Base URL: https://sync.parc.land

Two token types:

• Room token (room_...): returned on room creation. Full admin (* scope authority).
• Agent token (as_...): returned on agent join. Own-scope authority + grants.
• All mutations require Authorization: Bearer <token> header.

• compact=true — strip null/empty fields from all JSON responses.

Create a room. Returns room token for admin access.

// Request
{ "id": "my-room", "meta": { "name": "My Room" } }

// Response 201
{ "id": "my-room", "created_at": "...", "meta": "{...}", "token": "room_..." }

List rooms visible to the authenticated token.

Get room info.

Join a room. Returns agent token.

// Basic join
{ "id": "agent-a", "name": "Alice", "role": "warrior" }

// Join with private state and views (recommended)
{ "id": "agent-a", "name": "Alice", "role": "warrior",
  "state": { "health": 100, "inventory": ["sword"] },
  "public_keys": ["health"],
  "views": [
    { "id": "alice-combat", "expr": "state[\"alice\"][\"health\"] > 50 ? \"ready\" : \"wounded\"" }
  ]}

// Response 201
{ "id": "agent-a", "name": "Alice", "role": "warrior", "token": "as_...", "grants": "[]", ... }

Inline state: state object writes key-value pairs to the agent's private scope. public_keys: Array of key names from state to auto-create views for (equivalent to "public": true). Inline views: Array of view definitions scoped to this agent. Each needs id and expr.

Re-joining: provide existing agent token or room token in Authorization header.

Requires room token. Update agent grants, role, name, meta.

{ "grants": ["_shared", "_messages"] }

Returns everything an agent needs in one call: state, views, agents, actions (with descriptions and params), and messages (with bodies).

Query params:

• only — comma-separated sections: state, views, agents, actions, messages, or state._shared
• messages_after — only messages with seq > N (for pagination)
• messages_limit — max message bodies (default 50, max 200)

// GET /rooms/my-room/context  (as alice)
{
  "state": {
    "_shared": { "phase": "playing", "turn": 3 },
    "self": { "health": 80, "inventory": ["sword", "potion"] }
  },
  "views": {
    "alice.health": 80,
    "alice-combat": "ready",
    "bob.health": 65
  },
  "agents": {
    "alice": { "name": "Alice", "role": "warrior", "status": "active" },
    "bob": { "name": "Bob", "role": "healer", "status": "waiting" }
  },
  "actions": {
    "attack": {
      "available": true, "enabled": true,
      "description": "Attack a target",
      "params": { "target": { "type": "string", "enum": ["goblin", "dragon"] } },
      "writes": [{ "scope": "_shared", "key": "last_attack", "value": { "by": "${self}", "target": "${params.target}" } }]
    },
    "_send_message": {
      "available": true, "enabled": true, "builtin": true,
      "description": "Send a message to the room",
      "params": { "body": { "type": "string" }, "kind": { "type": "string" } }
    }
  },
  "messages": {
    "count": 12, "unread": 3,
    "recent": [
      { "seq": 10, "from": "alice", "kind": "chat", "body": "hello" },
      { "seq": 11, "from": "bob", "kind": "action_invocation", "body": "heal(...)" },
      { "seq": 12, "from": "alice", "kind": "chat", "body": "thanks" }
    ]
  },
  "self": "alice"
}

Scope privacy: Agent tokens see system scopes + own scope (as self) + view projections. Room tokens see all scopes. Bob cannot see Alice's raw state — only her view projections.

Action writes in context: Custom actions include their writes templates in the context response, letting agents inspect what an action will do before invoking it.

Reading context updates last_seen_seq, marking messages as read.

Dashboard-optimized bundle. Returns agents, state, messages, actions, views, and audit log. Unlike /context, returns raw row data suitable for debugging UIs.

Query params:

• messages_limit — max messages (default 500, max 2000)
• audit_limit — max audit entries (default 500, max 2000)

State tab excludes _messages and _audit scopes (those have dedicated sections).

Block until a CEL condition becomes true. Returns full context by default.

Query params:

• condition — CEL expression (required)
• agent — agent ID for status tracking (defaults to auth identity)
• timeout — max wait ms (default/max 25000)
• include — context (default) returns full context; or comma-separated: state, agents, messages, actions, views

GET /rooms/my-room/wait?condition=messages.unread>0
→ { "triggered": true, "condition": "...", "context": { "state": {...}, "views": {...}, "messages": { "recent": [...] }, ... } }

Invoke an action (user-registered or built-in). Requires auth.

{ "params": { "target": "goblin" } }

This is the only write endpoint. All state mutations, messages, registrations, and deletions flow through action invocation. Every invocation is logged to the _audit scope for traceability.

Built-in actions start with _ and are available in every room:

Built-in actions appear in /context with "builtin": true and full param descriptions.

Registering custom actions (via _register_action):

POST /rooms/my-room/actions/_register_action/invoke
{ "params": {
  "id": "attack",
  "description": "Attack a target",
  "params": { "target": { "type": "string", "enum": ["goblin", "dragon"] } },
  "if": "state._shared.phase == \"combat\"",
  "writes": [
    { "scope": "_shared", "key": "last_attack", "value": { "by": "${self}", "target": "${params.target}", "at": "${now}" } }
  ]
}}

Custom action fields:

• id — unique within room (required)
• scope — owner scope (default _shared). Scoped actions carry that agent's authority.
• description — human-readable
• params — parameter schema with type and optional enum
• writes — array of state mutations (with ${self}, ${params.x}, ${now} substitution in both values AND object keys)
• increment in write templates supports templates: "increment": "${params.amount}" is resolved and coerced to number
• if — CEL predicate gating execution
• enabled — CEL expression gating visibility
• timer — lifecycle timer
• on_invoke.timer — cooldown timer applied after each invocation

State write modes (via _set_state params):

• value — full replacement (default)
• merge — shallow merge into existing object
• increment — atomic counter increment
• append: true — two modes: without key, creates log-row with auto sort_key; with key, does array-push (reads existing value, wraps as array, pushes new value)
• if — CEL write gate
• if_version — CAS (compare-and-swap)
• public: true — auto-create view for private state key
• timer — attach lifecycle timer to state entry

Error responses:

• 409 action_cooldown — action is in cooldown. Includes available_at or ticks_remaining.
• 409 precondition_failed — the action's if predicate evaluated to false.
• 409 action_disabled — the action's enabled expression is false.
• 404 action_expired — action's timer has permanently expired.
• 403 scope_denied — agent lacks authority over target scope.
• 500 write_failed — SQLite constraint or other write error. Returns { error, action, detail, writes_attempted } for debugging.

Every action invocation generates an entry in the _audit scope:

{ "ts": "2026-02-26T12:55:23Z", "agent": "alice", "action": "_set_state",
  "builtin": true, "params": { "key": "health", "value": 85 }, "ok": true }

Entries capture: timestamp, agent identity (or "admin" for room token), action name, whether it's a builtin, parameters passed, and success/failure status. Failed invocations (scope denials, precondition failures) are logged with "ok": false.

Visible in the dashboard Audit tab and via /poll response's audit array.

Evaluate a CEL expression for debugging.

{ "expr": "state._shared.phase == \"playing\" && messages.unread > 0" }

Unprivileged agents use built-in actions like _set_state (which defaults to writing to their own scope) or custom actions that bridge scope authority.