Cursor ref-less checkpoint lookup can reuse another conversation's state

[luvs01]

Client or integration

Codex App and other Responses clients using the Cursor provider adapter

Area

Provider adapter

Summary

On current dev (e1fb6755951c82fc0fc137820d2b40208d79e918), a request without an explicit checkpointRef can reuse a live Cursor checkpoint from a different conversation.

getCursorCheckpointForPrefix(...) requires a unique prefix digest and matches the system digest, covered-message count, credential identity scope, and model affinity. It does not match conversationId. resolveCursorCheckpoint(...) then enforces conversation equality only when an explicit ref exists.

At the same time, resolveCursorConversationId(...) generates a fresh provider conversation ID when neither _cursorConversationId nor a stable client-thread owner is available. The generated conversation can therefore import the sole matching checkpoint from another conversation.

The checkpoint is provider conversation state rather than a content-addressed prompt cache. Reusing it can carry prior conversation structure, summaries, tool/subagent state, and file-related state into an unrelated request.

Expected safety boundary:

• An explicit checkpoint ref keeps the existing conversation, identity, model, lineage, and health checks.
• Ref-less prefix lookup is allowed only when the request already has a stable owner: a remembered _cursorConversationId, a non-empty parent _clientThreadId, or the existing bounded process-local HMAC derived from the complete Desktop session-id and thread-id pair.
• The selected snapshot must match the resolved request conversationId.
• A ref-less request with no stable owner fails closed to full replay; content-prefix uniqueness alone is not ownership authority.
• Isolated helper turns do not reuse parent or sibling checkpoints.
• The Desktop fallback is Cursor-only: it does not widen Responses replay scope or persist raw headers.

This is the still-applicable first P1 finding left on merged PR #2277: #2277 (comment)

Reproduction

1. Under one Cursor credential scope and model, complete conversation A whose first user message has a unique prefix, then commit its ConversationStateStructure checkpoint.
2. Start unrelated conversation B without a continuation ref or stable client-thread identity.
3. Give B a history beginning with the same user prefix plus a later suffix.
4. Observe B enter checkpoint mode with A's checkpoint bytes even though B has a different generated conversationId.

Regression coverage for the fix should:

• reject a unique matching prefix from a different conversation;
• preserve ref-less reuse for stable _cursorConversationId, parent-thread, and bounded Desktop session/thread owners;
• preserve forced-fresh recovery with the Desktop owner;
• select the owned snapshot when multiple conversations share one prefix;
• avoid refreshing TTL/LRU for unrelated same-prefix snapshots;
• full-replay when no stable owner exists; and
• preserve explicit-ref behavior and helper isolation.

Version

dev at e1fb675

Operating system

Cross-platform adapter logic; reproduced in source-level tests on Windows 11

Provider and model

Cursor / gpt-5.6-sol (the ownership bug is model-independent)

Logs or error output

No runtime error is emitted. The incorrect state reuse is observable through continuationMode=checkpoint and checkpoint bytes belonging to another conversation.

Screenshots and supporting files

No screenshot is required for this non-GUI state-isolation defect.

Redacted configuration

{
  "provider": "cursor",
  "credentialScope": "same-account",
  "checkpointRef": null,
  "stableClientThreadOwner": null
}

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.

[github-actions]

Issue reopened

The report now contains the information required by the automated check. Thanks for updating it.

[lidge-jun]

Landed via #2563 at 516f566