mirror of
https://github.com/QwenLM/qwen-code.git
synced 2026-05-20 09:24:03 +00:00
d3cbe7614e
* fix(core): preserve read-before-write state across idle microcompaction (#4239) Idle microcompaction blanked old tool outputs and then wiped the entire file-read tracking cache to keep the read_file "unchanged" fast path from pointing at content no longer in history. That wipe also destroyed the read-before-write record, so after any idle break the model was falsely told a file it had already read "has not been read in this session" and forced into a redundant re-read of an unchanged file before it could edit it. Replace the blanket wipe with per-file targeting: when microcompaction blanks a read/edit/write result, only that file's fast-path eligibility is disarmed; the on-disk fingerprint and "seen this session" markers are preserved, so edits/overwrites of an unchanged file proceed without a re-read while a file that actually changed on disk is still correctly caught. If a blanked result's file path cannot be resolved back to its cache entry, fall back to the old blanket wipe so a stale "unchanged" placeholder can never be served. * refactor(core): harden microcompaction fast-path eviction (PR #4243 review) - buildCallIdToFilePath now accumulates paths per callId (Map<string, string[]>) so a reused/duplicate functionCall.id disarms ALL candidate files instead of silently dropping one — over-disarming costs at most a redundant re-read, whereas keeping the wrong file armed would resurrect the issue #4239 dangling-placeholder hazard. - Add a guard-rail comment tying COMPACTABLE_TOOLS to FILE_PATH_TOOLS. - Document why readResidentInHistory is intentionally not reset on the fingerprint-drift branch. - Add tests: duplicate-id disarm, and a mixed on-disk/ghost evicted batch forcing the safe blanket wipe. |
||
|---|---|---|
| .. | ||
| channels | ||
| cli | ||
| core | ||
| sdk-java | ||
| sdk-python | ||
| sdk-typescript | ||
| vscode-ide-companion | ||
| web-templates | ||
| webui | ||
| zed-extension | ||