[WRONG BRANCH] fix(cursor): avoid quadratic replay cost by tracking entry index and bounding replay window

[luvs01]

Motivation

• The external-model Cursor replay deduper performed entries.indexOf(...) for every duplicate, creating attacker-controlled O(k·r) CPU behavior that could stall the server event loop.
• The replay traversal processed the full rawMessages before pruning, so large attacker-controlled histories could trigger the quadratic path before any size limits applied.

Description

• Replace array search with a tracked index: lastReplayEntry is replaced by lastReplayEntryIndex and duplicate replacements update the known index directly to make each dedupe operation O(1).
• Introduce CURSOR_EXTERNAL_REPLAY_MESSAGE_LIMIT = 4096 and limit the synchronous replay construction to the newest bounded suffix for external wire models so replay processing is bounded before later wire-size pruning.
• Reset the tracked index on user/developer turns to preserve the same-collapse semantics and keep the existing repetition-marking behavior intact.
• Add a unit test bounds replay construction before processing an oversized history to tests/cursor-repetition-breaker.test.ts that verifies oversized histories outside the bounded replay window do not trigger the severe repetition note.

Testing

• Ran the focused tests with bun test tests/cursor-repetition-breaker.test.ts and all 6 tests passed.
• Ran bun run typecheck (TS check) which completed without errors.
• Attempted a full bun run test; the repository-wide suite ran many tests and reported unrelated failures/timeouts in other subsystems, so only the focused regression and typecheck were relied on to validate this scoped fix.

---

Codex Task

[github-actions]

✅ Deterministic PR hygiene checks passed.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: a33da723-8212-4fe9-9386-57f42202a368

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[github-actions]

⏳ DRAFT

• wrong target branch (main); retarget to dev.

What to do

• Retarget this PR to dev — all contributions go to dev.

Its title has been prefixed with [WRONG BRANCH].
This pull request was already a draft. Its draft status will be preserved after every issue above is resolved.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d16b150861

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Align the capped replay window to a complete turn

When more than 4,096 messages follow the most recent historical user message, this cutoff starts on an assistant/tool-result entry. The later complete-turn selector requires a user root, so with the history constructed by the new test (old turn, 4,101 assistant messages, then the active user), it rejects the orphaned suffix and sends only the system root; the external model receives the active “continue” action with none of the conversation being continued. Anchor the bounded suffix at a user/developer boundary, or explicitly retain the preceding user message, so limiting construction cannot erase all replay context.

Useful? React with 👍 / 👎.