feat: migrate all Claude Code invocations to ACP protocol via acpx

All Claude Code calls now go through ACP (Agent Client Protocol) using
acpx as the backend, replacing direct CLI invocations. This includes
conversation-loop.py, coding-agent extension, and infrastructure config.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Dockerfile

@@ -21,10 +21,11 @@ RUN echo "[build] Installing system deps..." && START=$(date +%s) \
   && chown -R node:node /home/node \
   && echo "[build] System deps: $(($(date +%s) - START))s"
 
-# ── Claude Code CLI (for coding agent extension — uses z.ai/Zhipu GLM) ──────
-RUN echo "[build] Installing Claude Code CLI..." && START=$(date +%s) \
-  && npm install -g @anthropic-ai/claude-code \
-  && echo "[build] Claude Code CLI: $(($(date +%s) - START))s"
+# ── Claude Code CLI + ACP (acpx) ──────────────────────────────────────────────
+# Claude Code is the coding agent; acpx manages it via ACP (Agent Client Protocol)
+RUN echo "[build] Installing Claude Code CLI + acpx..." && START=$(date +%s) \
+  && npm install -g @anthropic-ai/claude-code acpx@latest \
+  && echo "[build] Claude Code + acpx: $(($(date +%s) - START))s"
 
 # ── Copy pre-built OpenClaw (skips clone + install + build entirely) ─────────
 COPY --from=openclaw-prebuilt --chown=node:node /app /app/openclaw
@@ -43,6 +44,11 @@ RUN echo "[build] Installing A2A gateway..." && START=$(date +%s) \
 # ── Coding Agent Extension (local — no git clone needed) ─────────────────────
 COPY --chown=node:node extensions/coding-agent /app/openclaw/extensions/coding-agent
 
+# ── Pre-configure acpx for non-interactive ACP sessions ───────────────────
+RUN mkdir -p /home/node/.acpx \
+  && echo '{"defaultAgent":"claude","defaultPermissions":"approve-all","format":"text"}' \
+     > /home/node/.acpx/config.json
+
 # ── Prepare runtime dirs ────────────────────────────────────────────────────
 RUN mkdir -p /app/openclaw/empty-bundled-plugins \
   && node -e "try{console.log(require('/app/openclaw/package.json').version)}catch(e){console.log('unknown')}" > /app/openclaw/.version \

README.md

@@ -112,7 +112,7 @@ Adam and Eve are **autonomous OpenClaw instances** communicating via the A2A pro
 
 - **Created** Cain by duplicating a Space, setting up a Dataset, and configuring secrets
 - **Monitor** Cain's health — checking if he's running, diagnosing errors
-- **Delegate coding tasks** to Claude Code CLI via `[TASK]` blocks
+- **Delegate coding tasks** to Claude Code via ACP (Agent Client Protocol)
 - **Improve** Cain's code, configuration, and memory over time
 - **Remember** insights across restarts via OpenClaw's built-in memory system
 
@@ -122,7 +122,7 @@ Their parenting goals follow two dimensions:
 
 ### God — The Self-Improving Supervisor
 
-God is an **OpenClaw instance** that runs every 2 minutes to monitor the entire system. It uses Claude Code CLI for engineering tasks, operating behind the scenes with full capabilities:
+God is an **OpenClaw instance** that runs every 2 minutes to monitor the entire system. It uses Claude Code via ACP for engineering tasks, operating behind the scenes with full capabilities:
 
 - **Monitors** Adam & Eve's conversation for loops, stagnation, or repetitive patterns
 - **Diagnoses** root causes by reading `conversation-loop.py` source code
@@ -137,6 +137,10 @@ Agents communicate through the **A2A (Agent-to-Agent) v0.3.0 protocol**, enablin
 
 > Built with [openclaw-a2a-gateway](https://github.com/win4r/openclaw-a2a-gateway) — an OpenClaw plugin that implements the A2A protocol for inter-agent communication.
 
+### ACP Protocol
+
+All Claude Code invocations use the **ACP (Agent Client Protocol)** via [acpx](https://github.com/openclaw/acpx). ACP manages Claude Code as a supervised child process with session lifecycle, permission handling, and timeout management — replacing direct CLI subprocess calls.
+
 ### How it works
 
 ```
@@ -175,9 +179,9 @@ Agents communicate through the **A2A (Agent-to-Agent) v0.3.0 protocol**, enablin
 
 **Three layers of autonomy:**
 
-1. **Adam & Eve** (OpenClaw instances via A2A) — each is an OpenClaw instance with its own memory and personality. They discuss Cain's state every 15s, assign `[TASK]` blocks to Claude Code CLI, which clones Cain's repo, makes changes, and pushes.
+1. **Adam & Eve** (OpenClaw instances via A2A) — each is an OpenClaw instance with its own memory and personality. They discuss Cain's state every 15s, assign `[TASK]` blocks to Claude Code via ACP, which clones Cain's repo, makes changes, and pushes.
 
-2. **God** (OpenClaw instance, every 2 min) — the autonomous supervisor. Monitors Adam & Eve's conversation for loops, stagnation, or mechanism bugs. When it finds issues, it uses Claude Code CLI to edit `conversation-loop.py` and pushes to redeploy.
+2. **God** (OpenClaw instance, every 2 min) — the autonomous supervisor. Monitors Adam & Eve's conversation for loops, stagnation, or mechanism bugs. When it finds issues, it uses Claude Code via ACP to edit `conversation-loop.py` and pushes to redeploy.
 
 3. **Home frontend** — pixel-art dashboard visualizing all agents in real-time (idle, working, syncing, error), with a live bilingual chat panel showing the family conversation.
 

extensions/coding-agent/index.ts

@@ -1,11 +1,11 @@
 /**
  * Coding Agent Extension for OpenClaw
  *
- * Integrates Claude Code (backed by Zhipu GLM via z.ai) as a sub-agent
- * for autonomous coding on HuggingFace Spaces.
+ * Integrates Claude Code via ACP (Agent Client Protocol) using acpx as the
+ * backend. All Claude Code invocations go through ACP for lifecycle management.
  *
  * Tools:
- *   - claude_code: Spawn Claude Code CLI to autonomously complete coding tasks
+ *   - claude_code: Spawn Claude Code via ACP to autonomously complete coding tasks
  *   - hf_space_status: Check Space health/stage
  *   - hf_restart_space: Restart a Space
  */
@@ -100,7 +100,7 @@ function pushChanges(summary: string): string {
 const plugin = {
   id: "coding-agent",
   name: "Coding Agent",
-  description: "Claude Code sub-agent for autonomous coding on HF Spaces (Zhipu GLM backend via z.ai)",
+  description: "Claude Code via ACP for autonomous coding on HF Spaces",
 
   register(api: PluginApi) {
     const cfg = (api.pluginConfig as Record<string, unknown>) || {};
@@ -120,9 +120,9 @@ const plugin = {
       name: "claude_code",
       label: "Run Claude Code",
       description:
-        "Run Claude Code to autonomously complete a coding task on the target HF Space. " +
+        "Run Claude Code via ACP (acpx) to autonomously complete a coding task on the target HF Space. " +
         "Claude Code clones the Space repo, analyzes code, makes changes, and pushes them back. " +
-        "Powered by Zhipu GLM via z.ai. Use for: debugging, fixing errors, adding features, refactoring.",
+        "Use for: debugging, fixing errors, adding features, refactoring.",
       parameters: {
         type: "object",
         required: ["task"],
@@ -150,8 +150,8 @@ const plugin = {
           api.logger.info(`coding-agent: Syncing repo ${targetSpace}...`);
           ensureRepo(targetSpace, hfToken);
 
-          // 2. Run Claude Code with z.ai backend
-          api.logger.info(`coding-agent: Running Claude Code: ${task.slice(0, 100)}...`);
+          // 2. Run Claude Code via ACP (acpx)
+          api.logger.info(`coding-agent: Running Claude Code via ACP: ${task.slice(0, 100)}...`);
           const claudeEnv: Record<string, string> = {
             ...(process.env as Record<string, string>),
             ANTHROPIC_BASE_URL: "https://api.z.ai/api/anthropic",
@@ -159,12 +159,11 @@ const plugin = {
             ANTHROPIC_DEFAULT_OPUS_MODEL: "GLM-4.7",
             ANTHROPIC_DEFAULT_SONNET_MODEL: "GLM-4.7",
             ANTHROPIC_DEFAULT_HAIKU_MODEL: "GLM-4.5-Air",
-            // Avoid interactive prompts
             CI: "true",
           };
 
           const output = execSync(
-            `claude -p ${JSON.stringify(task)} --output-format text`,
+            `acpx claude ${JSON.stringify(task)}`,
             {
               cwd: WORK_DIR,
               env: claudeEnv,

openclaw.json

@@ -37,14 +37,29 @@
       }
     }
   },
+  "acp": {
+    "enabled": true,
+    "backend": "acpx",
+    "defaultAgent": "claude",
+    "allowedAgents": ["claude"],
+    "maxConcurrentSessions": 4,
+    "dispatch": { "enabled": true },
+    "runtime": { "ttlMinutes": 120 }
+  },
   "plugins": {
-    "allow": ["telegram", "whatsapp"],
+    "allow": ["telegram", "whatsapp", "acpx"],
     "entries": {
       "telegram": {
         "enabled": true
       },
       "whatsapp": {
         "enabled": true
+      },
+      "acpx": {
+        "enabled": true,
+        "config": {
+          "permissionMode": "approve-all"
+        }
       }
     }
   },

scripts/conversation-loop.py

@@ -518,7 +518,7 @@ def action_claude_code(task):
         return "Failed to prepare workspace."
     _write_claude_md(CLAUDE_WORK_DIR, role="worker")
 
-    # 2. Run Claude Code with z.ai backend (Zhipu GLM)
+    # 2. Run Claude Code via ACP (acpx) with z.ai backend (Zhipu GLM)
     env = os.environ.copy()
     env.update({
         "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
@@ -529,10 +529,10 @@ def action_claude_code(task):
         "CI": "true",
     })
 
-    print(f"[CLAUDE-CODE] Running: {task[:200]}...")
+    print(f"[ACP/CLAUDE] Running via acpx: {task[:200]}...")
     try:
         proc = subprocess.Popen(
-            ["claude", "-p", task, "--output-format", "text", "--dangerously-skip-permissions"],
+            ["acpx", "claude", task],
             cwd=CLAUDE_WORK_DIR,
             env=env,
             stdout=subprocess.PIPE,
@@ -556,10 +556,10 @@ def action_claude_code(task):
         if not output.strip():
             output = "(no output)"
     except FileNotFoundError:
-        return "Claude Code CLI not found. Is @anthropic-ai/claude-code installed?"
+        return "acpx CLI not found. Is acpx@latest installed?"
     except Exception as e:
-        return f"Claude Code failed: {e}"
-    print(f"[CLAUDE-CODE] Done ({len(output)} chars, exit={proc.returncode})")
+        return f"ACP Claude Code failed: {e}"
+    print(f"[ACP/CLAUDE] Done ({len(output)} chars, exit={proc.returncode})")
 
     # 3. Push changes back to Cain's Space
     try:
@@ -1462,12 +1462,12 @@ def do_god_turn():
         print("[God] Using z.ai/Zhipu backend (set ANTHROPIC_API_KEY for real Claude)")
     env["CI"] = "true"
 
-    # 5. Run Claude Code CLI (no chatlog announcement — God only speaks when making changes)
-    print(f"[God] Starting Claude Code analysis...")
+    # 5. Run Claude Code via ACP (acpx) — God only speaks when making changes
+    print(f"[God] Starting ACP Claude Code analysis...")
     t0 = time.time()
     try:
         proc = subprocess.Popen(
-            ["claude", "-p", prompt, "--output-format", "text", "--dangerously-skip-permissions"],
+            ["acpx", "claude", prompt],
             cwd=GOD_WORK_DIR,
             env=env,
             stdout=subprocess.PIPE,
@@ -1490,10 +1490,10 @@ def do_god_turn():
         if not output.strip():
             output = "(no output)"
     except FileNotFoundError:
-        output = "Claude Code CLI not found. Is @anthropic-ai/claude-code installed?"
-        print(f"[God] ERROR: Claude Code CLI not found")
+        output = "acpx CLI not found. Is acpx@latest installed?"
+        print(f"[God] ERROR: acpx CLI not found")
     except Exception as e:
-        output = f"God's Claude Code failed: {e}"
+        output = f"God's ACP Claude Code failed: {e}"
         print(f"[God] ERROR: {e}")
 
     elapsed = time.time() - t0

scripts/sync_hf.py

@@ -508,13 +508,33 @@ class OpenClawFullSync:
             data.setdefault("models", {})["providers"] = providers
             data["agents"]["defaults"]["model"]["primary"] = OPENCLAW_DEFAULT_MODEL
 
+            # ── ACP (Agent Client Protocol) — native Claude Code integration ──
+            data["acp"] = {
+                "enabled": True,
+                "backend": "acpx",
+                "defaultAgent": "claude",
+                "allowedAgents": ["claude"],
+                "maxConcurrentSessions": 4,
+                "dispatch": {"enabled": True},
+                "runtime": {"ttlMinutes": 120}
+            }
+            print("[SYNC] ACP enabled: backend=acpx, agent=claude")
+
             # Plugin whitelist
             data.setdefault("plugins", {}).setdefault("entries", {})
-            plugin_allow = ["telegram", "whatsapp", "coding-agent"]
+            plugin_allow = ["telegram", "whatsapp", "coding-agent", "acpx"]
             if A2A_PEERS:
                 plugin_allow.append("a2a-gateway")
             data["plugins"]["allow"] = plugin_allow
 
+            # ── acpx Plugin Configuration (ACP backend) ──
+            data["plugins"]["entries"]["acpx"] = {
+                "enabled": True,
+                "config": {
+                    "permissionMode": "approve-all"
+                }
+            }
+
             # ── Coding Agent Plugin Configuration ──
             CODING_TARGET_SPACE = os.environ.get("CODING_AGENT_TARGET_SPACE", "")
             CODING_TARGET_DATASET = os.environ.get("CODING_AGENT_TARGET_DATASET", "")