kapso Files

Preflight: every verb checks what your config actually needs

research(), learn_knowledge(), evolve(), learn() and deploy() now verify their requirements before doing any work, and kapso doctor <verb> reports the same rows. One resolver serves both, so the doctor can never disagree with what a run will do.

Requirements are derived from your config rather than hardcoded:

• an all-codex config is never asked for claude
• learn_knowledge(skip_merge=True) is never asked for Neo4j or Weaviate
• learning.codify.target: local is never asked for gcloud
• learn asks for both CLIs, because its grading crew is deliberately cross-model

Every failed row names the config key that wants it and one copy-pasteable fix:

kapso evolve — preflight failed (1 of 8 requirements missing)

  [FAIL] claude CLI
         needed by  modes.GENERIC.coding_agent.model (+5 more) = claude-opus-5
         fix        npm install -g @anthropic-ai/claude-code

New CLI surface

:::bash
kapso doctor evolve            # one verb's requirements
kapso doctor learn --models    # + a one-token probe per {cli, model} pair

New config block

:::yaml
preflight:
  enabled: true            # static checks before every verb
  live_model_probe: false  # + one-token model probes; on for long unattended runs

Fixes

• kapso doctor --config mine.yaml read the packaged config for its bank rows, so it reported the wrong bank.
• Gate environment variables that Kapso injects at launch (KAPSO_BANK_*, EXPERIMENT_HISTORY_PATH, KG_INDEX_PATH) were reported as missing user setup on every clean machine.

Also in this release

• README: new headline and "What is KAPSO?", an Enterprise section, an examples/ index, KumoRFM-v2 and Leeroopedia reference links.
• benchmarks/posttrain and the dead archive/ directory are no longer in the tree; the expert-posttrain console script is withdrawn with it.