contextTop

top, but for your GitHub Copilot context. A live VS Code panel that shows what is filling your Copilot context window — system prompt, tools, files, history — in real time, and points out what to cut before context gets expensive or noisy.

contextTop is local-first: a Rust engine does the measurement, a VS Code panel draws it, and raw prompts or source text are never stored or sent anywhere.

• Live context dashboard — a streaming chart of context pressure by source, with gauges for tools/MCP, instruction files, open files, and terminals.
• Request breakdown — turns the one big "input tokens" number for a request into an honest split: system prompt vs. tool schemas vs. prompt vs. history/files, against the model's real budget.
• Per‑request metrics — time‑to‑first‑token, cache‑hit %, latency, output tokens, context growth per turn, and more — plotted over time.
• Tool‑efficiency "story" — names the tools loaded on every request, flags the ones you never call, and estimates the tokens you'd reclaim by turning them off.
• Honest labels — every number is tagged Observed / Estimated and Confirmed / Candidate, so visible material is never passed off as confirmed request context.

contextTop watches GitHub Copilot's own local activity, measures it in a Rust engine, and streams the result into a VS Code panel — nothing leaves your machine.

flowchart LR
  CC["GitHub Copilot Chat"]
  LOG[("Copilot debug logs<br/>main.jsonl + sidecars")]
  subgraph EXT["contextTop extension"]
    COL["Ambient collectors<br/>files · tools · instructions · selection"]
    TAIL["Diagnostic log tailer"]
    DASH["Webview dashboard"]
  end
  ENG["Rust engine<br/>measure · label · rank"]
  CC -->|"writes OTLP spans"| LOG
  LOG -->|"tail per request"| TAIL
  COL -->|"ingest candidates"| ENG
  TAIL -->|"request metrics"| ENG
  ENG -->|"snapshots · gauges · fixes"| DASH
  DASH -->|"manage tools · accept fix"| ENG

• VS Code 1.96+
• GitHub Copilot Chat installed and signed in (contextTop reads its local activity)

Each release ships a platform‑specific .vsix with the engine bundled inside — no source checkout or extra setup.

1. Pick your platform: darwin-arm64 (Apple Silicon Mac), darwin-x64 (Intel Mac), linux-x64 (Linux), or win32-x64 (Windows).
2. Download contexttop-<platform>.vsix from the latest Release (attached directly), or from a green CI run's artifacts (unzip the contexttop-<platform>-vsix artifact).
3. Install:

   code --install-extension contexttop-<platform>.vsix

   Or in VS Code: Extensions view → ⋯ → Install from VSIX….

See docs/TESTING.md — build the engine, compile the extension, and press F5.

1. Open the panel. Reload VS Code, then open the contextTop view in the bottom Panel area (beside Terminal) — or run contextTop: Open Fix from the Command Palette.
2. Use Copilot Chat as usual. Send a prompt or run an agent task.
3. Watch the dashboard. Within a second the chart moves and the cards fill in:
   • The Story card summarizes tool cost and what's unused.
   • The chart auto‑switches to Request mode on your first request; use the dropdown to plot all metrics, a single metric (cache hit %, TTFT, budget %…), or the source composition.
   • The breakdown bar shows where your tokens actually went.

There is nothing to configure to get started.

The panel is a stack of cards, top to bottom — the Story and Latest request cards answer "what is costing me and what can I cut?" at a glance.

flowchart TB
  T["Toolbar — chart mode · metric · time window"]
  S["Story — tools loaded, the unused ones, tokens to reclaim"]
  R["Latest request — input · cache-hit% · TTFT · latency · model"]
  B["Breakdown bar — system prompt · tools · prompt · other"]
  G["Gauges — peak · rate · tools · instructions · files · terminals"]
  C["Chart — per-source lines, or one chosen metric over time"]
  L["Sources table — top sources by tokens"]
  A["Analytics — turn duration · requests/turn · discovery counts"]
  T --> S --> R --> B --> G --> C --> L --> A

Want real screenshots here? Capture the panel from your installed build and drop the images in docs/media/, then reference them in this section — see docs/DASHBOARD.md for what each card shows.

All settings live under contextTop.* (Settings → search "contextTop").

Retention and storage caps (measurementRetentionDays, rollupRetentionDays, maxStorageMiB, …) are also configurable in the Settings UI.

Raw prompts, source text, file paths, and terminal content are never persisted or exported — they are processed transiently to compute counts, then dropped. Stored metrics are pseudonymous, bounded by age and size, and stay on your machine. Details: docs/arch/SIGNALS.md and docs/arch/METRICS.md.

Using contextTop

• docs/PRODUCT.md — the product experience and its boundaries.
• docs/DASHBOARD.md — every metric, the sidecar breakdown, and the extensible card model.
• docs/TOOL_STORY.md — the tool‑efficiency story and the limits of per‑call tool toggling.
• docs/TESTING.md — build, run, install, and release.

Architecture & technical specs — docs/arch/

• docs/arch/ARCHITECTURE.md — Rust‑first architecture and trust model.
• docs/arch/SIGNALS.md — the signal capability matrix and opt‑in tiers.
• docs/arch/METRICS.md — snapshot semantics, pipeline, retention, and chart contract.
• docs/arch/IPC.md — the versioned engine↔adapter protocol.
• docs/arch/IMPLEMENTATION_PLAN.md — milestones and acceptance criteria.

Every metric carries three labels — Measurement (Observed / Estimated / Unknown), Inclusion (Confirmed / Candidate / Unknown), and Coverage (Complete / Partial / Unknown). See docs/arch/METRICS.md.

Build from source, run the tests, and follow the CI‑parity checks in docs/TESTING.md. Use ./push.sh to commit + record the TODO + push as one unit, and scripts/release.sh <version> to cut a release.

MIT