fix: deduplicate message hashes across files, not just within each file#647
fix: deduplicate message hashes across files, not just within each file#647enzonaute wants to merge 1 commit intosteipete:mainfrom
Conversation
parseClaudeFile() resets its seenKeys set per file. Claude Code logs subagent messages to both the parent session JSONL and the subagent's own JSONL under <session>/subagents/. Since both files share the same messageId:requestId pairs, per-file dedup misses the cross-file duplicates and inflates the totals. On a real dataset (30 days, 538 files): - per-file dedup: 7.98B tokens (current behavior) - global dedup: 2.59B tokens (correct, matches ccusage) - ratio: 3.08x overcounting Fix: pass the seenKeys set through ClaudeScanState so it accumulates across all files in a single scan pass. The new existingSeenKeys parameter defaults to [] so existing call sites and tests are unaffected.
![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: 07c7558007
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".
| existingSeenKeys: state.globalSeenKeys) | ||
| state.globalSeenKeys = parsed.seenKeys |
There was a problem hiding this comment.
Passing globalSeenKeys into each parse makes later duplicate files persist with empty days, but that per-file attribution is then cached. If the file that originally “won” a duplicate key is deleted in a later refresh, stale cleanup subtracts its usage while the surviving duplicate file is skipped as unchanged, so those tokens vanish from totals until a forced rescan. This creates incorrect undercounting in normal log-rotation/deletion flows for Claude sessions with parent/subagent duplicates.
Useful? React with 👍 / 👎.
1 participant
parseClaudeFile() resets its seenKeys set per file. Claude Code logs subagent messages to both the parent session JSONL and the subagent's own JSONL under /subagents/. Since both files share the same messageId:requestId pairs, per-file dedup misses the cross-file duplicates and inflates the totals.
On a real dataset (30 days, 538 files):
Fix: pass the seenKeys set through ClaudeScanState so it accumulates across all files in a single scan pass. The new existingSeenKeys parameter defaults to [] so existing call sites and tests are unaffected.