Skip to content

v1.1.0 — first release

Choose a tag to compare

View all tags
@kadj-amoah kadj-amoah released this 23 May 10:42
v1.1.0
63e84f1

Changelog

All notable changes to Showrunner are documented here. Format follows Keep a Changelog; the project tracks loose semver — minor bumps for new capability, patch for fixes.

1.1.0 — 2026-05-23

First tagged release. Versioned 1.1.0 rather than 0.1.0 because the project has been tracked as "v1.1 — reliability hardening + provider-agnostic refactor" throughout development; this is the first cut where the pipeline works end-to-end against real Next.js targets and the LLM + TTS layers are swappable. No 1.0.0 was ever published.

Added

  • Provider-agnostic LLM layer (llm.default + per-stage llm.overrides). Four providers:
    • anthropic — Claude via the official SDK with structured outputs.
    • openai — gpt-4o-style models via response_format: json_schema with json_object fallback.
    • agent_bridge — spawn a headless CLI agent like claude -p --output-format json; no API key on file required. Supports spawn and file_poll modes.
    • custom — dynamic-import an operator-supplied module conforming to the LLMProvider interface.
  • Provider-agnostic TTS layer (voiceover.provider). Three providers:
    • elevenlabs — the only one that returns per-character alignment.
    • openai — tts-1-hd via the audio API.
    • custom — dynamic-import for in-house TTS pipelines.
  • alignment_strategy: required | best_effort — when a TTS provider doesn't return alignment, switch automatically to per-segment synthesis (one TTS call per segment) so audio doesn't get sliced across mid-word boundaries.
  • showrunner doctor — preflight checks before any run. Validates config syntax, provider env vars, ffmpeg + ffprobe on PATH, Playwright chromium binary, target URL reachability, free disk on every output dir, free memory + computed ffmpeg thread cap, and lifecycle script executability. Wired into run implicitly (escape with --skip-doctor).
  • showrunner record-actions with scroll capture. Drives a headed browser; captures click / input / change / keydown / submit / scroll events; coalesces them into a clean action stream. Scroll capture is 250 ms debounced with direction coalescing.
  • DOM preflight in the script stage — scrapes the live target's actionable selector inventory (≈60–80 entries on a typical marketing site) before asking the LLM for a manifest. The LLM is constrained to inventory selectors only, with a one-shot remediation retry if it strays.
  • Resource preflight in record and mux stages — checks free disk against estimated artifact sizes, computes an ffmpeg -threads cap from free RAM × resolution × quality preset (SHOWRUNNER_FFMPEG_THREADS env override available).
  • Structured ffmpeg errors — categorized as oom | no_space | codec_missing | permission | unknown, with per-category remediation hints.
  • --resolution flag on both init and run. Presets: low (854×480), standard (720p, the default), high (1080p), extreme (4K). On run it overrides both recording.viewport and output.resolution in one step so they stay matched and mux doesn't upscale.
  • init flags: --llm-provider, --tts-provider, --resolution. Scaffolds a demo.yaml with the right provider blocks, an .env.example listing only the env vars that combination actually needs (deduplicated when LLM + TTS both use OpenAI), and a real docs/PRD.md stub with section guides.
  • Legacy config auto-migration — pre-v0.1 demo.yaml files with flat voiceover.{voice_id, model, ...} fields and no llm block are normalized at load time. Existing configs keep working with zero edits.
  • validate --strict — exit nonzero on missing provider env vars (otherwise warns).
  • Captions — SRT + VTT emitted alongside the final MP4 when output.captions.enabled is set. Whole-segment cues are used as a fallback when per-character alignment isn't available.

Changed

  • engines.node bumped to >=20.6 (needed for process.loadEnvFile).
  • init defaults: script.vo_review_gate: false (was true), recording.viewport: 1280×720 (was 1920×1080), output.resolution: 1280x720. Each is opt-in to the larger/stricter value via a one-line edit.
  • tsup configured with skipNodeModulesBundle: true and target: 'node20'. Built dist/cli.js dropped from 1.75 MB to 247 KB and no longer crashes with Dynamic require of "tty" is not supported — npm link and npm i -g are viable now.
  • README rewritten for v0.1 — install path (with npm link), full prereqs (Node ≥ 20.6, ffmpeg, chromium), provider-choice tables, pipeline-per-stage table, daily-commands cookbook, troubleshooting section.

Fixed

  • understand --interactive no longer silently exits after the second question on piped stdin. Replaced readline/promises.question() with a queued 'line'-event reader that handles both TTY and pipe semantics.
  • understand --output <path> now correctly takes precedence over project.product_model in the config (was silently overridden).
  • Script-stage prompt now bans Tailwind arbitrary-value class names (.max-w-[1280px]) which aren't legal CSS and crashed Playwright with a SyntaxError. A post-validator catches violations and retries once with a remediation prompt before accepting the manifest.
  • looksUnstable() heuristic now catches Next.js compiled-CSS classes like __variable_3eb911 (leading underscores were previously slipping past).
  • mux output write now falls back to a timestamped sibling MP4 when the canonical path is locked by a media player, with a warning, instead of failing the run.

Known limitations (Tier 3 — coded but not exercised end-to-end)

  • OpenAI LLM + OpenAI TTS provider paths are wired but were not run with a real API key in v0.1 validation. The agent_bridge + ElevenLabs combination is the validated default.
  • agent_bridge file_poll mode is coded but only spawn mode was exercised.
  • Custom provider modules (LLM + TTS) — dynamic-import loader is wired but no reference implementation has been swapped in.
  • instrument, capture-auth, trace, preview, rerun-segment, print-vo, approve-vo commands are mostly thin wrappers but were not stress-tested in v0.1.
  • Background music mix and title-card logo / custom-font rendering are coded but were not exercised.
  • 1920×1080 mux remains RAM-sensitive on tight boxes; the thread-cap heuristic mitigates but the safe default is --resolution standard.

Validation artifact

A real end-to-end run against the Credstone Atlas marketing site produced output/ct-website-atlas_demo.mp4 (4.3 MB, H.264 + AAC, 1280×720, 48.2 s, with SRT + VTT) using:

  • agent_bridge LLM (via local claude CLI, no Anthropic key on file)
  • ElevenLabs TTS with alignment_strategy: required
  • DOM preflight + selector validator producing inventory-only manifest selectors
  • Full mux preflight + thread cap on a 5.9 GB RAM machine with 1.1 GB free at run time
Assets 2
Loading