AgentVitals
Real-time Claude Code & Codex usage, cost and rate limits — right inside VS Code & Cursor.
This extension was previously called “Claude Code Gauge.” Same extension, same
publisher, same settings — only the name changed. Nothing to reinstall, and your
claudeCodeGauge.* settings keep working untouched.
Stop switching to your browser to check rate limits. See your 5-hour session, 7-day weekly usage, today's token cost, what your agent actually did, which skill & branch cost how much, and how much your subagents spend — with burn rate predictions, tool usage breakdowns, per-skill cost attribution, and Git branch ROI — without leaving your editor. Works with both Claude Code and Codex, switchable from the sidebar.
What's New in v0.2.6
- New: why the prompt cache missed. The cache card lists last week's cache misses by cause (conversation changed, model changed, system prompt changed…) with an estimated extra cost each. Causes that don't record token counts are counted but not costed, so the total is a lower bound (
≥).
- New: rate-limit block history. Retries rejected by a usage limit are grouped into one block per window and reset time, and listed on the dashboard. The sidebar shows a chip for a block in the last week. Server errors (529/500) are listed separately, because they are not a usage limit.
- New: cost by pull request, from the PR links Claude Code records, including subagent work. A session linked to several PRs is marked shared.
- New: turn latency and hook overhead. Median, p90 and longest turn time, a daily median, and the average time each Stop hook adds per turn (script name only, no paths).
- New: thinking share, cost by effort level, and context compaction. A sidebar chip shows today's thinking-token share. The attribution card adds cost by effort level (high / medium / xhigh). The context gauge shows a chip when its session was compacted.
- New for Codex: credit balance, spend limit and block reason when Codex reports them; subagent work counted toward its parent session; readable plan names (Pro Lite instead of
PROLITE).
- Easier to read: red status markers now use a lighter red on dark themes, so a rate-limit block marker is as legible as the other markers.
v0.2.5
- Fixed: Claude Fable 5.1 and Opus 5.5 cost was overstated. Neither had its own price, so each borrowed its predecessor's — and because that counted as a match, no "price unknown" warning appeared. Fable 5.1 reads cache at a quarter of Fable 5's rate, so heavy cache use overstated it by roughly 30–90%. Both now have their own prices, derived from the costs Claude Code records in its logs.
- Fixed: subagent usage was missing. Claude Code writes each subagent's transcript to a separate file under the session folder, which was never read. Subagent cost is now in every total and in the dashboard's subagent breakdown. Responses copied from the parent session are still counted once, and stay attributed to the main session.
- New: cost by subagent type. The cost attribution card now shows what each kind of subagent spent — general-purpose, planner, reviewers and so on — with run counts and tokens on hover, following the 24h / 7d / All toggle. Untyped subagent work gets its own row, so the list always adds up to the subagent total.
- New: "⚠ price mismatch" marker. The dashboard's model breakdown now compares the extension's cost calculation against the cost Claude Code itself records, and flags models where the two disagree (details on hover). It catches a wrong price, not just a missing one.
v0.2.4
- The dashboard reads with fewer sentences. The explanatory lines under the cache hit rate and the Utilization Trend chart, and the warning banners for models missing from the price table, are now short status markers — Normal / Drop, Under pace / Over pace, ⚠ price unknown — with the full explanation on hover or keyboard focus. The Codex-only note about variable windows is gone (the window labels already show it), the Codex extra panel is now titled "Codex-only metrics", and the Git ROI card's footer note moved into the ≈ badge's tooltip.
- Fixed: the Long-term tab of Cost by Period showed today's cost. Its header figure took the last day of the selected range, so it always matched the Daily tab. It now shows the total for the selected range (30, 90 or 180 days).
v0.2.3
- Fixed: the status bar ignored the Claude/Codex switch. It kept polling and showing Claude's 5h/7d figures while Codex was selected — every other surface followed the switch, but the one always on screen did not. Under Codex the two items now hide rather than relabel: Codex reports a variable number of windows (one 30-day window on Free, 5h+7d on paid), which two fixed slots can't show without making up a label.
- Fixed: burn rate showed "0.00%/min" on long windows, which looks exactly like being idle. The rate was always per-minute, so 12% of a 30-day quota came out as zero. The unit now follows the window length — per minute, per hour or per day.
- Fixed: branch and file lists showed other projects' work. The branch chip, branch cost breakdown and Recently Edited Files were built from every session on the machine. They are now scoped to the open workspace.
- Fixed: the dashboard's plan badge could show the other agent's plan — e.g. the Claude tier with Codex selected, or the Codex plan lingering after switching back. It is now decided from the selected agent on every update.
- Fixed: the sidebar's Codex burn rows lost their "(proj N% left)" projection, shown — as on Claude's rows — only when the current pace would exhaust the window before reset.
- Codex windows on the dashboard now carry their own burn rate and safe-until time, like the sidebar has since v0.2.1 — the dashboard had been discarding each snapshot instead of keeping the history a burn rate needs.
- The Utilization Trend chart works under Codex now, one line per reported window instead of two hardcoded ones. The dashed "expected pace" line stays Claude-only, since it is derived from a five-hour window Codex has no equivalent of.
- The Cost by Period tabs are ordered Daily → Monthly → Long-term, so the two everyday views sit next to each other.
- Light themes are applied. The webviews had the dark palette hardcoded, so the light tokens shipped since v0.1.53 were unreachable. Both views now follow VS Code's theme — including high contrast — and switch live.
v0.2.2
- Daily, long-term and monthly cost are one section now. They were three separate blocks spread down the dashboard; comparing the same number across time windows meant scrolling past everything in between. They are now a single Cost by Period card with tabs — Daily / Long-term / Monthly. The long-term tab keeps its own 30/90/180-day range toggle, working exactly as before. No chart was removed.
- Fixed: a Codex log file replaced by a different file of the same size wasn't noticed. The reader would carry on from where it left off and never read anything before that point. It now checks the start of the file and re-reads from the beginning when it has changed.
- Internal: the sidebar's per-bucket history is now covered by tests that were impossible to write before, and the marketplace screenshot and demo GIF are re-shot from the real build by a single command.
v0.2.1
- A Codex usage-count bug fixed. If a subagent replayed a parent session's response, Codex's usage total could count that response twice — the same class of billing bug this extension already guards against for Claude Code. Duplicate responses are now recognized across files, not just within one.
- Codex now reads faster — and more safely. Every refresh used to re-read and re-parse the full Codex session history from disk; it now picks up only what's new since the last refresh, the same incremental approach Claude Code sessions already used. A record could have been silently lost if a refresh landed mid-write; the reader now only advances past complete lines.
- Fixed:
codex-mini-latest showed up as "Latest" instead of the model name.
- Fixed: the sidebar's per-bucket burn-rate row could show the wrong bucket's history after a Free→paid plan change or similar bucket-shape change. Each bucket's rate now tracks only its own history.
- Each Codex rate-limit bucket in the sidebar now shows its own burn-rate line, matching Claude's 5h/7d gauges.
v0.2.0
- Codex support. A Claude/Codex switcher in the sidebar header tracks either agent — gauges, today's cost, model/cache breakdown, Git branch ROI and the usage calendar all now read Codex's own session logs too. Rate-limit buckets are generated from what each session actually reports (a single 30-day window on Free, 5h+7d on paid plans), not hardcoded. What Codex's logs genuinely don't contain — per-skill cost, subagent breakdown, MCP attribution — is left out rather than guessed; in its place you get today's reasoning-token total, the model's real context-window size, and a plan badge. Three distinct empty states (not installed / not logged in / no sessions yet) each carry their own guidance instead of one generic "please log in."
- A visual pass on the dashboard and sidebar — clearer type hierarchy, the card count trimmed from 17 to 6, end-of-line readouts on every chart, and a bigger 5-hour gauge as the sidebar's opening view.
- The extension is now called AgentVitals. It used to be “Claude Code Gauge”. It is the same extension from the same publisher — you do not need to reinstall it, and every setting you have configured keeps working under its existing name.
v0.1.56
- Fixed: Fable rendered as a black wedge in the Model Breakdown donut, and its bar didn't show at all. Its colour was removed in v0.1.54 by a cleanup that searched for unused colours and concluded this one was unused — it wasn't; the chart looks it up by a name it assembles as it runs, which no search for that name can find. Every build check stayed green, so this shipped in both v0.1.54 and v0.1.55. The colour is back, and a new build rule now checks colours the way the code actually uses them, so a model can no longer end up without one.
- Fixed: changing a setting had no effect until you reloaded the window. The poll interval and credentials path were compared against a key that could never match, so the extension never noticed you'd changed them.
- Fixed: usage figures could stay frozen for as long as you kept working. The refresh waited for your logs to go quiet, but they don't go quiet while Claude Code is running — so the update arrived only after you stopped. It now refreshes immediately and then at a steady interval while you work. You can set that interval with
claudeCodeGauge.usageRefreshIntervalMs (default 15 seconds), and the refresh button ignores it — that button now updates usage too, not just rate limits.
- Fixed: charts replayed their animation on every background refresh, which looked like flickering rather than like data arriving.
v0.1.55
- Fixed: your costs were being reported as roughly 1–2% of what you actually spent. The price table hadn't been updated for the current model generation —
claude-opus-5 and claude-sonnet-5 were missing outright, and the fallback that was supposed to catch new models had never once fired, so nearly every record was priced at $0 while everything else looked normal. Prices are now taken from Anthropic's own per-model figures that Claude Code records in its logs (Opus 5 at $5/$25; Sonnet 5 at $2/$10 — cheaper than Sonnet 4.x), web search billing is counted, and a fixture keeps those numbers honest when prices change again.
- Fixed: the Model Breakdown pointed at the wrong model. Because shares were computed from cost, the missing prices meant whichever older model still had one took 100% of the chart while the model doing almost all the work showed 0%. When any model's price is unknown, shares now switch to a token basis and say so, the affected row reads "price unknown" instead of
$0.00, and the card warns you which models are involved — rather than showing a zero that looks like a measurement.
- Fixed: the commit retrospective credited your spending to other people's commits. On a shared repository it considered every colleague's commit from the last six months. It now only considers commits you authored (
claudeCodeGauge.retroCommitScope restores the old behaviour), and tells you when a repository has no user.email to filter by.
- New: long lists fold. Recently Edited Files, Recent Sessions and the branch list were unbounded and pushed the cards below them off-screen. They now show six rows with a "Show more (+N)" toggle that tells you how many are hidden.
- Fixed: on a narrow dashboard the Usage Calendar stopped short of today, cutting today's square off the right edge. The calendar scrolls itself to the right after every refresh so you see the most recent week — but that scroll was applied before the panel's layout had settled, so it landed 52px short and never corrected itself. Worse, the wrong position was then mistaken for "you scrolled into the past" and preserved on every later refresh. Dashboards wider than about 750px were unaffected, which is why this went unnoticed since v0.1.52. Scrolling back into history still works exactly as before.
- Fixed: the sidebar's mini calendar never scrolled to today at all. On a narrow sidebar (220px and below) it showed the oldest weeks and clipped today entirely. It now behaves like the dashboard.
- The rest of this release is internal, and mostly about why the above took three releases to notice: the test that had been catching it every day was never wired into the build. That's fixed, along with the checks that could report "all clear" while examining nothing.
v0.1.54
- Fixed: a handful of error messages shown in the sidebar/panel weren't HTML-escaped before being displayed. If the webview's messaging layer ever failed to start, its error text was inserted into the page without escaping first — a defensive hardening fix, not a reported exploit. (Completed as v0.1.53, which was never published on its own; it reached you in v0.1.54.)
- Everything else in v0.1.54 was internal and changed nothing you can see: unit-test coverage for the usage-rollup, cache and workspace-mapping services that had none; type checking turned on for the webview and test code, which had been silently excluded; an off-the-event-loop fix for a directory scan that ran on every refresh; and the 1882-line webview entry point split into four modules, checked against a structural snapshot of the rendered UI across both surfaces, two widths and two languages to confirm nothing moved.
v0.1.52
- Fixed: the Usage Calendar's month labels no longer drift away from the days they label. On a wide dashboard the heatmap's day cells stretched to fill the card while the month labels stayed put, so the labels ended up describing the wrong weeks — by as much as 704px on a 1600px-wide window, which made your most recent activity look like it belonged to a month far in the future. The calendar is now fixed-width on both the dashboard and the sidebar: cells stay the same size no matter how wide the panel gets, and the leftover space is simply left empty, the way contribution graphs normally work. If the panel is narrower than the calendar, it scrolls sideways as before.
- Fixed: month labels in Korean, Japanese and Chinese no longer break across two lines and overlap the top row of cells. The labels now stay on one line in every language, and English labels ("Aug", "Sep") no longer nudge the spacing out of alignment either.
v0.1.51
- Fixed: in a multi-root workspace, the context gauge could sit on a repo you weren't working in. The gauge only ever looked at the workspace's first folder — a limitation v0.1.48 shipped knowingly. In practice that meant opening three repos in one window, working in one of them, and watching the gauge report a different repo's session from a day and a half ago. It now considers every folder open in the window and shows the most recently active session among them.
- New: pick which session the gauge tracks. Click the
📁 workspace chip under the gauge to open a session picker listing every session seen in this workspace — repo, branch, how long ago it was active, its context usage, and model. Pick one to pin the gauge to it, which is useful when you're moving between several live sessions and don't want the reading jumping around.
- New: pinned sessions can't quietly go stale. Auto mode is still the default and follows your most recent activity. If you pin a session and it then goes quiet for more than 4 hours, the gauge turns amber and offers "Switch back to auto" — so a pinned reading can't become the very problem this release fixes. A pinned session that disappears entirely falls back to auto on its own.
v0.1.50
- Fixed: the session context gauge could over-report usage by up to 5×. The lookup table backing the gauge's denominator classified every current model as a 200K-token context window, even for accounts with the 1M window active — and the 100% cap hid just how far off that was (one workspace was actually at 292%, not 100%). The gauge now recognizes a 1M window two ways: if any request in your history ever exceeded 200K tokens for that model (which a 200K window couldn't physically produce), or if Claude Code itself recorded a
[1m] marker for that model in ~/.claude.json. Falls back to the old 200K assumption only when neither signal is available.
- Fixed (minor): the gauge's token count could nearly double on multi-call turns. It summed every API call within a turn instead of using only the final one, which is what actually reflects the context window's current occupancy.
- Fixed (minor): background/subagent sessions could no longer hijack the gauge. Sessions running in the background are now excluded when picking "your most recent session."
- Added: absolute token counts next to the percentage (
≈72K/1M), plus an age indicator — the gauge dims after 4 hours of inactivity so a stale reading doesn't look like a live one.
v0.1.49
- Fixed: the session context gauge never appeared at all on native Windows. v0.1.48's workspace scoping compared paths case-sensitively, but VS Code lowercases Windows drive letters (
c:\...) while Claude Code logs them as typed (C:\...) — so the match always failed and the gauge silently hid itself, on every native Windows install. Case is now folded only for Windows drive paths; POSIX paths are unaffected.
- Fixed: no way to tell that apart from a genuine "no session yet" state. The gauge now says "No session record for this workspace" when your machine has session history elsewhere but none for the current one, instead of hiding identically to the bug above.
v0.1.48
- Fixed: the session context gauge showed the wrong repo in multi-repo setups. It used to pick the most recently active session across all your projects — so if you had another repo open elsewhere, its context usage could show up in a repo you weren't even working in. It now scopes to the current workspace's first folder, and hides itself entirely if that workspace has no session yet, instead of silently falling back to a different repo's numbers. (In a multi-root workspace it always scopes to the first folder, not whichever one you're focused on — most multi-repo setups use one VS Code window per repo, which this fully covers.)
- New: a small repo chip under the context gauge (
📁 <folder name>) makes it explicit which workspace the percentage belongs to, with the full path on hover — since a session started from a subdirectory shows that subdirectory's name, not necessarily the repo root.
v0.1.47
- New: Session context gauge. A compact bar in the sidebar showing how full your latest session's context window is — computed from the most recent turn's input + cache tokens against that model's window. Approximate by nature (auto-compact isn't recorded in the logs), so it's labeled
≈ rather than presented as exact.
- New: MCP server attribution. Cost attribution now breaks down MCP usage per server (
mcp__<server>__*), ranked by call count. Share is deliberately call-count based, not cost based: a single assistant turn can mix MCP and non-MCP tools, so splitting cost per server would be false precision.
- New: 24h / 7d / All toggle on cost attribution. The skill, subagent, and MCP breakdowns now scope to the last 24 hours or 7 days instead of only all-time — see what's driving spend right now, not just cumulatively.
- Fixed: Burn Rate and Safe Until got stuck on "Collecting data…" whenever you paused. If your usage didn't move between two polls, the measured rate was
0 — which the cards misread as "no data yet" and could stay that way indefinitely. Idle is now a distinct state (0.00%/min · idle), separate from genuinely still collecting and from a 5h window reset. The rate is also averaged over the last 30 minutes rather than just the two most recent polls, so it no longer jumps on a single noisy sample.
v0.1.46
- Fixed: sidebar chip rows (model, tools, branch, monthly cost) no longer get cut off when you narrow the sidebar. They now wrap to a new line like every other section instead of overflowing off-screen.
- Fixed: faint stray divider line above the sidebar Usage Calendar.
v0.1.45
- New: Sidebar mini Usage Calendar. A compact, last-3-months version of the dashboard's GitHub-style heatmap, right below Overage usage — same coloring and hover tooltips, no legend, hidden entirely until you have usage history. Reuses the dashboard's exact cell size rather than shrinking it, sized so it fills the sidebar's default width without horizontal scrolling.
- Fixed: "Open Dashboard" button no longer sticks to the section above it. The sidebar's mount element was missing an explicit height, so the button collapsed up against Overage usage instead of anchoring to the bottom. It's now correctly pinned at the bottom, and the sidebar scrolls instead of clipping when its content (now including the calendar) is taller than the panel.
v0.1.44
- Fixed: Usage Calendar hid your newest weeks on narrow dashboards. The fixed 1-year grid was silently clipped on the right — exactly where your recent usage lives — making the calendar look empty despite active use. It now scrolls horizontally, opens anchored to today (GitHub-style), and remembers your scroll position across data refreshes.
- Fixed: rate-limit polling can no longer hang silently. Requests now time out after 15s and surface through the normal error path instead of leaking sockets while the gauge quietly goes stale.
- Polish: zero hardcoded colors (all UI now follows your VS Code theme, light or dark), cryptographically random CSP nonces, and fully async file I/O during usage refresh.
v0.1.43
- New: Usage Calendar. A GitHub-style contribution heatmap of your daily Claude Code cost, right below the Daily Cost card — hover any day for its cost/tokens, with a highlighted "today" cell. History now backfills in full on every refresh, so the calendar (and other history charts) fill in from day one instead of growing one day at a time.
- Security: hardened
credentialsPath against workspace-level hijacking. A malicious repo's .vscode/settings.json could previously point this setting at an arbitrary file; it's now locked to your user/global settings only (VS Code enforces this at the platform level, plus a code-level fallback).
Features
- StatusBar: Two independent items —
5H 🟦🟦⬜⬜⬜ 28% and 7D 🟦⬜⬜⬜⬜ 14% — emoji fill count based on utilization %; color (🟦🟨🟥), background, and font based on utilization thresholds (0–80 % blue · 80–90 % amber · 90–<100 % red Danger · 100 % red Blocked)
- Sidebar: Three labeled sections — Session (5h) · Weekly (7d) · Overage — each with
used% · left% display + status-colored progress bars (blue OK / amber Warning / red Danger·Blocked) + overall status badge inline with title
- Plan badge: Your subscription tier (e.g.
Max 5x) shown in the header — read from local credentials, no extra API call
- Burn Rate:
%/min consumption speed — estimated from session elapsed time on first open, then refined from poll history (averaged over the last 30 minutes, so a single noisy poll doesn't swing it). Distinguishes idle (0.00%/min · idle) from still-collecting and from a window reset, instead of showing "Collecting data…" for all three
- Safe Until: Predicted time when your 5h quota runs out at current burn rate
- Dashboard Panel: SESSION · WEEKLY · BURN RATE · SAFE UNTIL 4-card layout + utilization trend chart
- Trend Chart Scope: Toggle 30m / 2h / 24h view window directly on the chart
- Bottleneck highlight: The currently limiting window (5h or 7d) is outlined in amber so you instantly see what's constraining you
- Overage section: Progress bar + status chip for your overage (extra usage) quota — shows the overage rate-limit utilization only when active (amber), and a
DISABLED chip when overage is rejected/disabled instead of a misleading "0%". A help tooltip clarifies this is the overage rate-limit usage (consumed after your base 5h/7d quota is exhausted), distinct from the claude.ai "Usage Credits" dollar-spend figure
- Fallback banner: Inline warning when Claude throttles to reduced speed (e.g. 50%)
- 7d threshold badge: Red badge on the Weekly card when a usage threshold has been surpassed
- Rate-limit block history: Retries rejected by a usage limit, grouped into one block per window and reset time, on the dashboard, plus a sidebar chip for a block in the last 7 days. Server errors (529/500) are listed separately
- Threshold alerts: Native VS Code warning notification when usage exceeds your configured limit
- Auto-polling: Fetches latest rate limit headers from Anthropic API every 5 minutes
Token & Cost Analytics (local .jsonl)
- Today's usage: Sidebar shows "N tokens · ~$X.XX" — parsed directly from
~/.claude/projects/**/*.jsonl, no API call
- Model chip: Color-coded Fable / Opus / Sonnet / Haiku chip showing today's primary model in the sidebar
- Cache hit rate chip: Today's cache hit rate (e.g.
⚡ 72%) with saved cost in tooltip — pure local calculation
- 7-day cost bar chart: Dashboard panel shows daily spend for the past 7 days
- Model breakdown: Doughnut chart + bar list showing per-model cost share for today
- Cache efficiency: Hit rate KPI, cumulative saved cost, and 7-day sparkline in the dashboard
- Cache misses by cause: Last 7 days of prompt-cache misses grouped by cause, with an estimated extra cost each (a lower bound when a cause doesn't record token counts)
- Thinking share chip (sidebar): Today's thinking tokens as a share of output tokens — already part of output, so no extra cost
- Session history: Up to 20 recent sessions with start time, working directory, token count, and estimated cost
- LiteLLM pricing: Offline cost calculation using embedded model price snapshot (fable-5 / fable-5.1 / opus-5 / opus-5.5 / sonnet-5 / opus-4.5–4.8 / sonnet-4.5–4.6 / haiku-4.5, with legacy fallbacks), cross-checked against the costs Claude Code records
Language Support
- 4-language UI: Switch between 한국어 / English / 日本語 / 中文 via the compact dropdown in the sidebar header — all labels, section names, error messages, and burn-rate strings update instantly
- Full dashboard translation: Dashboard panel section headers, metric labels, chart titles, and empty states are all translated
- Real-time sync: Changing language in the sidebar instantly updates the dashboard panel without reopening — broadcast via extension messaging
- Auto-detection: Defaults to your system language (
navigator.language) on first install; persists your choice across sessions via extension globalState
- Open Dashboard button: Persistent button at the bottom of the sidebar — tactile depth styling with gradient and press effect — opens the full Dashboard Panel in one click
Action Insights — what Claude did (local .jsonl)
- Tool usage chips (sidebar):
Edit N · Write N · Bash N · Read N · Grep N · 🔍 N · 🌐 N · MCP N — today's tool call counts at a glance, with Read, Grep/Glob, WebFetch, and MCP (mcp__*) broken out from the old catch-all bucket
- Tool usage histogram (dashboard): Stacked bar chart of Edit / Write / Bash / Search per day for the last 7 days — spot heavy editing vs. execution sessions
- Turn latency & hook overhead (dashboard): Last 7 days of turns — count, median, p90 and a daily median — plus the average time each Stop hook adds, shown by script name only
- Context compaction chip (sidebar): How many times the gauge's session was compacted, with the last before → after size on hover
- Recently edited files (dashboard): Up to 20 files touched in recent sessions, ordered by last activity — filename + full path
Cost Attribution — where the cost went (local .jsonl)
- Cost by Skill (dashboard): Ranked bar list of cost per
attributionSkill (sh-dev-loop, ship, plan, research, …) — see which Claude Code skills drive your spend. Because Claude Code only stamps a skill on main-chain turns while a skill is actively loaded (~⅓ of cost-bearing turns), everything else — plain requests and work before/after a skill loads — is shown as a first-class "Outside skills" bucket rather than hidden, with a ≈ Partial badge. Subagent-delegated cost is surfaced separately below
- Subagent vs. main split (dashboard): Subagent consumption share, cost, and unique-agent count from
isSidechain/agentId — separate background subagent usage from your main session, broken down by subagent type (attributionAgent)
- MCP server breakdown (dashboard): Ranked list of MCP servers by call count, parsed from
mcp__<server>__<tool> tool names. Share is call-count based on purpose — one assistant turn can mix MCP and non-MCP tools, so a per-server cost split would be false precision
- Cost by effort level (dashboard): high / medium / xhigh, with responses that carry no effort level in their own row
- 24h / 7d / All scope toggle (dashboard): Re-scope the whole attribution section — skills, subagents, and MCP servers — to the last day or week instead of all time, to see what's driving spend right now
Long-term Cost Tracking (local persistence)
- CacheStore: Daily usage snapshots persist to
globalStorageUri/ccg-history.json — survives jsonl rotation so history accumulates across months
- Long-term trend chart (dashboard): Daily cost line chart with 30d / 90d / 180d scope toggle — see spending patterns across months
- Monthly cost bar chart (dashboard): Month-by-month cost aggregation — spot your most expensive periods
- This-month chip (sidebar):
◑ This month $X.XX / ≈$Y.YY — current month spend + projected end-of-month cost (linear extrapolation)
- Session context gauge (sidebar): How full your latest session's context window is — the most recent turn's input + cache-read + cache-creation tokens against that model's window. Approximate (
≈) because auto-compact isn't recorded in the logs; hidden entirely until there's any session history at all. Scoped to every folder open in the window (multi-root included) — a repo chip (📁 <folder name>, full path on hover) makes the source explicit, and if none of those folders has a matching session it says so explicitly instead of falling back to a different repo's numbers
- Session picker (sidebar): Click the repo chip to choose which session the gauge tracks. Lists every session seen in this workspace with its repo, branch, idle time, context usage, and model. Auto mode (the default) follows your most recently active session; pinning one holds the gauge on it — and if a pinned session goes quiet for over 4 hours the gauge turns amber and offers to switch back to auto, so a pin can't silently become a stale reading
Git Branch ROI (local .jsonl)
- Branch cost chip (sidebar):
⎇ main · $0.42 chip showing the active branch and its cumulative cost — parsed directly from gitBranch field in every jsonl entry, no Git API dependency
- Cost by pull request (dashboard): Sessions joined to PRs through the PR links Claude Code records, subagents included. Sessions linked to several PRs are marked shared, because adding those PRs together would count them twice
- Git ROI table (dashboard): Full branch breakdown — Branch · Cost · Tokens · Sessions · Last Active — sorted by cost so your most expensive branches surface first
- Cost by Commit — usage×git retrospective (dashboard): Extends branch ROI down to individual commits. Because git commits aren't recorded in session logs, cost is approximately attributed by
timestamp + cwd + branch — so the card is explicit about it: an ≈ Approximate badge, per-commit confidence dots, and a first-class "Other / Uncommitted" bucket (planning/research/debugging that hasn't been committed yet) instead of hiding it. Attributions persist SHA-keyed so they outlive the 30-day log window.
| |
