Channel Message Triage Router

Context

Slack messages (and other channel traffic) arrive at the gateway session with no filtering. Every message — Kent asking Nicoll about video trims, bot notifications, thread replies Joel doesn’t care about — consumes gateway context tokens and triggers agent processing. This wastes inference budget and pollutes the agent’s attention.

The gateway session is expensive (Sonnet-class model, 100k+ context). A cheap pre-filter would save 90%+ of wasted tokens on noise.

Decision

Add a triage router that classifies inbound channel messages before they reach the gateway session. Runs as an Inngest function in system-bus, not inside the gateway.

Classification Categories

Architecture

Channel → Inngest event → triage-router function → classified event → gateway
                              ↓
                     cheap LLM (haiku-class)
                     via inference-router catalog

• Model: Haiku-4.5 or cheapest capable model from inference-router catalog (ADR-0140). Use routeInference({ task: "triage", budget: "minimal" }).
• Inference: Via packages/system-bus/src/lib/inference.ts (pi sessions, not paid API keys).
• Debounce/throttle: Thread messages within a configurable window (default 60s) are batched into a single classification. Prevents 8 messages in a thread from triggering 8 LLM calls.
• Channel-specific rules: Hard rules that don’t need LLM — bot messages are always noise, DMs to Joel are always actionable, @mentions are always joel-mention. LLM only for ambiguous messages.

Configuration

~/.joelclaw/triage.yaml:

triage:
  enabled: true
  model: "haiku-4.5"  # or "auto" to let catalog decide
  
  debounce:
    thread_window_s: 60      # batch thread messages within this window
    channel_window_s: 30     # debounce rapid-fire channel messages
    max_batch_size: 10       # classify up to N messages at once
  
  channels:
    slack:
      enabled: true
      always_noise:
        - "#random"
        - "#general"
      always_deliver:
        - "DM"
      suppress_bots: true
    
    telegram:
      enabled: false  # Joel messages directly, no triage needed
    
    discord:
      enabled: true
      suppress_bots: true
  
  overrides:
    # Per-person overrides
    - match: { author: "kentcdodds" }
      default_category: "fyi"  # unless mentions Joel
    - match: { author: "zac" }
      default_category: "actionable"  # Zac's messages usually matter

Prompt Template

The triage prompt is minimal — just enough context:

Classify this Slack message. Reply with ONE word: noise, fyi, actionable.
 
Channel: #{channel_name}
Author: {author} (not Joel)
Thread: {is_thread_reply}
Message: {text}
 
Rules:
- noise: doesn't involve Joel, bot output, automated notifications
- fyi: interesting but not urgent, Joel can see in daily digest  
- actionable: needs Joel's attention or decision

Inngest Function

• Trigger: slack/message.received, discord/message.received
• Debounce: key: "triage-{channel}-{thread_ts}", period: "60s"
• Output: channel/message.triaged event with category field
• Gateway extension filters: only delivers actionable and joel-mention
• FYI messages accumulate in Redis, delivered as batch digest on heartbeat

Mention Nudge

When a joel-mention message is detected, start a response timer in Redis:

• SETEX slack:mention:{ts} 1800 {message_json} (30min TTL)
• Heartbeat check function scans for stale mention keys
• If Joel hasn’t responded (no message from Joel in that thread within window), nudge via Telegram: “You were mentioned by {author} in #{channel} 30min ago — {snippet}”
• Joel’s Slack app handles real-time notifications; this is the safety net for when he misses one
• Configurable window in ~/.joelclaw/triage.yaml under nudge.mention_timeout_m: 30

Consequences

• Gateway context savings: 80-90% reduction in noise messages
• Cost: ~$0.001 per classification with haiku-class model
• Latency: Adds 1-2s for debounce + classification, acceptable for non-urgent messages
• Risk: False negative on actionable — mitigated by joel-mention always-deliver rule and per-person overrides
• Observability: All classifications logged to OTEL via @joelclaw/telemetry

Related

• ADR-0140: Inference Router (model selection)
• ADR-0144: Gateway Hexagonal Architecture (channel interface)
• ADR-0131: Unified Channel Intelligence Pipeline (superseded approach)