fix(server): honor Cursor long-context selections

[emcee6798]

Problem

Cursor SDK requests omit RequestedModel.maxMode, so Grok 4.7 at 500K fails with AI Model Not Found, while other long-context selections can use the wrong tier. This is the bug reported in #15788 and originally fixed by @MatthewFeroz in #15884.

Change

This ports #15884 to packages/provider-cursor after the Cursor provider moved out of the server app. Selections above 300K set Max Mode; standard tiers remain unchanged. Credit for the original long-context fix belongs to @MatthewFeroz and the contributors to #15884; its commit remains authored by Matthew Feroz. This replaces #18685 to correct Git identity. Both commits now use emcee6798 as committer, and the test follow-up uses emcee6798 as author. The final code is identical to #18685 at 3958063ebe.

Scope and approval

This is the same focused obvious-bug fix as #15884, with no unrelated cleanup. Relevant feedback from the original PR is preserved by narrowly patching only agent.v1.RequestedModel, retaining existing values, and testing the real SDK request path. The separate default-option issue in #16149 and generic provider-error presentation are intentionally out of scope.

Verification

Verified on replacement head f50aad2422 before pushing, from packages/provider-cursor:

• vp test run src/server/CursorSdk.test.ts src/server/sdk.test.ts — 8 tests passed, including real SDK request checks for Grok 4.7 at 500K and 256K.
• Targeted format and lint checks passed.
• vp run typecheck passed with existing unrelated Effect suggestions.

The replacement commit trees and messages match #18685 exactly; only Git identity and commit metadata changed. No live authenticated Cursor turn was run on this replacement head. #15884 contains the original macOS and Linux live verification.

[macroscopeapp]

Approvability

Verdict: Approved at f50aad2

Macroscope's review found this PR approvable — The PR narrowly fixes serialization of explicitly selected Cursor long-context tiers while preserving standard requests and unrelated protobuf messages. Production behavior is confined to the affected model request, and dedicated tests cover the real SDK path and edge cases.

Notes:

• All code in this push has already been reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Path: .coderabbit.config.ts
• Review profile: CHILL
• Plan: Advanced
• Run ID: 0616a43e-5520-42b7-8401-253e5df9fa94

---

📥 Commits

Reviewing files that changed from the base of the PR and between 701fc83 and f50aad2.

---

📒 Files selected for processing (2)

• packages/provider-cursor/src/server/sdk.test.ts
• packages/provider-cursor/src/server/sdk.ts

---

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 7 remain after this review.

---

---

📝 Walkthrough

Walkthrough

The Cursor provider wraps protobuf initialization to set Max Mode on qualifying RequestedModel context tiers. New child-process tests check real SDK sends and serialized messages, including parameter preservation and unchanged input objects.

Changes

Cursor Max Mode context tiers

Layer / File(s)	Summary
RequestedModel initialization packages/provider-cursor/src/server/sdk.ts	The initPartial wrapper sets maxMode to true for RequestedModel messages when the parsed context tier exceeds 300,000 and maxMode is not already enabled.
SDK and serialization tests packages/provider-cursor/src/server/sdk.test.ts	Child-process probes test Max Mode values in live SDK sends and serialized messages. They also check preserved parameters, unchanged input objects, and the default value for an unrelated message type.

Priority: ➖ Normal

Estimated code review effort: 3 (Moderate) | ~20 minutes

Change: Bug fix

Suggested reviewers: juliusmarminge

---

Merge Risk | ⚪ Minimal · up to f50aa

Merge Risk: ⚪ Minimal · up to f50aa

The change appears ready to merge after normal checks. No actionable risk remains identified in the reviewed Max Mode behavior.

Security Architecture Review

Security architecture risk: 🔵 Low · up to f50aa

The change is limited to long-context request selection and does not show a new permission or access path. Compatibility risk is low, with authenticated operation and unusual failure recovery still unverified.

Retained concerns
No architecture-level concerns identified.

Security review details

Security Blast Radius

• inferred — The behavioral scope is qualifying RequestedModel construction through the patched runtime within a process, not just one test case or selection. The exact message-type guard limits the mutation, but it is not a per-user installation. No expanded credential, data-store, or environment authority was established by the inspected change.

Security Findings and Attack Paths

• inferred — The routed program-execution identifiers belong to a private test harness using an embedded program and two literal probe modes. Their source does not establish a new attacker-controlled command endpoint or a production control bypass.

Trust Boundaries and Controls

• observed — The inspected text-generation caller still requires an enabled provider and an API key before agent creation. Its workspace-isolation checks and agent options are unchanged in the full PR comparison. The new flag derivation occurs inside request construction rather than replacing those controls.

Resilience and Maintainability Implications

• inferred — The wrapper has no asynchronous step or per-request shared accumulator. Its added mutation is target-local and monotonic, so the inspected logic does not introduce a cross-request state-carrying transition. Original-initializer exceptions prevent the added tail from running; rollback of any original partial construction is not established.

---

Pre-merge checks | 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check		The title clearly and concisely describes the primary change: honoring Cursor long-context selections. It uses a valid conventional commit format.
Description check		The description includes all required sections. It explains the problem, implementation, scope, approval basis, verification commands, test results, and known verification limits.
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.

---

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

---

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

_{Comment @coderabbitai help to get the list of available commands.}

[rcatajar]

Tested the Max Mode rule from this PR locally. It works for the Grok 4.7 500K failure.

This nightly (0.0.46-nightly.20261014.3118, @cursor/sdk 1.0.35) does not include the loader hook, and the server is a packaged binary, so I could not apply sdk.ts itself. The local runtime still builds RequestedModel without maxMode in @cursor/sdk/dist/cjs/689.js (i.e(689) from local-executor). I set maxMode on that constructor call when the context param is above 300K, matching the threshold in this PR, and made the same edit in the ESM and bundled copies.

After restarting the server onto the patched runtime, Cursor runs of grok-4.7 with context=500k finished. Before the edit those runs died in about a second with AI Model Not Found Invalid parameters for registry model: "grok-4.7". Finished 500K runs after the restart include 2026-10-14T20:08:18Z and, on 2026-10-15, runs lasting about 17, 18, and 27 minutes. 256K still completes. I did not re-check a conversation checkpoint for a 1,000,000-token window.

The edit is wiped by the next nightly, since it lives in that runtime's node_modules. Shipping this PR is still the real fix.

[mazmaa]

Hi Team, any update?