CLI reference

Run kno <command> --help for the complete flag surface. kno doctor prints which providers, models, and goals this build supports and contacts nothing.

Command	What it does
kno init	Interactive wizard: writes a kno.yaml with your defaults
kno baseline	Run the agent over dev cases, score against the goal, persist every result; seals the holdout
kno mine	Turn production transcripts into an eval set (weak labels marked derived)
kno value	Inject each asset into the slices it could affect, re-measure against fresh controls, record Δ with a 95% CI
kno select	Build a portfolio under budget with a rejection log; Bonferroni-corrected decisions
kno export	Render selected assets into context / knowledge_base / tuning_set
kno report	Compose recorded stages into one page, verdicts, portfolio, gaps, caveats
kno doctor	Print supported providers, models, and goals (contacts nothing)
kno purge	Delete trace content; scores and costs remain, runs stay resumable

Configuration precedence

flag > env > kno.yaml > default. KNO_* environment variables mirror file settings; secrets never belong in kno.yaml, keys come from the environment or the OS keychain only.

Frequently used flags

Flag	Meaning
--evals <path>	Eval set: bare path = JSONL; langsmith:<name> / langfuse:<name> / braintrust:<name> / hf:<org>/<name>/<config>/<split>
--pool <path>	Candidate assets: bare path = JSONL; csv:<file>; md:<file-or-dir>; hf:<org>/<name>/<config>/<split>:<kind>
--agent <ref>	fake: (default), openai:<model>, anthropic:<model>, bedrock:<model-id>, vertex:<model-id>, exec:<command>
--base-url <url>	Any OpenAI-compatible endpoint
--max-cost-usd <n>	Hard budget cap; 0 = unlimited
--yes	Consent to the printed estimate (skips the interactive prompt)
--resume	Skip completed work, reconstruct spend, continue
--goal <name>	Scoring goal; default exact-match
--db <path>	SQLite state file; default kno.db in the working directory
--key-env host=VAR	Read a provider key from a named environment variable