chat-perf

Chat Performance Testing

When to use

• Before/after modifying chat rendering code (chatListRenderer.ts, chatInputPart.ts, markdown rendering)
• When changing the streaming response pipeline or SSE processing
• When modifying disposable/lifecycle patterns in chat components
• To compare performance between two VS Code releases
• In CI to gate PRs that touch chat UI code

Quick start

# Run perf regression test (compares local dev build vs VS Code 1.115.0):
npm run perf:chat -- --scenario text-only --runs 3
 
# Run all scenarios with no baseline (just measure):
npm run perf:chat -- --no-baseline --runs 3
 

Perf regression test

Script: scripts/chat-simulation/test-chat-perf-regression.js npm: npm run perf:chat

Launches VS Code via Playwright Electron, opens the chat panel, sends a message with a mock LLM response, and measures timing, layout, and rendering metrics. By default, downloads VS Code 1.115.0 as a baseline, benchmarks it, then benchmarks the local dev build and compares.

You don't always need a baseline. A baseline exists only for comparison (regression detection). If you just want the current build's numbers — profiling a single change, capturing traces/heap snapshots, or iterating on a scenario — pass --no-baseline to skip downloading and benchmarking the baseline entirely (roughly halves runtime). Baseline comparison is what turns raw measurements into a pass/fail verdict; without it you still get all the metrics, just no verdict.

Key flags

Comparing two remote builds

Comparing two local builds

Both --build and --baseline-build accept local paths to VS Code executables. This enables apples-to-apples comparisons between any two builds:

Local path baselines are never cached (the build may change between runs). Version string baselines are cached for reuse.

Build modes and mismatch detection

The tool classifies builds into three modes based on the executable path:

When test and baseline builds have different modes (e.g. dev vs release), the tool shows a warning and prompts for confirmation. Use --force or --ci to skip the prompt.

Using --production-build builds a local bundled package via gulp vscode for fair comparison against a release baseline. This eliminates dev-mode overhead while still testing your local changes.

Settings overrides

Use --setting, --test-setting, and --baseline-setting to inject VS Code settings into the launched instance. This is useful for A/B testing experimental features:

Precedence: --test-setting / --baseline-setting override --setting for the same key. Values are auto-parsed: true/false become booleans, numbers become numbers, everything else stays a string.

Resuming a run for more confidence

When results exceed the threshold but aren't statistically significant, the tool prints a --resume hint. Use it to add more iterations to an existing run:

--resume loads the previous results.json and its associated baseline-*.json, runs N more iterations for both builds, merges rawRuns, recomputes stats, and re-runs the comparison. The updated files are written back in-place. You can resume multiple times — samples accumulate.

Statistical significance

Regression detection uses Welch's t-test to avoid false positives from noisy measurements. A metric is only flagged as REGRESSION when it both exceeds the threshold AND is statistically significant (p < 0.05). Otherwise it's reported as (likely noise — p=X, not significant).

With typical variance (cv ≈ 20%), you need:

• n ≥ 5 per build to detect a 35% regression at 95% confidence
• n ≥ 10 per build to detect a 20% regression reliably

Confidence levels reported: high (p < 0.01), medium (p < 0.05), low (p < 0.1), none.

Exit codes

• 0 — all metrics within threshold, or exceeding threshold but not statistically significant
• 1 — statistically significant regression detected, or all runs failed

Scenarios

Scenarios are defined in scripts/chat-simulation/common/perf-scenarios.js and registered via registerPerfScenarios(). There are three categories:

• Content-only — plain streaming responses (e.g. text-only, large-codeblock, rapid-stream)
• Tool-call — multi-turn scenarios with tool invocations (e.g. tool-read-file, tool-edit-file)
• Multi-turn user — multi-turn conversations with user follow-ups, thinking blocks (e.g. thinking-response, multi-turn-user, long-conversation)

Run npm run perf:chat -- --help to see the full list of registered scenario IDs.

Metrics collected

• Timing: time to first token, time to complete, time to render complete (includes typewriter animation)
• Rendering: layout count, layout duration (ms), style recalculation count, forced reflows, long tasks (>50ms), long animation frame count and duration
• Memory: heap before/after, heap delta post-GC (informational, noisy for single requests)
• Extension host: heap before/after/delta via CDP inspector

Regression triggers vs informational metrics

Only these metrics trigger a regression failure (when they exceed the threshold with statistical significance):

• timeToFirstToken, timeToComplete — user-perceived latency
• layoutDurationMs — total layout time from the trace (the real layout cost)
• forcedReflowCount — forced synchronous layouts are always bad
• longTaskCount — main thread jank

These are reported but informational only (won't fail CI):

• longAnimationFrameCount — number of long animation frames; noisy and compositor-driven, so tracked for signal but does not gate. Real jank is gated via longTaskCount / layoutDurationMs.
• layoutCount — number of layout ops; inflated by CSS animations (compositor-driven, cheap). A build can do more but cheaper layouts, so gate on layoutDurationMs, not this count.
• recalcStyleCount — number of style recalcs; inflated by CSS animations (compositor-driven, cheap)
• timeToRenderComplete — includes typewriter animation tail
• Memory/heap metrics — too noisy for single-request benchmarks

Statistics

Results use IQR-based outlier removal and median (not mean) to handle startup jitter. The coefficient of variation (cv) is reported — under 15% is stable, over 15% gets a ⚠ warning. Baseline comparison uses Welch's t-test on raw run values to determine statistical significance before flagging regressions. Use 5+ runs to get stable results.

Memory leak check

Script: scripts/chat-simulation/test-chat-mem-leaks.js npm: npm run perf:chat-leak

Launches one VS Code session, sends N messages sequentially, forces GC between each, and measures renderer heap and DOM node count. Uses linear regression on the samples to compute per-message growth rate, which is compared against a threshold.

Key flags

What it measures

• Heap growth slope (MB/message) — linear regression over forced-GC heap samples. A leak shows as sustained positive slope.
• DOM node growth (nodes/message) — catches rendering leaks where elements aren't cleaned up. Healthy chat virtualizes old messages so node count plateaus.

Interpreting results

• 0.3–1.0 MB/msg — normal (V8 internal overhead, string interning)
• >2.0 MB/msg — likely leak, investigate retained objects
• DOM nodes stable after first message — normal (chat list virtualization working)
• DOM nodes growing linearly — rendering leak, check disposable cleanup

CI runs & pinpointing regressions

The perf + leak checks run in CI via the .github/workflows/chat-perf.yml workflow (a scheduled daily workflow_dispatch, plus manual dispatch). Each run benchmarks the current main as the test build against a fixed release baseline (from config.jsonc, e.g. 1.122.0). Because the baseline is fixed, the test median for a metric across successive daily runs traces main's trajectory over time — that's what lets you bisect when something regressed or went flaky.

Finding and reading historic runs

Artifacts per run (and retention — this matters for old runs)

Pinpointing where a metric regressed / went flaky

1. Bisect across dates. Pull chat-perf-summary/ci-summary.md from several runs spanning the window. Compare the test median (and ±stddev) for the suspect metric+scenario run-to-run — the day it jumps is when main changed. A metric that goes bimodal (e.g. a raw-run column showing two clusters like ~250 / ~900) is the flaky signature; a stable shift is a genuine regression.
2. Confirm it's real vs. measurement noise. Check the per-run raw tables (in ci-summary.md) — high ±stddev / bimodal values mean the median is being pulled around by a few outlier runs, not a uniform slowdown.
3. Deep-dive the cause. Download perf-results-<group> (has the trace.json per run) for a slow run and inspect what dominates the slow window. Trick that found the gc_stats artifact: sum main-thread RunTask durations between two code/chat/* marks (e.g. willCollectInstructions → didCollectInstructions) — if the window is ~0% busy, it's an async wait or a GC pause, not real work; then look at the largest X-phase events in that window (MajorGC, Layout, etc.).
4. Reproduce a suspect commit locally to bisect precisely: npm run perf:chat -- --build <sha> --baseline-build <ver> --runs 7 (or two commits directly via --build <shaA> --baseline-build <shaB>).

Tip: perf-summary-* (the machine-readable results.json with rawRuns) is deleted after 1 day, so for older runs rely on chat-perf-summary (raw tables, 30 days) or extract results.json from perf-results-* (also 30 days).

Architecture

Mock server

The mock LLM server (common/mock-llm-server.js) implements the full CAPI URL structure from @vscode/copilot-api's DomainService:

• GET /models — returns model metadata
• POST /models/session — returns AutoModeAPIResponse with available_models and session_token
• POST /models/session/intent — model router
• POST /chat/completions — SSE streaming response matching the scenario
• Agent, session, telemetry, and token endpoints

The copilot extension connects to this server via IS_SCENARIO_AUTOMATION=1 mode with overrideCapiUrl and overrideProxyUrl settings. The vscode-api-tests extension must be disabled (--disable-extension=vscode.vscode-api-tests) because it contributes a duplicate copilot vendor that blocks the real extension's language model provider registration.

Adding a scenario

1. Add a new entry to the appropriate object (CONTENT_SCENARIOS, TOOL_CALL_SCENARIOS, or MULTI_TURN_SCENARIOS) in common/perf-scenarios.js using the ScenarioBuilder API from common/mock-llm-server.js
2. The scenario is auto-registered by registerPerfScenarios() — no manual ID list to update
3. Run: npm run perf:chat -- --scenario your-new-scenario --runs 1 --no-baseline --verbose

Related skills

• heap-snapshot-analysis — When a perf regression or leak check identifies high memory growth, use the heap-snapshot-analysis skill to dig deeper. It can parse .heapsnapshot files, compare before/after snapshots, group object deltas, and trace retainer paths to find what keeps disposed objects alive. The chat-perf leak check measures overall heap slope; heap-snapshot-analysis finds the specific objects responsible.
• auto-perf-optimize — For launching VS Code, driving a scenario, and capturing heap snapshots or CPU profiles automatically before doing low-level analysis.*