Grade photos, camera raws, and videos with film looks and subject-aware crops from an AI agent, the command line, Python, or the browser.

When stdout is not a terminal, the CLI emits one JSON object per line with rounded floats. Multiple looks and crops run from a single decode.

$ silverbaths iphone-food-poke.jpg -l vintage-slide -c 1x1@smart -o graded
{"crop":null,"decode_seconds":0.036,"event":"output","frame":null,"height":2366,"input":"iphone-food-poke.jpg","kind":"image","look":"vintage-slide","output":"graded/iphone-food-poke_vintage-slide.jpg","seconds":0.835,"width":3155}
{"crop":"1x1","decode_seconds":0.036,"event":"output","frame":"1x1@0.0469,0,0.7968,1","height":2366,"input":"iphone-food-poke.jpg","kind":"image","look":"vintage-slide","output":"graded/iphone-food-poke_vintage-slide_1x1.jpg","seconds":0.6364,"width":2366}
{"errors":0,"event":"summary","seconds":1.5083,"skipped":0,"written":2}

silverbaths ops <name> and silverbaths schema recipe|look|plan|measurement print the parameter descriptions and JSON Schemas generated from the Rust types.

$ silverbaths ops halation --format human
halation: Coloured glow around (not on) bright areas, like light scattering in film.
  color      array of 3 numbers  required. Glow colour [R, G, B], e.g. [0.12, 0.035, 0.0].
  core       number              default 1.0. How much glow to remove on the bright source itself.
  spread     number              required. Glow radius is long_edge / spread (150 wide .. 400 tight).
  threshold  number              required. Luma threshold 0..1 above which pixels cast a halo.

Every error across the CLI, Python package, MCP server, and .sb checker carries a stable code and a hint stating what to run next.

{
  "event": "error",
  "code": "unknown_look",
  "phase": "typecheck",
  "span": { "file": "batch.sb", "line": 4, "col": 18, "json_pointer": "/bindings/0/value/look" },
  "message": "unknown look 'warm_negative'",
  "expected": "a valid look name (lowercase letters, digits, hyphens) or .toml path",
  "found": "warm_negative",
  "hint": "did you mean 'warm-negative'? Look names use hyphens, not underscores. Run `silverbaths looks list` for all looks.",
  "fix": { "line": 4, "col_start": 18, "col_end": 33, "replacement": "\"warm-negative\"" }
}

{
  "event": "diagnostic",
  "code": "L002",
  "phase": "lint",
  "severity": "warning",
  "message": "op 'grain' at step 4 is followed by spatial op 'halation' at step 5",
  "expected": "'grain' as the final op (or immediately before 'letterbox')",
  "hint": "bright grain speckles will trigger halation glow and get blurred. Move `grain(...)` after `halation(...)`."
}

silverbaths plan checks types, lint rules (L001..L012), input paths, and output collisions before decoding any media.

$ silverbaths plan carousel.sb
plan carousel.sb: 1 source, 4 resolves, 5 outputs (0 errors, 0 warnings)

  [1] open     photos/iphone-street-lyon.jpg (image, 4032x3024)
  [2] resolve  slide_1_cover     on [1] (teal-orange, 4x5@smart)
  [3] resolve  slide_2_pan_left  on [1] (without [vignette], 4x5@0.05,0,0.525,1)
  [4] resolve  slide_3_pan_right on [1] (without [vignette], 4x5@0.525,0,1,1)
  [5] resolve  story_9x16        on [1] (teal-orange, 9x16@0.29,0,0.71,1)
  [6] sheet    [2, 3, 4, 5]      -> out/carousel/_review_sheet.jpg (cell: 600)
  [7] encode   [2]               -> out/carousel/slide_1_cover.jpg (size: 2160)
  [8] encode   [3]               -> out/carousel/slide_2_pan_left.jpg (size: 2160)
  [9] encode   [4]               -> out/carousel/slide_3_pan_right.jpg (size: 2160)
 [10] encode   [5]               -> out/carousel/story_9x16.jpg (size: 2160)

summary: 1 decode shared across 1 sheet and 4 encodes

let raws = open_many("shoot/raw/*.{ARW,CR3,DNG}", recursive = true)
let cohort = anchor(raws, exposure = 0.70, white_balance = 0.50)

let matched = match_look(target = open("ref/still.jpg"), protect_skin = true)
  | halation(threshold = 0.78, spread = 260.0, color = [0.09, 0.03, 0.0])
  | grain(amount = 0.024, size = 1.4, chroma = 0.08, shadow = 0.3, mid = 2.0)