Gateway client fails against OpenClaw 2026.9.5: Control UI build check on connect, and sessions.create before the new agent registers

[FancyKat]

clawbench run cannot complete a task against an OpenClaw 2026.9.5 gateway (ec9c1a1). There are two separate incompatibilities in clawbench/client.py, found at ClawBench e3f8d25 (current main).

Setup

• ClawBench image built from e3f8d25 with --build-arg OPENCLAW_IMAGE= pointing at the ghcr.io/openclaw/openclaw:2026.9.5 index (linux/arm64).
• A loopback gateway in the same container:

  openclaw gateway run --allow-unconfigured --bind loopback --port 18789 --auth token --token "$TOKEN"
  

• The run:

  clawbench run --model ollama/gemma4:26b --task t1-fs-quick-note --runs 1
  

  (OPENCLAW_GATEWAY_TOKEN set; judge off.)

1. The handshake is rejected as a stale Control UI

RuntimeError: RPC connect failed: UNAVAILABLE - protocol mismatch: Control UI updated; reload this page to continue

Gateway log:

[ws] control ui build rejected ... clientBuild=legacy gatewayBuild=2026.9.5-ec9c1a13db89-2026-09-18T18-50-25.957Z; reload required

The client connects as {"id": "openclaw-control-ui", "mode": "ui"} with no buildId. The 2026.9.5 gateway's connect handler (resolveControlUiBuildMismatch) compares a Control UI client's buildId with its own build and closes the connection when they differ. A missing id counts as legacy. The protocol range (3–4) is fine: the gateway speaks v4.

Identifying as the CLI ({"id": "cli", "mode": "cli"}) gets past the handshake. That client then needs its device identity paired: its first connect answers NOT_PAIRED, and openclaw devices approve <requestId> resolves it. Setting CLAWBENCH_DISABLE_GATEWAY_DEVICE_IDENTITY=1 avoids the pairing, but the connection then lacks operator.admin, and agents.create fails with FORBIDDEN - missing scope: operator.admin.

2. sessions.create runs before the new agent exists

RuntimeError: RPC sessions.create failed: INVALID_REQUEST - Unknown agent id "clawbench-t1-fs-quick-note-run-0-541829"

Gateway log (timestamps):

16:44:43.767 res ✓ agents.create 78ms
16:44:43.808 res ✗ sessions.create ... Unknown agent id "clawbench-t1-fs-quick-note-run-0-541829"
16:44:43.952 [reload] config hot reload applied (... agents.entries.clawbench-t1-fs-quick-note-run-0-541829 ...)

On 2026.9.5, agents.create returns before the agent is registered. The agent becomes known when the config hot reload applies, about 200 ms later.

Suggested fixes

1. Connect as a backend or CLI client rather than the browser Control UI, or send a buildId the gateway accepts. Document the device-pairing step for token-authenticated harness clients, or have the worker approve its own pairing request, since it owns the gateway.
2. After agents.create, wait until the agent is registered before sessions.create: poll agents.list, or retry Unknown agent id briefly.

Workaround used

This local patch to clawbench/client.py, plus pairing the harness device with openclaw devices approve, made the smoke pass on 2026.9.5. The run scored 0.880 on t1-fs-quick-note with ollama/gemma4:26b (completion 1.00, trajectory 0.60, behavior 1.00).

-                    "id": "openclaw-control-ui",
+                    "id": "cli",
                     "version": __version__,
                     "platform": "linux",
-                    "mode": "ui",
+                    "mode": "cli",
@@ async def create_session(
-        response = await self._rpc("sessions.create", params)
+        for attempt in range(20):
+            try:
+                response = await self._rpc("sessions.create", params)
+                break
+            except RuntimeError as error:
+                if "Unknown agent id" not in str(error) or attempt == 19:
+                    raise
+                await asyncio.sleep(0.25)

[clawsweeper]

Codex review: this still needs some work. Reviewed October 4, 2026, 6:44 PM ET / 22:44 UTC.

Summary
Both failures remain present on current main. The open fix PR #92 owns the repair, and the discussion reports successful validation against OpenClaw 2026.9.5; keep this issue open until that PR merges.

Reproducibility: yes. the supplied run command and gateway logs identify both failures, and current source retains their triggers. This read-only review did not execute the reproduction.

Next step
Review and land or resolve #92 before closing this issue; a second repair PR would duplicate active work. Likely related people: scoootscooob and gchlebus are medium-confidence routing candidates from client history.

Review details

Best possible solution:

Use a correctly signed CLI identity with enforced pairing and bounded registration recovery for newly created agents.

Do we have a high-confidence way to reproduce the issue?

Yes: the supplied run command and gateway logs identify both failures, and current source retains their triggers. This read-only review did not execute the reproduction.

Is this the best way to solve the issue?

Yes: the linked repair targets client identity and narrowly bounded registration retries while preserving gateway authorization; its discussion reports validation of both recovery and permission denial.

AGENTS.md: not found in the target repository.

Codex review notes: model internal, reasoning medium; reviewed against e3f8d25a01f4.

Label changes

Label changes:

• add clawsweeper:linked-pr-open: Current issue advisory state selects this label.
• add clawsweeper:no-new-fix-pr: Current issue advisory state selects this label.

Label justifications:

• P1: The reported gateway incompatibilities prevent benchmark tasks from reaching execution on OpenClaw 2026.9.5.
• impact:auth-provider: Client identity and device-pairing admission prevent authenticated gateway access needed for agent creation.

Evidence reviewed

What I checked:

• Current main retains browser identity: The checkout and live main both resolve to e3f8d25. Connection metadata still uses openclaw-control-ui and ui mode without a buildId, and passes that identity into device signing. (clawbench/client.py:267, e3f8d25a01f4)
• Registration race remains exposed: Agent creation returns the RPC’s agent ID immediately; session creation makes a single RPC without registration polling or a targeted retry. The harness proceeds from agent creation directly to session creation. (clawbench/harness.py:348, e3f8d25a01f4)
• Existing test does not cover the failures: The WebSocket session smoke returns successful admission and session responses; it does not exercise stale Control UI rejection or delayed agent registration. (tests/test_client.py:27, e3f8d25a01f4)
• Open implementation candidate: GitHub confirms fix: restore gateway compatibility with OpenClaw 2026.9.5 #92 remains open and unmerged at head cc71bb4. Its supplied body explicitly links this issue and covers both failures. (cc71bb4323da)
• Reported real gateway validation: Gateway client fails against OpenClaw 2026.9.5: Control UI build check on connect, and sessions.create before the new agent registers #91 (comment) reports the original admission failure, registration failures for three new agents, patched recovery for all three, explicit device approval, paired reconnect, and continued operator.admin denial when identity is disabled. These are contributor-reported observations, not executions performed during this review.
• Feature-history routing: Client history identifies repeated work by scoootscooob and gateway timeout work by gchlebus; GitHub commit metadata verifies both handles. Exact line provenance could not be completed because blame and content-history searches required an unavailable historical blob and returned HTTP 403. Current files and commit-log history were inspected successfully. (clawbench/client.py, 0f1b45e4674b)

Likely related people:

• scoootscooob: Suggested for follow-up; no historical authorship or introduction is verified. (role: unverified routing candidate; confidence: low)
• gchlebus: Suggested for follow-up; no historical authorship or introduction is verified. (role: unverified routing candidate; confidence: low)

How this review workflow works

• ClawSweeper keeps one durable marker-backed review comment per issue or PR.
• Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
• A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
• PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
• Maintainers can also comment @clawsweeper review to request a fresh review only.
• Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
• Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
• Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.

[steipete]

Implemented in #92, awaiting code-owner review; leaving this issue open until the fix is merged.

The original client reproduced the stale Control UI admission failure against OpenClaw 2026.9.5 (ec9c1a1). Immediate session creation also reproduced the registration race for three newly created agents; the patched client recovered for all three. A fresh device with local auto-approval disabled received NOT_PAIRED, explicit approval enabled the run, reconnect reused the paired identity, and disabling identity still denied agents.create for missing operator.admin.

Four focused regressions failed before their fixes. Final AWS Crabbox validation: 548 tests passed, 5 skipped; Ruff, changed-file hooks, dependency consistency and wheel checks passed. Independent Codex review through P2 is clean. The live smoke used synthetic state and no model/provider calls: https://crabbox.openclaw.ai/portal/runs/run_225f0bbea4e6abe1c1db9f022a1b945a

Exact head cc71bb4323dace43a62bb2a40494440f8ba9db21 passes Python 3.11 and 3.12 CI.