v2.0.0

Migration needed: no runtime impact, version-only correction.

• 1.0.9 changed schema_version 2 -> 3 by removing the usage field from plan/file_change records (moved to a new usage record type). Per this project's own versioning rule (design doc 19.2), a schema_version change that isn't purely additive -- a field rename or removal -- requires a MAJOR bump of plugin.json, not a patch release. This release corrects that: no code changes from 1.0.9, version only

v1.0.9

Migration needed: no, but see the first item -- already-captured raw session data predating this release won't retroactively gain accurate token metrics if it's ever reprocessed.

• Token usage and turn counts were being lost or double-counted: usage was only attached to plan/file_change records, so a text-only assistant turn (most turns) had its usage silently dropped entirely, while a turn with 2+ matching tool_use blocks had its usage counted once per block. Every assistant turn now produces its own usage record, counted exactly once, independent of what tool_use blocks (if any) it also contains. schema_version bumped 2 -> 3 for this record-shape change
• A raw transcript chunk could be corrupted while splitting it into lines: str.splitlines() also breaks on Unicode line separators (U+2028, U+2029, U+0085, \v, \f), which can appear raw inside a JSON string value and are valid JSON. Switched to .split("\n"), the actual JSONL line delimiter
• git log --author matches its value as a regex, not a literal string. An unescaped email (almost always containing ., which matches any character) could over-match and pull another contributor's commits into the current user's worklog. The email is now escaped before being passed to --author
• Closed a race in the summarization lock: two SessionStart hooks firing at nearly the same instant, with no lock file yet, could both believe they'd acquired it and both summarize the same date. Lock creation is now atomic (O_CREAT|O_EXCL)
• notes/{date}.md is now written atomically (temp file + rename), like every other persisted file in this plugin, instead of a plain write that a crash (or the lock race above) could leave truncated or interleaved
• A session that spans midnight could permanently freeze a stale live/abandoned classification for its earlier day: the cache (and the outer "is this date done" check) only looked at that day's own captured shard, not the session's other shards across date folders, so a later-day fragment changing what's actually true for the earlier day never triggered a recompute. Both checks now consider all of a session's shards
• Archive files (archive/{date}.tar.gz) are now written atomically; previously a process killed mid-write left a permanently corrupt, permanently-skipped archive at that path (original data was never at risk either way)
• An invalid /worklog:search query (unbalanced quotes, a leading -, a bare operator) crashed with a raw traceback -- it was being retried 3 times as if it were lock contention, then re-raised uncaught. It now fails fast with a friendly message
• debug_log's rotation could itself raise on a torn multi-byte UTF-8 sequence from concurrent hook writes, defeating its own "logging must never be the reason a hook fails" guarantee. It now decodes tolerantly

v1.0.7

Migration needed: no.

• Added /worklog:list to show which dates have a worklog at a glance (most recent 14 by default, --all for full history), so you don't have to guess a date for /worklog:show or reach for /worklog:search just to see what's there. When more dates exist than are shown, the output points to --all or /worklog:search to find them