walkthru

Turn a sequence of application commands into an editable, re-renderable demo/tour artifact — and play that sequence while observers record video, draw visual cues, and narrate.

walkthru owns the representation (the Demo Document) and the playback/capture engine. It does not render the final video — it hands a validated artifact off to a renderer (the reelee ecosystem, Remotion, moviepy/ffmpeg, …). Owning representation, not pixels, is the load-bearing boundary of the whole design.

Two modes, one data model, one engine

• Generative — an author supplies a Demo Document (commands + annotations); walkthru plays it while recording, with optional cues (highlight, spotlight, hotspot, callout, synthetic cursor), pauses, and narration.
• Capture — a human operates the app manually; walkthru records the video and the underlying command stream and annotations, producing the same Demo Document.

The only difference between the modes is who fills in the document.

The core

play(demoDoc, executor, observers) -> Outcome

A pure higher-order function that walks the document and emits lifecycle events (onStepEnter → beforeCommand → afterCommand → onStepExit, onCueBegin/End, onNarration, …). It never records, renders, or speaks — every effect is an injected observer or port.

Quickstart

pip install walkthru

import asyncio
from walkthru import DemoDocument, Section, CommandStep, Command, Timing, play

doc = DemoDocument(
    id="demo",
    sections=[
        Section(
            id="s1",
            steps=[
                CommandStep(
                    id="step-1",
                    command=Command(id="app.open"),
                    timing=Timing(duration_ms=500),
                ),
                CommandStep(
                    id="step-2",
                    command=Command(id="app.save"),
                    timing=Timing(duration_ms=800),
                ),
            ],
        )
    ],
)


async def executor(command):  # your app's command bus (acture, Playwright, …)
    print("run", command.id)
    return {"ok": True}


asyncio.run(play(doc, executor))  # → runs step-1, step-2; returns an Outcome

Three runnable scripts in examples/ go further — a generative demo (commands + cues + narration, with JSON + WebVTT hand-off), a capture demo (record a human's actions into the same document and replay it), and a narration demo (segment → time → caption). All are pure-core, no optional deps:

python examples/generative_demo.py
python examples/capture_demo.py
python examples/narration_demo.py

Narration & captions

Narration is a track of editable text anchored to steps (the Descript model: text is the source of truth; audio and timing are regenerable). Turning authored text into timed, voiced narration with captions is a three-step, fully composable pipeline:

from walkthru import realize_narration, pace_steps_to_narration, resolve_timeline
from walkthru.adapters.export import narration_to_webvtt, narration_to_srt
from walkthru.adapters.gif import GifRenderTarget  # screencast + camera -> GIF
from walkthru.adapters.playwright import (
    PlaywrightCommandPlayer,  # run a document's commands against a real page
    install_synthetic_cursor,  # a visible pointer: screencasts do not record one
)
from walkthru.adapters.synth import MixingSynthesizer, mixing_duration_ms

# 1. Synthesize each segment and time it from its own audio (anchor.duration_ms ← measured clip).
synth = MixingSynthesizer(
    voice_query="narrative_story", out_dir="tts"
)  # [synth] extra → ElevenLabs
realized = await realize_narration(doc, synth=synth, measure_ms=mixing_duration_ms)

# 2. (optional) Hold each beat at least as long as its line is spoken — narration-led pacing.
paced = pace_steps_to_narration(realized.document, policy="max")

# 3. Captions fall straight out of the resolved timeline — WebVTT (web-native) and SRT (universal).
open("demo.vtt", "w").write(narration_to_webvtt(paced))
open("demo.srt", "w").write(narration_to_srt(paced))

Because each segment is synthesized and measured independently, the beat boundaries fall out of resolve_timeline with no alignment math. Synthesis is the Synthesizer port and the audio assembler is an injected seam, so the whole flow is testable with fakes and pulls in no media stack; MixingSynthesizer (the [synth] extra) is the hosted-voice tier. See examples/narration_demo.py for an offline run (no API key) and DECISIONS.md §D8/§D12.

Ecosystem-biased, ecosystem-independent

The core and ports depend on nothing from our ecosystem. Integration with acture (the command layer), reelee (the renderer), and zodal ships as optional adapters. You can pip install walkthru / npm i acture-walkthru and use the core with your own adapters.

Packages

Status

The Python side is implemented: the Demo Document schema (the SSOT), the pure play/capture engine, dependency-free JSON + WebVTT + SRT export, segmented narration (synthesize → time → pace → assemble), and the first adapters (Playwright locator/recorder, the mixing/ElevenLabs voice synthesizer, and the reelee Ken Burns render target). The TypeScript package currently ships the schema seam — a Zod validator codegened from the Pydantic SSOT — and its live capture/play engine over acture lands next.

See PLAN.md for the implementation plan, DECISIONS.md for design decisions and deviations from the brief, and the repo's enhancement issues for the running development journal. Source documents live in misc/docs/.

License

MIT — see LICENSE.