Uh oh!
There was an error while loading. Please reload this page.
Agent-first rework: staged toolchain, workflows + skills, hyperframes runtime, audio engine - #23
Draft
ennsharma wants to merge 19 commits into
Draft
Agent-first rework: staged toolchain, workflows + skills, hyperframes runtime, audio engine#23ennsharma wants to merge 19 commits into
ennsharma wants to merge 19 commits into
Conversation
showrunner.json carries a project's workflow/runtime/style/canvas; workflow manifest.yaml (package data) declares stages, produced artifacts, named checks, and numeric constraints.
Projects replace throwaway work dirs: showrunner.json + the runtime scaffold (remotion template, preset tokens, asset dirs). --no-install skips npm for offline/test use.
Pure validator returning findings (errors + warnings) driven by the workflow spec's constraints; storyboard validate exposes it with --json output and a nonzero exit on errors.
Provider construction moves to providers/factory.py (Pipeline delegates). tts synthesizes per-scene narration into the project, records measured durations in narration.json, persists stretched scene durations back to storyboard.json, and skips silent scenes.
Music picking (music/selection.py) and work-dir staging with the ducking envelope (music_staging.py) are now callable outside the Pipeline; compose generates Root.tsx from storyboard + measured narration durations.
Runs every check the workflow's stages declare (storyboard rules, narration coverage, scene lint + whole-project tsc, compose freshness) and writes check.json with a content fingerprint of the gated artifacts.
render refuses without a fresh passing check.json (content fingerprint must match); --force bypasses. preview dispatches to the project's runtime.
Rule Zero routing (AGENTS.md), a compact skills index, the explainer workflow run book, and storyboard/scene-code craft skills extracted from the production rules. A surface test keeps the index complete and pins skill-referenced CLI commands to real ones.
Single-file HTML compositions rendered/gated through the pinned hyperframes CLI: contract-valid skeleton at scaffold, render/preview dispatch, and the runtime's own check --json mapped into showrunner check findings.
Type-driven pieces: storyboard as beat contract, optional narration, directly-authored index.html gated by the runtime check. Adds the html-composition craft skill (timing contract + determinism rules).
librosa when installed, else a numpy comb-scoring estimator (fractional beat grids, ±1-frame tolerance, octave correction). Powers beat-synced clip timing; new [audio] extra.
Image reels cut to a music bed: music.json (track + analyzed beat grid) is a gated stage, and the check verifies every clip's data-start sits on a beat within one frame.
ffmpeg loudnorm wrapper (default -14 LUFS / -1.5 dBTP); copies the video stream untouched when mastering a rendered mp4.
README and CLAUDE.md now document the staged, gated, agent-driven flow as the primary interface (legacy one-shot mode retained). Repo-wide lint cleanup of stale imports/vars.
The real check --json nests findings under sections (lint.findings[] with severity/code/fixHint), not top-level errors[]. Surface error findings with their fix hints; warnings stay advisory. Found producing a real carousel reel.
Conditions (agent command × toolchain access) run headless on an identical brief in isolated workspaces; the harness stages the toolchain, strips it from bare conditions' PATH, collects out/final.mp4 + cost/turns, and builds a side-by-side report.html with a quality-rubric scorecard + results.json.
An inherited ANTHROPIC_API_KEY silently overrides the machine's subscription login and bills every run at per-token API prices. Child agents now get it stripped by default; conditions can opt back in with use_api_key: true.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this is
Inverts showrunner's control surface: the coding agent is now the director, and showrunner is the studio it drives. The fixed embedded pipeline is decomposed into staged, gated CLI commands; workflows are declarative contracts; craft knowledge ships as repo skills. The agent authors the storyboard and the visuals itself — no embedded LLM calls anywhere in the new path.
Phase 1 — inversion
showrunner newscaffolds a persistent project (showrunner.jsonmanifest, runtime scaffold).storyboard validate→tts(measured narration durations, stretch-back) →scene validate(design-system lint + tsc) →compose(timeline + music bed + ducking) →check→render.showrunner check: manifest-driven quality gate. Each workflow stage names its check; the report fingerprints gated artifacts andrenderrefuses missing/failed/stale checks (--forceto bypass).src/showrunner/workflows/<name>/manifest.yaml(stages, artifacts, checks, numeric constraints) — machine-read, not prose.AGENTS.md(Rule Zero),skills/INDEX.mdrouter, explainer workflow run book, storyboard + scene-code craft skills extracted from the production prompts. A surface test pins skills to real CLI commands.create/refineunchanged and green.Phase 2 — second runtime + audio
check --jsonmapped intoshowrunner checkfindings.musiccheck verifies the track, the analyzed beat grid, and that every clip start sits on a beat within one frame.showrunner music analyze: BPM + beat grid (librosa when installed, numpy comb-scoring fallback with octave correction); new[audio]extra.showrunner audio master: ffmpeg loudnorm to -14 LUFS / -1.5 dBTP, video stream copied untouched.Verification
workflows,newon both runtimes,storyboard validateon a kinetic project.🤖 Generated with Claude Code
https://claude.ai/code/session_01RLwStc1sMxYtjo77Qg6jww