How shimo4228/jev-skill-router works
jev-skill-router is a Claude Code prompt-hook experiment: it gathers the available skill roster, asks Jev to rank candidates, then logs the result and, in inject mode, provides a hint. Claude Code still makes the final skill choice. A separate evaluation script compares logged suggestions with session activity. The graph follows documented workflow and sampled calls; some router-to-client and plugin-hook wiring is supported by the README rather than a shown call site.
Main entry points: route.py Prompt route · router.py Skill ranking · jev_client.py Jev API client · shadow_join.py Shadow join
Components 13
Hook workflow
- Hook registration
hooks/hooks.json - Prompt route
scripts/route.py - Installed skills
- Roster builder
scripts/roster.py - Prompt hint
scripts/route.py
Jev judgment
- Skill ranking
scripts/router.py - Jev API client
scripts/jev_client.py
Outcome measurement
- Decision log
scripts/decision_log.py - Shadow join
evals/shadow_join.py - Skill-use outcome
evals/shadow_join.py
Other components
- Claude Code
- TypeSafe Jev
- Session transcripts
Connections 15
- Claude Code → Hook registration: submits prompt
README.md - Hook registration → Prompt route: runs hook
- Prompt route → Roster builder: builds roster
scripts/route.py - Installed skills → Roster builder: scans skills
scripts/roster.py - Prompt route → Skill ranking: requests suggestion
scripts/route.py - Skill ranking → Jev API client: asks Jev
README.md - Jev API client → TypeSafe Jev: sends questions
scripts/jev_client.py - TypeSafe Jev → Jev API client: returns probabilities
scripts/jev_client.py - Skill ranking → Prompt route: returns suggestion
scripts/route.py - Prompt route → Decision log: builds and appends row
scripts/route.py - Prompt route → Prompt hint: formats inject hint
README.md - Prompt hint → Claude Code: adds hint
README.md - Decision log → Shadow join: reads decisions
README.md - Session transcripts → Shadow join: reads skill calls
evals/shadow_join.py - Shadow join → Skill-use outcome: joins and counts
evals/shadow_join.py