git ai usage#1434
Open
clarete wants to merge 91 commits into
Open
Conversation
|
|
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
clarete
force-pushed
the
lincoln/git-ai-usage
branch
2 times, most recently
from
84412f5 to
88ce909
Compare
Adds a new `git-ai activity` CLI command that shows aggregated stats from locally persisted metric events (Committed, Checkpoint, SessionEvent). - Schema migration 2→3: adds local_events table (never cleared, unlike the upload queue) with indexes on event_id and ts - store_local_events() in telemetry_worker fires before upload attempts, persisting interesting events regardless of network status - local_stats module aggregates events in memory: AI/human line splits, per-tool breakdowns, unique session counts, files touched - git-ai activity [--period 7d|30d|all] [--json] (default: 30d) - Excluded from daemon-required commands; reads local DB standalone Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…tted Commits with no AI lines were dropping their human_additions because the early-return guard fired before the accumulation. Moved the accumulation above the guard. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Displays AI acceptance rate (committed AI lines / checkpoint AI lines) below the per-tool breakdown. Only shown when checkpoint history is sufficient to produce a meaningful ratio (<= 100%). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…t-ai activity - Activity over time: bar chart bucketed by day (<=7d), week (30d), or month (all) with empty buckets filled in so the chart is always contiguous - AI coding velocity: each bucket shows AI lines + commit count so output doubles as a velocity trend across the period - Time of day: 24-slot sparkline (local time) showing when AI-assisted commits land, using Unicode block chars for density Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Weekly buckets now display "May 18 – May 24" instead of just "May 18" so users can see their work falls within the bucket rather than thinking it refers to a specific day. Activity over time section moved below Sessions and above Time of day. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace Commits/Checkpoints/Sessions headers with a top-level AI vs Human percentage bar followed by stacked AI and Human sections. Surfaces the same data with clearer framing — no internal "checkpoint" terminology. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Preserve the model from tool_model_pairs instead of stripping it, so the breakdown distinguishes sonnet/opus/etc. Trims the redundant tool prefix from the model name (claude::claude-sonnet-4-6 -> "claude · sonnet-4-6"). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Surface how much committed code was confidently attributed to AI or known-human vs left "untracked". Computed as (ai + human) / total git diff additions. Answers the "can I trust these numbers" question by exposing the size of the unattributed holes in the data. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace the single summarized "Attributed" bar with per-bucket coverage in the Activity over time chart, so the trend (improving/degrading) is visible rather than one aggregate number. Buckets now accumulate diff additions and attributed lines across all commits; AI-lines bar and commit count remain AI-only for consistency. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Extract per-message token usage (input/output/cache read/cache write) from SessionEvent transcript JSON, grouped by model, with an estimated USD cost from a built-in Anthropic pricing table (labeled estimate). Dedups by assistant message id keeping field-wise max, since the incremental transcript reader re-emits the same message and streaming partials report lower token counts than the final message. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Codex transcripts use a different schema than Claude: cumulative per-session totals on token_count events (payload.info.total_token_usage) rather than per-message usage. Parse them via a session-keyed accumulator that keeps the running max, mapping codex field semantics onto ours (input_tokens includes cached, so non-cached input is the difference; cached -> cache_read; no cache-creation concept). Model name is captured from the separate payload.model event. Adds GPT-5 family pricing estimate. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
For each session, check whether a commit landed within 4 hours of the session's last observed event. Surfaces shipped / abandoned / yield% inline with the Sessions count in the AI section. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Accumulates AI lines committed per weekday (Mon-Sun) from committed event timestamps and renders a sparkline heatmap below the hourly time-of-day chart. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Track the timestamp of each token-usage event so the cost can be split into this-week vs last-week halves. Renders as a gray annotation under Est. cost: 'This week ~$X.XX · Last week ~$Y.YY ↑/↓ N% vs last week'. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Compute cache_read / (cache_read + cache_creation) per model and display it inline with the per-model token line, e.g. 'cache 97% hit'. A low ratio means context is being discarded and re-warmed frequently. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Track checkpoint AI lines per tool alongside committed AI lines, then compute acceptance rate (committed / checkpoint) per tool. Displayed inline with the per-tool commit breakdown, e.g. 'claude · sonnet-4-6: 912 (claude 72% accept)'. The global acceptance rate is kept. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
A user with checkpoint events but no commits or sessions in the window was falsely getting the no-data error even though checkpoint stats would render in the output. Now checks checkpoints.ai_lines_added and checkpoints.human_lines_added before declaring the result empty. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…pers The DB lock-acquire + get pattern was duplicated verbatim in compute_activity, compute_session_list, and compute_repo_summaries. Extracted into two free functions so the lock-poison error string and the acquire pattern live in one place. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Was defined identically inside both compute_activity and compute_session_list. Now a single module-level const with a doc comment explaining its purpose. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The two-call trim_start_matches chain appeared at four separate sites. Now a single strip_protocol(&str) -> &str function handles it. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The dashboard CTA line was written in both branches of the repo_filter if/else. Move it out so it prints unconditionally, with only the --repo tip gated on the filter being absent. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The LocalEventRecord construction closure was copy-pasted across all three query branches (NULL filter, LIKE filter, no filter). Now defined once as map_row and shared by all three. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
model_tokens is now keyed by shorten_model() at insertion, so the model string in the iteration variable is already shortened. The shorten_model call at the by_model push site was a no-op. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The codex→TokenAccum field mapping (subtract cached from input, zero cache_creation) was inlined in both build_token_summary and compute_session_list. Now lives once as an impl method with a doc comment explaining the field semantics. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Three sites in print_terminal independently computed filled/empty block counts and formatted them. Now handled by ratio_bar(value, max, width); bar(pct, width) becomes a thin wrapper over it. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…rings bucket_key and fill_buckets both independently produced daily/weekly/ monthly label strings. A change to any format required two edits. Now handled by bucket_label(date, granularity) called from both sites. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Repositories block: replace block-bar + dot-separated stats with aligned tabular columns; singular/plural for commit/session labels - Model breakdown lines now include "lines" unit after the count - Acceptance rate shows a range (e.g. 56–81%) when multiple tools have valid data, falling back to the overall rate for single-tool - WoW block: drop "↑ new this week" label when last week had no spend (redundant); fix "$-0.00" display by formatting near-zero as "$0" - Add format_cost() helper: rounds to whole dollars for amounts ≥ $10, keeps cents below that threshold; applied to all cost display sites Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Pre-compute max widths for tool names, line counts, model names, token counts, and costs so each column lines up across rows. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously each repo triggered a separate SQL scan via compute_activity, giving O(n × repos) fetches. Now one fetch covers all repos and grouping is done in memory via compute_activity_from_records. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Remove get_distinct_repo_urls, is_backfilled, mark_backfilled, and get_existing_commit_shas -- all had no callers - Fix days_ago() u32 truncation: clamp before casting so values above u32::MAX do not silently wrap - Inline store_local_events into flush_metrics chunk loop so events are walked once; interesting events are serialized once per flush instead of in a separate pre-pass Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds opportunistic pruning to insert_local_events: at most once per day a DELETE removes rows with ts older than 30 days, keeping the table bounded to roughly the window used by git-ai usage (max 60d period). Last-prune timestamp is persisted in schema_metadata. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Events predating the repo_url column (migration 2->3) have no repo identity and were mapping to repo_url="", inflating repos.len() and incorrectly triggering the multi-repo display path for single-repo users. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…1d/3d) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously a failed DELETE left the timestamp unwritten, causing prune_local_events_if_due to retry on every subsequent flush. Wrapping both writes in a transaction makes them atomic: either both succeed or neither does, so a failure advances the timestamp and avoids a retry loop. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
committed * 100 could overflow u32 (wraps in release, panics in debug) for tools with >42.9M committed AI lines. Widening to u64 before multiplication eliminates the risk. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
SessionRecord, compute_session_list, extract_claude_user_text, extract_codex_user_text, and normalize_title had no call sites. Removing them eliminates ~300 lines and a diverging copy of the token-accumulation and yield-classification logic. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…usage Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
clarete
force-pushed
the
lincoln/git-ai-usage
branch
from
88ce909 to
5870c0d
Compare
- Use u64 arithmetic for all percentage multiplications in usage.rs to prevent u32 overflow on large line counts (same fix pattern as the prior acceptance-rate fix) - Fix spark_char overflow: value*8 now computed in u64 before casting - Emit u32::MAX instead of silently dropping tools with no checkpoint events in the query window, so the ">100% (incomplete data)" signal is surfaced as intended - Consolidate the two separate DB fetches in handle_usage into a single compute_all() call so overall stats and per-repo breakdown always reflect the same event snapshot Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
svarlamov
force-pushed
the
lincoln/git-ai-usage
branch
from
5870c0d to
3a78249
Compare
If two processes race on the schema migration, the second ALTER TABLE attempt would fail with "duplicate column name". Catch that specific error so both processes proceed normally. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
git-ai usage — local AI usage statistics
This PR adds a new git-ai usage subcommand that aggregates locally recorded metric events into a human-readable (or JSON) summary of AI coding activity, without requiring a network connection or account.
usage
How it works
The telemetry worker already flushes metric events to a local SQLite DB (
metric_events). This PR adds a second table,local_events, that stores a filtered copy of the three event types relevant to usage stats (Committed,Checkpoint,SessionEvent) as they are flushed. A 30-day retention window is enforced with a once-per-day prune pass.git-ai usagereads directly fromlocal_events— no DB writes, no network calls.Global View
Repository View