Live demos

Two recordings for the two product surfaces — the console and the language. Both run the real surfaces against DeepSeek; the agent demo also hits the live web through the host search tool. They are generated from versioned VHS tapes, not hand-edited terminal captures.

Agent: natural-language commissioning

The agent recording is a free-language, multi-turn session:

1. Raise the session cost budget (/budget) so the live path has room.
2. Ask in natural language to RUN news_search for this week's open-source AI models with a short linked brief — the brain commissions the workspace machine and requests host capability consent for search (or a turn-budget continue) once; the tape answers always yes so further session confirms stay unattended.
3. After the sourced brief, chain a follow-up turn that distills a one-line takeaway — no slash commands involved for the work itself.

Read the terminal transcript.

Language: gates, tools, and the reasoning loop

The language recording walks the recommended CLI path on react.mkl:

1. mklang check — schema + semantic OK.
2. mklang lint --strict — zero findings.
3. mklang run — a real reason → act (calc host tool) → observe loop. Gates route on natural-language conditions, the tool observation re-enters the context, the step budget bounds the loop, and finalize escalates to the reasoning tier. The result is the arithmetic answer (153).

Read the terminal transcript.

More flows — the console's stdlib fan-out, over:/call: composition, HITL suspend/resume, and keyless scenario tests — are covered in the guides and the runnable examples/.

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 runs through the Demo assets GitHub Actions workflow, on the first of each month and on demand, because it performs live provider calls. It skips the render entirely when nothing is stale. The workflow opens or updates a review PR, where a human reviewer confirms readability, pacing, accuracy, and the absence of sensitive information.

Two of those hashes are checked differently, and the difference is deliberate:

• Asset hashes block. demo_assets.py check-drift fails if a published recording does not match the manifest — the guarantee that nothing here was edited by hand.
• Source hashes do not. A moved source means these recordings might be out of date; only a person or a re-render can say whether they are. check-drift reports it and passes, demo_assets.py staleness answers the same question for the scheduled workflow, and demo_assets.py manifest re-pins offline when the answer is "these are still accurate".

What replaces the blocking source check is a claim the schedule can keep: the recordings are never older than MAX_AGE_DAYS. The manifest's generated_at moves only when an asset actually changes, so re-pinning cannot quietly reset that clock.