Read Cursor tokens and line changes from modern state schema

[1chooo]

Summary

• read the cumulative context window usage (contextWindowStatusAtCreation.tokensUsed) on Cursor user bubbles as a cumulative input token counter, reusing the existing monotonic delta logic
• compute AI line changes by diffing the before/after file snapshots referenced by beforeContentId/afterContentId in edit_file_v2/edit_file tool results (stored under composer.content.<hash> keys), used when streamingContent/code_edit is absent
• include $.contextWindowStatusAtCreation and $.modelInfo rows in the SQLite query filter
• add regression coverage modeled on the current Cursor schema (zero tokenCount placeholders, noCodeblock edit params, content snapshot blobs)

Root Cause

Recent Cursor versions (1.7.x) stopped persisting per-bubble token data entirely: every bubbleId: row now carries a {"inputTokens":0,"outputTokens":0} placeholder and no usage objects exist anywhere in state.vscdb, so #1496/#1499 could no longer recover tokens — the data source moved. Likewise edit_file_v2 params no longer include streamingContent ("noCodeblock":true), so AI line changes were always zero. The remaining local signals are the cumulative context window token usage on user bubbles and the before/after file content snapshots referenced by edit tool results.

Note: Cursor no longer persists output token counts locally, so only input/context tokens can be reported for now.

Validation

• go test ./pkg/ai -count=1
• ran the parser against a live Cursor 1.7.x state.vscdb: previously inputTokens=0 lineChanges=0 across 254 heartbeats in 24h; with this change inputTokens=462037 lineChanges=800

Fixes wakatime/vscode-wakatime#489.

🤖 Generated with Claude Code

[alanhamlett]

Can't merge, this breaks several things:

• pkg/ai/cursor.go:198 contextWindowStatusAtCreation.tokensUsed we need real Cursor test data to confirm this is cumulative state and doesn't reset with compaction, throwing off the token counts.
• pkg/ai/cursor.go:1107 reports snapshot-backed line changes as len(after) - LCS, which can only count added/modified lines. Line change counters should be diffs too, including deletions. Deletion-only AI edits become 0 here instead of a negative line-change value, unlike the existing unified-diff path.
• pkg/ai/cursor.go:1118 adds an O(beforeLines * afterLines) LCS pass for every snapshot edit, up to 3,000 x 3,000 comparisons per edit across the 5,000-row Cursor scan window. That is too expensive for sync-time parsing, and the large-file fallback is also inaccurate because it only reports positive net line-count deltas, so large same-size rewrites or deletion-heavy edits are reported as 0.

[1chooo]

Hi @alanhamlett can you please approve the ci/cd?

[codecov]

Codecov Report

❌ Patch coverage is 86.75676% with 49 lines in your changes missing coverage. Please review.
✅ Project coverage is 89.64%. Comparing base (14891b1) to head (8cd96b2).

Files with missing lines	Patch %	Lines
pkg/ai/cursor.go	88.61%	25 Missing and 12 partials ⚠️
pkg/ai/ai.go	73.33%	7 Missing and 5 partials ⚠️

@@             Coverage Diff             @@
##           develop    #1503      +/-   ##
===========================================
- Coverage    89.69%   89.64%   -0.06%     
===========================================
  Files          417      417              
  Lines        20809    21079     +270     
===========================================
+ Hits         18665    18896     +231     
- Misses        1301     1324      +23     
- Partials       843      859      +16     

Flag	Coverage Δ
unittests	89.64% <86.75%> (-0.06%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
pkg/ai/ai.go	91.73% <73.33%> (-0.97%)	⬇️
pkg/ai/cursor.go	90.52% <88.61%> (-1.48%)	⬇️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[1chooo]

@alanhamlett just make a new push for fixing codecov

[1chooo]

@alanhamlett can you check with this fix because Cursor right now can not count AI usage anymore...

[alanhamlett]

Agent specific code doesn't belong in the ai.go file.