  │  │
                     │  └───────┬────────┘  │
                     │          │           │
                     │   ┌──────┴──────┐    │
                     │   ▼             ▼    │
                     │ auto         queued  │
                     │ _resolved   for human│
                     └────┬─────────────┬───┘
                          │             │
              proceeds ◀──┘             └──▶  attn next  →  YOU
              silently                        (1-3 cards, ranked)

Decision card you actually see (ASCII):

billing-api · ~30s requested · expires in 25m

PICK BACKFILL STRATEGY FOR USERS_PROFILE TABLE

A. Inline backfill in the migration
   Evidence: lock window <30s on test data, atomic with schema change
   Next: run UPDATE in same txn
   Risk: lock contention on prod; rollback safe

B. Background job (Sidekiq), schema first
   Evidence: safer for live traffic, observed pattern across team
   Next: queue 1000 batches of 50k rows
   Risk: schema drift if job lags > 1h

C. Maintenance window tonight
   Evidence: zero risk, but delays release by 12h
   Next: schedule 02:00 deploy
   Risk: pager wakes you up if it slips

Agent default: B    Council prediction: B    (entropy=0.971)
Why this reached you:
  - urgency=now (reversibility=costly)
  - expected loss is high (estimated $5000)
  - council split between A and B

Reply:  A / B / C / override: <text> / skip

Install — Claude Code plugin (recommended)

/plugin marketplace add sderosiaux/claude-plugins
/plugin install attention-router@sderosiaux-claude-plugins

That's it. The plugin ships:

Component	Role
MCP server (stdio)	Exposes ask_human, wait_for_decision, get_pending tools to Claude
ask-human skill	Teaches Claude when to invoke ask_human (vs. resolving the fork itself)
SessionStart hook	Idempotently launches the daemon on 127.0.0.1:7777 if not already up
CLI (attn)	Your side: attn next, attn batch, attn decide, attn status

Requirements

• Node ≥ 20 in PATH
• An Anthropic API key as THE_ANTHROPIC_API_KEY (Claude Code overrides ANTHROPIC_API_KEY=dummy, so the daemon's hook reads THE_ANTHROPIC_API_KEY first). Without a key, the council falls back to MockProvider (deterministic but useless beyond smoke tests).
• No MCP_TIMEOUT tweak needed since v0.2.0. ask_human polls briefly (30s) and returns; the plugin's Stop hook (asyncRewake: true) then waits in the background and wakes the agent with the decision when the human responds — even hours later. If you previously set MCP_TIMEOUT=14400000, you can drop it.

Your daily flow