feat: subagent attribution + optional ccusage integration (5h windows, burn rate, multi-agent)#140
Closed
john988 wants to merge 8 commits into
Closed
feat: subagent attribution + optional ccusage integration (5h windows, burn rate, multi-agent)#140john988 wants to merge 8 commits into
john988 wants to merge 8 commits into
Conversation
…s table) Re-port of the subagent-attribution feature onto v1.4.0. Adds per-turn is_subagent / agent_id columns and an agents dispatch table (agent_id -> agent_type, captured from the parent's toolUseResult), populated in both the full and incremental scan paths. Subagents are detected via isSidechain / agentId / a transcript path under a subagents/ directory. Schema changes are additive (ALTER columns + new table), so existing DBs migrate in place. Tests: 124 pass (8 new in tests/test_subagent.py). Verified on real transcripts: 11,428 subagent turns and 438 agent dispatches detected.
Dashboard: get_dashboard_data now returns subagent_by_type and top_dispatches (JOIN turns->agents, acompact-* -> auto-compact, else unknown). Adds a "Subagent Tokens by Type" stacked bar chart, a "Top Subagent Dispatches" table, a Subagent Tokens stat card, and an agent-type colour palette — all filtered by the existing model + range controls. Dynamic values are escaped via esc() (no innerHTML injection). CLI: `today` and `stats` print subagent token/turn summaries (included in totals). Tests: 130 pass (+6 for dashboard/CLI subagent data). Verified on real transcripts (75 type-rows, 50 dispatches) and inline JS passes `node --check`.
…daily totals Adds ccusage_bridge.py wrapping the optional `ccusage` CLI as a data source: detect_runtime() (probes npx.cmd first so Windows shell=False works), a UTF-8 / LOG_LEVEL=0 subprocess runner that never raises, pure transforms (blocks_to_rows / daily_to_rows mapping ccusage's camelCase + nested burnRate/projection), and idempotent upserts. New SQLite tables billing_windows and ccusage_daily_cache (additive; kept separate from the native turns/sessions tables so the two sources never conflict). cmd_scan now enriches via ccusage when available and degrades gracefully when not. Tests: 140 pass (+10 hermetic bridge tests, no Node required). Verified end-to-end against real ccusage 20.0.9: 129 billing windows (1 active, 762k tok/min, 140 min remaining) + 34 daily rows.
…data
ccusage_bridge gains compute_p90_limit() (90th-percentile of your completed 5h
window totals — a personal "typical heavy window" baseline, since Anthropic's
hard token limit isn't exposed; same approach as Claude-Code-Usage-Monitor) and
summarize_billing() (active window + window count + P90 + recent history, or
{available: False} so the UI can show an install prompt). get_dashboard_data
returns a "billing" key, guarded so a bridge issue can never break the page.
Tests: 145 pass (+5). Verified on real data: 129 windows, P90 99M, active
window 131% of baseline at 798k tok/min.
Adds a "Current 5h Billing Window" card above the charts: a progress bar of the active window's tokens vs the P90 baseline (green/amber/red at 80%/100%), plus burn rate, time remaining, projected end-of-window tokens/cost, and cost so far. Falls back to a slim ccusage/Node install prompt when billing data is absent — the rest of the dashboard is unaffected. Verified: inline JS passes node --check; 145 tests pass.
Bridge ingests per-source daily totals for non-Claude agent CLIs (CCUSAGE_EXTRA_SOURCES) into ccusage_daily_cache; Claude Code is excluded here (covered natively) so it's never double-counted. get_dashboard_data exposes ccusage_daily (per-source, sans the unified 'ccusage-all'), and a new "Other Agent CLIs (via ccusage)" chart shows non-Claude token usage by source, range-filtered, hidden when there's nothing to show. Tests: 145 pass. Verified end-to-end: real ccusage run surfaced codex usage (665.8M tokens, ~$694) alongside the native Claude Code views.
…release v1.5.0 P6 — single source of pricing: - Move PRICING + get_pricing + calc_cost into pricing.py, imported by cli.py and served via /api/data. The dashboard JS reads rawData.pricing at runtime (its embedded table is now only a cold-start fallback), so the Python and JS tables can no longer drift. P7 — limit events + transparency: - Detect "Claude AI usage limit reached" events into a new limit_events table, gated on isApiErrorMessage so ordinary text mentioning a limit isn't misdetected (parse_jsonl_file now returns limit events; both scan paths upsert). - Footer notes that figures are transcript-derived estimates that may not match Anthropic billing, and that native vs ccusage numbers are shown separately. Release: bump to v1.5.0 across scanner.VERSION, CHANGELOG, and the extension package.json (parity test enforces all three). Tests: 150 pass (+5 limit-event tests); cli.calc_cost is pricing.calc_cost; inline JS passes node --check.
…hboard cache) - Scanner: the no-growth branch now also updates processed_files.lines, so a transcript rewritten shorter (compaction) isn't skipped forever and later appends are still ingested. - CLI: `today` / `week` compute their window in UTC to match UTC transcript timestamps (drop now-unused `date` import). - Dashboard: cache the /api/data payload keyed on DB path + mtime so the 30s poll doesn't re-run every query when nothing changed; getRangeBounds now uses UTC date math (week/month/relative) for consistency with the data. Tests: 154 pass (+4). Inline JS passes node --check.
john988
added a commit
to john988/claude-usage
that referenced
this pull request
Jun 17, 2026
Scanner shrink-skip, today/week UTC dates, and dashboard /api/data caching are resolved; note the residual stale-turn limitation on compaction.
phuryn
added a commit
that referenced
this pull request
Jun 21, 2026
…/data The subagent attribution split adds an `agents`-table query to get_dashboard_data. The /api/data handler called it with no argument, so it used the default db_path frozen to the module global at def time — ignoring a monkey-patched dashboard.DB_PATH and reading the real ~/.claude/usage.db. On a pre-migration DB (no `agents` table yet) that raised "no such table: agents", failing TestDashboardHTTP locally (CI passed only because no real DB exists there). Pass DB_PATH explicitly so the handler resolves the module global at call time — same contract already documented for /api/rescan. Behaviour is unchanged for real users; tests that patch dashboard.DB_PATH now work. Follow-up to the subagent-attribution split (PR #140, thanks @john988). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
phuryn
added a commit
that referenced
this pull request
Jun 21, 2026
Bump scanner.VERSION + vscode-extension/package.json to 1.5.0 and add the v1.5.0 — TBD CHANGELOG heading covering the subagent-attribution split (#140, @john988) and Docker support (#143, @RafikFarhad). Minor bump: both are non-breaking user-visible features. Date stays TBD and the DEV->main release merge is left to the maintainer per the documented release flow. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Owner
|
Thanks for this, @john988 — and for re-porting the subagent work onto current The subagent attribution half shipped in v1.5.0: The optional ccusage integration we're passing on for now — it adds an external Node/ Closing since the attribution piece has landed and the ccusage piece is out of scope for now — not a quality judgment at all, the work was excellent. Thank you! 🙏 |
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
2 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.
Summary
Two things, both built on top of
v1.4.0:ccusageintegration (WRAP mode) — when Node/npx+ccusageare present, we ingest its 5-hour billing windows, burn rate, projections, and per-source totals for other agent CLIs. Everything degrades gracefully when they're not.The native, stdlib-only tool keeps working exactly as before with zero new required dependencies —
ccusageis strictly optional enrichment.What's new
Scanner (
scanner.py)turns.is_subagent+turns.agent_idcolumns and anagentsdispatch table (additive migrations — existing DBs upgrade in place, no rebuild).isSidechain/agentId/ asubagents/transcript path; dispatch identity (agentType, stats) captured from the parenttoolUseResult. Wired into both the full and incremental scan paths.ccusage bridge (
ccusage_bridge.py, new, optional)detect_runtime()(probesnpx.cmdfirst so Windowsshell=Falseworks), a UTF-8 /LOG_LEVEL=0subprocess runner that never raises, and idempotent upserts.billing_windows+ccusage_daily_cache, kept separate from the native tables so the two sources never conflict.compute_p90_limit()(P90 of your own completed 5h windows — a personal baseline, since Anthropic's hard limit isn't exposed; same approach as Claude-Code-Usage-Monitor).Dashboard (
dashboard.py)esc().CLI (
cli.py)today/statsprint subagent token + turn summaries.scanenriches via ccusage when available; never blocks or fails a native scan.Verification
python -m unittest discover -s tests), including 33 new tests covering subagent parsing/scan, the dashboard subagent/billing data, the ccusage transforms/upserts/P90, and the CLI summaries. The bridge tests are hermetic (no Node required).node --check.ccusagerun (into a throwaway DB):Notes / limitations
python cli.py dashboardto see the new cards locally.PRICING. The two sources are shown side by side, never summed together.