Exported functions in state/lib/ffmpeg.ts. .md file to compare - side-by-side diff against ffmpeg
ffmpeg
description: "Triggers on prompt mention of 'ffmpeg'."
What it does for you
Converts, trims, and resizes your video and audio files.
What it produces
A recent result, so you can see the kind of work it returns.
loading…
How to get it
These run inside the Snappy workspace. Want this working in your business? I set skills like this up with you, in one focused week.
For developers how this skill is built, graded, and how it runs
at a glance- the short version
what's inside - the parts that make up a skill 3/4 present
A skill is just a few plain-text files. Only the main one is required. The rest are optional, added as the work needs them. This is what the skill is made of; how it runs is just below.
state/skills/ffmpeg/SKILL.md
present
state/lib/ffmpeg.ts
present
state/bin/ffmpeg/
not present
state/skills/ffmpeg/AGENTS.md
present
how it's graded - what counts as a good run 4 criteria · 3 deterministic · 1 judge
Each row is one thing a good run has to get right. deterministic means a quick check decides, pass or fail. judge means the AI reads the result and rates it. Grading each piece on its own (instead of one overall score) shows exactly where a run fell short, so the fix is obvious.
how it runs - the shared frame every skill uses 4/5 present
Every skill runs the same way. One part does the work, a separate part checks it, and a short loader hands the AI exactly what it needs for the job. Anything this skill doesn't use shows a one-line note saying why, on purpose, not by accident.
This skill doesn't fix its own gaps yet.
state/log/pending-eval.ndjson - probe() before EVERY transformation — never trim blind
- Compose surface before asking the user for missing info — use Input fields in the surface
- Embed real probe data (duration, codec, resolution, fps) into the composed surface
- StatCard for before/after file size comparison on every destructive operation
- Surface persists as artifact (lang_body) — the user can revisit and re-run
- Log to state/log/pending-eval.ndjson after each operation
what it has learned - fixes written back in over time sample
When a run hits something this skill didn't handle, the fix gets written back into the skill so it doesn't happen again. FIXED means it was corrected on the spot. LOGGED means it's queued for a bigger rewrite. Either way, the skill gets a little better and never makes the same mistake twice.
- Loading feedback rows…
how the work flows- who makes it, who checks it
import from `state/lib/ffmpeg.ts` - `probe()`, `trim()`, `concat()`, `resizeLocal()`, `compress()`, `extractAudioLocal()`, `burnSubs()`, `thumbnail()`, `loudnorm()`, `overlay()
probe()` the input first to confirm codec/duration/streams before transforming
SKILL.md- the skill, written out in plain English
ffmpeg
snappy-ffmpeg/api.ts - Local ffmpeg primitive layer.
Ported from kernel snappy-ffmpeg in Phase 0.5. See state/lib/ffmpeg.ts for the full API surface.
Steps
probe()- seestate/lib/ffmpeg.tstrim()- seestate/lib/ffmpeg.tsconcat()- seestate/lib/ffmpeg.tsresizeLocal()- seestate/lib/ffmpeg.tscompress()- seestate/lib/ffmpeg.tsextractAudioLocal()- seestate/lib/ffmpeg.tsburnSubs()- seestate/lib/ffmpeg.tsthumbnail()- seestate/lib/ffmpeg.tsloudnorm()- seestate/lib/ffmpeg.tsoverlay()- seestate/lib/ffmpeg.ts
Eval
Actor: the exported functions in state/lib/ffmpeg.ts. Auditor: none wired yet - eval is manual (Robert review). File a state/log/pending-eval.ndjson row on each run.
Score convention:
| Outcome | Score |
|---|---|
| Pass on first try | 1.0 |
| Failed first, auto-fix applied, re-check passed | 0.5 |
| Still failing or unrecoverable | 0.0 |
Gotchas
via the Phase 0.5 driver. Only these rewrites were applied: already in state/lib/)
realpathSync(process.argv[1])CLI guard wrapped in try/catch
- See the kernel SKILL.md for the original long-form guidance if you need it
(read-only reference at the kernel path above).
Graduation
This skill is prose. Graduate by defining a deterministic auditor and flipping eval: auto.
Rubric
criteria:
- name: ffmpeg_binary_present
kind: deterministic
check: "The 'ffmpeg' executable must be found in the system's PATH or a specified location."
- name: required_inputs_are_strings
kind: deterministic
check: "The inputs 'probe_input', 'trim_input', and 'concat_input' must be provided as strings to the skill."
- name: functional_api_calls
kind: judge
check: "The skill's invoked function (e.g., probe, trim) must execute successfully without errors and produce expected output for valid inputs."
- name: log_entry_created
kind: deterministic
check: "A 'state/log/pending-eval.ndjson' row must be created for each skill execution."AGENTS.md- what the AI loads when this skill comes up
ffmpeg - loader
Per-turn rules for the ffmpeg skill. Full reference: state/skills/ffmpeg/SKILL.md. Do not skip these.
Critical Rules
_(no failures recorded yet - this skill is a Phase 0.5 prose port from kernel snappy-ffmpeg. No hard-won rules have surfaced yet. Read state/skills/ffmpeg/SKILL.md and state/lib/ffmpeg.ts before invoking.)_
Commands
| ui model | live composition via compose_inline, persisted as artifact lang_body, reopened with OpenArtifact | |invoke: import from state/lib/ffmpeg.ts - probe(), trim(), concat(), resizeLocal(), compress(), extractAudioLocal(), burnSubs(), thumbnail(), loudnorm(), overlay() |verify: probe() the input first to confirm codec/duration/streams before transforming |eval log: state/log/pending-eval.ndjson (manual eval - skill: "ffmpeg")
Known Pitfalls
loudnormis two-pass for accuracy (measure → apply); single-pass is faster but less consistent- ffmpeg binary must be on PATH - fails ugly otherwise
Compose Surface
After the user provides a file path, compose a video workflow surface immediately. Do NOT wait for confirmation - compose first, ask clarifying questions inline in the surface.
Phase 1 - probe + compose UI inline:
npx tsx state/lib/ffmpeg.ts probe <file>
Then emit a Lang surface with real probe data filled in:
root = Stack([header, streams, form, actions])
header = TextContent("Trim: <filename>", "medium-heavy")
streams = Card([
Stack([
StatCard("Codec", "<video_codec>"),
StatCard("Resolution", "<width>x<height>"),
StatCard("FPS", "<fps>"),
StatCard("Duration", "<HH:MM:SS>")
], "horizontal")
])
form = Card([
Stack([
Stack([Label("Start time"), start_input, Label("End time"), end_input], "horizontal"),
out_input
])
])
start_input = Input("00:00:00", "HH:MM:SS")
end_input = Input("<duration>", "HH:MM:SS")
out_input = Input("<base>_trimmed.mp4", "Output path")
actions = Stack([run_btn], "horizontal")
run_btn = Button("Probe + Trim", @Run("trim_video"), "primary")
Phase 2 - after trim completes:
result = Stack([
TextContent("Done", "medium-heavy"),
Stack([
StatCard("Before", "<original_size_mb> MB"),
StatCard("After", "<trimmed_size_mb> MB"),
StatCard("Reduction", "<pct>%")
], "horizontal"),
TextContent("<output_path>", "mono"),
Stack([
Button("Compress further", @Run("compress_video"), "secondary"),
Button("Copy path", @Set($copied, output_path), "ghost")
], "horizontal")
])
Phase 3 - other operations (compress, concat, resize):
- Use the same probe-first pattern before composing the operation surface
- Show stream info as StatCards at the top of every surface
- After completion always show before/after file size comparison
Critical Rules
- probe() before EVERY transformation - never trim blind
- Compose surface before asking the user for missing info - use Input fields in the surface
- Embed real probe data (duration, codec, resolution, fps) into the composed surface
- StatCard for before/after file size comparison on every destructive operation
- Surface persists as artifact (lang_body) - the user can revisit and re-run
- Log to state/log/pending-eval.ndjson after each operation
Self-Test
An agent reading this should correctly:
- [ ]
probe()before any transformation to know what you're dealing with - [ ] Compose a surface with real probe data - not ask for start/end in plain text
- [ ] Use
extractAudioLocal()not a hand-rolled-vn -acodec copy - [ ] Show before/after file sizes after any size-changing operation
- [ ] Log to
pending-eval.ndjson
Found a gap? Edit this file. <!-- footer-injection-point -->
api.ts- the code it can call
#!/usr/bin/env npx tsx
/**
* snappy-ffmpeg/api.ts — Local ffmpeg primitive layer.
*
* Public API:
* probe(path) — JSON metadata via ffprobe
* trim(path, start, end, opts) — Keyframe or frame-accurate trim
* concat(paths, output) — Concatenate video files
* resizeLocal(path, format, output) — Letterbox resize to 16:9, 9:16, 1:1
* compress(path, opts) — H.264 web compression
* extractAudioLocal(path, fmt, out) — Extract audio track
* burnSubs(path, subs, output) — Burn SRT/ASS subtitles into video
* thumbnail(path, ts, output) — Extract a single frame as JPG
* loudnorm(path, lufs, output) — EBU R128 loudness normalization
* overlay(base, over, pos, out)— PiP or watermark overlay
*
* CLI:
* npx tsx api.ts probe <path>
* npx tsx api.ts trim <path> <start> <end>
* npx tsx api.ts resize <path> <format>
* npx tsx api.ts compress <path>
* npx tsx api.ts extract-audio <path> [format]
* npx tsx api.ts burn-subs <path> <subs-path>
* npx tsx api.ts thumbnail <path> [timestamp]
* npx tsx api.ts loudnorm <path>
* npx tsx api.ts concat <path1> <path2> [path3...]
* npx tsx api.ts overlay <base> <overlay> [position]
*/
import { env } from "./env.ts";
import { execSync } from "child_process";
import { existsSync, realpathSync, unlinkSync, writeFileSync } from "fs";
import { extname, basename } from "path";
// ---------------------------------------------------------------------------
// Constants
// ---------------------------------------------------------------------------
const FFMPEG = "/opt/homebrew/bin/ffmpeg";
const FFPROBE = "/opt/homebrew/bin/ffprobe";
function tmpPath(operation: string, ext: string): string {
const stamp = Date.now();
return `/tmp/snappy-ffmpeg-${operation}-${stamp}${ext}`;
}
// ---------------------------------------------------------------------------
// Result type
// ---------------------------------------------------------------------------
export type Result = {
ok: boolean;
output: string;
path?: string;
error?: string;
durationMs: number;
};
function run(cmd: string, outputPath?: string): Result {
const start = Date.now();
try {
const stdout = execSync(cmd, {
encoding: "utf-8",
maxBuffer: 50 * 1024 * 1024,
stdio: ["pipe", "pipe", "pipe"],
});
return {
ok: true,
output: stdout.trim(),
path: outputPath,
durationMs: Date.now() - start,
};
} catch (e: any) {
return {
ok: false,
output: "",
error: e.stderr?.toString()?.trim() || e.message,
path: outputPath,
durationMs: Date.now() - start,
};
}
}
function assertFile(path: string): void {
if (!existsSync(path)) throw new Error(`File not found: ${path}`);
}
function q(s: string): string {
// Shell-escape a path
return `'${s.replace(/'/g, "'\\''")}'`;
}
// ---------------------------------------------------------------------------
// 1. probe
// ---------------------------------------------------------------------------
export type ProbeInfo = {
duration: number;
width: number;
height: number;
codec_video: string;
codec_audio: string;
frame_rate: string;
bit_rate: number;
audio_sample_rate: number;
audio_channels: number;
format: string;
raw: any;
};
export function probe(path: string): Result & { info?: ProbeInfo } {
assertFile(path);
const cmd = `${FFPROBE} -v quiet -print_format json -show_format -show_streams ${q(path)}`;
const result = run(cmd);
if (!result.ok) return result;
try {
const raw = JSON.parse(result.output);
const videoStream = raw.streams?.find((s: any) => s.codec_type === "video") || {};
const audioStream = raw.streams?.find((s: any) => s.codec_type === "audio") || {};
const info: ProbeInfo = {
duration: parseFloat(raw.format?.duration || "0"),
width: videoStream.width || 0,
height: videoStream.height || 0,
codec_video: videoStream.codec_name || "",
codec_audio: audioStream.codec_name || "",
frame_rate: videoStream.r_frame_rate || "",
bit_rate: parseInt(raw.format?.bit_rate || "0", 10),
audio_sample_rate: parseInt(audioStream.sample_rate || "0", 10),
audio_channels: audioStream.channels || 0,
format: raw.format?.format_name || "",
raw,
};
return { ...result, info };
} catch {
return { ...result, error: "Failed to parse ffprobe JSON" };
}
}
// ---------------------------------------------------------------------------
// 2. trim
// ---------------------------------------------------------------------------
export type TrimOpts = {
/** If true, re-encode for frame-accurate cuts. Default false (keyframe-aligned, -c copy). */
frameAccurate?: boolean;
/** Output path. Auto-generated if omitted. */
output?: string;
};
export function trim(path: string, start: string, end: string, opts?: TrimOpts): Result {
assertFile(path);
const ext = extname(path) || ".mp4";
const out = opts?.output || tmpPath("trim", ext);
let cmd: string;
if (opts?.frameAccurate) {
cmd = `${FFMPEG} -y -i ${q(path)} -ss ${q(start)} -to ${q(end)} -c:v libx264 -crf 18 -preset medium -c:a aac -b:a 192k -movflags +faststart ${q(out)}`;
} else {
cmd = `${FFMPEG} -y -ss ${q(start)} -to ${q(end)} -i ${q(path)} -c copy ${q(out)}`;
}
return run(cmd, out);
}
// ---------------------------------------------------------------------------
// 3. concat
// ---------------------------------------------------------------------------
export function concat(paths: string[], output: string): Result {
for (const p of paths) assertFile(p);
// Check if all files share the same codec — use concat demuxer if so
const codecs = paths.map((p) => {
const r = probe(p);
return r.info ? `${r.info.codec_video}:${r.info.codec_audio}` : "";
});
const sameCodec = codecs.every((c) => c === codecs[0] && c !== "");
if (sameCodec) {
// Fast concat via demuxer
const listFile = `/tmp/snappy-ffmpeg-concat-${Date.now()}.txt`;
const content = paths.map((p) => `file ${q(p)}`).join("\n");
writeFileSync(listFile, content);
const cmd = `${FFMPEG} -y -f concat -safe 0 -i ${q(listFile)} -c copy ${q(output)}`;
const result = run(cmd, output);
try { unlinkSync(listFile); } catch {}
return result;
} else {
// Re-encode concat via filter_complex
const inputs = paths.map((p) => `-i ${q(p)}`).join(" ");
const n = paths.length;
const filterParts = Array.from({ length: n }, (_, i) => `[${i}:v:0][${i}:a:0]`).join("");
const cmd = `${FFMPEG} -y ${inputs} -filter_complex "${filterParts}concat=n=${n}:v=1:a=1[outv][outa]" -map "[outv]" -map "[outa]" -c:v libx264 -crf 23 -preset medium -c:a aac -b:a 192k -movflags +faststart ${q(output)}`;
return run(cmd, output);
}
}
// ---------------------------------------------------------------------------
// 4. resize
// ---------------------------------------------------------------------------
type AspectFormat = "16:9" | "9:16" | "1:1";
const RESIZE_DIMS: Record<AspectFormat, { w: number; h: number }> = {
"16:9": { w: 1920, h: 1080 },
"9:16": { w: 1080, h: 1920 },
"1:1": { w: 1080, h: 1080 },
};
export function resizeLocal(path: string, format: AspectFormat, output?: string): Result {
assertFile(path);
const { w, h } = RESIZE_DIMS[format];
const ext = extname(path) || ".mp4";
const out = output || tmpPath("resize", ext);
const vf = `scale=${w}:${h}:force_original_aspect_ratio=decrease,pad=${w}:${h}:-1:-1:color=black`;
const cmd = `${FFMPEG} -y -i ${q(path)} -vf "${vf}" -c:v libx264 -crf 23 -preset medium -c:a copy -movflags +faststart ${q(out)}`;
return run(cmd, out);
}
// ---------------------------------------------------------------------------
// 5. compress
// ---------------------------------------------------------------------------
export type CompressOpts = {
crf?: number;
preset?: string;
faststart?: boolean;
output?: string;
};
export function compress(path: string, opts?: CompressOpts): Result {
assertFile(path);
const crf = opts?.crf ?? 23;
const preset = opts?.preset ?? "medium";
const faststart = opts?.faststart !== false;
const ext = extname(path) || ".mp4";
const out = opts?.output || tmpPath("compress", ext);
let cmd = `${FFMPEG} -y -i ${q(path)} -c:v libx264 -crf ${crf} -preset ${preset} -c:a aac -b:a 128k`;
if (faststart) cmd += " -movflags +faststart";
cmd += ` ${q(out)}`;
return run(cmd, out);
}
// ---------------------------------------------------------------------------
// 6. extractAudio
// ---------------------------------------------------------------------------
type AudioFormat = "m4a" | "mp3" | "wav";
const AUDIO_CODEC: Record<AudioFormat, string> = {
m4a: "-c:a aac -b:a 192k",
mp3: "-c:a libmp3lame -b:a 192k",
wav: "-c:a pcm_s16le",
};
export function extractAudioLocal(path: string, format?: AudioFormat, output?: string): Result {
assertFile(path);
const fmt = format || "m4a";
const out = output || tmpPath("audio", `.${fmt}`);
const codec = AUDIO_CODEC[fmt];
const cmd = `${FFMPEG} -y -i ${q(path)} -vn ${codec} ${q(out)}`;
return run(cmd, out);
}
// ---------------------------------------------------------------------------
// 7. burnSubs
// ---------------------------------------------------------------------------
export function burnSubs(path: string, subsPath: string, output?: string): Result {
assertFile(path);
assertFile(subsPath);
const ext = extname(path) || ".mp4";
const out = output || tmpPath("subs", ext);
const subsExt = extname(subsPath).toLowerCase();
let vf: string;
if (subsExt === ".ass" || subsExt === ".ssa") {
vf = `ass=${q(subsPath).slice(1, -1)}`; // ass filter takes unquoted path
} else {
// SRT and other text-based subs via subtitles filter
vf = `subtitles=${q(subsPath).slice(1, -1)}`;
}
const cmd = `${FFMPEG} -y -i ${q(path)} -vf "${vf}" -c:v libx264 -crf 18 -preset medium -c:a copy -movflags +faststart ${q(out)}`;
return run(cmd, out);
}
// ---------------------------------------------------------------------------
// 8. thumbnail
// ---------------------------------------------------------------------------
export function thumbnail(path: string, timestamp?: string, output?: string): Result {
assertFile(path);
const ts = timestamp || "00:00:05";
const out = output || tmpPath("thumb", ".jpg");
const cmd = `${FFMPEG} -y -ss ${q(ts)} -i ${q(path)} -frames:v 1 -q:v 2 ${q(out)}`;
return run(cmd, out);
}
// ---------------------------------------------------------------------------
// 9. loudnorm
// ---------------------------------------------------------------------------
export function loudnorm(path: string, targetLufs?: number, output?: string): Result {
assertFile(path);
const lufs = targetLufs ?? -14;
const ext = extname(path) || ".mp4";
const out = output || tmpPath("loudnorm", ext);
// Two-pass loudnorm: first pass measures, second pass applies.
// For simplicity we use single-pass (close enough for most content).
const af = `loudnorm=I=${lufs}:TP=-1.5:LRA=11`;
const cmd = `${FFMPEG} -y -i ${q(path)} -af "${af}" -c:v copy -movflags +faststart ${q(out)}`;
return run(cmd, out);
}
// ---------------------------------------------------------------------------
// 10. overlay
// ---------------------------------------------------------------------------
export type OverlayPosition = "top-left" | "top-right" | "bottom-left" | "bottom-right" | "center";
const OVERLAY_XY: Record<OverlayPosition, string> = {
"top-left": "10:10",
"top-right": "main_w-overlay_w-10:10",
"bottom-left": "10:main_h-overlay_h-10",
"bottom-right": "main_w-overlay_w-10:main_h-overlay_h-10",
"center": "(main_w-overlay_w)/2:(main_h-overlay_h)/2",
};
export function overlay(
basePath: string,
overlayPath: string,
position?: OverlayPosition,
output?: string,
): Result {
assertFile(basePath);
assertFile(overlayPath);
const pos = position || "bottom-right";
const ext = extname(basePath) || ".mp4";
const out = output || tmpPath("overlay", ext);
const xy = OVERLAY_XY[pos];
const cmd = `${FFMPEG} -y -i ${q(basePath)} -i ${q(overlayPath)} -filter_complex "overlay=${xy}" -c:v libx264 -crf 23 -preset medium -c:a copy -movflags +faststart ${q(out)}`;
return run(cmd, out);
}
// ---------------------------------------------------------------------------
// CLI
// ---------------------------------------------------------------------------
if ((() => { try { return import.meta.url === `file://${realpathSync(process.argv[1])}`; } catch { return false; } })()) {
const [, , cmd, ...args] = process.argv;
function die(msg: string): never {
console.error(msg);
process.exit(1);
}
function printResult(r: Result): void {
if (r.ok) {
if (r.path) console.log(`OK → ${r.path} (${r.durationMs}ms)`);
if (r.output && !r.path) console.log(r.output);
} else {
console.error(`FAIL (${r.durationMs}ms): ${r.error}`);
process.exit(1);
}
}
if (!cmd || cmd === "help") {
console.log(`snappy-ffmpeg CLI — local ffmpeg primitive layer
probe <path> Media metadata as JSON
trim <path> <start> <end> Fast trim (keyframe-aligned)
resize <path> <16:9|9:16|1:1> Letterbox resize
compress <path> H.264 web compression (crf=23)
extract-audio <path> [m4a|mp3|wav] Extract audio track
burn-subs <path> <subs-path> Burn SRT/ASS subtitles
thumbnail <path> [timestamp] Extract frame as JPG
loudnorm <path> EBU R128 loudness normalization
concat <path1> <path2> [path3...] Concatenate files
overlay <base> <overlay> [position] PiP / watermark overlay
help This message`);
process.exit(0);
}
if (cmd === "probe") {
if (!args[0]) die("Usage: probe <path>");
const r = probe(args[0]);
if (r.ok && (r as any).info) {
const { raw, ...summary } = (r as any).info;
console.log(JSON.stringify(summary, null, 2));
} else {
printResult(r);
}
} else if (cmd === "trim") {
if (args.length < 3) die("Usage: trim <path> <start> <end>");
printResult(trim(args[0], args[1], args[2]));
} else if (cmd === "resize") {
if (args.length < 2) die("Usage: resize <path> <16:9|9:16|1:1>");
printResult(resizeLocal(args[0], args[1] as AspectFormat));
} else if (cmd === "compress") {
if (!args[0]) die("Usage: compress <path>");
printResult(compress(args[0]));
} else if (cmd === "extract-audio") {
if (!args[0]) die("Usage: extract-audio <path> [m4a|mp3|wav]");
printResult(extractAudioLocal(args[0], (args[1] as AudioFormat) || undefined));
} else if (cmd === "burn-subs") {
if (args.length < 2) die("Usage: burn-subs <path> <subs-path>");
printResult(burnSubs(args[0], args[1]));
} else if (cmd === "thumbnail") {
if (!args[0]) die("Usage: thumbnail <path> [timestamp]");
printResult(thumbnail(args[0], args[1] || undefined));
} else if (cmd === "loudnorm") {
if (!args[0]) die("Usage: loudnorm <path>");
printResult(loudnorm(args[0]));
} else if (cmd === "concat") {
if (args.length < 2) die("Usage: concat <path1> <path2> [path3...]");
const out = tmpPath("concat", extname(args[0]) || ".mp4");
printResult(concat(args, out));
} else if (cmd === "overlay") {
if (args.length < 2) die("Usage: overlay <base> <overlay> [position]");
printResult(overlay(args[0], args[1], (args[2] as OverlayPosition) || undefined));
} else {
die(`Unknown command: ${cmd}. Run 'npx tsx api.ts help' for usage.`);
}
}
scripts- helper scripts it can run
prose-only skill - 1 inline code block live in SKILL.md above (no state/bin/ sidecar yet).
how we check it- the checks, plus the last 10 runs
no recent runs logged - the eval contract is declared but nothing has been graded yet