diff --git a/.github/workflows/endpoint-audit.yml b/.github/workflows/endpoint-audit.yml
new file mode 100644
index 0000000..ffefa64
--- /dev/null
+++ b/.github/workflows/endpoint-audit.yml
@@ -0,0 +1,143 @@
+name: Endpoint Audit
+
+on:
+  push:
+    branches: [ '**' ]
+  pull_request:
+
+# Needed so the audit can post its report as a comment on the PR.
+permissions:
+  contents: read
+  pull-requests: write
+
+jobs:
+  # Verify the client doesn't call endpoints the agent-server no longer exposes.
+  # The audit diffs the client's HTTP surface against the live server's OpenAPI
+  # spec. See scripts/endpoint-audit.mjs.
+  endpoint-audit:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    env:
+      AGENT_SERVER_PORT: 8010
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Use Node.js 20.x
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20.x
+          cache: 'npm'
+
+      - name: Install dependencies
+        run: npm ci
+
+      # Single source of truth: package.json's config.agentServerImage.
+      - name: Resolve agent-server image
+        run: echo "AGENT_SERVER_IMAGE=$(node -p "require('./package.json').config.agentServerImage")" >> "$GITHUB_ENV"
+
+      - name: Start agent-server (ground-truth OpenAPI spec)
+        run: |
+          docker pull "${AGENT_SERVER_IMAGE}"
+          docker run -d --name agent-server -p ${AGENT_SERVER_PORT}:8000 \
+            -e LOG_JSON=true "${AGENT_SERVER_IMAGE}"
+          for i in $(seq 1 30); do
+            curl -sf http://localhost:${AGENT_SERVER_PORT}/health && break
+            echo "waiting for agent-server ($i)..."; sleep 2
+          done
+
+      - name: Endpoint audit (gates on client/server mismatch)
+        run: npm run audit:endpoints
+
+      - name: Upload audit report
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: endpoint-audit
+          path: .audit/
+          # .audit/ is a dot-directory; upload-artifact@v4 skips hidden paths by default.
+          include-hidden-files: true
+
+      # Post the report as a single, self-updating comment on the PR. Runs even
+      # when the audit gate fails (the report is written before the gate), so
+      # reviewers see the mismatches that failed the check.
+      - name: Comment audit report on PR
+        if: always() && github.event_name == 'pull_request'
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const fs = require('fs');
+            const reportPath = '.audit/endpoint-audit.json';
+            if (!fs.existsSync(reportPath)) {
+              core.warning(`No audit report at ${reportPath}; skipping PR comment.`);
+              return;
+            }
+            const r = JSON.parse(fs.readFileSync(reportPath, 'utf8'));
+            const list = (items) =>
+              items && items.length ? items.map((i) => `- \`${i}\``).join('\n') : '_none_';
+            const marker = '<!-- endpoint-audit-report -->';
+            const NO_BACKEND = '(no known backend)';
+            const offContract = r.mismatch || [];
+            const byBackend = r.byBackend || {};
+            const status = offContract.length
+              ? `❌ ${offContract.length} off-contract call(s) — not on the agent-server`
+              : '✅ All client calls are on the agent-server';
+            const backendSections = Object.keys(byBackend)
+              .sort()
+              .map((name) => {
+                const heading =
+                  name === NO_BACKEND
+                    ? `#### ⛔ ${name} — served by no backend we can see (${byBackend[name].length})`
+                    : `#### ↗️ served by \`${name}\` (${byBackend[name].length})`;
+                return [heading, '', list(byBackend[name]), ''].join('\n');
+              })
+              .join('\n');
+            const classifiers =
+              r.classifiers && r.classifiers.length ? ` · classifiers: ${r.classifiers.join(', ')}` : '';
+            const summary = [
+              '| Category | Count |',
+              '| --- | ---: |',
+              `| ❌ Off-contract (not on agent-server) | ${offContract.length} |`,
+              ...Object.keys(byBackend)
+                .sort()
+                .map((name) =>
+                  name === NO_BACKEND
+                    ? `| &nbsp;&nbsp;⛔ no known backend | ${byBackend[name].length} |`
+                    : `| &nbsp;&nbsp;↗️ served by \`${name}\` | ${byBackend[name].length} |`
+                ),
+              `| ➕ Missing API (agent-server has, client lacks) | ${(r.missingApi || []).length} |`,
+              `| agent-server endpoints | ${r.agentServer} |`,
+              `| client endpoints | ${r.client} |`,
+            ].join('\n');
+            const body = [
+              marker,
+              '## Endpoint audit',
+              '',
+              `**${status}**${classifiers}`,
+              '',
+              summary,
+              '',
+              `### ❌ Not on agent-server (gated, ${offContract.length})`,
+              '',
+              offContract.length ? backendSections : '_none_',
+              `### ➕ Missing API — agent-server exposes it, client does not implement (${(r.missingApi || []).length})`,
+              '',
+              list(r.missingApi),
+            ].join('\n');
+            const { owner, repo } = context.repo;
+            const issue_number = context.issue.number;
+            const comments = await github.paginate(github.rest.issues.listComments, {
+              owner,
+              repo,
+              issue_number,
+            });
+            const existing = comments.find((c) => c.body && c.body.includes(marker));
+            if (existing) {
+              await github.rest.issues.updateComment({ owner, repo, comment_id: existing.id, body });
+            } else {
+              await github.rest.issues.createComment({ owner, repo, issue_number, body });
+            }
+
+      - name: Stop agent-server
+        if: always()
+        run: docker rm -f agent-server || true
diff --git a/.github/workflows/integration-tests.yml b/.github/workflows/integration-tests.yml
index e3b0541..e16ab31 100644
--- a/.github/workflows/integration-tests.yml
+++ b/.github/workflows/integration-tests.yml
@@ -16,7 +16,6 @@ env:
   AGENT_SERVER_PORT: 8010
   HOST_WORKSPACE_DIR: /tmp/agent-workspace
   AGENT_WORKSPACE_DIR: /workspace
-  AGENT_SERVER_IMAGE: ghcr.io/openhands/agent-server:1.29.0-python
 
 jobs:
   integration-test:
@@ -36,6 +35,10 @@ jobs:
       - name: Install dependencies
         run: npm ci
 
+      # Single source of truth: package.json's config.agentServerImage.
+      - name: Resolve agent-server image
+        run: echo "AGENT_SERVER_IMAGE=$(node -p "require('./package.json').config.agentServerImage")" >> "$GITHUB_ENV"
+
       - name: Build package
         run: npm run build
 
@@ -162,6 +165,10 @@ jobs:
       - name: Install dependencies
         run: npm ci
 
+      # Single source of truth: package.json's config.agentServerImage.
+      - name: Resolve agent-server image
+        run: echo "AGENT_SERVER_IMAGE=$(node -p "require('./package.json').config.agentServerImage")" >> "$GITHUB_ENV"
+
       - name: Build package
         run: npm run build
 
diff --git a/.gitignore b/.gitignore
index cf3977d..e5e4da9 100644
--- a/.gitignore
+++ b/.gitignore
@@ -93,3 +93,6 @@ jspm_packages/
 
 # TernJS port file
 .tern-port
+
+# endpoint-audit runtime coverage logs
+.audit/
diff --git a/AGENTS.md b/AGENTS.md
index 4321006..8b84bf8 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -304,7 +304,8 @@ Integration tests are in `src/__tests__/integration/` and require a running agen
 export LLM_API_KEY="your-api-key"
 export LLM_MODEL="anthropic/claude-sonnet-4-5-20250929"
 
-# Start agent-server in Docker (software-agent-sdk v1.29.0)
+# Start agent-server in Docker (software-agent-sdk v1.29.0;
+# canonical pin: package.json -> config.agentServerImage)
 docker run -d --name agent-server -p 8010:8000 \
   -v /tmp/agent-workspace:/workspace \
   ghcr.io/openhands/agent-server:1.29.0-python
@@ -336,7 +337,7 @@ Required GitHub secrets:
 
 ### CI Image Version
 
-- The integration workflow pins `ghcr.io/openhands/agent-server:1.29.0-python`, which corresponds to the `software-agent-sdk` release `v1.29.0`.
+- The agent-server image is defined **once** in `package.json` under `config.agentServerImage`. The `integration-tests.yml` and `endpoint-audit.yml` workflows read it from there at runtime, so bump the version in that single place (and the local-setup snippets above).
 - Keep the TypeScript client tests strict against that released server image rather than adding compatibility fallbacks for older prerelease builds.
 
 ## Agent Behavior Guidelines
diff --git a/README.md b/README.md
index 17eafc6..fea9fb7 100644
--- a/README.md
+++ b/README.md
@@ -377,7 +377,8 @@ Integration tests require a running agent-server in Docker with a mounted worksp
    chmod 777 /tmp/agent-workspace
    ```
 
-2. Start the agent-server container (software-agent-sdk v1.29.0):
+2. Start the agent-server container (software-agent-sdk v1.29.0; the canonical
+   image pin lives in `package.json` under `config.agentServerImage`):
 
    ```bash
    docker run -d \
diff --git a/endpoint-audit.config.json b/endpoint-audit.config.json
new file mode 100644
index 0000000..e8d70e4
--- /dev/null
+++ b/endpoint-audit.config.json
@@ -0,0 +1,23 @@
+{
+  "_comment": "Endpoint audit config. The client is GATED against the agent-server only: any client call not on the agent-server fails the gate. Additional `classify` specs are fetched live purely to label each gated call with the backend that actually serves it (cloud) vs no known backend at all; if a classify spec is unreachable its labels just degrade to 'no known backend'. See scripts/endpoint-audit.mjs.",
+  "specs": [
+    {
+      "name": "agent-server",
+      "role": "gate",
+      "url": "http://localhost:8010/openapi.json"
+    },
+    {
+      "name": "cloud",
+      "role": "classify",
+      "url": "https://app.all-hands.dev/openapi.json"
+    }
+  ],
+  "clientGlobs": ["src/client", "src/conversation", "src/workspace"],
+
+  "ignoreServerOnly": ["/v1/"],
+
+  "gate": {
+    "mismatch": true,
+    "missingApi": false
+  }
+}
diff --git a/package.json b/package.json
index 1649d11..e02e4bd 100644
--- a/package.json
+++ b/package.json
@@ -41,6 +41,9 @@
       }
     }
   },
+  "config": {
+    "agentServerImage": "ghcr.io/openhands/agent-server:1.29.0-python"
+  },
   "scripts": {
     "build": "tsc && node ./scripts/copy-json-assets.mjs && node ./scripts/rewrite-relative-imports.mjs",
     "lint": "eslint src/**/*.ts",
@@ -56,6 +59,7 @@
     "test:integration:llm": "jest --config jest.integration.config.cjs --runInBand --testPathIgnorePatterns=\"deterministic-api.integration.test.ts|bash.integration.test.ts\"",
     "test:git-dependency": "node ./scripts/smoke-test-git-dependency.mjs",
     "test:all": "npm run test && npm run test:integration && npm run test:git-dependency",
+    "audit:endpoints": "node ./scripts/endpoint-audit.mjs",
     "dev": "tsc --watch",
     "clean": "rimraf dist",
     "prepare": "npm run clean && npm run build",
diff --git a/scripts/endpoint-audit.mjs b/scripts/endpoint-audit.mjs
new file mode 100644
index 0000000..3823104
--- /dev/null
+++ b/scripts/endpoint-audit.mjs
@@ -0,0 +1,185 @@
+#!/usr/bin/env node
+/**
+ * Endpoint audit — reconciles the client's HTTP surface against the
+ * agent-server OpenAPI spec:
+ *
+ *   - The client is GATED against the agent-server spec only. Any endpoint the
+ *     client calls that the agent-server does NOT expose fails the gate — this
+ *     is an agent-server client, so off-contract calls are surfaced as errors
+ *     rather than silently allowlisted.
+ *   - To explain WHERE each off-contract call actually goes, extra `classify`
+ *     specs (e.g. the cloud app at app.all-hands.dev) are loaded and used only
+ *     to label each gated endpoint with the backend that serves it, or to mark
+ *     it as served by NO known backend (genuinely unsupported).
+ *   - missing API — agent-server exposes an endpoint the client does NOT
+ *     implement (informational, not gated).
+ *
+ * Specs are fetched live from each backend (preferred) or read from a committed
+ * fallback file. A `gate` spec with neither is a hard error; a `classify` spec
+ * that is unavailable is skipped (its endpoints just go unlabeled).
+ *
+ * Exit code is non-zero on gating violations (see config.gate).
+ */
+import fs from 'node:fs';
+import path from 'node:path';
+
+const ROOT = process.cwd();
+const cfg = JSON.parse(fs.readFileSync(path.join(ROOT, 'endpoint-audit.config.json'), 'utf8'));
+
+const VERBS = ['get', 'post', 'put', 'patch', 'delete'];
+
+// "/api/x/${id}" or "/api/x/{name}" -> "/api/x/{}", strip query + trailing slash
+const norm = (verb, p) =>
+  `${verb.toUpperCase()} ${
+    p
+      .split('?')[0]
+      .replace(/\$?\{[^{}]*\}/g, '{}')
+      .replace(/\/+$/, '') || '/'
+  }`;
+
+// ---------------------------------------------------------------------------
+// 1. Backend specs: agent-server (gate) + others (classify)
+// ---------------------------------------------------------------------------
+async function loadSpec(spec) {
+  if (spec.url) {
+    try {
+      const res = await fetch(spec.url, { signal: AbortSignal.timeout(8000) });
+      if (res.ok) return await res.json();
+      console.warn(`  ! ${spec.name}: ${spec.url} -> HTTP ${res.status}, falling back to file`);
+    } catch (e) {
+      console.warn(`  ! ${spec.name}: ${spec.url} unreachable (${e.message}), falling back to file`);
+    }
+  }
+  if (spec.file && fs.existsSync(path.join(ROOT, spec.file)))
+    return JSON.parse(fs.readFileSync(path.join(ROOT, spec.file), 'utf8'));
+  if (spec.role === 'classify') {
+    console.warn(`  ! ${spec.name}: no reachable url and no fallback file — skipping (classify only)`);
+    return null;
+  }
+  throw new Error(`spec "${spec.name}": no reachable url and no fallback file`);
+}
+
+const specToSet = (doc) => {
+  const set = new Set(); // normalized "VERB /path"
+  for (const [p, methods] of Object.entries(doc.paths ?? {}))
+    for (const verb of Object.keys(methods)) if (VERBS.includes(verb)) set.add(norm(verb, p));
+  return set;
+};
+
+const gate = new Set(); // union of all `gate` specs (the agent-server contract)
+const classifiers = []; // [{ name, set }] for labelling off-contract calls
+for (const spec of cfg.specs) {
+  const doc = await loadSpec(spec);
+  if (!doc) continue;
+  const set = specToSet(doc);
+  if (spec.role === 'classify') classifiers.push({ name: spec.name, set });
+  else for (const k of set) gate.add(k);
+}
+
+// ---------------------------------------------------------------------------
+// 2. Client surface: static extraction of every endpoint the client can call
+// ---------------------------------------------------------------------------
+function walk(dir) {
+  const out = [];
+  for (const e of fs.readdirSync(dir, { withFileTypes: true })) {
+    const full = path.join(dir, e.name);
+    if (e.isDirectory()) {
+      if (e.name === '__tests__' || e.name === 'node_modules') continue;
+      out.push(...walk(full));
+    } else if (e.name.endsWith('.ts')) out.push(full);
+  }
+  return out;
+}
+
+const VERB_CALL = /\.(get|post|put|patch|delete)\s*(<[^>]*>)?\(/;
+const PATH_LIT = /[`'"](\/(?:api|server_info|alive|health|ready)[^`'"]*)[`'"]/;
+
+const client = new Set();
+for (const glob of cfg.clientGlobs) {
+  for (const file of walk(path.join(ROOT, glob))) {
+    const lines = fs.readFileSync(file, 'utf8').split('\n');
+    lines.forEach((ln, i) => {
+      // direct `.get/post/...(` calls, plus generic `.request({ method, url })`
+      let verb = ln.match(VERB_CALL)?.[1];
+      let win = 3;
+      if (!verb && /\.request\s*\(/.test(ln)) {
+        const block = lines.slice(i, i + 6).join('\n');
+        verb = block.match(/method:\s*['"`](GET|POST|PUT|PATCH|DELETE)['"`]/i)?.[1];
+        win = 6;
+      }
+      if (!verb) return;
+      const p = lines
+        .slice(i, i + win)
+        .join('\n')
+        .match(PATH_LIT)?.[1];
+      if (p) client.add(norm(verb, p));
+    });
+  }
+}
+
+// ---------------------------------------------------------------------------
+// 3. Diff + classify
+// ---------------------------------------------------------------------------
+const sorted = (it) => [...it].sort();
+const ignoredApi = (k) => (cfg.ignoreServerOnly ?? []).some((pre) => k.includes(pre));
+
+const NO_BACKEND = '(no known backend)';
+const classify = (k) => classifiers.find((c) => c.set.has(k))?.name ?? NO_BACKEND;
+
+// Every client call the agent-server does NOT expose is off-contract -> gated.
+const mismatch = sorted([...client].filter((k) => !gate.has(k)));
+const byBackend = {}; // backend name -> [endpoints]
+for (const k of mismatch) (byBackend[classify(k)] ??= []).push(k);
+
+const missingApi = sorted([...gate].filter((k) => !client.has(k) && !ignoredApi(k)));
+
+// ---------------------------------------------------------------------------
+// 4. Report
+// ---------------------------------------------------------------------------
+const section = (title, items) => {
+  console.log(`\n${title} (${items.length})`);
+  for (const k of items) console.log(`  ${k}`);
+};
+console.log(
+  `\n=== Endpoint audit — agent-server=${gate.size} client=${client.size}` +
+    ` classifiers=[${classifiers.map((c) => c.name).join(', ') || 'none'}] ===`
+);
+console.log(`\n❌ NOT ON AGENT-SERVER — off-contract client calls (${mismatch.length})`);
+for (const name of Object.keys(byBackend).sort()) {
+  const tag = name === NO_BACKEND ? `${name} ⛔` : `served by: ${name}`;
+  console.log(`  ${tag} (${byBackend[name].length})`);
+  for (const k of byBackend[name]) console.log(`    ${k}`);
+}
+section('➕ MISSING API — agent-server has it, client does not implement', missingApi);
+
+fs.mkdirSync(path.join(ROOT, '.audit'), { recursive: true });
+fs.writeFileSync(
+  path.join(ROOT, '.audit/endpoint-audit.json'),
+  JSON.stringify(
+    {
+      agentServer: gate.size,
+      client: client.size,
+      classifiers: classifiers.map((c) => c.name),
+      mismatch,
+      byBackend,
+      missingApi,
+    },
+    null,
+    2
+  )
+);
+
+// ---------------------------------------------------------------------------
+// 5. Gate
+// ---------------------------------------------------------------------------
+const g = cfg.gate ?? {};
+const violations = [];
+if (g.mismatch !== false && mismatch.length)
+  violations.push(`${mismatch.length} off-contract (not on agent-server)`);
+if (g.missingApi && missingApi.length) violations.push(`${missingApi.length} missing API`);
+
+if (violations.length) {
+  console.error(`\n❌ endpoint audit failed: ${violations.join(', ')}`);
+  process.exit(1);
+}
+console.log('\n✅ endpoint audit passed');