Project Time Tracker

Combines five data sources → reconciles → produces HTML dashboard + Markdown report.

Scripts

All scripts live in scripts/ next to this SKILL.md. Run them with python3:

claude_messages.py, codex_messages.py, and cursor_messages.py each output two timestamp arrays: timestamps (user prompts only — for prompt-count metrics) and activity_timestamps (every timestamped event, including tool-call/tool-result cycles and assistant replies — use this one for reconciliation). See "Use activity_timestamps, not timestamps" in references/reconciliation.md for why this distinction matters — using prompt-only timestamps for hour reconciliation silently drops real work time.

Use reconcile.py instead of hand-rolling the merge in Python. It's a deterministic implementation of the algorithm below — run it on the JSON files the other scripts already produced rather than re-deriving the merge/sessionize/intersect logic inline each time. Hand-rolling it live is exactly how a real bug happened once: two different gap-threshold runs (15min vs 30min "engine time") got compared against each other as if they were the same base, producing a nonsensical number. reconcile.py fixes the gap threshold and presence window as explicit flags so every run is reproducible and comparable.

Workflow

1. Determine scope

Ask for or infer:

• Project directory/directories (git repos)
• Date range (first commit → last commit, or user-specified)
• Output location for HTML + markdown files

Auto-discover sub-repos: By default, scan the project directory for .git folders in subdirectories (not just the root). Each parent of a .git directory is a sub-repo to analyze.

# Find all git repos under the project directory
find /path/to/project -name ".git" -type d 2>/dev/null | sort

This produces a list like:

/path/to/project/frontend/.git
/path/to/project/backend/.git
/path/to/project/libs/shared/.git

Each of these (minus the /.git suffix) is a repo to run git_sessions.py, claude_messages.py, and codex_messages.py on. Also run these scripts on the root project directory itself (for Claude/Codex messages sent from the root, which is common when using monorepo-style workflows).

2. Extract data

Run all five scripts on every discovered repo. For git, Claude, Codex, and Cursor, run per sub-repo. For WakaTime, use the multi-project discovery approach described below.

# Git sessions — run per sub-repo
python3 git_sessions.py /path/to/project/frontend --since 2026-01-15 --until 2026-02-02
python3 git_sessions.py /path/to/project/backend --since 2026-01-15 --until 2026-02-02

# Claude Code — run per sub-repo AND the root directory
python3 claude_messages.py --project-path /path/to/project
python3 claude_messages.py --project-path /path/to/project/frontend
python3 claude_messages.py --project-path /path/to/project/backend

# Codex CLI — same as Claude
python3 codex_messages.py --project-path /path/to/project
python3 codex_messages.py --project-path /path/to/project/frontend
python3 codex_messages.py --project-path /path/to/project/backend

# Cursor IDE — same as Claude/Codex
python3 cursor_messages.py --project-path /path/to/project
python3 cursor_messages.py --project-path /path/to/project/frontend
python3 cursor_messages.py --project-path /path/to/project/backend

WakaTime multi-project discovery: WakaTime often tracks sub-directories as separate projects (e.g., a monorepo at my-project/ may have WakaTime projects named my-project, frontend, backend, shared). A single --project query will miss the others.

1. First, run wakatime_fetch.py without --project to get the full project list for the date range:

   python3 wakatime_fetch.py --start 2026-01-15 --end 2026-02-02
   # Returns: { "projects": [{"project": "my-project", "hours": 9.2}, {"project": "frontend", "hours": 5.1}, ...] }
   

   Copy

2. Filter the returned projects list for names matching any of:

   • The root project directory basename (e.g., my-project)
   • Any sub-repo directory basename (e.g., frontend, backend)
   • Any intermediate directory basename that contains a sub-repo (e.g., libs)

3. Fetch intervals for each matching project:

   python3 wakatime_fetch.py --start 2026-01-15 --end 2026-02-02 --project my-project
   python3 wakatime_fetch.py --start 2026-01-15 --end 2026-02-02 --project frontend
   python3 wakatime_fetch.py --start 2026-01-15 --end 2026-02-02 --project backend
   

   Copy

4. Combine all intervals from all matching WakaTime projects into a single list for reconciliation.

Why this matters: In a project with 4 sub-repos, a single --project query captured only 9h of the actual 26.5h of WakaTime data. The other 17.5h was tracked under sub-directory project names.

Folder move detection: If claude_messages.py, codex_messages.py, or cursor_messages.py return 0 results, check the output for alternate_paths. If present, ask the user:

"No Claude/Codex history found at /current/path, but found sessions for project-name at /old/path. Was this project moved? Should I include that history too?"

If confirmed, re-run with --project-path /old/path and merge timestamps from both paths.

See references/folder-move-detection.md for full detection logic and edge cases.

3. Reconcile hours

Merged total = best estimate (git ∪ Claude ∪ Codex ∪ Cursor ∪ WakaTime intervals, no double-counting). Run reconcile.py on the JSON files from step 2 rather than hand-rolling the merge:

python3 reconcile.py \
  --git git_backend.json --git git_frontend.json \
  --claude claude_backend.json --claude claude_frontend.json \
  --codex codex_backend.json --codex codex_frontend.json \
  --cursor cursor_backend.json \
  --waka waka.json \
  --since 2026-06-29 --until 2026-08-19 \
  --tz Europe/Berlin --daily

(--git/--claude/--codex/--cursor/--waka are all repeatable — pass one per sub-repo for multi-repo projects; reconcile.py merges them together.)

This does, deterministically:

1. Take git's raw commit timestamps, and Claude/Codex/Cursor's activity_timestamps (not timestamps — see below) as point events.
2. Sessionize each source independently via the gap threshold (--gap-hours, default 0.5h; use 0.25 for a stricter/conservative estimate).
3. Union WakaTime's ready-made [start, end] intervals in alongside the sessionized point sources.
4. Merge everything that overlaps or sits within the gap threshold into final sessions.
5. Sum raw durations — no +0.5h buffer, no 0.5h floor (legacy behavior; see reconciliation.md if you need it, e.g. for a timestamps-only source with no activity-level granularity).
6. If --presence-window-min > 0 (default 10): also compute a supervised vs. unattended split — how much of the merged engine time falls within that window of an actual user prompt or git commit, vs. how much is the agent running with no human check-in nearby. See "Supervised vs. unattended time" below.
7. If --daily: bucket per day using an 08:00 local-time boundary (--day-boundary-hour, default 8) so an overnight session doesn't get split across two days in the table.

Use activity_timestamps, never timestamps, for engine-time reconciliation (this is what reconcile.py does automatically). timestamps is prompt-only and will make an actively-worked session look like a single 0-duration point whenever the agent kept running after the last recorded prompt. activity_timestamps includes tool-call/tool-result cycles and assistant replies, so it reflects when the agent was actually still working. See references/reconciliation.md for the concrete example (a "0-duration" session that turned out to be ~50 minutes of continuous work) and why the gap threshold dropped from the old 1.5h default to 0.5h.

Don't compare engine_hours across different --gap-hours runs as if they're the same measurement — a 15min-threshold total and a 30min-threshold total are two different quantities. If you need both, run reconcile.py twice and label each result with its threshold.

Supervised vs. unattended time

engine_hours measures when the agent was active — not necessarily when the human was watching. If a task ran autonomously for 20+ minutes with no new prompt, that time still counts as engine time but may not reflect real hands-on-keyboard effort. reconcile.py's supervised split addresses this: it builds a "presence window" of ± --presence-window-min minutes around every real user prompt or git commit, and reports how much of engine_hours falls inside vs. outside that window.

The window size is a real sensitivity knob, not a solved parameter — on one real project, ±5min gave 72% supervised, ±30min gave 99%. ±10min is the recommended default: long enough to cover a normal "read the output, think, reply" pause, tight enough that a genuine away-from-keyboard stretch still shows up as unattended. Always report which window was used alongside the number.

Why merge matters: AI agent prompts (Claude/Codex/Cursor) often appear minutes before/after git commits in the same work session. WakaTime captures IDE keystrokes that may fall between commits. A user might research with Claude, use Cursor's AI, write code (WakaTime), then commit (git) — all one session. Union of all five sources captures the true session boundaries without double-counting.

• The merged total replaces "git-only" as the primary estimate
• WakaTime hours shown for reference (active keystrokes only, always lower)

See references/reconciliation.md for the full underlying algorithm reconcile.py implements, the day-boundary bucketing rationale, and the legacy buffer/floor formula.

4. Generate HTML dashboard

Write a single-file HTML with inline Chart.js (CDN). Dark theme (#0a0a0a bg, #1a1a1a cards).

Required sections:

1. Stat cards — Merged total (git∪claude∪codex∪cursor∪waka), Git estimate, WakaTime, Sessions, Commits, Claude prompts, Codex prompts, Cursor prompts
2. Daily activity chart — Overlapping bars (git + WakaTime + merged) + AI prompts line on secondary axis
3. Gantt timeline — UTC horizontal bars; git, Claude, Codex, and Cursor as separate colored datasets on same chart (separate swimlane rows when they overlap on same day)
4. Data table — Session | Time (UTC) | Active | Est. | WakaTime | Claude | Codex | Cursor | Commits

Chart.js essentials:

Chart.defaults.color = '#888';
Chart.defaults.borderColor = '#2a2a2a';
// Overlapping bars: same barPercentage/categoryPercentage on both datasets, different opacity
// Mixed chart: type:'bar' on container, each dataset has its own type + yAxisID
// Gantt floating bars: data: [[startH, endH]], indexAxis: 'y'
// Claude as line on secondary axis:
{
  type: 'line',
  yAxisID: 'yClaude',
  // right-side axis, max ~100, different color
}
// All charts: responsive: true, maintainAspectRatio: false

Save as <project-dir>/work-hours-analysis.html.

5. Generate Markdown report

# [Project] - Work Hours Analysis
## Summary
**Estimated Total Working Hours: Xh** (based on commit timing analysis)
**Supervised: Xh (Y%) · Unattended: Xh** (±10min presence window — see Methodology)

| Date | Sessions | Time Range | Git Est. | WakaTime | Commits | Project |
...

## Timeline
- Start: [date], End: [date], Duration: N days

## Per-Project Breakdown
[sub-project sections with commit/session counts]

## Charts
### Daily Activity (ASCII)
Jan 15  ██░░░░░░░░░░░░░░░░░░  0.5h
...
### Project Distribution
project-a  ████████████████████  45%  (~18h)
...

## Methodology
- Session Detection: events within 0.5h gap = same session (git commits + Claude/Codex/Cursor activity_timestamps, not just prompts)
- Hour Estimate: Σ(session_duration), no buffer, no floor — activity_timestamps already captures real start/end
- Supervised vs. Unattended: hours within ±10min of a real user prompt/commit count as supervised; the rest is the agent running with no recent human check-in
- Why Git > WakaTime: WakaTime only tracks active IDE typing; git includes thinking/research/AI prompting

ASCII bar: blocks = round(hours / max_hours * 20), █ filled, ░ empty, 20 cols wide.

Save as <project-dir>/total_hours.md.

Validation Checklist

Before running

• WakaTime API key — ~/.wakatime.cfg exists and contains api_key = waka_... under [settings]. Run cat ~/.wakatime.cfg to verify. If missing, the WakaTime script will fail silently or with an auth error.
• Git repo accessible — the project directory is a git repo with commits (git log --oneline -5 /path/to/repo returns results). If not, git_sessions.py will return 0 sessions.
• Claude history exists — ~/.claude/history.jsonl is present and non-empty, OR ~/.claude/projects/ contains session files for the project. If both are missing, Claude hours will be 0.
• Cursor data accessible — the Cursor state database exists at the platform-specific path (macOS: ~/Library/Application Support/Cursor/User/globalStorage/state.vscdb, Windows: %APPDATA%\Cursor\User\globalStorage\state.vscdb, Linux: ~/.config/Cursor/User/globalStorage/state.vscdb). If missing, Cursor hours will be 0. The script auto-detects the platform and reads SQLite databases in read-only mode.
• Date range is valid — --since is before --until; the range covers dates when work actually happened.

After generation

• HTML loads without errors — open work-hours-analysis.html in a browser; all charts render; no JS console errors.
• Total hours match — the "Merged total" stat card in HTML equals the "Estimated Total Working Hours" in total_hours.md (allow ±0.01h for rounding).
• Date range matches input — the first and last dates in the data table and the Timeline section match the requested --since/--until values.
• Session count non-zero — at least one source contributed sessions. If all sources return 0, something is wrong (wrong path, wrong project name, date range outside project history).

Examples

Example 1: Single repo, standard report

Trigger phrase: "How many hours did I spend on the api-server project this month? Generate the full report."

Actions:

1. Determine scope: project at /Users/alice/code/api-server, date range inferred as 2026-02-01 → 2026-02-23 (current month to today).

2. Extract data:

   python3 git_sessions.py /Users/alice/code/api-server --since 2026-02-01 --until 2026-02-23
   python3 wakatime_fetch.py --start 2026-02-01 --end 2026-02-23 --project api-server
   python3 claude_messages.py --project-path /Users/alice/code/api-server
   python3 codex_messages.py --project-path /Users/alice/code/api-server
   python3 cursor_messages.py --project-path /Users/alice/code/api-server
   

   Copy

3. Reconcile: git_sessions.py returns 14 sessions (22.5h), WakaTime returns 11.2h, Claude returns 47 prompts across 9 days, Codex returns 0, Cursor returns 12 prompts across 3 days. Merged total after union + gap merging: 27.1h.

4. Generate work-hours-analysis.html and total_hours.md in /Users/alice/code/api-server/.

Result: "You spent approximately 27.1 hours on api-server in February 2026 (14 git sessions, 47 Claude prompts, 12 Cursor prompts, WakaTime reference: 11.2h active typing). Report saved to /Users/alice/code/api-server/work-hours-analysis.html."

Example 2: Multi-repo project with folder move

Trigger phrase: "Calculate the total dev time for the marketplace project — it has a frontend and backend repo. Also I think I renamed the folder at some point."

Actions:

1. Determine scope: two repos at /Users/bob/marketplace-backend and /Users/bob/marketplace-frontend, user specifies date range 2025-11-01 → 2026-01-31.

2. Extract data:

   python3 git_sessions.py /Users/bob/marketplace-backend --since 2025-11-01 --until 2026-01-31
   python3 git_sessions.py /Users/bob/marketplace-frontend --since 2025-11-01 --until 2026-01-31
   python3 wakatime_fetch.py --start 2025-11-01 --end 2026-01-31 --project marketplace
   python3 claude_messages.py --project-path /Users/bob/marketplace-backend
   python3 codex_messages.py --project-path /Users/bob/marketplace-backend
   python3 cursor_messages.py --project-path /Users/bob/marketplace-backend
   

   Copy

3. claude_messages.py returns 0 results with alternate_paths: ["/Users/bob/old-market/backend"].

4. Ask user: "No Claude history found at /Users/bob/marketplace-backend, but found sessions for backend at /Users/bob/old-market/backend. Was this the previous location? Should I include that history?"

5. User confirms. Re-run: python3 claude_messages.py --project-path /Users/bob/old-market/backend. Merge timestamps from both runs.

6. Merge git commit timestamps from both repos (backend + frontend) with Claude's activity_timestamps, sort, re-merge within the 0.5h gap threshold. Reconcile with all sources.

Result: "Total estimated time: 84.7h across 3 months (backend + frontend combined, including Claude history from the old path /Users/bob/old-market/backend)."

Troubleshooting

WakaTime API key missing or invalid

Symptom: wakatime_fetch.py exits with an auth error, HTTP 401, or KeyError: 'api_key'.

Fix:

1. Check if the config exists: cat ~/.wakatime.cfg
2. If missing, create it:

   [settings]
   api_key = waka_xxxx...
   

   Copy
3. Get your key from wakatime.com/settings/api-key.
4. If the key exists but returns 401, it may be expired or revoked — generate a new one.

Also check: The --project flag matches the project name exactly as WakaTime recorded it (case-sensitive). You can verify project names in the WakaTime dashboard under Projects.

Git returns 0 sessions

Symptom: git_sessions.py returns "sessions": [] or "total_hours": 0.

Possible causes and fixes:

Claude, Codex, or Cursor returns 0 sessions (no alternate_paths)

Symptom: Both timestamps: [] and alternate_paths: [].

Possible causes and fixes:

alternate_paths found but user says the path is wrong

Symptom: The script reports alternate paths but the user says none of them are the old project location.

Fix: Fall back to --filter <project-name> which does a substring match on directory names rather than an exact path match. Be aware this may pick up unrelated projects with similar names — review the session list with the user before merging.

python3 claude_messages.py --filter marketplace

HTML chart renders blank or shows NaN

Symptom: The HTML file opens but charts are empty or show "NaN" values.

Possible causes:

• Reconciliation produced null or None values that were serialized into the JS data arrays.
• Timestamps were not converted to UTC before writing to HTML (local epoch vs UTC epoch mismatch).
• Chart.js CDN failed to load (offline environment).

Fix:

1. Open browser DevTools console — the specific JS error pinpoints the issue.
2. Verify all epoch timestamps are UTC floats, not strings.
3. For offline environments, download Chart.js and embed it inline: <script>/* chart.js source */</script>.

Total hours mismatch between HTML and Markdown

Symptom: The HTML stat card shows a different merged total than total_hours.md.

Cause: The reconciliation was run twice independently and produced slightly different results (e.g., due to floating-point rounding, or one file used stale data).

Fix: Run reconciliation once, store the result in a variable, and write the same computed value to both output files. Do not recompute independently for each output.

Notes

WakaTime auth and config: The script reads ~/.wakatime.cfg automatically. Always pass --project for per-project data. WakaTime tracks active keystrokes only — always lower than git estimate. Heavy Claude Code usage creates a large gap between WakaTime and actual effort. See references/data-sources.md for full auth setup, config format, and limitations.

Codex CLI data source: codex_messages.py reads ~/.codex/sessions/YYYY/MM/DD/rollup-*.jsonl. Each file has session_meta (with payload.cwd + payload.id — can recur mid-file for long-lived/resumed sessions, so cwd is tracked continuously rather than assumed fixed), event_msg entries with payload.type == "user_message" for actual prompts, and many other timestamped response_item/event_msg entries (tool calls, patches, reasoning) marking agent execution between prompts. Output includes timestamps (prompts only), activity_timestamps (every timestamped event — use this for reconciliation), and alternate_paths for folder-move detection. See references/data-sources.md for full file structure.

Claude Code data sources: claude_messages.py uses two sources: ~/.claude/history.jsonl (primary; project field = abs path, timestamp in ms, prompts only) and ~/.claude/projects/<encoded>/*.jsonl (session files; cwd field, type=="user" and type=="assistant" entries, ISO timestamps). Encoded dir name format: /Users/foo/bar → -Users-foo-bar. Always prefer --project-path over --filter. Output includes timestamps (real user prompts only) and activity_timestamps (user prompts + tool_result entries + assistant entries — use this for reconciliation). See references/data-sources.md for full field reference.

Cursor IDE data sources: cursor_messages.py reads from Cursor's SQLite databases (.vscdb files). Primary source is the platform-specific state.vscdb (macOS: ~/Library/Application Support/Cursor/User/globalStorage/, Windows: %APPDATA%\Cursor\User\globalStorage\, Linux: ~/.config/Cursor/User/globalStorage/) — the cursorDiskKV table contains composerData:{sessionId} entries (with workspaceUri for project matching) and bubbleId:{sessionId}:{messageId} entries (with per-message timestamps, type=1 = user). Fallback source is workspace-level state.vscdb files under workspaceStorage/*/, with workspace.json mapping each workspace to its project folder. Databases are opened read-only. The script auto-detects the platform. Output format matches Claude/Codex scripts: timestamps (user bubbles only) + activity_timestamps (every bubble/message, any role — use this for reconciliation) + alternate_paths for folder-move detection. See references/data-sources.md for full field reference.

Reconciliation algorithm (updated 2026-08-19): Run via reconcile.py, not hand-rolled. Gap threshold is 0.5h by default (0.25h for a stricter estimate), applied to activity_timestamps — not timestamps — for Claude/Codex/Cursor, and to git's raw commit timestamps. All sources converted to UTC epoch float intervals. Intervals merged if gap ≤ threshold. Per-interval estimate is the raw duration (end - start), no buffer, no floor. The old 1.5h threshold + max(duration + 0.5h, 0.5h) formula is legacy — it existed to compensate for reconciling against prompt-only timestamps, which undercounts real work whenever the agent kept executing after the last recorded prompt (confirmed on a real project: an apparently-isolated prompt had 1,343 agent-execution events in the following 50 minutes, invisible to timestamps alone). Still available via git_sessions.py's default (non---raw) mode for sources without activity-level granularity. reconcile.py also computes a supervised-vs-unattended split by default (--presence-window-min, default 10) — see "Supervised vs. unattended time" in the workflow section above. See references/reconciliation.md for the full underlying pseudocode, the session detection helper function, and the day-boundary bucketing helper for daily breakdowns.

Folder move detection: claude_messages.py, codex_messages.py, and cursor_messages.py all scan known history for matching project names when 0 results are found at the provided path. Returns alternate_paths list. If non-empty, ask user to confirm, re-run with old path, merge timestamps. See references/folder-move-detection.md for detection logic and edge cases.

Multi-repo projects: By default, scan for .git subdirectories to auto-discover all sub-repos. Run git_sessions.py, claude_messages.py, codex_messages.py, and cursor_messages.py on each sub-repo plus the root directory. Use WakaTime multi-project discovery to find all matching WakaTime project names. Merge all session arrays, re-sort by date, recompute daily totals and grand total.