Query-daemon repeated-read measurement — 2026-08-09

This non-gating Slice 9 report measures repeated JSON reads on the Codira checkout. It is evidence for the optional warm service, not a timing assertion.

Method

• Commit: e9347c79d20e893819a31a698ac8c0b385c64909
• Backend: SQLite; embedding engine: ONNX on CPU
• Workload: ctx, emb, plugins, and caps with query query daemon
• Samples: three direct and three warm invocations per command
• Warm process: one repository-fixed IPC daemon at generation 27
• Peak process RSS: 292,488 KiB
• hyperfine was unavailable; the committed benchmark script used perf_counter and saved raw JSON outside the repository worktree.

Results

Command	Direct mean (ms)	Warm mean (ms)	Improvement
ctx	883.373	709.320	1.25x
emb	194.581	22.407	8.68x
plugins	359.586	2.577	139.54x
caps	1007.862	31.694	31.80x

The warm-session shutdown path resets embedding runtime caches. No unbounded growth was observed within this three-run campaign; that is not a long-duration leak proof.

Decision and reproduction

The repeated-read benefit is material for embedding search, plugin diagnostics, and capabilities, and positive for context retrieval. Enable the service for repeated SQLite reads when the documented local-process memory footprint is acceptable; direct execution remains deterministic and is retried after IPC failure. DuckDB replacement/reopen behavior is covered by parameterized warm-runtime tests, but operators should repeat this campaign under their DuckDB configuration before generalizing these SQLite timings.

uv run codira index
uv run python scripts/benchmark_query_daemon.py --runs 10

The script refuses to overlap an existing endpoint and writes a timestamped artifact to .artifacts/benchmarks/campaigns/ unless --output is supplied.