feature-testinglisted

# Feature Testing (Dogfood Verification) Libi's primary user is the **agent**. Unit tests prove logic; this skill verifies the *experience* — does the agent actually pick the right tool, follow the right skill, and produce a sensible result when the developer's change lands? ## When this skill applies Apply after editing any of: - MCP tool definitions or schemas (`mcp/tools/*.ts`, `mcp/tools/schemas.ts`, `mcp/server.ts`) - Skill content (`mcp/skills/<name>/SKILL.md`) - Agent instructions (`mcp/templates/instructions.md`, `mcp/instructions.ts`, `mcp/registry/instruction-builder.ts`) - REST or SSE routes the agent reads or writes (`app/api/*`) - Any `libi.*` behavior that the user could trigger via chat If you finish a task in any of these areas without running this flow, the work is not actually verified — only the unit tests are. ## The flow 1. **Check whether the user's dev server is in test mode** - Read `~/.libi/logs/server.log` (last ~50 lines) — look for "Loaded N sessions" AFTER a recent restart, AND for the TEST MODE banner in the agent instructions output - OR query `GET /api/settings/mcp-servers` — `fal-ai` will still be listed (it's present as the fake), but the server's spawn command will point at `mcp/dev/fake-fal/`. In test mode, `LIBI_TEST_MODE=1` is visible in the process environment and the agent's instructions contain a TEST MODE banner. You can confirm by checking `~/.libi/test-mode/fal-calls.jsonl` — it will exist (and grow) only while the fake