feat(evm-wallet-experiment): OpenClaw LLM via Docker Model Runner in Compose

grypez · grypez · commit 7ff75b335b1a · 2026-04-06T12:35:38.000-04:00
- Declare top-level Compose models (ai/qwen3.5:4B-UD-Q4_K_XL) and bind to kernel-away-bundler-7702; remove in-stack llm proxy + entrypoint-llm-proxy.mjs.

- setup-openclaw requires DMR-injected LLM_URL/LLM_MODEL, openai-completions, /v1 base normalization; trim docker-e2e-stack-constants llm defaults.

- package.json: interactive compose without empty profile, openclaw gateway via node path, up --build, docker:interactive:reset-openclaw.

- context_size + runtime_flags --ctx-size 32768; MAINTAINERS troubleshooting (configure, volume persistence).

Made-with: Cursor
diff --git a/packages/evm-wallet-experiment/docker/Dockerfile.kernel-base b/packages/evm-wallet-experiment/docker/Dockerfile.kernel-base
@@ -67,4 +67,6 @@ RUN mkdir -p /logs /run/ocap
 FROM kernel AS interactive
 
 # OpenClaw loads local plugins as TypeScript via jiti (no extra TS runner in the image).
+# `package.json` docker:interactive:setup starts the gateway via
+# `node /usr/local/lib/node_modules/openclaw/openclaw.mjs` (global bin PATH is unreliable under `docker exec`).
 RUN npm install -g openclaw@2026.4.1
diff --git a/packages/evm-wallet-experiment/docker/MAINTAINERS.md b/packages/evm-wallet-experiment/docker/MAINTAINERS.md
@@ -1,14 +1,14 @@
 # Docker stack — maintainer notes
 
-Local E2E stack for `@ocap/evm-wallet-experiment`: Anvil + deployed contracts, Pimlico Alto, and two kernel containers (`home`, `away`). See `package.json` scripts (`docker:compose`, `test:e2e:docker`, etc.).
+Local E2E stack for `@ocap/evm-wallet-experiment`: Anvil + deployed contracts, Pimlico Alto, and six kernel containers (three `kernel-home-*` / `kernel-away-*` pairs). See `package.json` scripts (`docker:compose`, `test:e2e:docker`, etc.).
 
 ## Startup order
 
 Compose encodes this dependency chain:
 
 1. **`evm`** becomes healthy when `/run/ocap/contracts.json` exists (written only after `deploy-contracts.mjs` finishes).
 2. **`bundler`** waits on that file, reads `EntryPoint`, then starts Alto.
-3. **`home` and `away`** wait on **both** `evm` and **`bundler` healthy** so wallet setup does not race Alto boot.
+3. **Kernel services** wait on **both** `evm` and **`bundler` healthy** so wallet setup does not race Alto boot.
 
 If you add a service that kernels need before they are ready, extend `depends_on` and healthchecks accordingly.
 
@@ -48,7 +48,6 @@ Host-side scripts (e.g. `yarn docker:setup:wallets`) use the workspace **`tsx`**
 
 - **`evm`**: File-based (`contracts.json`). The image itself does not define `HEALTHCHECK`; Compose is the source of truth.
 - **`bundler`**: JSON-RPC `eth_supportedEntryPoints` must return a **non-empty** array. If Alto changes RPC surface, adjust the probe in `docker-compose.yml`.
-- **`llm`**: HTTP GET `/` on the proxy; **5xx** (e.g. upstream unreachable) marks the service unhealthy.
 
 ## Kernel image build (`Dockerfile.kernel-base`)
 
@@ -59,11 +58,18 @@ Host-side scripts (e.g. `yarn docker:setup:wallets`) use the workspace **`tsx`**
 
 `docker-compose.yml` embeds **well-known Anvil private keys** for Alto. That is intentional for an isolated local chain. **Do not reuse this pattern** for any network that is exposed or shared.
 
-## Interactive profile
+## Interactive stack (`docker-compose.interactive.yml`)
 
-- **`llm`** defaults `LLM_UPSTREAM` to `http://host.docker.internal:8080`. On **Linux**, `host.docker.internal` may be missing unless you add `extra_hosts` or another reachability strategy; document any project-standard workaround here when you add one.
-- **`docker-compose.interactive.yml`** overrides `away` (OpenClaw + LLM). Ensure the **`interactive`** profile is used when you expect those services.
+- Requires **Docker Compose v2.38+** and [**Docker Model Runner**](https://docs.docker.com/ai/model-runner/) enabled.
+- Merges over **`kernel-away-bundler-7702`**: **`interactive`** image target (OpenClaw) and Compose [**`models`**](https://docs.docker.com/ai/compose/models-and-compose/) binding **`llm`** → **`ai/qwen3.5:4B-UD-Q4_K_XL`** with **`context_size: 32768`** (DMR’s default 4096 is too small for OpenClaw + tool prompts). Larger context uses more RAM.
+- Run **`yarn docker:compose:interactive`**. Pull the model first if needed: **`docker model pull ai/qwen3.5:4B-UD-Q4_K_XL`**.
+- **`compose.interactive.yml`** sets **`context_size`** and **`runtime_flags: ['--ctx-size','32768']`** so llama.cpp does not stay at the 4096 default. If requests still fail with 4096, run on the host: **`docker model configure --context-size 32768 ai/qwen3.5:4B-UD-Q4_K_XL`**, then recreate the stack.
+- OpenClaw UI history lives under **`$HOME/.openclaw`** in the away container (on the **`ocap-run`** named volume). Rebuilding images does **not** clear it. Use **`yarn docker:interactive:reset-openclaw`**, then **`yarn docker:interactive:setup`**, or **`docker compose … down -v`** (wipes **all** kernel/contract state on that volume — use with care).
+
+## Optional **`ollama`** profile
+
+- Service **`ollama`** uses profile **`ollama`** for an in-stack Ollama server. Interactive OpenClaw uses **DMR via Compose `models`**, not this service, unless you change the stack.
 
 ## Ports and conflicts
 
-Published ports include **8545**, **4337**, **11434** (profile), and **UDP 4001/4002**. They can clash with other stacks on the host; use Compose [profiles](https://docs.docker.com/compose/profiles/) or alternate port mappings if needed.
+Published TCP ports include **8545** (Anvil), **4337** (bundler). Kernels publish **UDP 4011–4032** (QUIC, three pairs). The **`ollama`** profile does not publish **11434** to the host by default in the current compose file; check `docker-compose.yml` if that changes. Use alternate mappings if these clash with other stacks.
diff --git a/packages/evm-wallet-experiment/docker/docker-compose.interactive.yml b/packages/evm-wallet-experiment/docker/docker-compose.interactive.yml
@@ -1,18 +1,40 @@
-# Compose override for interactive use (OpenClaw + LLM).
+# Compose override for interactive use (OpenClaw + Docker Model Runner).
+#
+# Requires Docker Compose v2.38+ and Docker Model Runner enabled.
+# Model: https://hub.docker.com/r/ai/qwen3.5/tags — tag 4B-UD-Q4_K_XL (digest prefix 6dfe4f74414).
 #
 # Usage:
 #   yarn docker:interactive:up
+#   (`up --build` so `kernel-away-bundler-7702` uses the `interactive` image with OpenClaw.)
 #
 # Then run wallet + OpenClaw setup:
 #   yarn docker:interactive:setup
+#
+# DMR injects `LLM_URL` and `LLM_MODEL` into `kernel-away-bundler-7702` (short `models: - llm` syntax).
+# OpenClaw is configured for the OpenAI-compatible `/v1` API (`openai-completions`).
+#
+# OpenClaw and shell access use compose service `kernel-away-bundler-7702` (not `away`).
+#
+# If you still see "4096 tokens" from the runner, `context_size` in Compose may be ignored
+# (known DMR/llama.cpp edge cases). Use `runtime_flags` below and/or on the host:
+#   docker model configure --context-size 32768 ai/qwen3.5:4B-UD-Q4_K_XL
+#
+# OpenClaw chat/session data lives on volume `ocap-run` under this service's HOME
+# (`.../kernel-away-bundler-7702/.openclaw`), not in the image — `docker build` does not
+# clear it. Run `yarn docker:interactive:reset-openclaw` (or `docker compose down -v`) to wipe.
 services:
   kernel-away-bundler-7702:
     build:
       target: interactive
-    environment:
-      - LLM_BASE_URL=${LLM_BASE_URL:-http://llm:11434/v1}
-      - LLM_MODEL=${LLM_MODEL:-glm-4.7-flash}
-      - LLM_API_TYPE=${LLM_API_TYPE:-openai-completions}
-    depends_on:
-      llm:
-        condition: service_healthy
+    models:
+      - llm
+
+models:
+  llm:
+    model: ai/qwen3.5:4B-UD-Q4_K_XL
+    # DMR llama.cpp defaults to 4096; OpenClaw + wallet tools exceed that.
+    context_size: 32768
+    # Explicit llama.cpp ctx (Compose `context_size` alone is sometimes not applied).
+    runtime_flags:
+      - '--ctx-size'
+      - '32768'
diff --git a/packages/evm-wallet-experiment/docker/docker-compose.yml b/packages/evm-wallet-experiment/docker/docker-compose.yml
@@ -51,30 +51,6 @@ services:
       timeout: 5s
       retries: 20
 
-  llm:
-    profiles: [interactive]
-    build:
-      context: ../../../
-      dockerfile: packages/evm-wallet-experiment/docker/Dockerfile.kernel-base
-      target: kernel
-    environment:
-      - LLM_UPSTREAM=${LLM_UPSTREAM:-http://host.docker.internal:8080}
-    networks: [ocap-e2e]
-    ports: ['11434:11434']
-    entrypoint:
-      - node
-      - /app/packages/evm-wallet-experiment/docker/entrypoint-llm-proxy.mjs
-    healthcheck:
-      test:
-        - CMD
-        - node
-        - -e
-        - >-
-          fetch("http://127.0.0.1:11434/").then((r)=>process.exit(r.status<500?0:1)).catch(()=>process.exit(1))
-      interval: 3s
-      timeout: 5s
-      retries: 10
-
   ollama:
     profiles: [ollama]
     image: ollama/ollama
diff --git a/packages/evm-wallet-experiment/docker/docker-e2e-stack-constants.json b/packages/evm-wallet-experiment/docker/docker-e2e-stack-constants.json
@@ -1,10 +1,4 @@
 {
-  "llm": {
-    "upstreamOllama": "http://ollama:11434",
-    "baseUrl": "http://llm:11434",
-    "model": "qwen2.5:0.5b",
-    "apiType": "ollama"
-  },
   "anvilChainId": 31337,
   "openclawPluginPathContainer": "/app/packages/evm-wallet-experiment/openclaw-plugin"
 }
diff --git a/packages/evm-wallet-experiment/docker/docker-e2e-stack-constants.mjs b/packages/evm-wallet-experiment/docker/docker-e2e-stack-constants.mjs
@@ -13,7 +13,6 @@ const jsonPath = join(dir, 'docker-e2e-stack-constants.json');
 
 /**
  * @typedef {object} DockerStackConfig
- * @property {{ upstreamOllama: string, baseUrl: string, model: string, apiType: string }} llm - Defaults for LLM proxy and OpenClaw provider.
  * @property {number} anvilChainId - Chain ID for the local Anvil stack in Docker E2E.
  * @property {string} openclawPluginPathContainer - Wallet plugin path inside the kernel container.
  */
diff --git a/packages/evm-wallet-experiment/docker/entrypoint-llm-proxy.mjs b/packages/evm-wallet-experiment/docker/entrypoint-llm-proxy.mjs
diff --git a/packages/evm-wallet-experiment/docker/setup-openclaw.mjs b/packages/evm-wallet-experiment/docker/setup-openclaw.mjs
@@ -7,10 +7,11 @@
  * This replaces the fragile sequence of `openclaw onboard` +
  * `openclaw config set` calls that broke on every OpenClaw release.
  *
- * Env vars (all optional, with Docker-friendly defaults):
- *   LLM_BASE_URL   — LLM provider base URL (default: http://llm:11434)
- *   LLM_MODEL      — Model ID (default: qwen2.5:0.5b)
- *   LLM_API_TYPE   — OpenClaw API type: ollama | openai-completions | … (default: ollama)
+ * Expects Docker Compose **models** + Docker Model Runner to inject:
+ *   LLM_URL   — OpenAI-compatible API base (normalized to end with `/v1` below)
+ *   LLM_MODEL — Model id for requests
+ *
+ * @see https://docs.docker.com/ai/compose/models-and-compose/
  *
  * Usage:
  *   node /app/packages/evm-wallet-experiment/docker/setup-openclaw.mjs
@@ -25,9 +26,33 @@ import { dockerConfig } from './docker-e2e-stack-constants.mjs';
 const home = process.env.HOME || '/run/ocap/away';
 const ocDir = resolve(home, '.openclaw');
 
-const llmBaseUrl = process.env.LLM_BASE_URL || dockerConfig.llm.baseUrl;
-const llmModel = process.env.LLM_MODEL || dockerConfig.llm.model;
-const llmApiType = process.env.LLM_API_TYPE || dockerConfig.llm.apiType;
+const llmUrlRaw = process.env.LLM_URL;
+const llmModel = process.env.LLM_MODEL;
+
+if (!llmUrlRaw || !llmModel) {
+  throw new Error(
+    'openclaw-setup: LLM_URL and LLM_MODEL must be set. ' +
+      'Use docker-compose.interactive.yml with top-level `models:` and Docker Model Runner enabled.',
+  );
+}
+
+/**
+ * OpenClaw `openai-completions` expects a base URL whose paths resolve under `/v1/...`.
+ *
+ * @param {string} url - Injected model runner URL (may omit `/v1`).
+ * @returns {string} Normalized base URL ending with `/v1` (no trailing slash after v1).
+ */
+function openAiCompletionsBaseUrl(url) {
+  const trimmed = url.replace(/\/+$/u, '');
+  if (trimmed.endsWith('/v1')) {
+    return trimmed;
+  }
+  return `${trimmed}/v1`;
+}
+
+const llmBaseUrl = openAiCompletionsBaseUrl(llmUrlRaw);
+/** DMR serves an OpenAI-shaped API alongside Ollama compatibility. */
+const llmApiType = 'openai-completions';
 const providerId = 'llm';
 const pluginPath = dockerConfig.openclawPluginPathContainer;
 
@@ -109,7 +134,7 @@ const config = {
 const cfgPath = resolve(ocDir, 'openclaw.json');
 writeFileSync(cfgPath, JSON.stringify(config, null, 2));
 console.log(
-  `[openclaw-setup] config written: ${providerId}/${llmModel} (api: ${llmApiType})`,
+  `[openclaw-setup] config written: ${providerId}/${llmModel} (api: ${llmApiType}, base: ${llmBaseUrl})`,
 );
 
 // -- Write auth profiles for the provider --
diff --git a/packages/evm-wallet-experiment/package.json b/packages/evm-wallet-experiment/package.json
@@ -57,16 +57,17 @@
     "test:node:peer-e2e": "yarn build && node --conditions development test/e2e/run-peer-e2e.mjs",
     "test:node:spending-limits": "yarn build && node --conditions development test/e2e/run-spending-limits-e2e.mjs",
     "docker:compose": "docker compose -f docker/docker-compose.yml",
-    "docker:compose:interactive": "docker compose -f docker/docker-compose.yml -f docker/docker-compose.interactive.yml --profile interactive",
+    "docker:compose:interactive": "docker compose -f docker/docker-compose.yml -f docker/docker-compose.interactive.yml",
     "docker:build": "yarn docker:compose build",
     "docker:build:force": "yarn docker:compose build --no-cache",
     "docker:up": "yarn docker:compose up",
     "docker:down": "yarn docker:compose down",
-    "docker:interactive:up": "yarn docker:compose:interactive up",
+    "docker:interactive:up": "yarn docker:compose:interactive up --build",
     "docker:interactive:down": "yarn docker:compose:interactive down",
     "docker:setup:wallets": "yarn tsx test/e2e/docker/setup-wallets.ts",
-    "docker:interactive:setup": "yarn docker:setup:wallets && yarn docker:compose:interactive exec away node /app/packages/evm-wallet-experiment/docker/setup-openclaw.mjs && yarn docker:compose:interactive exec -d away openclaw gateway && echo 'OpenClaw configured + gateway started. Shell in with: yarn docker:compose:interactive exec away bash'",
-    "docker:delegate": "yarn docker:compose cp docker/create-delegation.mjs home:/app/packages/evm-wallet-experiment/docker/create-delegation.mjs && yarn docker:compose exec home node --conditions development /app/packages/evm-wallet-experiment/docker/create-delegation.mjs",
+    "docker:interactive:setup": "yarn docker:setup:wallets && yarn docker:compose:interactive exec kernel-away-bundler-7702 node /app/packages/evm-wallet-experiment/docker/setup-openclaw.mjs && yarn docker:compose:interactive exec -d kernel-away-bundler-7702 node /usr/local/lib/node_modules/openclaw/openclaw.mjs gateway && echo 'OpenClaw configured + gateway started. Shell in with: yarn docker:compose:interactive exec kernel-away-bundler-7702 bash'",
+    "docker:interactive:reset-openclaw": "yarn docker:compose:interactive exec kernel-away-bundler-7702 rm -rf /run/ocap/kernel-away-bundler-7702/.openclaw && echo 'OpenClaw state removed from ocap-run volume. Run: yarn docker:interactive:setup'",
+    "docker:delegate": "yarn docker:compose cp docker/create-delegation.mjs kernel-home-bundler-7702:/app/packages/evm-wallet-experiment/docker/create-delegation.mjs && yarn docker:compose exec kernel-home-bundler-7702 node --conditions development /app/packages/evm-wallet-experiment/docker/create-delegation.mjs",
     "docker:logs": "yarn docker:compose logs -f --tail=100",
     "test:e2e:docker": "vitest run --config vitest.config.docker.ts"
   },

Original file line number	Diff line number	Diff line change
`@@ -1,10 +1,4 @@`
`1`	`1`	`{`
`2`		`- "llm": {`
`3`		`- "upstreamOllama": "http://ollama:11434",`
`4`		`- "baseUrl": "http://llm:11434",`
`5`		`- "model": "qwen2.5:0.5b",`
`6`		`- "apiType": "ollama"`
`7`		`- },`
`8`	`2`	`"anvilChainId": 31337,`
`9`	`3`	`"openclawPluginPathContainer": "/app/packages/evm-wallet-experiment/openclaw-plugin"`
`10`	`4`	`}`