Free keyword prefilter to cut per-delegation router classifier cost
@JAORMX is already working on this.
Since Sep 7, 2026.
- Dominant language
- Go
- Stars
- 152
- Forks
- 16
- Avg merge
- 14h 48m
- Merged PRs (30d)
- 536
Description
Documented follow-up from PR #87 (https://github.com/stacklok/mecatl/pull/87) — not in scope for the model-selection slice.
Current state
The subagent model router (ADR 0031) classifies every delegation through a tool-less one-turn classifier on the router slot. Every plain delegation under a routed session pays a full classifier round-trip before the child runs — even when the task prompt obviously matches a category (e.g. a one-word "fix the typo" → small).
Work
Add a free keyword prefilter that short-circuits the classifier call for unambiguous task prompts, cutting per-delegation classifier cost. Keep it operator-gated (--subagent-model-router empty = OFF, byte-identical when unconfigured) and fail-soft: on any ambiguity, fall through to the full classifier. Decide-once semantics preserved.
Related: the cumulative classifier spend is itself unbounded by --max-run-tokens today — see #92 for that brake fix.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.