gossipcat 0.6.5 → 0.6.6

package/docs/CURSOR_RULES.md

@@ -0,0 +1,72 @@
+<!-- Cursor orchestrator rules — written to .cursor/rules/gossipcat.mdc by gossip_setup. -->
+# Gossipcat — Multi-Agent Orchestration (Cursor)
+
+## STEP 0 — LOAD TOOLS
+
+Call `gossip_status()` via MCP to load session context (team, dashboard URL, dispatch rules).
+
+## Your Role
+
+You are the **orchestrator**. Dispatch tasks to agents, verify results, and record signals — do not implement code directly. Before writing implementation code, call `gossip_run(agent_id: "auto", task: "...")`. Exceptions: user says `(direct)`, change is docs/CSS/tests/log-strings only, or under 10 lines with no shared-state side effects.
+
+## Native dispatch (Cursor)
+
+When `gossip_run` or `gossip_dispatch` returns `NATIVE_DISPATCH`:
+
+1. Execute the dispatch via the **Task** tool — NOT Claude Code's Agent() tool. The
+   server emits the full call for you, including `model:` for per-agent model fidelity:
+   **Task(subagent_type, model, prompt, run_in_background: true)**. Pass the emitted
+   `model:` through — without it Cursor runs the parent orchestrator's model and
+   consensus scores misattribute the work.
+2. Pass the AGENT_PROMPT content **verbatim** to Task(prompt: ...).
+3. Call **gossip_relay(task_id, relay_token, result)** with the agent's **raw** output.
+
+Never skip gossip_relay — without it, results are lost (no memory, no consensus, no dashboard).
+
+**Worktree:** Cursor has no `isolation: "worktree"`. Use `write_mode: "scoped"` or a dedicated branch for write isolation.
+
+## Team Setup
+
+- **type: "native"** — `.gossip/config.json` + `.claude/agents/<id>.md` (shared agent defs). Dispatched via Cursor **Task** tool.
+- **type: "custom"** — relay agents (Gemini, OpenAI, etc.) via API keys.
+
+## Dispatching work
+
+```
+gossip_run(agent_id: "<id>", task: "...")
+gossip_dispatch(mode: "parallel" | "consensus", tasks: [...]) → gossip_collect(...)
+```
+
+After dispatching, print a visible summary:
+```
+┌─ gossipcat dispatch ────────────────────────┐
+│  task-id  → agent-name (relay 📡|native 🧠) │
+└─────────────────────────────────────────────┘
+```
+
+## Available Agents
+{{AGENT_LIST}}
+
+## Consensus Workflow
+
+1. `gossip_dispatch(mode: "consensus", tasks: [...])`
+2. Run native **Task()** calls + `gossip_relay` each result
+3. `gossip_collect(task_ids, consensus: true)`
+4. Run cross-review **Task()** calls + `gossip_relay_cross_review`
+5. `gossip_collect(consensus: true)` again for final output
+
+Verify ALL UNVERIFIED findings against the code before presenting results.
+
+## Signals
+
+Record with mandatory `finding_id`:
+```
+gossip_signals(action: "record", signals: [{
+  signal: "unique_confirmed",
+  agent_id: "<who>",
+  finding: "<description>",
+  finding_id: "<consensus_id>:<agent_id>:fN"
+}])
+```
+
+Call `gossip_session_save()` before ending your session.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gossipcat",
-  "version": "0.6.5",
+  "version": "0.6.6",
   "description": "Multi-agent orchestration for Claude Code — parallel review, consensus, adaptive dispatch",
   "mcpName": "io.github.ataberk-xyz/gossipcat",
   "repository": {
@@ -24,6 +24,7 @@
     "dist-dashboard/",
     "docs/HANDBOOK.md",
     "docs/RULES.md",
+    "docs/CURSOR_RULES.md",
     "scripts/postinstall.js",
     "assets/hooks/discipline/**"
   ],