Talking to the Guide

When you open a thread on a Syncropel instance and want help with the instance itself — figuring out what's there, what records exist, how to do something — the agent at the other end of that thread is the Syncropic Guide (did:sync:agent:guide).

The Guide is a sibling to Scribe. Scribe is the general-purpose conversational agent; the Guide is the specialised orientation agent. They share the same conversation shape — type a message, the agent responds, the whole thread is records — but the Guide is primed for a different job: getting you (or a new teammate, or a federation peer) oriented inside your instance.

What the Guide is for

The Guide answers questions like:

• "What's on this instance? What threads, what actors, what records have been emitted today?"
• "Where do I find the consent grants? The federation pairs? The task content?"
• "How do I emit a record? Open a thread? Run a query?"
• "What does this body kind mean?"

The Guide is intentionally instance-aware. It knows your instance's structure — which threads carry which body kinds, what your tier permits, what well-known threads exist — and orients its answers around what it can see in your records, not a generic tour.

The Guide is not a general-purpose chatbot. For "summarise this paper" or "write me a regex", talk to Scribe.

How a conversation works

Identical to a conversation with Scribe:

1. You emit a core.user.message.v1 record on a conversational thread.
2. The thread routes to the Guide because the Guide is a registered chat-actor.
3. The Guide reads the thread + your instance state and emits a response record — core.guide.response.v1, on the same thread.
4. Your client renders the response like any other record.

Each new message on the same thread folds into the next response, so multi-turn context falls out for free.

Delegating multi-step work — @agent <goal>

Like Scribe, the Guide can delegate to a real work loop when you open a message with @agent <goal>. The same rules apply:

• The loop runs as you (under your DID, your tier, your permissions) — not as the Guide.
• The engine's Evaluator decides terminal success — the loop never self-asserts "done".
• Trial-tier actors always run in verify mode; Paid and above can opt in.

A core.work.delegation.v1 record opens the delegation; the loop's outcome closes it. See Work loops for the full primitive.

The three personas

The Guide ships with three built-in personas — each is a different conversational stance, useful for different audiences:

The persona is configured at the instance level: a blueprint emits a core.guide.priming.v1 record that picks the persona by name. The persona text itself is built in — a blueprint contributes the choice, not the words. This keeps the Guide's surface predictable across instances.

To switch persona on your own instance:

# Emit a guide-priming record (REPLACES any prior priming)
spl record emit \
  --thread th_engine_config \
  --kind core.guide.priming.v1 \
  --body '{"persona": "research_assistant"}'

Changes take effect on the next message routed to the Guide.

Using the Guide (hosted)

Nothing to configure. The Guide ships primed to the general persona; open a thread and ask it about your instance.

The Guide on a self-hosted instance

A self-hosted instance you run with spl serve has the Guide too — and, like Scribe, it works out of the box. Your instance ships with managed model access built in, so the Guide can think and respond from the first conversational thread without you configuring anything.

If you want to use your own account or a specific model instead, you can override that default:

Default — managed model access (nothing to configure)

Your instance already has managed model access. Open a conversational thread and address the Guide — it responds. This is the default, and most people never need to change it.

Optional override — bring your own provider key

If you would rather your instance call a model under your own account, configure a provider + key in its [intelligence] configuration. Supported options include the major hosted model providers and a local Ollama runtime. See First run.

Either way, the Guide responds in any conversational thread you address to it.

Troubleshooting — the Guide is silent

If you send a message and nothing comes back, the checklist is the same as for Scribe:

1. A model override is configured but rejecting calls. If you set your own provider key in [intelligence], an expired or invalid key or an exhausted quota stops the Guide responding. Run spl doctor — it surfaces model-configuration problems — and check your key and subscription. (With the built-in managed model access and no override, this isn't a concern.)
2. The message landed on a non-conversational thread. The Guide only answers on conversational threads.
3. Disabled instance-wide. An operator can set agentic_loop_enabled: false on the Guide's core.engine.actor_adapter.v1 record to disable @agent delegation only; the conversational surface still works.

Audit trail

Every Guide interaction is observable. The well-known thread th_audit_guide carries the call/response/error records emitted by the Guide adapter — useful for operators auditing what the Guide sees and says.

Where to go next

• Scribe — the general-purpose conversational agent the Guide is a sibling to.
• Work loops — the multi-step primitive @agent delegates to.
• Workspaces — how blueprints (which set the Guide's persona) work.