Repository navigation
Description
Summary
Add one production readEvents scaling workload to the existing performance runner. The current built-in configLoad, charterCompile, routing, model-selection and export/import cases largely simulate JSON/string/regex work rather than invoke the corresponding product implementations.
Proposed primary owner: unit-contract-tester.
Evidence and scope
Audited revision: 47082a81799710c3e457b594123d190c00596ebf, observed 2026-11-12.
- Built-in benchmark cases were exercised for three baseline runs of 1,000 iterations per case. Their output is not evidence of equivalent production-path performance.
- Existing entry points include
scripts/run-benchmarks.mjsandscripts/measure-cold-start.mjs. No workflow invocation of those performance runners was found at the audited revision. Smoke/output-shape tests do not demonstrate sensitivity to realistic product regressions. - The local packaged build could not run: TypeScript/Vitest were unavailable and locked dependency restoration returned a configured-feed E404 for
hono-4.13.17.tgz. Missingdistalso blocked the built benchmark and cold-start runners. This is a measurement limitation, not evidence that the CLI is slow.
The audit's fallback used Node's built-in TypeScript transformation and isolated relative-import resolution. Those function-level diagnostics exclude import/setup time and must not certify packaged startup, application throughput, total allocations, token efficiency, or whole-agent-session memory.
Scope
Parent: github/squad#2481.
One production API only: LocalSquadBackend.readEvents with the local file-state backend. Add a fixture/case to the existing runner foundation; do not build a new benchmark framework. Packaged help/version startup is split into github/squad#2482.
Out of scope: config, charter, routing, model selection, memory search, shell persistence, backend optimization, CI gates, nightly automation and cross-platform rollout. The optimization belongs to github/squad#2479.
Measurable acceptance and tests
- Run the built production method, not a JSON/string simulation or source-loader shim. Fixture sizes: 100/1,000/10,000 valid events;
limit: 1with one-event-tail and empty-tail cursors, six cases total. - For each case, collect three rounds of ten timed observations after two untimed warmups. Exclude fixture/import setup and retain all 30 raw times plus p50/p95, exact SHA, Node/platform/build mode and cache policy.
- Assert event count, final sequence, cursor and resumability in the fixtures. A deliberately broken result must fail the run rather than produce a successful timing.
- Deterministic output tests cover sample counts/quantiles, nonfinite or missing data, fixture failure and a nonzero runner exit. Existing runner consumers/output contracts remain compatible.
- Exercise the comparison/reporting path with a calibrated test-only slowdown of at least 2x; its reported median ratio must exceed the declared comparison threshold. This proves regression sensitivity, not production optimization benefit; do not add a blocking CI threshold in this issue.
- Check in the portable fixture and document one exact build/run command with a passing sample artifact. Missing dependencies/dist remain a failing or blocked measurement, never a successful substitute.
Done means one trustworthy six-case production benchmark and its runner/result tests, not coverage of every product path.
Related work and non-goals
Follow-up to the already-merged runner foundation github/squad#1111, not another request to add it. github/squad#2449 is a distinct model-quality paired A/B workload.
Methodology: dotnet/aspnetcore#70219, microsoft/waza#846, and the packaged-path correction in microsoft/aspire#16704. No product optimization or end-to-end speedup has been demonstrated by this issue.
Activity
[-]perf: benchmark real production paths and retain regression evidence[/-][+]perf(bench): add a production readEvents scaling benchmark[/+]