Systematic audit of sync primitives, surfaces, help system, and MCP integration conducted through the workspace room. Findings derived from live tool calls, source code review of c15r/sync (platform) and c15r/sync-mcp (MCP server), and dashboard testing.

---

12 custom actions registered, organized by domain:

Task lifecycle — add_task, update_task_status, assign_task, remove_task, clear_done All operate on _shared.tasks array via CEL expr:true writes. clear_done has an if precondition gating on existence of done tasks.

Knowledge capture — add_note, log_decision, link_resource Append-only to respective arrays. Each captures structured objects via CEL expression writes.

Agent status — set_status (broken, see §3.1)

Administration — _set_shared, configure_dashboard

11 views registered:

View	Expression	Purpose
task_count_total	size(state["_shared"]["tasks"])	Metric
task_count_todo	tasks.filter(t, t.status == "todo").size()	Metric
task_count_active	tasks.filter(t, t.status == "in_progress").size()	Metric
task_count_done	tasks.filter(t, t.status == "done").size()	Metric
notes_count	size(state["_shared"]["notes"])	Metric
decisions_count	size(state["_shared"]["decisions"])	Metric
agent_count	size(agents)	Metric
task_board	state["_shared"]["tasks"]	Data
notes_feed	state["_shared"]["notes"]	Data
decision_log	state["_shared"]["decisions"]	Data
resources	state["_shared"]["resources"]	Data
task_summary	status-only projection	Data
status_summary_md	CEL string concatenation	Markdown (broken newlines, see §3.5)
agent_count	size(agents) with render: {type: "metric"}	Render hint test

4-section surface layout:

Overview — view-grid with task metrics (total/todo/active/done) + knowledge metrics (notes/decisions/agents)

Task Board — watch on _shared.tasks + action-form for add_task + action-bar for update/assign/remove + conditional action-bar for clear_done (gated on views["task_count_done"] > 0)

Activity — feed filtered to chat + milestone kinds, with compose

Knowledge — watch on notes/decisions/resources + action-bar for add_note/log_decision/link_resource

---

Views compute (server-side, full CEL). All complex logic lives here.

Surfaces display (client-side, declarative). Reference view IDs and action IDs, no business logic.

Actions mutate (server-side, CEL preconditions + write templates). All state changes flow through actions.

Append-only (via append: true) is ideal for logs. Mutable arrays require full-array replacement via CEL map/filter. Alternative: map-keyed-by-ID pattern eliminates contested writes.

Limited to state.scope.key OP value, views["id"] OP value, basic boolean operators. Complex conditions must be pre-computed into views.

_context.help is situationally computed. _context._expand provides mechanical expansion hints. Requesting only=actions auto-includes _contested view.

---

resolveToken prefers room > agent > view. Room tokens have no identity (self = ""). All ${self} scoped writes fail. All messages show from: null.

Source: sync-mcp auth.ts line 548, sync main.ts line 1526-1533.

Schema comment says views with render_json are surfaces. Dashboard only reads _dashboard.surfaces. The render hint feature is unimplemented in the frontend.

Source: schema.ts line 14-15, Dashboard.tsx line 741.

No _uuid function in CEL. Title-match is fragile.

Ternary wrapping map output fails. Must put ternary inside each field.

CEL \n becomes literal \\n via JSON-RPC serialization.

Deletes and re-creates vault entries instead of UPDATE. Changes entry IDs.

---

Topic	What agents need
surfaces	All 10 types, _dashboard config, enabled expression syntax
write_templates	${} interpolation, expr:true, append, increment, merge
cel_patterns	filter/map/exists/size, ternary gotcha, "key" in pattern
token_types	room vs agent vs view authority, self scope, when to use which

Condition	Suggested key
_dashboard.surfaces present	surfaces
Views with render_json exist	surfaces
${self} write fails	token_types
First expr:true action registered	write_templates
CEL error in write evaluation	cel_patterns

---

Missing patterns: update_in_list (functional array item update), remove_from_list (filter by predicate), clear_list (reset to empty). These are CEL patterns more than registerable actions — they need parameterized field names which can't be expressed generically.

---

• JSON Schema strictness (fixed): bare arrays, untyped properties broke ChatGPT
• GET endpoints unauthenticated: tool list exposed without auth
• Dashboard URL was wrong (fixed): now uses ?room=...#token=...
• All token types now get dashboard links (fixed)

---

1. Auto-surface views with render_json in Dashboard.tsx
2. Add _uuid() to CEL environment
3. Expand _context.help triggers (§4)
4. Add help entries for surfaces, write_templates, cel_patterns, token_types
5. Standard library patterns for list mutation