Live demos

These six recordings exercise the real mklang CLI and Textual console. All but the deterministic test demo run against DeepSeek; the search and agent demos also hit the live web through the host search tool. They are generated from versioned VHS tapes, not hand-edited terminal captures.

CLI: check, lint, run

The CLI recording checks and lints a machine before running it with a live provider. Read the terminal transcript.

Console: interactive run

The console recording starts a clean workspace, lists the commissionable machines with /machines, runs std_self_consistency — a five-sample fan-out with majority vote — on the classic "9.11 vs 9.9" trap, and inspects the session with /session. Read the terminal transcript.

Agent: natural-language commissioning

The agent recording is a free-language, multi-turn session: the first request makes the agent brain commission news_search (live web via the host search tool) and report a sourced brief; the follow-up turn chains on that context to distill a one-line takeaway — no slash commands involved. Read the terminal transcript.

HITL: suspend and resume

The HITL recording runs expense_approval with --hitl: an escalate gate suspends the run and writes a checkpoint, then mklang resume injects the manager's reply and completes it. Read the terminal transcript.

Search: chained states with live web

The search recording runs news_search end to end: the machine plans a query, calls the host search tool (Tavily), judges whether the notes are usable, and finalizes a sourced brief — a chained multi-state flow with real side effects. Read the terminal transcript.

Tests: deterministic scenarios

The test recording checks triage.mk, then runs its scripted scenarios with mklang test — the LLM, judges, and tools are all scripted, so this command needs no provider or API key. Read the terminal transcript.

Reproducibility and review

The canonical sources are demos/tapes/ and scripts/demo_assets.py. The pinned toolchain installs and verifies JetBrains Mono before rendering. The script renders WebM recordings, derives compact GIF previews, validates dimensions, duration, size, transcripts, and secret leakage, then records exact source, toolchain, and asset hashes in manifest.json.

Regeneration is intentionally manual through the Demo assets GitHub Actions workflow because it performs live provider calls. The workflow opens or updates a review PR; automated checks catch source drift, while a human reviewer confirms readability, pacing, accuracy, and the absence of sensitive information.