Tools, commands, and artifacts

CLI commands and shell workflows

The keating executable handles setup, runtime launch, project operations, and artifact generation. Inside the interactive shell, slash commands expose teaching workflows and session operations. The command registry explicitly marks shell-only and CLI-only entries, so the two interfaces are related but not identical.

keating setup
keating doctor
keating shell

# Within the interactive shell:
/diagnose fractions
/learn fractions
/quiz fractions
/outputs

Produce teaching artifacts

Outputs are project-scoped and can be inspected independently of the conversation. The authoritative directory helpers place plans, maps, verifications, quizzes, and animations below .keating/outputs, or the selected learner profile's corresponding root.

• plan <topic>: generate a deterministic lesson plan.
• map <topic>: generate an editable Mermaid concept map.
• verify <topic>: generate a fact-checking checklist to complete before teaching.
• quiz <topic>: generate retrieval practice material.
• animate <topic>: generate a Hyperframes animation bundle; this path includes model-backed generation.

keating plan derivative
keating map derivative
keating verify derivative
keating quiz derivative

Review the learner and the run

These commands expose different evidence. Engagement, subjective feedback, and quiz performance should remain distinguishable when you build downstream reports.

• feedback <up|down|confused> [topic] records a feedback signal, optionally with a comment.
• learner-state shows the learner profile and session history.
• timeline shows engagement across topics; due shows topics due for review.
• policy shows the active teaching policy.
• trace [query] browses debugging traces and artifacts; /outputs browses artifacts from the shell.

Export data deliberately

The CLI can export fine-tuning data and import supported fine-tune JSONL as session data. Export produces material for downstream workflows; it does not itself start model training or publish learner data. Review what is included before sharing it.

keating export --finetune --source=all --format=both

Select an execution mode explicitly

keating web serves the production Nitro build. Browser-only mode runs supported agent work on the learner's device. Host mode is opt-in, project-root-confined trusted-machine execution; it is not a security sandbox. Remote mode connects to a separately configured execution service. Cloud mode uses the configured cloud endpoint.

The webmcp command exposes Keating tools over Streamable HTTP. Its bind address and access boundary are separate from the normal browser application's model execution mode.

keating web --browser-only-agent 3000
keating webmcp 3001 --host=127.0.0.1

Distinguish the improvement commands

• bench summarizes historical recorded learner evidence and its limits.
• evolve writes unvalidated policy proposals without replacing the active policy.
• prompt-eval and prompt-evolve evaluate or revise prompt templates; prompt snapshots are diagnostics, not learner outcome evidence.
• teaching-bench executes training episodes without release holdout access.
• auto-improve runs a bounded skill proposal through independent validation and holdout gates.
• learning-check records revision-linked learner assessments separately from synthetic experiments.