fix(anthropic): bound terminal guard buffering

[luvs01]

Motivation

• Prevent an unbounded-memory DoS introduced by the Anthropic terminal guard that retained every non-terminal streaming event in-memory until a terminal done was observed.
• Ensure the guard still detects short no-tool announcements but stops analysing/retaining once a stream is too large or a real tool call appears.

Description

• Add bounded retention limits: MAX_RETAINED_EVENTS and MAX_RETAINED_CONTENT_CHARS, and track retained text/content size while streaming.
• Introduce an analysisEnabled flag that is cleared when any bound is exceeded or a tool_call_start is seen, and clear the per-turn seen buffer when analysis is disabled so the guard stops growing memory usage.
• Only run analyzeTerminalTurn when analysisEnabled is true; otherwise fall back to pass-through behavior and do not attempt continuations.
• Add a regression test stops retaining events once a streamed answer cannot be a short announcement that verifies oversized/non-terminal streams are still forwarded to the client but not retained for analysis or continuation.

Testing

• Ran the focused suite bun test tests/terminal-guard.test.ts, which passed (15 tests, 0 failures).
• Ran type checking with bun run typecheck, which succeeded.
• Ran the privacy scan with bun run privacy:scan, which succeeded.
• Ran the full test suite bun run test; a pre-existing unrelated assertion in tests/service.test.ts failed, so the full suite was not green, but the terminal-guard focused tests and relevant regressions all passed.

Summary by CodeRabbit

• Bug Fixes

  • Improved handling of oversized streamed responses.
  • Streams involving tool calls or excessive content now continue safely without automatic continuation analysis.
  • Prevented unnecessary buffering once response limits are exceeded.

• Tests

  • Added regression coverage for forwarding oversized streamed answers without triggering continuation.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[github-actions]

✅ Deterministic PR hygiene checks passed.

[coderabbitai]

📝 Walkthrough

Walkthrough

Changes

Terminal guard bounds

Layer / File(s)	Summary
Retention limits and analysis state src/server/responses/terminal-guard.ts	The terminal guard adds maximum retained-event and retained-content limits. It tracks retention counters and runs terminal analysis only while analysis remains enabled.
Bounded retention and regression coverage src/server/responses/terminal-guard.ts, tests/terminal-guard.test.ts	Tool calls and retention limits disable analysis and clear buffered events. The regression test verifies that an oversized stream forwards all 2,003 events without auto-continuation.

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

Suggested reviewers: ingwannu, lidge-jun, wibias

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: bounding buffering in the Anthropic terminal guard.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/fix-unbounded-streaming-in-terminal-guard

---

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]

✅ READY

• this PR is ready for review.

Hygiene

✅ Deterministic PR hygiene checks passed.

[luvs01]

상류 dev 기준으로 필요한 수정만 이식한 PR을 게시했습니다: lidge-jun#4739

기존 상류의 heartbeat/tool-delta 통과 처리, 단일 타임스탬프, OpenAI Chat opt-in 동작을 보존하면서 이벤트 수·콘텐츠 보존량 제한과 44개 회귀 검증 사례를 추가했습니다. 게시 커밋: 469e7ab21689d5cf938ea9cf85ef36f9ff977572.

Node 기반 격리 검증은 63 통과 / 0 실패 / 1개 기존 bridge 통합 검증 제외입니다. Bun 전체 검증·타입 검사·CI 통과를 주장하지 않으며, 상류 PR은 해당 검증과 리뷰가 끝나기 전까지 초안 상태입니다. 이 원본 PR은 닫거나 병합하지 않았습니다.

[luvs01]

Closing this tracking PR. The terminal guard content retention work was reimplemented against current upstream dev and is being carried forward there, so keeping a second copy of the same change open here only splits review and rebase effort. Nothing in this branch is lost: the implementation, its regression coverage and the documentation all moved with it, and the remaining verification is happening on the branch that is actually up for review.