Path-search performance

The unreleased 1.0.0 candidate stores one parent link per queued BFS node instead of copying its complete path history. It reconstructs only the winning path and returns zero-hop results before constructing adjacency. Direction, hop limits, edge evidence and output ordering are unchanged.

Measured comparison

Baseline: d94a8f4. Runtime: Bun 1.3.14, macOS arm64. The existing retrieval harness now includes MCP path queries and a 32,767-node four-way tree. Each case has ten warmups; the second comparison reverses implementation order. Normalization is included in path timings. Sanitized measurements and source/harness fingerprints retain p50/p95 values and matching path-payload hashes.

These cases show roughly 12–20% lower median path latency, not a universal speedup. Ordinary two-hop MCP path p50 was 24.48 → 25.07 ms initially and 24.61 → 24.63 ms in the replay: no improvement is established there. Other retrieval medians were similar; tail variation remains visible in the report. These path measurements establish no peak-memory, package-size, cold-start or end-to-end agent improvement.

Package loading

The optional type worker initially duplicated the existing parser in a second bundle. Shared chunks now reuse it; the tokenizer vocabulary is also loaded separately. This adds no runtime dependency. Builds clear generated output before bundling so stale hashed chunks cannot enter the package.

A development comparison of non-split and shared-chunk builds measured Node 26.8.1 process wall time for --version, using ten warmups and 30 samples per build in alternating order with a warm filesystem cache. Median/p95 fell from 114.10/117.07 ms to 60.23/63.57 ms; both returned 1.0.0. This is one startup case, not a general command or cold-cache speedup.

The corresponding local archives were 1,313,355 and 1,219,546 bytes, both including the type module. The shared-chunk archive SHA-256 is 66b15f8e170575b822a5647d26a047547244f4aa6291b30f55262b57a6a7955d; its clean-consumer checks passed on Node 22.12.0, 24.22.1 and 26.8.1 with Bun entry points. These are development checkpoints, not a final frozen artifact or npm publication. To repeat the loading comparison, build the same entry points with and without Bun’s --splitting into separate directories, preserve all chunks, and compare identical command output before timing.

Reproduce

Completed CLI build review

A same-source comparison at 85eb225 covers a synthetic 200-file static-import chain plus a selected-type fixture. Both builds use Bun 1.3.14 with identical entry points/minification; only --splitting differs. Node 26.8.1 commands run sequentially in alternating variant order, with three warmups and twenty measured samples per variant. Sanitized measurements, entry hashes and equivalent-result hashes retain the complete percentiles, output sizes and OS maximum-RSS observations.

Query outputs and selected type contracts/input identity match between variants. Refresh assertions require successful regeneration and correct change detection; they are not full-payload equivalence checks. Output byte sizes are unchanged. Darwin time -l maximum-RSS medians fall from about 95–117 MB to 53–77 MB on the non-type cases; type inspection remains about 98 MB. This is OS-reported maximum resident memory, not aggregate simultaneous process-tree memory or a memory guarantee. Timing includes process creation and the measurement wrapper; the OS filesystem cache is warm even for artifact-free refresh.

To reproduce, build the same revision twice into separate output directories, omitting --splitting for the baseline. Supply the matching package metadata alongside each build. Generate files src/m0.ts through src/m199.ts, each exporting value and statically importing the next file except the last. Repeat node dist/cli.js version/refresh/context/impact/path commands under /usr/bin/time -l; select file:src/m100.ts and use outgoing m100.ts → m104.ts for the four-hop case. For the type case, use an explicit strict/no-emit ES2022/Bundler project with Ready and commit(value: Ready): Ready, and the same project-installed TypeScript 7.0.2. Alternate build order, assert results before comparing times, and keep raw machine-specific logs private. These measurements establish no universal or end-to-end agent speedup.

Path-search harness

From a built development checkout with its frozen dependencies installed, use a new temporary directory. These are maintainer benchmarks, not product prerequisites.

DOTDOG_BENCH_DIR="$(mktemp -d)"
git worktree add --detach "$DOTDOG_BENCH_DIR/before" d94a8f4
ln -s "$PWD/node_modules" "$DOTDOG_BENCH_DIR/before/node_modules"
bun scripts/benchmark-retrieval.ts --tokenizer packages/dotdog/node_modules/gpt-tokenizer --implementation "$DOTDOG_BENCH_DIR/before" --nodes 32767 --samples 50 --output "$DOTDOG_BENCH_DIR/before.json"
bun scripts/benchmark-retrieval.ts --tokenizer packages/dotdog/node_modules/gpt-tokenizer --nodes 32767 --samples 50 --output "$DOTDOG_BENCH_DIR/after.json"

Repeat in reverse order and compare result hashes before timings. The harness also checks exact neighborhoods, retained evidence, changed-graph reloads and missing queries. Its old assertion that a depth-limited neighborhood was not truncated was stale; it now requires explicit truncation without changing the expected returned neighborhood.

An independent all-pairs-distance regression exercises cyclic directed graphs, incoming/outgoing/any traversal, verb filters, zero-hop cases, hop limits, path continuity and edge evidence. This supplements the synthetic timing cases; it does not prove bug freedom or optimality. Broader optimization and final release gates remain in #335 and the release plan.