claude-code-swarm 0.3.22 → 0.3.23
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.claude-plugin/marketplace.json +1 -1
- package/.claude-plugin/plugin.json +1 -1
- package/package.json +3 -3
- package/plan.md +0 -214
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "claude-code-swarm",
|
|
3
3
|
"description": "Spin up Claude Code agent teams from openteams YAML topologies with optional MAP (Multi-Agent Protocol) observability and coordination. Provides hooks for session lifecycle, agent spawn/complete tracking, and a /swarm skill to launch team configurations.",
|
|
4
|
-
"version": "0.3.
|
|
4
|
+
"version": "0.3.23",
|
|
5
5
|
"author": {
|
|
6
6
|
"name": "alexngai"
|
|
7
7
|
},
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "claude-code-swarm",
|
|
3
|
-
"version": "0.3.
|
|
3
|
+
"version": "0.3.23",
|
|
4
4
|
"description": "Claude Code plugin for launching agent teams from openteams topologies with MAP observability",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"exports": {
|
|
@@ -24,7 +24,7 @@
|
|
|
24
24
|
},
|
|
25
25
|
"peerDependencies": {
|
|
26
26
|
"agent-inbox": "*",
|
|
27
|
-
"opentasks": "
|
|
27
|
+
"opentasks": ">=0.1.1",
|
|
28
28
|
"swarmkit": "*"
|
|
29
29
|
},
|
|
30
30
|
"peerDependenciesMeta": {
|
|
@@ -55,7 +55,7 @@
|
|
|
55
55
|
"devDependencies": {
|
|
56
56
|
"agent-inbox": "^0.1.9",
|
|
57
57
|
"minimem": "^0.1.1",
|
|
58
|
-
"opentasks": "^0.
|
|
58
|
+
"opentasks": "^0.1.1",
|
|
59
59
|
"skill-tree": "^0.1.5",
|
|
60
60
|
"vitest": "^4.0.18",
|
|
61
61
|
"ws": "^8.0.0"
|
package/plan.md
DELETED
|
@@ -1,214 +0,0 @@
|
|
|
1
|
-
# Refactor: Replace custom `swarm.*` events with MAP SDK primitives
|
|
2
|
-
|
|
3
|
-
## Goal
|
|
4
|
-
|
|
5
|
-
Eliminate all custom `swarm.*` event types by using the MAP SDK's built-in `AgentConnection` methods (`spawn()`, `done()`, `updateState()`, `updateMetadata()`, `send()`) and the server's automatic event emission (`agent_registered`, `agent_state_changed`, etc.). This means MAP clients only need to subscribe to standard MAP event types — zero swarm-specific integration.
|
|
6
|
-
|
|
7
|
-
## Guiding principle
|
|
8
|
-
|
|
9
|
-
**No new message types.** All swarm-specific context goes into `metadata` on agents and `payload` on messages — never into custom event type strings.
|
|
10
|
-
|
|
11
|
-
---
|
|
12
|
-
|
|
13
|
-
## Current state → Target state mapping
|
|
14
|
-
|
|
15
|
-
| Current custom event | Target SDK primitive | Auto server event |
|
|
16
|
-
|---|---|---|
|
|
17
|
-
| `swarm.agent.registered` (sidecar broadcasts) | `conn.spawn({ name, role, parent, scopes, metadata })` | `agent_registered` |
|
|
18
|
-
| `swarm.agent.unregistered` (sidecar broadcasts) | `spawnedAgent.done({ exitReason })` or sidecar tracks + calls `conn.callExtension(...)` to unregister | `agent_unregistered` |
|
|
19
|
-
| `swarm.agent.spawned` (emitEvent broadcast) | Replaced by `spawn()` above — no separate event needed | `agent_registered` |
|
|
20
|
-
| `swarm.agent.completed` (emitEvent broadcast) | Replaced by agent `done()` above | `agent_state_changed` → `stopped` |
|
|
21
|
-
| `swarm.task.dispatched` (emitEvent broadcast) | `conn.send({ scope }, { type: "task.dispatched", ... })` — typed payload in regular message | `message_sent` |
|
|
22
|
-
| `swarm.task.completed` (emitEvent broadcast) | `conn.send({ scope }, { type: "task.completed", ... })` | `message_sent` |
|
|
23
|
-
| `swarm.task.status_completed` (emitEvent broadcast) | `conn.send({ scope }, { type: "task.completed", ... })` with richer payload | `message_sent` |
|
|
24
|
-
| `swarm.turn.completed` (emitEvent broadcast) | `conn.idle()` + `conn.updateMetadata({ lastStopReason })` | `agent_state_changed` → `idle` |
|
|
25
|
-
| `swarm.subagent.started` (emitEvent broadcast) | `conn.spawn({ role: "subagent", metadata: { agentType, sessionId } })` | `agent_registered` |
|
|
26
|
-
| `swarm.subagent.stopped` (emitEvent broadcast) | `spawned.done({ exitReason })` | `agent_state_changed` → `stopped` |
|
|
27
|
-
| `swarm.teammate.idle` (emitEvent broadcast) | `agent.updateState("idle")` (via sidecar) | `agent_state_changed` → `idle` |
|
|
28
|
-
| `swarm.sessionlog.sync` (broadcast fallback) | `conn.callExtension("trajectory/checkpoint", ...)` — keep existing; fallback sends as regular message payload instead of custom event type | `trajectory.checkpoint` |
|
|
29
|
-
|
|
30
|
-
---
|
|
31
|
-
|
|
32
|
-
## Implementation steps
|
|
33
|
-
|
|
34
|
-
### Step 1: Refactor `sidecar-server.mjs` — replace `register`/`unregister` with `spawn`/`done`
|
|
35
|
-
|
|
36
|
-
**File:** `src/sidecar-server.mjs`
|
|
37
|
-
|
|
38
|
-
Replace the `register` command handler:
|
|
39
|
-
- **Before:** `conn.send({ scope }, { type: "swarm.agent.registered", ... })` — broadcasts a custom message
|
|
40
|
-
- **After:** `conn.spawn({ agentId, name, role, scopes, metadata })` — uses SDK primitive; server auto-emits `agent_registered`
|
|
41
|
-
- Store spawned agent references in a `Map<agentId, AgentSpawnResult>` instead of just a `Set<agentId>`
|
|
42
|
-
|
|
43
|
-
Replace the `unregister` command handler:
|
|
44
|
-
- **Before:** `conn.send({ scope }, { type: "swarm.agent.unregistered", ... })` — broadcasts custom message
|
|
45
|
-
- **After:** Use `conn.callExtension("agents/unregister", { agentId, reason })` or track spawned agent IDs and send a deregistration request. Since the sidecar is the parent, it can manage child agent lifecycle.
|
|
46
|
-
- Actually, the simplest approach: call `conn.send()` with method `"map/agents/unregister"` as a request, or use the lower-level approach of sending a state update. Since the sidecar owns the spawned agents, we'll call `conn.callExtension("map/agents/unregister", { agentId, reason })`.
|
|
47
|
-
- Remove from tracked agents map
|
|
48
|
-
|
|
49
|
-
Replace the `trajectory-checkpoint` fallback:
|
|
50
|
-
- **Before:** Falls back to `conn.send({ scope }, { type: "swarm.sessionlog.sync", ... })`
|
|
51
|
-
- **After:** Falls back to `conn.send({ scope }, { type: "trajectory.checkpoint.fallback", checkpoint })` — still a regular message but with a standardized payload shape, not a custom event type. Or even simpler: just use `{ type: "trajectory.checkpoint", ... }` as the message payload since this is informational.
|
|
52
|
-
|
|
53
|
-
Keep `emit` command handler for now (backward compat) but mark as deprecated — it's the generic "send arbitrary payload" path.
|
|
54
|
-
|
|
55
|
-
Keep `state` and `ping` handlers unchanged.
|
|
56
|
-
|
|
57
|
-
### Step 2: Refactor `map-events.mjs` — replace builders with SDK-aligned builders
|
|
58
|
-
|
|
59
|
-
**File:** `src/map-events.mjs`
|
|
60
|
-
|
|
61
|
-
This is where the core change happens. Replace the custom event builders with functions that produce either:
|
|
62
|
-
- **Sidecar commands** (for `spawn`/`done`/state updates) — structured for the sidecar socket protocol
|
|
63
|
-
- **Message payloads** (for task lifecycle) — sent via `conn.send()` as regular MAP messages
|
|
64
|
-
|
|
65
|
-
New functions:
|
|
66
|
-
|
|
67
|
-
```javascript
|
|
68
|
-
// Agent lifecycle — produce sidecar commands
|
|
69
|
-
export function buildSpawnCommand(agentName, matchedRole, teamName, hookData) → { action: "spawn", agent: { ... } }
|
|
70
|
-
export function buildDoneCommand(agentName, matchedRole, teamName) → { action: "done", agentId, reason }
|
|
71
|
-
export function buildSubagentSpawnCommand(hookData, teamName) → { action: "spawn", agent: { role: "subagent", ... } }
|
|
72
|
-
export function buildSubagentDoneCommand(hookData, teamName) → { action: "done", agentId, reason }
|
|
73
|
-
|
|
74
|
-
// State updates — produce sidecar commands
|
|
75
|
-
export function buildStateCommand(agentId, state, metadata?) → { action: "state", agentId, state, metadata? }
|
|
76
|
-
|
|
77
|
-
// Task lifecycle — produce message payloads (sent via "emit" or "send")
|
|
78
|
-
export function buildTaskDispatchedPayload(hookData, teamName, matchedRole, agentName) → { type: "task.dispatched", ... }
|
|
79
|
-
export function buildTaskCompletedPayload(hookData, teamName, matchedRole, agentName) → { type: "task.completed", ... }
|
|
80
|
-
export function buildTaskStatusPayload(hookData, teamName, matchedRole) → { type: "task.completed", ... }
|
|
81
|
-
```
|
|
82
|
-
|
|
83
|
-
Remove `emitEvent()` wrapper for agent lifecycle events — those go directly as sidecar commands.
|
|
84
|
-
Keep `emitEvent()` only for task-related message payloads (these are still sent as MAP messages to scope, which is fine).
|
|
85
|
-
|
|
86
|
-
### Step 3: Refactor `map-hook.mjs` — update action handlers
|
|
87
|
-
|
|
88
|
-
**File:** `scripts/map-hook.mjs`
|
|
89
|
-
|
|
90
|
-
`handleAgentSpawning()`:
|
|
91
|
-
- **Before:** `sendToSidecar({ action: "register", agent: {...} })` + `emitEvent(buildSpawnEvent)` + `emitEvent(buildTaskDispatchedEvent)`
|
|
92
|
-
- **After:** `sendToSidecar(buildSpawnCommand(...))` + `emitEvent(buildTaskDispatchedPayload(...))`
|
|
93
|
-
- The spawn command replaces both the register and the spawn event — one operation, server auto-emits the event
|
|
94
|
-
|
|
95
|
-
`handleAgentCompleted()`:
|
|
96
|
-
- **Before:** `sendToSidecar({ action: "unregister", ... })` + `emitEvent(buildCompletedEvent)` + `emitEvent(buildTaskCompletedEvent)`
|
|
97
|
-
- **After:** `sendToSidecar(buildDoneCommand(...))` + `emitEvent(buildTaskCompletedPayload(...))`
|
|
98
|
-
|
|
99
|
-
`handleTurnCompleted()`:
|
|
100
|
-
- **Before:** `sendToSidecar({ action: "state", state: "idle" })` + `emitEvent(buildTurnCompletedEvent)`
|
|
101
|
-
- **After:** `sendToSidecar(buildStateCommand(sidecarId, "idle", { lastStopReason }))` — just a state+metadata update, server auto-emits `agent_state_changed`
|
|
102
|
-
|
|
103
|
-
`handleSubagentStart()`:
|
|
104
|
-
- **Before:** `emitEvent(buildSubagentStartEvent)`
|
|
105
|
-
- **After:** `sendToSidecar(buildSubagentSpawnCommand(...))` — spawn in MAP for observability
|
|
106
|
-
|
|
107
|
-
`handleSubagentStop()`:
|
|
108
|
-
- **Before:** `emitEvent(buildSubagentStopEvent)`
|
|
109
|
-
- **After:** `sendToSidecar(buildSubagentDoneCommand(...))`
|
|
110
|
-
|
|
111
|
-
`handleTeammateIdle()`:
|
|
112
|
-
- **Before:** `sendToSidecar({ action: "state" })` + `emitEvent(buildTeammateIdleEvent)`
|
|
113
|
-
- **After:** `sendToSidecar(buildStateCommand(agentId, "idle"))` — state change only, server auto-emits
|
|
114
|
-
|
|
115
|
-
`handleTaskCompleted()`:
|
|
116
|
-
- **Before:** `emitEvent(buildTaskStatusCompletedEvent)`
|
|
117
|
-
- **After:** `emitEvent(buildTaskStatusPayload(...))`
|
|
118
|
-
|
|
119
|
-
### Step 4: Refactor `map-connection.mjs` — update `fireAndForget` and trajectory fallback
|
|
120
|
-
|
|
121
|
-
**File:** `src/map-connection.mjs`
|
|
122
|
-
|
|
123
|
-
`fireAndForget()`:
|
|
124
|
-
- Keep as-is but now it sends typed message payloads (for task events) rather than custom event types
|
|
125
|
-
|
|
126
|
-
`fireAndForgetTrajectory()`:
|
|
127
|
-
- **Before:** Fallback broadcasts `{ type: "swarm.sessionlog.sync", ... }`
|
|
128
|
-
- **After:** Fallback sends `{ type: "trajectory.checkpoint", checkpoint: {...} }` as a regular message payload — clients looking for trajectory data can filter on this
|
|
129
|
-
|
|
130
|
-
### Step 5: Update `sidecar-server.mjs` command handler — add `spawn`/`done` commands
|
|
131
|
-
|
|
132
|
-
**File:** `src/sidecar-server.mjs`
|
|
133
|
-
|
|
134
|
-
Add new command cases:
|
|
135
|
-
|
|
136
|
-
```javascript
|
|
137
|
-
case "spawn": {
|
|
138
|
-
if (conn) {
|
|
139
|
-
const result = await conn.spawn({
|
|
140
|
-
agentId: command.agent.agentId,
|
|
141
|
-
name: command.agent.name,
|
|
142
|
-
role: command.agent.role,
|
|
143
|
-
scopes: command.agent.scopes,
|
|
144
|
-
metadata: command.agent.metadata,
|
|
145
|
-
});
|
|
146
|
-
registeredAgents.set(command.agent.agentId, result);
|
|
147
|
-
}
|
|
148
|
-
respond(client, { ok: true });
|
|
149
|
-
break;
|
|
150
|
-
}
|
|
151
|
-
|
|
152
|
-
case "done": {
|
|
153
|
-
if (conn) {
|
|
154
|
-
// Unregister the spawned child agent
|
|
155
|
-
try {
|
|
156
|
-
await conn.callExtension("map/agents/unregister", {
|
|
157
|
-
agentId: command.agentId,
|
|
158
|
-
reason: command.reason || "completed",
|
|
159
|
-
});
|
|
160
|
-
} catch {
|
|
161
|
-
// Agent may already be gone
|
|
162
|
-
}
|
|
163
|
-
registeredAgents.delete(command.agentId);
|
|
164
|
-
}
|
|
165
|
-
respond(client, { ok: true });
|
|
166
|
-
break;
|
|
167
|
-
}
|
|
168
|
-
```
|
|
169
|
-
|
|
170
|
-
Change `registeredAgents` from `Set` to `Map` in `scripts/map-sidecar.mjs`.
|
|
171
|
-
|
|
172
|
-
Keep old `register`/`unregister` commands working (deprecated) for backward compat during rollout.
|
|
173
|
-
|
|
174
|
-
### Step 6: Update tests
|
|
175
|
-
|
|
176
|
-
**Files:**
|
|
177
|
-
- `src/__tests__/map-events.test.mjs` — Update for new function signatures and return shapes
|
|
178
|
-
- `src/__tests__/sidecar-server.test.mjs` — Add tests for `spawn`/`done` commands, update `register`/`unregister` tests
|
|
179
|
-
|
|
180
|
-
### Step 7: Update `CLAUDE.md` documentation
|
|
181
|
-
|
|
182
|
-
Update the "MAP hooks" section to reflect:
|
|
183
|
-
- No custom `swarm.*` event types
|
|
184
|
-
- Agent lifecycle uses `spawn()`/`done()` SDK primitives
|
|
185
|
-
- Task lifecycle uses typed message payloads
|
|
186
|
-
- Turn lifecycle uses `updateState()` + `updateMetadata()`
|
|
187
|
-
- Clients subscribe to standard MAP events only
|
|
188
|
-
|
|
189
|
-
---
|
|
190
|
-
|
|
191
|
-
## What stays the same
|
|
192
|
-
|
|
193
|
-
- **Sidecar architecture** — persistent process with socket IPC, unchanged
|
|
194
|
-
- **Hook wiring** — same hooks, same scripts, same conditions in `hooks.json`
|
|
195
|
-
- **Trajectory checkpoints** — same `callExtension("trajectory/checkpoint")` with broadcast fallback
|
|
196
|
-
- **Inbox system** — unchanged (read/clear/format/write)
|
|
197
|
-
- **Role matching** — unchanged (`roles.mjs`)
|
|
198
|
-
- **Config system** — unchanged (`config.mjs`)
|
|
199
|
-
- **Bootstrap flow** — unchanged (`bootstrap.mjs`)
|
|
200
|
-
- **Fire-and-forget recovery** — same sidecar → recovery → direct pattern
|
|
201
|
-
|
|
202
|
-
## What changes
|
|
203
|
-
|
|
204
|
-
- **No custom event types** — everything uses MAP SDK primitives or typed message payloads
|
|
205
|
-
- **Sidecar registers agents properly** — `conn.spawn()` instead of broadcasting fake events
|
|
206
|
-
- **Server emits lifecycle events automatically** — clients see `agent_registered`, `agent_state_changed`, etc.
|
|
207
|
-
- **Task events are messages, not events** — sent via `conn.send()` with typed payloads
|
|
208
|
-
- **Cleaner client integration** — subscribe to standard MAP events, no `swarm.*` parsing needed
|
|
209
|
-
|
|
210
|
-
## Risk / rollback
|
|
211
|
-
|
|
212
|
-
- The refactor is fully backward-compatible at the config level (`.claude-swarm.json` unchanged)
|
|
213
|
-
- Old `register`/`unregister` sidecar commands still work during transition
|
|
214
|
-
- If `conn.spawn()` fails (e.g., server doesn't support it), we can fall back to the old broadcast approach
|