Routing Tuning

Use this skill to make free-text routing reliable and evidence-backed.

Procedure

1. Inspect the target story's room and intent definitions.

2. Write 10-15 cases under stories/<story>/intents/.

   • Include target-route positives.
   • Include near-miss negatives that should route elsewhere.
   • Include slots for text-capturing intents.

3. Run kitsoki test intents in a cheap mode first:

   go run ./cmd/kitsoki test intents stories/dev-story/app.yaml \
     --intents stories/dev-story/intents/landing_proposal_routing.yaml \
     --dry-run

4. If live model tuning is explicitly requested, run one pass against the target profile:

   go run ./cmd/kitsoki test intents stories/dev-story/app.yaml \
     --intents stories/dev-story/intents/landing_proposal_routing.yaml \
     --harness claude \
     --profile codex-native \
     --runs 1 \
     --max-cost 5 \
     --json .artifacts/<topic>/routing.json \
     --emit-recording .artifacts/<topic>/routing.recording.yaml

5. Summarize with:

   python3 tools/routing-tuning/intent_report_summary.py .artifacts/<topic>/routing.json

6. Tune story intent examples/descriptions/priorities. Do not tune prompts around a bad story boundary.

7. Re-run. Use --runs 3 only for a final confidence pass after --runs 1 passes.

Session Mining

Use session mining to source future cases:

• phrases typed at hubs or broad conversational rooms;
• user corrections after a wrong route;
• broad work turns whose result shows they should have entered a specialized room;
• specialized-room entries that should have stayed in work.

Then convert mined phrases into intent fixtures and cite the source transcript or trace when helpful.

See docs/testing/routing-tuning.md for the full runbook.