fix(sessions): pick the latest custom-title from anywhere in Claude session files - #6809
fix(sessions): pick the latest custom-title from anywhere in Claude session files#6809SailingLoong wants to merge 2 commits into
Conversation
…ession files Claude Code appends custom-title entries inline at rename time, so the entry can sit far above the 30-line tail window that parse_session samples. After a rename followed by more conversation, the Session Manager silently fell back to the first user message and the new name never showed up. Sample head/tail lines are kept for positionally stable metadata (session id, cwd, first message, last activity, summary). Custom titles now come from a dedicated full-file scan that keeps the last non-empty entry and only JSON-parses lines passing a cheap substring prefilter. Fixes farion1231#6787
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6d51aa51ba
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".
|
|
||
| // Claude Code appends custom-title entries inline at rename time, so they | ||
| // can sit far above the tail window sampled above; scan the whole file. | ||
| let custom_title = scan_last_custom_title(path); |
There was a problem hiding this comment.
For users with many or large Claude histories, opening or refreshing the session manager now reads the total contents of every transcript because claude::scan_sessions calls this once per discovered file, and list_sessions waits for that worker before returning any sessions. This replaces the previous bounded head/tail I/O with work proportional to the entire retained history—including potentially huge tool-result lines—even for files that have never been renamed, so the page can become extremely slow and I/O-heavy. Please cache indexed titles using file metadata or otherwise avoid an unconditional full-file pass during every listing.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Fair concern — the full scan is the price of correctness here (title events have no guaranteed position, which is the bug this PR fixes). Addressed the cost side in 2e26836: titles are now cached per file keyed on mtime + size, and since Claude Code only ever appends to these transcripts, unchanged metadata means the cached title is still current and the full read is skipped on every refresh. First listing still pays the full scan for files that carry titles; the substring prefilter keeps that pass cheap for the (majority of) transcripts that never had one.
Tests: a chmod-based test proves the read is skipped on a hit (stat stays readable, the full read fails, the title still comes back), and a growth test proves the entry is invalidated on append. Full suite: 2683 passed.
Listing sessions now rescans every transcript in full because the latest custom-title can sit anywhere in the file. Claude Code only appends to these transcripts, so unchanged mtime + size means the last scanned title is still current and the full read can be skipped. - stat failures (missing/unreadable metadata) fall back to an uncached full read, so behavior outside the listing path is unchanged - cache-hit test proves the read is skipped (chmod 000 keeps stat working but fails the full read); growth test proves invalidation
2 participants
Hi 👋 Small Session Manager fix from the open-issue pile.
Fixes #6787
Problem
Claude Code appends
custom-titleentries to the session JSONL inline at rename time, so where the entry lands in the file depends on when the rename happened.parse_sessiononly samples the first 10 and last 30 lines (read_head_tail_lines(path, 10, 30)), so once ~30 more lines are appended after a rename — resuming the session, a long follow-up exchange — the entry falls out of the tail window and the title silently falls back to the first user message. Renamed sessions stop showing their name, as reported in #6787.Title events genuinely have no positional home: on a real session file here, title entries are spread from line 7 to line 2131 out of 2132 total. Any fixed sampling window is the wrong tool for them.
Fix
Tests
parse_session_prefers_the_latest_custom_title_above_the_tail_window: two renames early in the file + 40 trailing lines → the latest name wins. Revert-verified: it fails onmainwithfirst messageand passes with the fix.cargo test(14 test binaries, 0 failed),cargo fmt --checkandcargo clippy --all-targets -- -D warningsclean. No frontend files touched;pnpm typecheck/pnpm format:checkpass on the identical frontend tree.Checklist / 检查清单
pnpm typecheckpasses / 通过 TypeScript 类型检查pnpm format:checkpasses / 通过代码格式检查cargo clippypasses (if Rust code changed) / 通过 Clippy 检查（如修改了 Rust 代码）Happy to adjust.