Repository navigation
v1.38.1
Maintenance release since 1.38.0: a determinism fix in the benchmark harness, a retry for the flaky setup-bun CI step, and a broad test-coverage pass across the MCP servers and benchmark artifacts.
Fixed
scripts/run-benchmark.ts:bootstrapCIsorts a copy of its input ratios before seeding the mulberry32 resampler, so the reported 95% confidence interval no longer depends on measurement order.aggregateis exported for direct testing.- CI
fast checks:setup-bunretries once when thebun-version: latestresolve fails transiently, instead of failing the run outright (previously took down a main-push run).
Added
scripts/bin-shebang.node-test.mjs: anode --testsuite asserting every packagebinentry exists and uses the#!/usr/bin/env nodeshebang.- Test coverage, roughly 2,100 lines across 18 files:
ashlr__search_replace_regexbyte-exactness for literal CR/CRLF, Unicode and$$escapes through the direct call, the pureapplyRegexpath and the tool handler; benchmark v2 JSON validity, docs/site artifact sync, site entries and shared-values parity; malformed v2 manifest handling treated as absent; lazy-routerinitialize/tools/listalignment with the generated catalog; tool-catalog name uniqueness, non-emptiness and MCP-safety; pipe grep preserving genuine no-match results; orchestrate stub and LLM executors through DI seams; gracefulJSON.parsefailure surfaces on untrusted input; summarize-marker warning comments and deprecations; token measurement on empty input; and auto-tag/CI reconciliation invariants.