learnctl 0.32.8

Consumer CLI for the learnctl catalog — install and study hands-on courses.

learnctl

A rich Textual TUI host for the learning-coach skill. learnctl reads a Package the skill generates, calls coach.py for all deterministic state, runs declarative lab checks on the host, and walks the learner through the protocol turn-by-turn with a live terminal UI.

Built on the principle the skill enforces: the LLM is a build-time author, not a run-time dependency. v1 runs the entire free path with no model wired up — discover, resume, active-recall warm-up, prerequisite check, primer reading, hands-on task, declarative verifier, hint ladder, solution, scoring, advance. The two metered LLM turns (interview and the opt-in diagnose) are defined slots that v2 will fill via a pluggable provider.

Install

pip install -e .
export LEARNCTL_COACH=/path/to/learning-coach/scripts/coach.py

Commands

learnctl discover              # list topics under ./learning + progress
learnctl status <topic>        # one-shot dashboard (no TUI)
learnctl study   <topic>       # launch the TUI session

A <topic> is a slug (docker), a folder path (learning/docker), or a state.json path. --root <path> overrides the workspace root for slug lookup.

Keyboard shortcuts (TUI)

The control bar is docked to the bottom of the screen so it can never scroll out of view, and every action has a keyboard shortcut as a fallback:

Key	Action	Available when
r	Reveal answer / Run check	recall reveal, task step, verify retry
h	Show next hint	verify failed and rungs remain
s	Show solution	verify failed
g / f / m	Got it / Fuzzy / Missed (self-rate recall)	after recall reveal
c	Continue / Dismiss	most other turns
1 / 2 / 3	Score 60 / 75 / 90 (manual interview)	after the interview
q	Quit	always

Architecture

Five layers, bottom-up:

Layer	File	Responsibility
engine	engine.py	Typed subprocess wrapper around coach.py. The only layer that touches state. Parses the ---JSON--- block from each command into typed results. Subprocess (not import) on purpose — validates the host contract any language could implement.
package	package.py	Read-only loader for path.json + scenario.json + Markdown. Pure file I/O, no logic.
verifier	verifier.py	Executes the declarative verify.checks on the host: command and script checks run via subprocess with timeouts; file checks read the filesystem. Evaluates expect (contains/regex/equals/exit_code/json with path matching, including json-lines). Honors per-check os restrictions. v1 runs directly on the host; container isolation is a future opt-in.
session	session.py	The protocol state machine. Walks WARMUP → PREREQS → PRIMER → TASK → VERIFY → INTERVIEW → SCORE → DONE, emitting Turn objects with type, say, data, expects. The active-recall trio (recall_prompt → recall_reveal → recall_grade) is enforced here — the answer is withheld from the prompt envelope by construction, so no host can leak it early. Host-agnostic and unit-testable without a terminal.
tui	tui.py	Textual app. Renders say as Markdown in a scrolling transcript, shows the right control buttons for each expects, streams hint rungs into a log, and resumes the session with the chosen response.

Plus __main__.py for the CLI entry point. The session has a stub for a future provider layer (interview / diagnose) — has_llm=False in v1 keeps those turns manual.

Tests

LEARNCTL_COACH=/path/to/coach.py python3 tests/test_foundation.py
LEARNCTL_COACH=/path/to/coach.py python3 tests/test_session.py
LEARNCTL_COACH=/path/to/coach.py python3 tests/test_tui.py

• test_foundation.py — 8 unit tests for the engine, package, and verifier layers including the expect matchers and the Docker-style json-lines path.
• test_session.py — drives the full free-path session end-to-end and asserts the turn sequence.
• test_tui.py — runs the Textual app headlessly via app.run_test(), clicking every button along the path and asserting the final mastery score.

All three currently pass against the real coach.py from the skill.

What's intentionally not in v1

• LLM provider. The interview turn collects a score manually (score buttons in the TUI); the diagnose button falls back to revealing the solution. v2 will add an interface for interview() / diagnose() / ask() with a null implementation for the offline case and concrete providers (Anthropic, etc.) for hosts that wire one up.
• Container sandbox. The verifier executes on the host with per-check timeouts; v2 can opt-in to a Docker-based sandbox. The skill's authoring safety contract (scoped, non-destructive, transparent, reversible) is the current safety boundary.
• Generation commands. learnctl consumes packages; generating them is the skill's job. The CLI doesn't wrap init/add-scenario/add-item because those are author-side commands a chat agent or another tool runs.

Cost model

Zero LLM calls per session in v1 — the entire free path is deterministic and free. v2's metered turns (interview, optional diagnose) match the skill's design: ~1 model call per stage plus opt-in diagnoses when novel-stuck.