Ayaan Zaidi obviyus

Connector sign-in in Settings

Before: 9d9c673c92e4ab5c6fd0d989c7160c447144d4ff. Browser and installed-package after: 229f2b7fe6fb5c35e865b614867f9e49207f2262. Final source: a73d2cd4dc674f75feefe9387c15aa2f6679a355. The final revision only changes tests; production bytes are unchanged. Integrated main: aac3b2773c7f80bede88e8ce67417ed72a3eb33d.

Browser behavior

The baseline's configured example-docs row offers only openclaw mcp login example-docs. The candidate's real Settings action runs the existing browser wizard and reports Authentication saved. The sanitized images are crops of these actual rendered screens, not mockups.

Cancel followed by retry exposed stale cancellation feedback in an earlier candidate. The corrected shared controller passes the same browser sequence and eight controlled event-order cases across MCP and Models. A separate source run saves authentication, acquires the connector through the real runtime and calls its tool before and after an owned Gateway restart. Restar

OpenClaw Full tool default — browser evidence

Captured from a real loopback Gateway with isolated configuration, workspace and home directories. No provider credentials, real channels or inference sessions were configured. Images show configured tool availability, not successful tool execution. Gateway authentication was disabled only in this loopback test fixture.

Onboarding

The same supported command ran on baseline 9ca2d1dfcd161ed862a30749f40486e95d6459f1 and the candidate:

pnpm openclaw onboard --non-interactive --accept-risk --mode local \

OpenClaw model support proof

Before and after the model support repair in openclaw/openclaw#150128.

The same synthetic custom Google provider uses gemini-2.5-flash with the alias Paid Gemini. The repaired resolver supplies the native model name, context size, and thinking metadata. Explicit model definitions and settings retain their own values.

Before: revision 673efe0e1b9.
After: revision 8afb413ec9a.

Both captures use the packaged Control UI and an isolated Gateway.

Forwarded message attribution

Before and after screenshots for openclaw/openclaw#138579.

The first message has a sender asserted only in text. The second has a recorded source session.

Complete subagent answer evidence

An isolated Gateway serves Chat against a deterministic local provider. The parent spawns a child with delete cleanup and yields. Child completion is held until the recorded yield. An 8,638-character answer loses its required tail on the baseline; the corrected parent receives all 11,758 escaped characters and retains them in a later turn after child deletion.

A public nested case gives the middle child a distinct conclusion; that complete answer reaches the root. Earlier aggregate, nested-yield, and ordinary completion cases pass. Native children cannot call the message tool under the existing hard deny; that live case is excluded. The registered maybeWakeRequesterAfterAllChildrenSettled regression uses capped visible terminal evidence and a complete >4,096-character exact-run transcript with a trailing same-run NO_REPLY row. It fails on the original implementation and passes here. The existing producer contract is documented by src/agents/embedded-agent-runner/run-entry

	Protected credential request delivery

	Before: Telegram did not receive the protected form link. The request remained pending and expired without storing a credential. The screenshot shows the form opened directly from the Gateway question record for inspection, not through Telegram.
	After: Telegram received the protected form link. Opening it showed a masked API key field. Submitting a synthetic credential completed the request and returned only metadata with the edited allowed hosts. Reloading showed Answered with no input or Submit button.

	The screenshots use synthetic data. No actual credential is shown.