fix(server): failed Cursor turns show Cursor's reason#16146

Problem

When a Cursor turn fails, the thread shows "Provider turn failed.", and a delegate_task parent gets the same text, even though Cursor reported the reason. Over 7 days on one machine, 7 of 59 Cursor runs failed this way. The reasons Cursor sent in run.completed were:

• "Model Blocked Please ask your admin to enable access to Claude Fable 5."
• "AI Model Not Found Invalid parameters for registry model: "grok-4.7""
• "[resource_exhausted] Error"
• "[unknown] [canceled] This operation was aborted"

A parent can't tell a blocked model from a capacity limit, so it retries blindly. There are two causes:

• Run ends with an error: CursorAdapterV2 passes the run's error to makeProviderFailure only as cause, and makeProviderFailure doesn't read a cause's text.
• SDK throws: CursorAgentSdkRunnerError.message is only Cursor Agent SDK <method> failed., which is what a failed session start shows.

Change

• A run that ends with status: "error" fails the turn with the run's own error.message, and keeps error.code when Cursor sends one.
• cursorSdkFailureMessage (in CursorAgentSdk.ts) builds the failure text for a runner error from the SDK error's message, plus its code and HTTP status when they say something; the adapter passes it to makeProviderFailure. CursorAgentSdkRunnerError.message itself stays generic, per the repo's error conventions. The generic codes "error" and "unknown" are dropped, and so is a code the message already contains. Example: Cursor Agent SDK agent.create failed: Invalid User API Key (HTTP 401).
• A run that throws fails the turn with that message.

The text still goes through makeProviderFailure's redaction and 4,096-character bound. No contract or client change.

Scope and approval

This is a very small, focused fix for an obvious bug, so it qualifies without a prior issue: T3 already receives Cursor's reason and drops it before the user or the parent agent sees it. Nothing changes beyond which text is shown, and the change stays in the two Cursor adapter files. Other adapters already pass the provider's text (Codex payload.error.message, OpenCode 2 event.data.error.message).

How this differs from the open PRs:

• #14969 makes session-open and thread-load failures read through to the first error underneath, for every provider, and adds ACP's data. For Cursor that first error is CursorAgentSdkRunnerError, whose message is still only "Cursor Agent SDK agent.create failed." without this PR. It doesn't touch failures after a run started, which are the cases above. The two PRs don't overlap and complement each other.
• #15569 changes only the provider status message for plan_required in CursorProvider.ts, not turn failures.

Related: #15075 (plan_required), #15901 (Windows run.start failed; its runner error would now carry the SDK's sandbox message, not tested on Windows), #15447 (generic failure while the run continues).

Verification

Focused tests:

• packages/provider-cursor/src/server/adapter.test.ts, fails the turn with the error Cursor's run ended with: a run result { status: "error", error: { message: "Model Blocked …" } } fails the turn with that message, class provider_error.
• adapter.test.ts, fails the turn with the SDK error that broke the run: run.wait throws a runner error around { message: "Too many requests", code: "resource_exhausted", status: 429 }. The turn fails with Cursor Agent SDK run.wait failed: Too many requests (resource_exhausted, HTTP 429), class transport_error.
• CursorAgentSdk.test.ts, names the SDK's reason, code and HTTP status in a runner error: covers an added code, a code already in the message, the generic error code, and no cause.
• All three fail against main (expected 'Provider turn failed.' to equal 'Model Blocked …', expected 'Cursor Agent SDK agent.create failed.' to equal …).

Live, with T3 from source and @cursor/sdk 1.0.31. Each cell is the thread's terminal error item:

Not reproduced live: [resource_exhausted] and [canceled]. Both arrive the same way, as run.completed with status: "error" and error.message, which the first test covers.

Rebased onto main @ 2025430708 on 2026-10-11 to fix CI: the new adapter test now builds the adapter the way main's Effect-conventions refactor does (CursorAgentSdkRunner service, TestProviderHost.layer, McpProviderSessions.layer). The provider-cursor tests pass, except 4 Cursor shell spawn guard tests in CursorSdk.test.ts that fail in my local sandbox (they spawn processes and touch no file this PR changes); fmt, lint and both typechecks pass. Before that, rebased onto main @ b707eeb052 on 2026-10-09 (Cursor moved into packages/provider-cursor; the adapter test's CursorAgentSdk import conflicted, and the new failure test now builds the adapter with a ProviderHost instead of ServerConfig, as main's tests do); the focused tests (26/26), lint, fmt and typecheck (vp run --filter t3 typecheck and provider-cursor's tsc) pass.

Model: Claude Opus 5.5 (1M). Harness: Claude Code in T3 Code.