# Evidence Router

> Use when choosing an unfamiliar external evidence source and project source or references are insufficient — cold capability map for MCPs, references, and web.

- Skill: `ryan-brosas/evidence-router` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add ryan-brosas/evidence-router`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ryan-brosas/evidence-router/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: ryan-brosas (https://skillmd.com/u/ryan-brosas)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ryan-brosas/evidence-router

---


# Evidence capability map

Cold reference. **Evidence priority:** current project source/tests → project
`reference/` / `reference/web/` when relevant → one applicable cold-foundation
capsule or operational skill on a named gap → external source/docs/web only for
that gap. No fixed tool chain.

| Need | Capability |
|---|---|
| Active or owned repo orientation / neighborhoods | Direct source + Fovea (`fovea_sketch` → `fovea_focus` → `fovea_dwell`) |
| IDE-aware types / semantic refactor | `ide` profile: MCP Steroid / JetBrains (`steroid_*`) + source |
| Runtime behavior | tests, debugger, runtime output |
| Named inspiration repo | project `reference/<repo>/` + source/tests |
| Live site capture | `reference/web/<site>/` or `web-reference` |
| Stack patterns from prior implementations | `skills/*-foundation` + cited source |
| Code graph / cross-repo pattern index | `code-graph` profile: Codebase Memory → verify candidate in source |
| Past attempts / lessons | Project-scoped session history (`/recall-session`); OpenViking only as an optional rebuildable cache |
| Current library docs | official docs / Context7 |
| External facts / advisories | Exa or read-only fetch |
| GitHub repo overview | DeepWiki → repository |

Name one uncertainty, select its primary source, and escalate only when that
source is insufficient. Model output is not evidence until verified against
source/tests/runtime. MCP capabilities are selected per task from
`../../mcp/profiles.json`; each profile selects at most one server and `minimal`
connects none. For execution or model escalation, see `execution-router` and
`model-resolution`.

## References

- `references/openviking.md`, retrieval, capture, and degraded-path boundaries