Performance

anneal has a per-run timing tracer, exposed as /perf and written into the session. The source note, docs/PERFORMANCE.md, includes the baseline and the comparison harness against opencode.

What is measured

• Time to first token and total run time.
• Tool call latency, by tool.
• Compaction and summary costs.
• Prompt assembly and memory injection.
• The improvement loop: propose, probe and compose per candidate.

The probe records steps, failures and token counts, because “better” has to mean something measurable. A candidate with no measurable gain is skipped and the reason is recorded in the run.

Reproducing

The harness in testing/perf/ runs the same task against anneal and opencode, with a local model, and compares the traces. It is intentionally boring: a fixed task, a fixed model, a fixed machine, and the raw numbers saved.

cd testing/perf
./ab.sh

Do not trust a marketing claim that is not in this note. If the number changes, the note changes first.