Alibaba's open-code-review project shipped version 1.13.13 on 19 November 2026 with prebuilt binaries for Linux, macOS and Windows, according to the GitHub release notes. The release addresses two reliability bugs in the shared LLM loop and ChatGPT OAuth login flow alongside documentation and project-page updates.1
How the context budget fix works
Previously, three-zone compression could keep recent conversation rounds against nearly the entire warning threshold, then add the fixed prompt and a new summary on top, pull request #1993 explains. That accounting error could trigger StopCompression even when compressing more history would have let the session continue. The fix charges the fixed system and user prefix, including earlier summaries, when selecting active rounds. It also reserves space for the new summary and its wrapper before retaining complete recent rounds.3
In a local recorded-history replay with a stubbed summary response, the pull request reports that the unchanged code produced a 160,173-token rebuilt context against a 160,000-token threshold. Under the same replay conditions, the fixed reconstruction used 131,291 tokens, with 143,897 tokens reserved when allowing the full summary cap. The pull request notes that this replay checks budgeting and history retention rather than summary quality.3
What changed in ChatGPT OAuth
In the ChatGPT authentication package, a login attempt could finish while the local loopback server was still writing the browser callback response, according to pull request #1996. Closing the server immediately left the CLI signed in while the browser reported a connection error. The update uses a shutdown context independent of the login context to drain active responses and force-closes remaining connections if graceful shutdown fails. Cleanup does not change the login result or wait for the browser opener to exit, the pull request states.2
Limits and unchanged behavior
The 60% and 80% compression thresholds and aggregate --max-tokens-budget semantics remain unchanged in v1.13.13, pull request #1993 states. Compression requests may now use a smaller output cap, particularly with small configured context limits. Unrecoverable context overflow still stops explicitly rather than silently dropping uncompressed history. Compression-request input bounding and suppression of repeated rejected requests remain separate work, according to the pull request.3