🐻‍❄️ Bear in Mind — Copilot Token Meter

Track your reported Copilot plan allowance and combined account usage in your sidebar, with a dashboard for tokens, credits, speed and quality signals. Select a chat to compare its cost and reported prompt context. The polar bear's ice is a usage metaphor, not a measurement of energy consumption, CO2 or real ice loss.

Bear in Mind visualises usage. It cannot cap spending, throttle requests or change your bill, including on unlimited plans.

Install

VS Code Marketplace

Install from the Marketplace listing: search for Bear in Mind — Copilot Token Meter by obrocki in VS Code's Extensions view, or run:

code --install-extension obrocki.bear-in-mind

VSIX

Download a .vsix from Releases, then run:

code --install-extension bear-in-mind-0.7.2.vsix

Use the downloaded filename if it differs. Open the Iceberg activity-bar icon. Every merge also updates the rolling dev build. A new package version is released automatically after all main CI platforms pass. The release attaches an archive-checked VSIX; this is project-side packaging verification, not Marketplace certification. Automated Marketplace publishing additionally requires the repository's VSCE_PAT secret.

Requirements and compatibility

• VS Code desktop 1.95 or newer on Windows, macOS or Linux. This is not a web extension for vscode.dev or github.dev.
• Copilot usage readings require GitHub Copilot Chat and the corresponding local telemetry or chat-history data. Account allowance additionally requires an explicit connection and authorized GitHub sign-in; local metering works without that account connection.
• Use current stable VS Code and Copilot Chat for the fullest telemetry support. SQLite-backed trace readings and indexed chat titles require an extension host that provides node:sqlite. Older hosts can still read JSON-lines telemetry and transcript credits, but SQLite-only readings are unavailable.
• Copilot identity capture requires the supported github.copilot.chat.otel.captureIdentity setting in VS Code 1.140 or newer and separate consent. Missing identity is not inferred.
• Data files must be accessible to the extension host. Remote extension-host placement can change which machine's telemetry and history are available.

Bear in Mind is an independent personal project by Dawid Obrocki, not an official GitHub or Microsoft product. It is not sponsored or endorsed by either company. Product names identify compatible services, not sponsorship or endorsement. The optional account connection uses an unofficial GitHub API that can change or fail. See SECURITY.md for the local data read, network requests, consent and disconnect behavior, and License for reuse permissions.

The dashboard

Open Iceberg: Open Token Dashboard or the sidebar's Cost, Speed, Quality view.

Section	Signals
Cost	Combined account usage and reported plan allowance; selected-session cost and prompt context. Local model-call tokens and credits, cache/reasoning subtotals, cache-read share and feed rate. Retained-trace credits by model, repository, user, caller and reasoning effort. Not a bill.
Speed	Elapsed session duration, agent-invocation latency, model-call latency, legacy first token and canonical first stream chunk (kept separate), output throughput and slow tools.
Quality	Edit acceptance, code survival, pull requests, tool success and response feedback. Not a correctness score.

Token share is not credit share. Retained-trace credits expose small helper calls. Repository and branch come from agent spans; with identity capture on, credits are also grouped by the reported user.name. Calls without matching agent context remain unattributed.

The dashboard follows VS Code light, dark and high-contrast themes; the habitat artwork keeps its polar-night palette. Retained token diagnostics show reporting coverage, preserve measured zeroes and leave unreported fields unknown. Cache-read share uses only valid paired input/cache reports, not an incomplete cache subtotal divided by every call's input.

SDK-harness github.copilot.nano_aiu appears separately as SDK root-invocation credits only when retained parent metadata confirms the invocation root. Child/nested totals are excluded; missing or inconsistent ancestry stays unclassified. These different-grain credits are never added to VS Code model-call credits, the meter or transcript Session Cost. A model's cost multiplier is not credits or money.

With All sessions (the default), the gauge uses GitHub's reported remaining Copilot plan allowance across sessions and surfaces, not one chat's context or the sum of local observations. Run Iceberg: Refresh Copilot Account Usage… or Refresh account usage… in the dashboard to explicitly connect account usage and authorize VS Code's GitHub sign-in. Once connected, the extension refreshes once a minute; it never prompts automatically. An existing GitHub sign-in grant alone does not connect account usage.

The account connection uses the unofficial copilot_internal/user API used by VS Code. GitHub's documented billing usage APIs do not expose the same plan allowance. The quota response supplies the plan, remaining percentage, included allowance when reported, and reset date; no plan limits are hard-coded. Counts derived from percentages are labeled approximate. Legacy premium-request quotas and free-chat quotas retain their own units rather than becoming AI credits. Pooled/unlimited plans have no per-user denominator, so their account gauge stays unscaled; reported pooled credits can still be displayed.

Selecting a comparison session switches the gauge to that session's reported prompt allowance, when available. Missing, failed, stale (over 15 minutes old) or reset account quotas and missing/stale prompt limits leave the ice unscaled, unless you explicitly set a personal iceberg.tokenBudget. Account errors are visible, and a missing pinned session never switches to another chat or account scope. Neither gauge is a spending cap. The synthetic demo never changes usage.

Understanding the readings

The dashboard and Copilot report different units and scopes:

Reading	Meaning
Copilot account usage and allowance	GitHub quota API values across sessions and surfaces, with reported plan and reset date. Requires authorized GitHub sign-in; unofficial API, not an invoice.
Copilot Session Cost	Compare with transcript credits for the same session. Uses VS Code's max(sum(turn credits), reported session credits) formula and includes existing history.
Local tokens	Persisted metric/span usage since the meter start, reconciled by per-dimension maximum, plus Copilot Agent Host harness calls only the Agent Host recorded and labeled manual reports. Not a session or billing-period total.
Trace credits	Reported on unique model-call spans. Coverage is shown; missing credits are unknown. Diagnostics only, never added to transcript credits or the meter.
Gauge with All sessions	Remaining account allowance from GitHub, kept separate from local tokens and credits. Pooled/unlimited plans have no percentage denominator.
Context Window and selected-session gauge	The native window includes prompt and completion tokens against the selected model's full window. The selected-session gauge uses a trace's reported prompt allowance, not that full window.

Use Iceberg: Select Comparison Session (or Select session… in the dashboard) to pin a session, or choose All sessions to return to combined account usage. Session names come from the same local chat-history index used by VS Code, including generated titles and renames; IDs remain searchable. Both JSON and JSONL transcript formats are supported. No session is selected automatically, and this does not detect the active editor chat. The dashboard also shows local sessions observed since the 1st, separately from account usage. These session totals can include earlier history; they are not a billing-period usage report. Transcript credits take precedence; trace-credit fallbacks are labeled and never added to them. Trace data covers at most seven retained days and may be incomplete. These are local observations, not an account balance or invoice. Use Iceberg: Telemetry Diagnostics for totals, coverage, meter start and gauge basis.

VS Code has no supported third-party API for quota or active-chat detection. The unofficial GitHub quota API can change or fail; its failures do not become zero-usage claims. Disable account requests with iceberg.accountUsage.enabled: false. No local transcripts or telemetry are sent, credentials and account quota data are not persisted, and local metering works without the connection. See the ROI research for telemetry details.

Model climate

The weather around the bear shows the pace of the model in use, while the ice keeps showing how much allowance is left. The same 75% allowance can sit under snow, a thaw or, for a model running on your machine, an aurora:

Light model · snowy	Mid model · mild	Heavy model · thawing	Local model · aurora
Slow clouds, heavy snow, fresh snow on the ledges, a pinprick sun	Today's sky, light snow, a small sun	A big, low sun, a warm sky, heat shimmer, meltwater and drips	A clear aurora night, a crescent moon, glinting ice and a happy bear

• Model size is measured, not looked up. It is the model's average reported AI credits per call across retained trace-store calls; calls without credits are excluded, not counted as zero. A fixed scale from 0.05 to 25 credits per call gives a model the same weather every time. Per-call credits also grow with prompt size, so this is the model's pace as you use it, not a price list.
• Thinking effort adds sun rays (4 to 12) and spins them faster. The prompt limit (max_prompt_tokens) widens the sun's halo, the cloud cover and the horizon haze.
• With All sessions, the latest call in any chat session sets the weather. A pinned chat uses only its own latest call. Title, progress and subagent helper calls are ignored.
• The scene stays neutral, exactly as before, without trace-store spans, with fewer than three credit-reporting calls for a cloud model, 30 minutes after the latest call, during the demo, or with iceberg.climate.enabled off.

The HUD line names the band, the model and its average; its tooltip adds the sample size, effort and prompt limit. For a local model it says where the model runs instead. The figures in these screenshots are illustrative. Like the ice, the weather is a usage metaphor, not energy, emissions or a bill.

Local models

When the latest chat call goes to a model endpoint on this machine (Ollama, Foundry Local, LM Studio or any other server on localhost, 127.0.0.1 or ::1), the weather leaves the credit scale. The sky clears to an aurora night with a crescent moon instead of a sun, the ice glints and stops dripping, and the bear is happy: hearts, a wagging tail and rolls in the snow. The ice still shows the plan allowance.

• Locality is reported, not guessed. It comes only from the server.address Copilot records on each chat call, never from zero credits (some cloud models report zero), model names or bring-your-own-key status.
• A model server elsewhere on your network counts only when you list its exact host name in iceberg.climate.localHosts, for example ["gpu-box.lan"].
• No credit-reporting calls are needed. The 30-minute, demo and iceberg.climate.enabled rules above still apply.
• The extension sees only the endpoint, so a local proxy that forwards to a cloud service looks the same. Local compute still uses your machine's power.

Connecting the telemetry

Run Iceberg: Connect Copilot Telemetry…, choose a source, reload the window, then send a chat request. Allow time for the export to arrive.

Source	Provides	Exporter impact
Local trace store (agent-traces.db, plus agent-host-traces.db for the Copilot Agent Host harness)	Model-call tokens, optional credits, exact timings, reported prompt limits, session repository/branch and user, callers, reasoning effort and tool-span status.	Runs alongside an existing OTLP collector.
JSON-lines file feed	Token metrics, quality signals and (in current Copilot) serialized spans with session details.	Replaces the OTLP exporter. The command warns first.
Both	All available signals.	Same replacement caveat as the file feed.

VS Code 1.141+ can run chats in the Copilot Agent Host harness, which is becoming the default. Its telemetry comes from VS Code's Agent Host, not Copilot Chat, so choosing the trace store also turns on the Agent Host's local store (chat.agentHost.otel.dbSpanExporter.enabled). Connect skips this when the Agent Host already exports traces over OTLP protobuf or gRPC, because its store would keep them from that collector. The Agent Host reads the setting only when it starts: quit and reopen VS Code, or run Developer: Restart Local Agent Host, which interrupts running Copilot Agent Host harness chats.

Bear in Mind reads only Copilot runtime traces from that store (spans stamped with github.copilot.*), not Claude or Codex harness sessions. Calls recorded only there are metered in their own ledger and added to the reconciled Local figure, because they never reach the Local metric feed; a call also recorded by a Local source reconciles with it. Their github.copilot.nano_aiu appears as SDK root-invocation credits, never as VS Code model-call credits.

Older file-feed spans may serialize as {}; use the trace store for exact timings and session credits on those builds. Quality appears after actions such as accepting/rejecting an edit or rating a response. An empty section is waiting for signals, not asking you to reconnect.

Attributing usage to a user

Copilot Chat can add your identity to its telemetry (github.copilot.chat.otel.captureIdentity, VS Code 1.140+). It is off by default, so after you pick a source, Connect asks separately whether to turn it on. When on, agent invocation spans carry user.name (your GitHub account) and telemetry resources carry process.user.name and host.name. Bear in Mind keys user attribution on user.name: model calls inherit it from the latest agent span at or before the call through session, conversation, parent-session or trace keys, without crossing conflicting native chat IDs; calls before any such span stay unattributed. The trace store keeps span attributes only, so the OS user and host name are not available from it.

SDK enduser.pseudo.id is an opaque analytics identity, shown separately from GitHub-account labels. Explicit resource user.name values retain configured provenance; neither kind is a verified organisation join. Canonical actor/device keys are derived, and team/cost-centre assignments need configured attributes or governed enrichment. The extension does not infer a person from an OS user, machine or repository owner.

Identity capture also reaches an OTLP collector that stays connected, and the prompt says so. If replacing a collector fails, Connect leaves identity capture off. COPILOT_OTEL_CAPTURE_IDENTITY outranks the setting, so when it is set Connect does not offer the choice. An organisation policy (CopilotOtelCaptureIdentity) overrides both. The setting covers only the Local chat harness; Copilot Agent Host harness sessions get identity only through managed telemetry.capture.identity (VS Code 1.141+), and only when the bundled runtime adds user.name. Iceberg: Telemetry Diagnostics shows whether capture is requested and how many retained calls are attributed to a user.

Disconnecting and restoring defaults

Choose Restore defaults and disconnect… in the connect picker or dashboard, or run Iceberg: Restore Defaults and Disconnect…. A confirmation lists changes first. Restore then:

• returns each Copilot telemetry setting it changed, including identity capture and the Agent Host trace store, to your previous user value (or the default), leaving any you changed yourself afterwards alone;
• resets Bear in Mind's user settings, but not workspace settings;
• disconnects account usage, immediately clears the displayed account quota, and resets the saved connection opt-in. Requests do not resume after reload until you explicitly reconnect, even if the GitHub sign-in grant still exists;
• deletes its meter history, session pin and its own storage folder, which holds the default feed file. The folder is kept while Copilot is still set to write there, and a feed at a custom iceberg.otel.feedPath is never deleted. Other open windows stop saving their copies and start fresh when reloaded.

Copilot's trace stores and transcripts are untouched. Reload so Copilot Chat applies the restored settings (the Agent Host applies them when it next restarts), or choose Uninstall Bear in Mind. Connections made before 0.6.3 have no saved previous values; only settings still matching Bear in Mind's values can be reset. Account disconnection takes effect even if you dismiss the reload prompt. Your GitHub sign-in is preserved for other extensions; use Refresh Copilot Account Usage… to connect again.

Quality uses cumulative metrics where available and documented events otherwise; the two are never added together. When tool-call data is absent, retained execute_tool span statuses provide Tool success · traces. Only OK and ERROR count; missing and UNSET statuses are unknown.

How tokens get counted

Only chat spans count; span IDs deduplicate file and SQLite copies, and parent invoke_agent totals are excluded. Metrics and spans may overlap, so the meter uses their per-dimension maximum, not their sum. Coverage may be incomplete. Transcripts supply session metadata and reported credits only; token snapshots are ignored and not metered. Manual reports are separate and labeled. @iceberg uses reported telemetry, not estimates from visible text. Metering stays local; see SECURITY.md.

API

The exported API and iceberg.report command share one implementation and accept { input?, output? }. A numeric argument remains supported as input tokens. Report only usage the automatic watchers do not already see.

const extension = vscode.extensions.getExtension('obrocki.bear-in-mind');
const api = await extension?.activate();
api?.reportUsage({ input: 1200, output: 340 });

The equivalent command is vscode.commands.executeCommand('iceberg.report', { input: 1200, output: 340 }).

Member	Contract
reportUsage(usage)	Adds a report; missing fields are zero. Non-positive/non-finite values are ignored; positive values are rounded.
getUsage()	Current totals, optional budget (0 = none), animation health (0–1), source (otel or none), basis, context, model climate (absent when neutral), manual/legacy counts and metric/span drift. Only display a percentage for a measured/custom basis, never unavailable.
onDidChangeUsage(listener)	Usage snapshots; returns a disposable subscription.

Types: IcebergApi, UsageReport, UsageSnapshot, Climate. The meltdown demo never changes usage. @iceberg is counted only by automatic telemetry.

Commands and settings

Search the Command Palette for Iceberg to open the habitat/dashboard, connect or disconnect telemetry, inspect diagnostics/stats, name the bear or toggle the meltdown demo.

Common settings:

Setting	Default	Purpose
iceberg.tokenBudget	0	No default target. Positive values opt into a personal fallback, not a Copilot limit.
iceberg.trackCopilotChat	true	Read transcript session metadata and reported credits; ignore token snapshots.
iceberg.otel.enabled	true	Read local telemetry.
iceberg.otel.authoritative	true	Include new telemetry usage in the meter; off makes it display-only.
iceberg.otel.feedPath	""	Override the feed path; otherwise follow Copilot's configuration.
iceberg.otel.tracesDbPath	""	Override trace-store discovery.
iceberg.climate.enabled	true	Let the weather show the pace of the model in use; off keeps the neutral scene.
iceberg.climate.localHosts	[]	Extra model-server host names that count as local (aurora weather). Loopback always counts.

Use Settings → Iceberg for polling, input/output counting, animation, pixel scale, status bar and bear name. Turning telemetry off does not erase counted usage.

Troubleshooting

• No data: reload after connecting, send a chat request, then run Iceberg: Telemetry Diagnostics. Check View → Output → Iceberg. Transcript credit comparison requires a VS Code build that persists usage metadata.
• Quality is empty: an active feed can have token data before any quality actions occur. Trace-store-only setups show tool success only; edit, survival and feedback signals need a file feed.
• No user breakdown: turn on github.copilot.chat.otel.captureIdentity (or reconnect and accept the identity prompt), reload, and send a request from the Local harness. A policy can deny capture; diagnostics show what was observed.
• Weather stays neutral: the model climate needs trace-store spans, a chat call in the last 30 minutes, and at least three calls of that model that reported credits. Check Iceberg: Telemetry Diagnostics for trace-store activity and the dashboard's credits by model for reported credits.
• No aurora for a local model: the chat call must report its endpoint in server.address, and a non-loopback server must be listed exactly in iceberg.climate.localHosts. Models added by other extensions may not report an endpoint, so they stay neutral.
• Collector stopped: clear github.copilot.chat.otel.outfile to restore OTLP, or run Iceberg: Restore Defaults and Disconnect…; use the local trace store alongside it.
• Duplicate menus: remove old local.iceberg-copilot or obrocki.iceberg-copilot installs, then reload. iceberg.* settings remain valid.

Research: ROI on AI-assisted development

The ROI research compares telemetry across Copilot surfaces and proposes an attribution model. The AI attribution canvas renders it with read-only local coverage. Neither is part of the VSIX; reported usage alone does not prove monetary return.

An ordered, invented work-to-billing example illustrates the model, not live account data. Billing is a separate check; work costs are estimated.

Development

npm ci
npm test
npm run vsix

The VSIX build typechecks, bundles and verifies the archive. Press F5 for an Extension Development Host. See CONTRIBUTING.md for architecture, accounting invariants and release steps.

License

This project is licensed under the standard MIT License. You may use, copy, modify, merge, publish, distribute, sublicense and sell it, including for commercial use and as part of proprietary software. You do not have to publish your modifications.

Keep the copyright and permission notice in copies or substantial portions of the software. The existing notice names Dawid Obrocki, not Microsoft or GitHub. The software is provided as is, without warranty. The independent-project statement above clarifies project status; it does not add restrictions to the MIT license.

MIT © Dawid Obrocki