Roadmap

Status

Feature phase complete (2026-08-19). Tiers 1 and 2 shipped; Tier 3 is the MCP surface, which is built. ADRs current through 0016.

Everything below is the record of what got built and — as usefully — what got cut and why.

Tier 1 — Core editor

Folder picker → justified gallery (embedded-preview thumbnails; RAW+JPEG pairs collapse to one)
Edit view: GPU live preview, RGB histogram, hold-\ before/after
Develop panel: exposure, contrast, highlights, shadows, whites, blacks, white balance (temp/tint), vibrance, saturation, clarity, sharpening, noise reduction, vignette
Crop + straighten
Sidecar persistence (ADR 0002), undo/redo
Copy/paste edits across photos (⌘⇧C / ⌘⇧V)
Export: JPEG / HEIF / 16-bit TIFF, full resolution, quality control

Tier 2 — Differentiators

Tier 3 — AI via MCP (ADR 0008: no embedded agent)

Local MCP server: list_photos, get_edit, set_edit (live in UI), open_photo, get_preview, export · discovery at ~/.chiaro/mcp.json · repo .mcp.json
Agent presence pill + soft input lock while an agent edits
Portrait mask refinement (Mask slider: grow/shrink subject boundary)
Albums: cut by owner decision (2026-08-19)
SAM-assisted focus selection: built, tested, cut — SAM 2.1 tiny masks weren't good enough (owner decision 2026-08-19)
Clean up (LaMa inpainting): built, tested, cut — fills weren't shippable (owner decision 2026-08-19, ADR 0012). Cloud provider layer possible post-release. Positioning stands: AI-driven editing, not generative content
Auto-culling: not an in-app feature — an agent workflow over MCP (get_preview + set_starred; agent picks the keepers)

Tier 4 — Grading (shipped, ADR 0015)

Both were found by auditing the render pipeline while writing the editing skill, not by feature comparison.

Colour grading by tonal zone: shadow/mid/highlight hue and strength plus balance, as scalar rail rows rather than wheels, since a wheel is a two-dimensional control and ADR 0005 committed to one value at a time
Monochrome mixer: converts using the colour mixer's per-band luminance as channel weights, which the old pipeline order made unreachable
RAW decode parameters: CIRAWFilter already exposes detailAmount, moireReductionAmount and split luminance/colour noise reduction, and we set none of them. Decode-time detail and noise handling beats the post-demosaic filters we ship today, and moiré is unfixable afterwards. Design: sharpness and noiseReduction stay the only controls a user sees and route to the decode-time parameters for RAW files, rather than adding a parallel set — one set of controls implemented at the right point is what Lightroom and Capture One do. Colour noise reduction and moiré reduction are added as RAW-only rows, hidden for JPEGs. Accepted cost: an existing sidecar's sharpness on a RAW file renders slightly differently afterwards, since decode-time detail is a different operation and not merely a better one. Taken deliberately while pre-1.0, because the alternative is six overlapping detail controls forever

Backlog (post-1.0)

Ranked, with the reasoning that put them here rather than in a tier.

Luminance-range masking on existing locals — shipped. Two scalars (lumaLow/lumaHigh) added to a local, so "only the shadows inside this radial" is expressible. Composed with the masks already built instead of adding a concept, and stays agent-drivable.
Grain — shipped. grain and grainSize in Effects. The missing piece for a film look; the skill no longer has to admit it cannot deliver one.
chiaro CLI. A thin client of the same localhost server the app already runs — same EditState, live rendering when the window is open. MCP stays the session surface (presence, typed schemas, in-band previews, agents without a shell); the CLI is the throughput surface: dependent calls chain in one shell invocation and a whole shoot batches in one loop, where MCP costs a model turn per call. Deliberately not in 1.0 — the launch story is MCP.
Thumbnail virtualization. Library.loadThumbnails() decodes every photo up front and Photo.thumbnail never evicts, so a 5,000-photo folder holds several GB. The justified gallery needs every aspect ratio before it can lay out rows, which is why plain laziness does not work. Plan: read aspect from image metadata only (CGImageSourceCopyProperties, no decode, milliseconds per file) so layout is complete immediately; decode thumbnails only for rows near the viewport, with a bounded LRU (~500 resident) and eviction. Held out of 1.0 because it touches the gallery the product shots depend on, days before launch, and needs testing on a real large folder.
Accessibility. VoiceOver labels on icon-only buttons (cheap, an hour) and a decision on Dynamic Type against the fixed-width rail and toolbar.
Dataless cloud files. Detect placeholders (isUbiquitousItem plus the download status keys, and the File Provider equivalents for OneDrive and Dropbox) and offer to download rather than showing an empty tile.
A per-folder "keep edits on this Mac" switch. The cloud-folder rule in ADR 0007 is a path heuristic; a NAS mounted as a plain folder or a synced folder outside ~/Library/CloudStorage needs a way to opt in. This is the first real candidate for a Settings window, which 1.0 does not have.
Healing and clone. Wanted by every photographer, and a real gap. Note ADR 0012 already cut generative inpainting on quality evidence, so this means classical patch-based healing, not a model.
Share pixel-statistics primitives between AutoEnhance and StatsSampler — shipped (PixelStats). Both walked pixels to compute percentiles, clip fractions, saturation means and a gray-world average, with constants that had already diverged: 96x96 sRGB versus full-resolution Display P3, clip thresholds of 0.98/0.02 versus 254/255, and different neutral-pool gates. Nothing cross-checked one against the other, which is precisely why the inverted white balance in temp survived from the first commit until f19516d. Full unification was skipped, since AutoEnhance needs a fast downsample for interactive latency, but the percentile helper and the clip and saturation predicates are now one implementation with one set of constants.
Re-tune AutoEnhance's temp strength. Its multiplier and clamp were hand-tuned while the pipeline rendered temperature backwards, so they have never been validated under the correct sign. Note the clamps are load-bearing rather than a backstop: unclamped the formula reaches about ±67 against a clamp of ±20, and a 5% channel imbalance already asks for half the maximum correction, so do not raise the multiplier without rechecking where the clamp bites.
Move the grading kernel off Core Image Kernel Language — shipped (ColorGradeCube). CIColorKernel(source:) had been deprecated since macOS 10.14 and failed silently: a source typo returned nil, and grading shipped as a permanent no-op because of exactly that. The fix is a computed CIColorCube, matching what HSLCube already does, since zone grading is a pure function of RGB and expresses fine as a 3D LUT.

Considered and deliberately not planned:

Freehand brush masks. A painted mask is a raster, so it either bloats the sidecar with base64 or breaks ADR 0002's one-readable-file model. It is also the least agent-authorable control in photography, which cuts directly against what this app is for. Would need an ADR overturning 0002.
Per-channel RGB curves. Duplicates Tier 4's zone grading with a two-dimensional control per channel, and would mean maintaining two grading systems.
Camera calibration profiles. Apple's RAW engine exposes no colour-matrix override, so there is nothing to build against.

FilesExpand file tree

ROADMAP.md

Latest commit

History