The phase 4a docs were written before the last two fix rounds landed. They
now describe the code at 8c2b9f1b: a sequence left as it was inside a moving
folder is taken along instead of keeping that folder in place; a held plan on
an offline archive keeps its plan and fails as offline; render filing answers
503 folder_moved; a rebase never moves the rows of the sequence that took the
old folder (with its known limitation); every writer refuses a folder another
sequence holds (SequenceFolders.Holds). Three DECISIONS rows added, six
updated; five fix rounds in STATUS and the request.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 31d479c5f4dd0b7076b39331246815e95895a1fa)
169 KiB
Chronos — Build Brief for Claude Code
This document is the complete specification and standing instruction set for building the Chronos time-lapse ecosystem from scratch. It is written to be read by Claude Code at the start of every session (via
CLAUDE.md, see Appendix A) and to be the single source of truth for scope, architecture, constraints and delivery order.Working name: Chronos. Three shipped programs — Chronos (Windows desktop client), Chronos Server (Linux/Docker and Windows), Chronos Uploader (Windows) — sharing one engine. Target audience: photographers, astro/landscape shooters, construction and weather-cam operators, hobbyists worldwide.
0. How to work on this project (operating rules for Claude Code)
0.1 Kickoff message
Paste this as the first message in Claude Code, from an empty repository that contains only
this file at docs/BRIEF.md and the prototype dump at docs/reference/chronos-prototype.md:
Read docs/BRIEF.md completely before doing anything. It is the full specification for this
repository and it overrides your defaults. Then:
1. Create CLAUDE.md from Appendix A of the brief, and docs/DECISIONS.md, docs/CHANGELOG.md.
2. Execute Phase 0 exactly as described in §22, stopping at its phase gate.
At every phase gate, report: what was built, how to run it, test results on Windows and
Linux, open decisions. Do not ask questions before the gate unless a decision is
irreversible; otherwise decide, record it in docs/DECISIONS.md, and continue.
0.2 Standing rules
- Zero third-party packages, everywhere. No
<PackageReference>in any.csproj, including tests, tools and build scripts. Anything that is not the .NET BCL, a Microsoft shared framework that ships with the .NET SDK/runtime (Microsoft.NETCore.App,Microsoft.AspNetCore.App,Microsoft.WindowsDesktop.App, the Windows SDK targeting pack), or an operating-system API reached by P/Invoke or COM interop, is implemented in-house. CI fails the build if aPackageReferenceappears (§19.2). The prototype indocs/reference/chronos-prototype.mdproves this is workable: it decodes images with WIC, encodes video with Media Foundation, parses EXIF/TIFF/XMP with its own parsers, and writes its own MP4 container. - No external processes. No FFmpeg, no ExifTool, no ImageMagick, no PowerShell scripts at runtime. The only OS executables/dialogs the product may invoke are listed in §8.6 (format-volume dialog, shell "open with", browser launch for links).
- Cross-platform core, Windows-only presentation.
Chronos.Core,Chronos.Data,Chronos.ProtocolandChronos.Servertargetnet10.0and must build and pass tests on Linux. Windows-only code lives in*.Windowsprojects or the desktop apps. - English everywhere in the codebase: identifiers, comments, commit messages, docs. The prototype is written in Italian; port the algorithms, not the language.
- Every user-visible string is a resource. No literal UI text in code or XAML. English is the source language; Italian is the first translation; all other languages follow in Phase 8 (§16).
- UI copy rules (mandatory): labels ≤ 3 words when possible; no helper text or paragraphs anywhere in the UI; explanations live only in tooltips (≤ 2 sentences); every actionable control has a tooltip; empty states are one sentence plus one button; errors are one sentence saying what happened plus one action.
- Decide, record, continue. When the brief is silent, choose the option that is
simplest, safest for user data, and consistent with the constraints; write one line in
docs/DECISIONS.md(date, decision, alternative, reason). Ask the human only at phase gates, or before any irreversible or data-destroying design choice. - User data is sacred. Never delete, move or format anything without (a) a verified hash match of every copied file and (b) an explicit user action. Every destructive operation is dry-run-able and shows its plan first.
- Definition of Done for every phase: CI green on Gitea Actions (§19.1: the Linux
jobs; the
windowsjob once a Windows runner is registered, until then the Windows-only tests run on a Windows PC; amended 2026-10-23, R3); all tests green; no new warnings;docs/DECISIONS.mdanddocs/CHANGELOG.mdupdated; a short "how to try it" note in the phase report; localization completeness check passes forenandit. - Commit style: Conventional Commits (
feat(core): …,fix(server): …), small commits, one concern each. Never commit generated binaries, fixtures downloaded by scripts, or secrets. - Reference prototype:
docs/reference/chronos-prototype.mdis a Repomix dump of the previous single-program prototype (C#/.NET 10, WinForms, zero packages). Its engine (metadata parsing, deflicker, holy-grail transitions, optical flow, motion blur, stabilization, virtual camera, time remap, stacking, Media Foundation encoding, MP4 muxing, self-tests) was validated on real material. Port those algorithms faithfully into the new structure, with English names and unit tests; do not copy its UI (WinForms) or its file layout. Search it with grep; do not try to read it whole.
1. Product vision
Chronos turns folders of photos (RAW, JPEG, HEIC…) into finished time-lapse videos and keeps a lifetime archive of shoots organised, verified and searchable — on one PC or on a shared server.
- Import memory cards with per-card rules, session detection, renaming and hash verification; import directly into a local archive or straight to a server.
- Analyse every sequence deeply (metadata + pixels) and let the advisor set frame timing, motion blur, deflicker, transitions and export presets — with a reason for every choice, visible in a tooltip.
- Render with the in-house engine and GPU encoders (H.264/HEVC/AV1 via Media Foundation) to purpose-built presets, from 720p previews to 8K 10-bit archives.
- Archive hundreds of thousands of photos across years and cameras, on local or network drives, with an index that can always be rebuilt from the files themselves.
- Share through Chronos Server: central archive, users with logins, uploads from the tiny Uploader, sync of projects, server-side jobs, render delegation to GPU-equipped clients, and a web admin panel — deployable as a Docker container (Unraid template included) or on Windows.
1.1 Goals (measurable)
- A first-time user imports a card, gets a recommended preset, and renders a 1080p preview in under five interactions from launch.
- A library of 500,000 photos in 2,000 sequences: Library screen opens in < 1 s, scrolls at 60 fps, filters respond in < 200 ms, full index rebuild from sidecars in < 10 min.
- Rendering quality equals or exceeds the prototype's measured results (§20.4 regression thresholds), with hardware encoding used whenever a hardware MFT exists.
- Every feature reachable without reading any documentation; every control explained by its tooltip; UI available in ≥ 30 languages with RTL support.
- Server: 10 concurrent clients uploading at line speed; resumable uploads; zero data loss on crash (append-only ingest + manifest commit).
1.2 Non-goals (v1)
- Audio tracks (P2: mux an external AAC file), HDR/PQ output, colour-managed display.
- macOS/Linux desktop clients (server only), mobile apps, browser-based editing.
- Cloud object storage backends (S3 etc.), multi-server federation.
- General RAW development beyond what the render pipeline needs (no Lightroom clone).
- GPU compute for image processing (CPU SIMD only in v1; keep interfaces open).
- Real-time tethered capture, camera control.
- Plugin system.
2. Hard constraints and allowed native surface
2.1 Language and platform decision
C# on .NET 10 LTS for every component. Reasons: one toolchain for a cross-platform
server and a first-class Windows desktop; WPF and ASP.NET Core ship with the SDK (no
packages); mature BCL for crypto, HTTP/2, WebSockets, JSON, compression and SIMD
(System.Numerics.Vector<T>, Vector256); direct P/Invoke/COM interop to WIC, Media
Foundation and winsqlite3.dll; the validated engine is already C#. Alternatives were
rejected: Python (stdlib GUI is dated, no GPU encode, no PostgreSQL driver), C++ (UI and
server cost), Rust/Go (no standard GUI, no SQLite/PostgreSQL in std), Java (JavaFX is
external).
Desktop UI: WPF with the built-in Fluent theme (Application.ThemeMode, available in
WPF since .NET 9, refined in .NET 10) for a Windows 11 look with dark/light/system modes and
system accent colour, plus in-house custom FrameworkElements for charts, filmstrip,
curve editors and the virtualized photo grid. If the Fluent theme proves inadequate for any
control, restyle it with an in-house ResourceDictionary — never a package. Window chrome:
WindowChrome + DwmSetWindowAttribute (dark title bar, Mica/backdrop where available).
2.2 Allowed native/OS surface (Windows)
| Need | API (all OS components, no packages) |
|---|---|
| Image decoding incl. RAW/HEIC/WebP/AVIF when codecs installed | WIC (IWICImagingFactory… via COM interop) |
| Video encoding, hardware MFTs (NVENC/QSV/AMF) | Media Foundation (MFStartup, MFTEnumEx, IMFTransform, ICodecAPI) |
| Local database | winsqlite3.dll (ships with Windows 10 1709+ / Windows 11) via P/Invoke |
| Volume identity, device arrival | GetVolumeInformationW, GetDriveTypeW, WM_DEVICECHANGE, SetupAPI if needed |
| Format a card | SHFormatDrive (OS format dialog) |
| GPU name/VRAM for encoder capability display | DXGI (CreateDXGIFactory1, IDXGIAdapter::GetDesc) |
| Taskbar progress, toasts | ITaskbarList3; toasts via Windows SDK projection (net10.0-windows10.0.19041.0 targeting pack) or in-house tray balloon |
| MTP cameras (P2) | Windows Portable Devices (COM) |
| Window chrome | dwmapi.dll |
Linux (server): BCL only. No libsqlite3, no libva, no FFmpeg. PostgreSQL is reached
through the in-house wire-protocol client (§6.3).
2.3 What this implies (and how the brief resolves it)
- RAW decoding without libraw is tiered (§8.2): full in-house decode for DNG/TIFF/ JPEG/PNG and a growing set of native RAW formats; embedded full-size JPEG previews as a universal fallback (every major RAW format embeds one); on Windows, WIC with the free Microsoft "Raw Image Extension" gives full decode of everything else.
- Video encoding on Linux has no OS encoder API; the server produces pure-managed outputs (MJPEG-AVI, image sequences, GIF, APNG) itself and delegates MP4/H.264/HEVC/AV1 jobs to render nodes — Chronos clients that volunteer their GPU (§14.7). An in-house intra-only H.264 encoder is a P2 universal fallback (§11.6).
- Testing uses an in-house runner (
Chronos.Testing, §20.1), not xUnit/NUnit. - HEIC/HEIF cannot be decoded in-house (HEVC intra decoder is out of scope): Windows clients decode via WIC (HEIF + HEVC extensions); the server stores originals and relies on proxies produced by a Windows client or render node.
3. Repository layout
chronos/ (Gitea repository BalboSoft/Chronos; amended 2026-10-23, R2)
├── CLAUDE.md operating rules (Appendix A)
├── Directory.Build.props version, LangVersion, nullable, warnings-as-errors, analyzers
├── Directory.Build.targets zero-package guard, resource generation hooks
├── Chronos.slnx
├── global.json pins .NET 10 SDK
├── src/
│ ├── Chronos.Core/ net10.0 — engine: Metadata, Imaging, Codecs, Analysis, Motion, Video (container/muxers), Archive model, Naming, Localization runtime
│ ├── Chronos.Core.Windows/ net10.0-windows — WIC decoder, Media Foundation encoder, DXGI, device APIs, winsqlite3 binding
│ ├── Chronos.Data/ net10.0 — IDatabase abstraction, SQL dialects, migrations, repositories, Postgres wire client (Chronos.Data.Postgres), SQLite driver surface (implemented by Core.Windows binding)
│ ├── Chronos.Protocol/ net10.0 — API DTOs, JSON contexts (source-generated), chunked upload protocol, sync manifests, WebSocket event schema, version negotiation
│ ├── Chronos.Server/ net10.0 — ASP.NET Core minimal API, auth, job engine, storage service, admin web UI (embedded static files), Docker entrypoint
│ ├── Chronos.App/ net10.0-windows — WPF desktop client
│ ├── Chronos.Uploader/ net10.0-windows — WPF uploader (tray-capable)
│ ├── Chronos.Setup/ net10.0-windows — in-house installer/uninstaller (self-extracting, Add/Remove Programs, shortcuts, file associations)
│ └── Chronos.Tools/ net10.0 — CLI: selftest, diagnose, make-icon, i18n export/check, fixtures fetch, changelog
├── tests/
│ ├── Chronos.Testing/ in-house test framework + runner (attributes, asserts, fixtures, JUnit XML)
│ ├── Chronos.Core.Tests/
│ ├── Chronos.Data.Tests/ includes PostgreSQL integration tests (skipped without CHRONOS_TEST_PG)
│ ├── Chronos.Server.Tests/ in-process Kestrel on a random port
│ ├── Chronos.App.Tests/ view-model tests + XAML lint (tooltip/resource rules) + screenshot smoke (`--capture`)
│ └── fixtures/ tiny checked-in samples; larger CC0 RAW samples fetched by `Chronos.Tools fixtures`
├── branding/ icon sources (SVG), generated ICO/PNG sets
├── i18n/ .resx sources (en) + translations, glossary, machine-translation notes
├── deploy/
│ ├── docker/Dockerfile, docker-compose.yml, entrypoint
│ └── unraid/chronos.xml
├── .gitea/workflows/ ci.yml, release.yml, docker.yml, perf.yml (amended 2026-10-23, R2)
└── docs/ BRIEF.md (this), DECISIONS.md, CHANGELOG.md, ARCHITECTURE.md (generated per phase), reference/chronos-prototype.md
Project references: App → Core, Core.Windows, Data, Protocol. Server → Core,
Data, Protocol. Uploader → Core (hashing, metadata light), Protocol. Core and
Data never reference anything Windows-specific; CI builds them on Linux.
4. Architecture
4.1 Component diagram
┌────────────────────────────────────────────────────────────────┐
│ Chronos (WPF client) │
│ Library · Edit · Import · Archive manager · Jobs │
│ Local mode: SQLite + local/network archive │
│ Server mode: REST/WS client, proxies cache, render node │
└───────┬───────────────────────────────────────┬────────────────┘
│ HTTPS REST + WebSocket │ Memory cards (USB/SD)
▼ ▼
┌───────────────────────────────────┐ ┌──────────────────────────┐
│ Chronos Server │ │ Chronos Uploader (WPF) │
│ Kestrel · API v1 · Auth · Jobs │◄───────────│ queue · watch folders │
│ Storage service · Admin web panel │ HTTPS │ resumable chunk upload │
│ Pure-managed encoders │ └──────────────────────────┘
└──────┬──────────────────┬─────────┘
│ │
▼ ▼
PostgreSQL (Docker) Archive volume (/archive: originals, sidecars, proxies, renders)
or SQLite (Windows)
4.2 Shared engine
Chronos.Core is the same assembly on client and server: metadata parsing, image codecs,
analysis advisor, render pipeline, container writers. Only the decoder/encoder providers
differ per platform, behind two interfaces:
public interface IImageDecoderProvider { bool CanDecode(ImageFormatInfo f); DecodedImage Decode(DecodeRequest r); }
public interface IVideoEncoderProvider { IReadOnlyList<EncoderCapability> Capabilities { get; } IVideoEncoderSession Open(EncodeSettings s); }
Registered providers, in priority order: client → WicDecoderProvider, then
ManagedDecoderProvider; server → ManagedDecoderProvider only. Encoders: client →
MediaFoundationEncoderProvider (hardware first, software H.264 fallback), then
ManagedEncoderProvider (MJPEG/AVI, sequences, GIF, APNG); server → managed only, plus
delegation to render nodes.
4.3 Data flow: card → archive → render
Card ──scan(light metadata)──► Import plan (sessions, names) ──copy+hash──► originals/
──light analysis──► sequence.json sidecar + DB rows ──queue──► deep analysis job
──► proxies + thumbs ──► advisor recommendations + suggested presets
User opens sequence ──► settings (project JSON) ──render──► exports/ + DB render row
Frames are streamed to the server by the client or uploader through chunked, resumable, hash-verified uploads; the server commits the frame into the archive layout and updates the sequence manifest atomically (write temp → fsync → rename).
4.4 Modes of the client
| Mode | Storage | Index | Who renders |
|---|---|---|---|
| Local | local or network drive archive | SQLite | this PC |
| Server-connected | server archive; local proxies/thumbs cache | server DB via API; local SQLite caches only | server (managed formats), render nodes (MP4), or this PC after downloading originals |
| Mixed | both; each sequence has an origin and a sync state |
both | as above |
5. Domain model
Primary entities (identical concepts in SQLite and PostgreSQL; ids are UUIDv7 stored as
TEXT/uuid; timestamps UTC ISO-8601/timestamptz):
- Archive — id, name, root path, layout templates, schema version, created, proxy policy.
- Sequence — id, archive id, title, folder (relative), camera make/model/serial, lens,
first/last capture, frame count, median interval, total bytes, status flags
(
analyzed_light,analyzed_deep,proxies_built,thumbs_built,rendered_count,verified_at), hero frame id, tags, rating, notes (short), settings JSON (project), analysis report JSON (compact) + path to full report, origin (local/server), sync state, version counters (settings_version,frames_version), owner user id (server). - Frame — id, sequence id, index, relative path, file name, size, sha256, captured at (with sub-second and source: EXIF/XMP/filesystem), exposure time, f-number, ISO, focal length, exposure bias, white balance, orientation, width, height, format, embedded preview info (offset/length), light stats (mean log-luminance, clip fractions) once analysed, flags (missing, corrupt, duplicate, excluded).
- Render — id, sequence id, preset name, settings JSON, output relative path, codec, width, height, fps, duration, bytes, created, produced by (client id / server / node).
- Job — id, kind (scan, rebuild, verify, light-analyse, deep-analyse, proxies, thumbs, render, import, upload-commit, maintenance; backup, cleanup, and the read-only duplicates and orphans reports; empty-trash, stored as 18, which erases exactly the trash entries of a confirmed plan, erases nothing when the trash changed since, and is never scheduled: amended 2026-10-26, R6; move, stored as 14, a Reorganize that applies exactly a confirmed plan of new folders, titles and file names to sequences of the archive, its target, and holds the rest of it when paused, cancelled or failed: amended 2026-10-26, R7), target ids, state (queued, running, paused, done, failed, cancelled), progress (done/total, fps, eta), log tail, created by, node id, attempts, error, schedule id (the schedule that queued it; amended 2026-10-26, R10).
- Schedule (amended 2026-10-26, R10) — id, system key (set, and unique, on a host's
system schedules, §14.6), name, action (scan, verify, thumbnails, proxies, analysis,
maintenance, backup, cleanup, duplicates, orphans), target (
current: the host's archive), parameters JSON, recurrence (text, no cron:every:<n>m|h|dfrom 5 minutes to a day, 1–168 hours or 1–366 days;daily:HH:MM;weekly:mon,thu@HH:MM;monthly:<1–28>@HH:MM; wall-clock times of the host's zone), catch-up, priority of its jobs, enabled, next run (none while disabled), last run, last job id, last outcome (queued, skipped-pending, skipped-offline, skipped-nothing, unavailable, failed:), created by, created, updated. - FrameRemoval (amended 2026-10-26, R6) — the tombstone of a photo removed from a
sequence: sequence id, sha256, replacement sha256 (the converted file that replaced it,
R7), reason (deleted, replaced), removed at. Written from the sidecar's
removedlist (§7) at every index sync, so a rebuild keeps it, leaving out content a frame of the sequence holds again; it goes with its sequence row and outlives the trash entry. - MediaCard (client) — id, volume serial, filesystem label, capacity, nickname, camera identity (Make/Model/BodySerial from first file), profile JSON, first/last seen, import history (job ids).
- User (server) — id, username, display name, password hash (PBKDF2), role (admin, editor, uploader, viewer), permissions bitmask, quota bytes, used bytes, disabled, must-change-password, created, last login.
- Session/Token (server) — token hash, user id, client name, created, expires, last used, revoked.
- RenderNode (server) — id, user id, client name, capabilities JSON (codecs, hardware encoder names, max resolution), last heartbeat, state.
- Setting — key/value (typed JSON), scope (app, archive, user).
- Tag, SequenceTag — normalised for filtering.
- AuditLog (server) — who did what to which sequence/user, when.
Indexes: frame(sequence_id, index), frame(sha256), frame(captured_at),
sequence(archive_id, first_capture), sequence(camera_model), tags, jobs by state and by
schedule, schedules(enabled, next_run) (amended 2026-10-26, R10),
frame_removals(sequence_id, sha256) (the key) and frame_removals(sha256) (migration 0008;
amended 2026-10-26, R6).
Paging everywhere (keyset pagination on (first_capture, id)).
6. Data layer
6.1 Chronos.Data abstraction
public interface IDatabase : IAsyncDisposable {
Dialect Dialect { get; } // Sqlite | Postgres
Task<IDbSession> OpenAsync(CancellationToken ct); // connection from pool
}
public interface IDbSession : IAsyncDisposable {
Task<int> ExecuteAsync(string sql, params DbParam[] p);
IAsyncEnumerable<DbRow> QueryAsync(string sql, params DbParam[] p);
Task<IDbTransaction> BeginAsync();
Task BulkInsertAsync(string table, IReadOnlyList<string> columns, IAsyncEnumerable<object?[]> rows); // COPY on PG, multi-row INSERT on SQLite
}
Hand-written SQL in repositories (SequenceRepository, FrameRepository, …); no ORM.
Migrations: numbered .sql files per dialect in Chronos.Data/Migrations/{sqlite,postgres}/,
applied at startup inside a transaction, recorded in schema_migrations. Dialect
differences (types, RETURNING, upserts) are confined to SQL files and a small SqlDialect
helper.
6.2 SQLite (client, and Windows standalone server)
- Binding:
Chronos.Core.Windows/Sqlite/WinSqlite3.csP/Invoke towinsqlite3.dll(sqlite3_open_v2,prepare_v2,step,reset,finalize,bind_*,column_*,exec,errmsg,last_insert_rowid,changes,busy_timeout,libversion,extended_result_codes). UTF-8 marshalling by hand. Checksqlite3_libversion() ≥ 3.31at startup; probePRAGMA compile_optionsand never depend on FTS/JSON extensions. - Pragmas:
journal_mode=WAL,synchronous=NORMAL,foreign_keys=ON,temp_store=MEMORY,cache_size=-65536,mmap_size=268435456. - Location:
%LOCALAPPDATA%\Chronos\chronos.db(never Documents/OneDrive-synced folders — SQLite in a synced folder corrupts). Portable mode:portable.txtnext to the exe puts everything under.\data\. - Backups:
VACUUM INTOcopies intobackupsnext to the index, made by the Backup jobs of theindex-backupsystem schedule (§14.6: every day on the desktop, daily at 03:00 on a server, both with catch-up; the first start after an upgrade copies at once when the newest copy is a day old), by Back up now (Settings › Advanced) and Back up database (Archive › Maintenance), and of the index a restore replaces. Every copy prunes the folder to the largest Copies kept of the backup schedules (3 by default), so no schedule deletes copies another one keeps. Restore from Settings › Advanced. (Amended 2026-10-26, R10.)
6.3 PostgreSQL client (Chronos.Data.Postgres) — in-house wire protocol v3
- TCP + optional TLS via
SslStreamafterSSLRequest;sslmode= disable | prefer | require | verify-full. - Startup (
user,database,application_name=chronos-server,client_encoding=UTF8,TimeZone=UTC); authentication: SCRAM-SHA-256 (Rfc2898DeriveBytes.Pbkdf2, HMAC-SHA256, channel binding not required), MD5, cleartext, trust. - Simple query, extended query (Parse/Bind/Describe/Execute/Sync) with parameters in text
format except
bytea(binary), prepared-statement cache per connection, transactions,COPY … FROM STDINfor bulk inserts,LISTEN/NOTIFY(used for multi-instance job wake-ups, optional), cancellation viaCancelRequest. - Type mapping:
text, varchar, int4, int8, float8, bool, bytea, timestamptz, uuid, jsonb, numeric(as string); arrays not needed. - Connection pool (min/max, health check
SELECT 1, idle eviction), typedPostgresExceptioncarrying SQLSTATE, retry on transient errors (57P01, 08xxx). - Test suite runs against PostgreSQL 16 and 17 service containers in CI (§19).
7. Archive format on disk
The archive is the source of truth; the database is an index that can be rebuilt from it.
<ArchiveRoot>/
├── .chronos/
│ ├── archive.json { id, name, schemaVersion, created, layoutTemplates, proxyPolicy }
│ ├── index-snapshot.bin optional accelerator for rebuild (all sidecars concatenated + hash); ignored if stale
│ ├── locks/ lease files for concurrent writers (server + client on same share)
│ ├── journal/ rename-<plan>.json: a mass rename or Reorganize being applied (below; amended 2026-10-26, R7)
│ └── trash/ deleted sequences and photos, kept until Empty (below; amended 2026-10-26, R6)
├── 2026/
│ └── 2026-09-28 Canon EOS R5 - Harbour sunset/ folder = expanded FolderTemplate
│ ├── originals/ IMG_0001.CR2 … names = expanded FileTemplate
│ ├── proxies/ 0001.jpg … (JPEG, long edge per policy; optional, regenerable)
│ ├── exports/ 2026-09-28 Harbour sunset - 4K HEVC.mp4 …
│ └── .chronos/
│ ├── sequence.json sidecar manifest (below)
│ ├── analysis.json full deep-analysis report
│ ├── project.json render settings (the "project")
│ ├── thumbs.pack packed thumbnails (one file per sequence, indexed; avoids 100k tiny files on SMB)
│ ├── rename-files.json the file renames of a Reorganize a crash interrupted (below; amended 2026-10-26, R7)
│ └── *.pending.json journals of an ingest, a photo deletion or a restore a crash interrupted (below)
sequence.json (canonical, versioned, written atomically):
{ id, title, camera, lens, frames: [{ i, file, size, sha256, capturedAt, exposure, fNumber, iso, focal, bias, wb, orientation, w, h, format, previewOffset, previewLength, flags, source? }], tags, rating, notes, status, versions, createdBy, updatedAt, removed? }. Amended 2026-10-26,
R6: removed: [{ sha256, replacement?, reason: Deleted | Replaced, at }] holds the tombstones
of the photos removed from the sequence (§5 FrameRemoval), left out when there are none;
frames[].source: { sha256, format, file, size } names the file a converted frame was made
from, reserved for R7 (read and written, nothing acts on it yet). A build that does not know
either field keeps it and writes it back unchanged.
Trash (amended 2026-10-26, R6). Nothing in it is erased except by Empty (§12.5, §14.8):
trash/<id>[-<time>[-n]]/with<name>.jsonbeside it: a deleted sequence folder, whole (originals, proxies, exports, sidecars), and its record (the index row, renders and tombstones the sidecar may lack).trash/photos/<id>-<time>[-n]/originals/…with<name>.jsonbeside it: photos deleted from sequence<id>, each original with its XMP companion when no photo that stays shares it, and the record: each photo's frame record, each file's former place, size and SHA-256, the sequence's title and folder, who deleted them, when, and the kind (photos, or replaced for originals a conversion replaced, R7). A photo never hashed is hashed as it goes; one whose original was already missing goes with its record alone.trash/.erasing/: entries moved aside by Empty before they are deleted, so a listing never sees half an entry; leftovers are deleted by the next Empty.- In a sequence's
.chronos/,remove.pending.jsonis written before a photo deletion's first move andrestore.pending.jsonbefore a restore's; until the sidecar is written the entry's record is<name>.json.pending, which the trash never lists, restores or empties. The next writer of the sequence (anyReadLocked, a scan, a sequence delete) undoes a deletion the sidecar does not show yet (files back, record and folder gone) and completes one it shows, and settles an interrupted restore under the entry's lease (files back into the entry, or the record without the photos the sidecar lists again). A restore into a new sequence stages each original in<entry>/.restoring/(never the archive's staging) and puts it back on any failure.
Deleting photos writes the sidecar without them (frames version + 1) and with a tombstone of
each SHA-256 no photo that stays holds, and clears what covered the frames (light and deep
analysis, thumbnails, proxies), which follow as jobs; proxies are renamed and thumbs.pack
is re-keyed so the frames that stay keep theirs (while a thumbnail pass holds the pack, that
pass draws them again). Filing a photo into the sequence again (an import, a restore, an
explicit upload again, §14.4) drops its tombstone; emptying the trash does not.
Reorganize (amended 2026-10-26, R7; §12.5, §14.8) applies a confirmed plan step by step, a step being one sequence and the sequences nested in a folder that moves with it:
journal/rename-<plan>.jsonholds the plan (its steps, templates and the sequences it was made for), the steps completed and what they did; it is written before the first step and deleted after the last. A plan whose job was paused, cancelled or failed is marked held and waits for that job; any other is completed at the next archive open. A journal that cannot be read is logged and left where it is: the archive opens, the plan is not resumed and its job fails with the reason.- Each step takes the leases of its sequences, then their thumbnail leases, and checks under them, before anything of it moves, that the plan still fits: the sequences where the index had them, nothing indexed since inside a folder that moves or at its new place, no sidecar of a newer format to rewrite, the title the plan started from where it changes the title or names the folder or files after it, the folder and title the templates give from the first photo as it is now, and the file renames planned again from the sidecar (compared by digest). What no longer fits refuses the rest of the plan, except photos changed since, whose sequence keeps its names (§12.5).
- Then the folder moves (a change of case only through a temporary name, never onto another folder of that exact name), and in one transaction the index follows: the sequence's folder and title, its frames' paths and the paths of the exports that moved with the folder. Only then are files renamed; once its folder moved, a step always goes to its end.
- A sequence of the plan that an earlier step left as it was inside a folder that moves later (it changed after the plan, or the folder that was to carry it stayed) goes with that folder, under the same leases, keeping its title and file names. Before a held plan resumes, a folder that moves may hold only such sequences and those of the steps still to come; any other refuses the plan, which is then given up.
- File renames run under
rename-files.jsonin the sequence's.chronos/: every file first to a hidden temporary name beside it (.chronos-rename-<token>-<n>.tmp), then every file to its new name, so chains and swaps of names never meet a name another file holds and no file is ever written over; then the sidecar names the new files (frame order and ids unchanged, frames version + 1), then the index, then the journal goes. A failure while files move puts every file back. After a crash the next writer of the sequence (anyReadLocked, the plan's resume) completes the renames, safely run again from wherever it stopped: a new name taken meanwhile sends that photo and its companions back together to their old names, or to the first free_2,_3…, and the journal stays while a temporary file is left. - Proxies (named by frame index), thumbnails (by index and content hash), tombstones and trash records need nothing; a photo restored from the trash comes back under the name it had.
Rules: originals are immutable (never rewritten; edits are settings); every write of a
sidecar goes to a temp file, fsync, rename; the folder name is derived from templates at
import time and never auto-renamed afterwards: only Reorganize, an explicit user action
behind a dry run, moves and renames what is filed, updating sidecars and DB together
(amended 2026-10-26, R7). Network drives: UNC or mapped paths; long-path aware
(\\?\ prefix); no FileSystemWatcher reliance — periodic polling by mtime/size (the
Archive scan system schedule of each host, §14.6; amended 2026-10-26, R10) plus manual
rescan; offline detection with graceful read-only state.
7.1 Naming templates
FolderTemplate default {year}/{date} {camera}; FileTemplate default
{date}_{time}_{n:0000}. Tokens (localised descriptions in resources): year, month, day, date, time, hm, session, sessionstart, camera, lens, iso, shutter, aperture, n, nsession, name, ext, title. {token:format} passes the format to the .NET formatter. Illegal
characters are replaced, never passed through; templates are validated live in the UI with a
preview of the first three names.
One naming rule for import, upload and Reorganize (amended 2026-10-26, R7):
- A photo is named in its sequence with the sequence's camera and lens (its first photo's, or
the camera hint its creator gave): a second body in a whole-folder import, or a lens changed
during a session, does not show in the names.
{n}is its place in the sequence,{name}its name on the card,{ext}its file's extension (the card name's when the file lost the extension of a photo), and a frame recorded without a capture time takes the sequence's start (its first frame's time, else when it was made; never today's date). A name already given, or held by a file that stays, gets_2,_3… before the extension. Folder and title templates see the sequence's first photo as filing saw it. - So a Reorganize with the archive's own templates changes nothing filing named. It gives
other names, which its dry run lists, where filing could not know the sequence as it is now:
{session}of a card's second session on (an import numbers its sessions, a sequence is one),{n}after photos skipped mid-session as duplicates or filed out of capture order, and a folder or title from a photo that is no longer first. - Rules every template meets wherever it is typed or sent (Settings › Archive, Import,
Reorganize, the admin panel's Archive): a folder template may not write a level scans never
enter (starting with
., ororiginals,proxies,exports,.chronos); a file template must keep the photo's extension (appended, or after.{ext}:{n}{ext}is refused), tell the photos of a session apart (three sample photos two seconds apart must get three names) and hold no slash. A level a value such as a title would hide that way is written with a leading_. A file name longer than 200 characters is shortened before its.{ext}, never through it. A title template gives text, not a file name (Paris: nightstays), at most 200 characters. A templatearchive.jsonalready holds, saved before these rules, is kept through edits of other settings; changing it holds it to them, and a new archive meets them all.
8. Image pipeline
8.1 Supported input formats
| Family | Extensions | Metadata (in-house) | Pixels on Windows | Pixels on server (Linux) |
|---|---|---|---|---|
| JPEG | jpg jpeg jpe jfif | APP1 Exif/XMP, MPF | WIC or managed | managed (baseline + progressive) |
| PNG | png | eXIf, iTXt XMP | WIC or managed | managed |
| TIFF | tif tiff | IFDs | WIC or managed | managed (strips/tiles; none/PackBits/LZW/Deflate/JPEG) |
| DNG (smartphones: Apple ProRAW, Android, Samsung; DJI; Insta360; Leica; Pentax; Ricoh; GoPro files converted by Adobe software) | dng | IFDs, MakerNote | managed full decode (amended 2026-10-16, R4) | managed full decode |
| Canon | cr2 cr3 | cr2: TIFF; cr3: ISO-BMFF CMT1..4 |
WIC (Raw ext.) → managed CR2 (P1) → embedded preview | managed CR2 (P1) → embedded full-size JPEG (CR3 track 1) |
| Nikon | nef nrw | TIFF | WIC → managed NEF (P1) → preview | managed NEF (P1) → preview |
| Sony | arw srf sr2 | TIFF | WIC → managed ARW (P1) → preview | idem |
| Fujifilm | raf | RAF header + TIFF | WIC → managed RAF uncompressed (P1) → preview | idem |
| Olympus/OM, Panasonic, Pentax, Samsung, Hasselblad, Phase One, Leica | orf rw2 raw pef srw 3fr fff iiq rwl | TIFF-like | WIC → managed (P1/P2) → preview | preview (P1: managed decoders as feasible) |
| GoPro | gpr (DNG+VC-5; a DNG is GPR only when its raw image is VC-5, amended 2026-10-16, R4) + jpg sidecar | IFDs | preview/sidecar JPEG; VC-5 out of scope | sidecar JPEG |
| HEIF/HEIC/AVIF (iPhone, Samsung, DJI) | heic heif hif avif | ISO-BMFF Exif item |
WIC (HEIF + HEVC/AV1 extensions) | metadata only; pixels via client-made proxies |
| WebP | webp | RIFF EXIF chunk |
WIC (WebP ext.) | metadata only (P2: managed VP8L) |
| JPEG XL, JPEG XR, BMP, GIF (single frame) | jxl jxr bmp gif | minimal | WIC if a codec exists | not supported |
Windows codec availability is enumerated at startup (IWICImagingFactory:: CreateComponentEnumerator). Missing codecs are surfaced as a one-line notice with an
"Install" button opening the Microsoft Store page (Raw Image Extension, HEIF Image
Extension, HEVC Video Extensions, AV1 Video Extension, WebP Image Extension; verify product
IDs at implementation time). The format table in Settings › Formats shows, per format,
which tier will be used on this machine.
8.2 Decode tiers (deterministic, logged per frame)
- Native full decode (in-house or WIC) at working resolution, 16-bit linear light.
- Embedded preview (largest JPEG found: TIFF
JPEGInterchangeFormat/SubIFD strips, RAF preview, CR3 full-size JPEG track, MPF). MarkedPreviewOnlyon the frame; the sequence shows a chip with the count and the advisor caps output resolution to the preview size unless the user overrides. - Unavailable → frame flagged
Undecodable; excluded from render; listed in the sequence warnings.
The tier is what the frame can give. Analyses, thumbnails, proxies and the editor's preview take an embedded preview that is large enough when a RAW has one, by choice and without changing the frame's tier; renders always decode natively. WIC counts as a native RAW decoder only when the decoder it chose for the file is the Raw Image Extension (Windows' in-box DNG decoder reads only the preview). (Amended 2026-10-16, R4.)
The in-house DNG pipeline (server and client): CFA/linear DNG, Compression 1/7 (lossless
JPEG SOF3)/8 (Deflate)/34892 (lossy JPEG), LinearizationTable, BlackLevel (+ repeat
dims, delta), WhiteLevel, AsShotNeutral/AsShotWhiteXY, ColorMatrix1/2 +
CameraCalibration, AnalogBalance, Bradford adaptation to D65, demosaic AHD (default,
"perfect result" priority) with bilinear for Draft quality, highlight reconstruction (clip
- desaturate),
ActiveArea/DefaultCrop*,Opcodelists 1–3 (at leastGainMap,WarpRectilinear,FixVignetteRadial; unknown opcodes ignored with a warning),BaselineExposure. Output: linear Rec.709/sRGB primaries float planes.
8.3 Managed codecs (all in Chronos.Core/Codecs, pure C#, SIMD where it pays)
JpegDecoder(baseline, progressive, restart intervals, 4:4:4/4:2:2/4:2:0/4:1:1/gray, 12-bit optional, ICC ignored),JpegEncoder(baseline, quality 1–100, subsampling choice, optimised Huffman tables optional, Exif/XMP APP1 embedding).PngDecoder/PngEncoder(all colour types, 8/16-bit, Adam7 decode, filters, zlib viaSystem.IO.Compression.ZLibStream).TiffReader/TiffWriter(as prototype: 16-bit TIFF, linear DNG writer withPhotometricInterpretation=34892).GifEncoder(animated; median-cut palette + Floyd–Steinberg; frame differencing).ApngEncoder.AviMjpegWriter(RIFF AVI withMJPG, OpenDML index for > 2 GB).Mp4Muxer(port of prototype: streamingftyp/mdat/moov, 90 kHz timescale, variable durations,avcC/hvcC; addav1C+ OBU handling in P1),AnnexBParser.Bmffreader (HEIF/CR3/AVIF boxes),Riffreader (WebP/AVI),Rafreader.
8.4 Image buffers
Port the prototype's ImageBuffer/FrameBufferPool (pooled float planes, linear light),
FrameWindow (sliding window, disk spill under a memory budget, read-ahead), and
GeometryStage (crop/aspect/affine with Catmull-Rom). Memory is governed by the window
size and budget, never by sequence length; the self-tests assert this.
8.5 Thumbnails and proxies
- Thumbnail: 320 px long edge JPEG q80, packed per sequence in
thumbs.pack(header, index table, entries), with an LRU in-memory cache and async loading for the UI. - Proxy: JPEG at policy long edge (default 1920; options 1280/1920/2560/full), q92, Exif orientation applied, capture time copied into Exif. Proxies are optional and regenerable; they enable remote editing, server-side rendering of formats the server cannot decode, and fast scrubbing.
8.6 The only OS executables/dialogs the product may invoke
SHFormatDrive (format card), ShellExecute("open", url) (links, Store pages),
ShellExecute("explore", folder) (reveal in Explorer). Nothing else.
9. Metadata model and parsers (Chronos.Core/Metadata)
Port and extend the prototype's binary parsers: container detection by signature; JPEG
segment scan (APP1 Exif, APP1 XMP + extended XMP, APP2 MPF); TIFF/IFD walker with on-demand
window reads (never load a whole RAW to read tags), SubIFD selection of the main image,
0x02BC XMP; ISO-BMFF walker (meta/iinf/iloc Exif item, CR3 CMT boxes, PRVW/THMB,
track sample tables for the CR3 JPEG); RIFF; RAF; PNG chunks. MakerNote parsing only where
it adds value: Canon/Nikon/Sony/Fuji/Olympus/Panasonic body serial, shutter count, lens
name, exposure mode, image stabilisation flag, intervalometer hints, sRaw flags.
Fields extracted per frame: DateTimeOriginal + SubSecTimeOriginal + OffsetTimeOriginal
(fallback: CreateDate, GPS time, file mtime — recorded with TimestampSource), ExposureTime,
FNumber, ISO, FocalLength (+35 mm equiv), ExposureBias, ExposureProgram, MeteringMode,
WhiteBalance mode + Kelvin if present, Flash, Orientation, dimensions, Make/Model/BodySerial,
Lens, GPS (lat/lon/alt), Artist/Copyright, Rating/Label from XMP, Software. All parsers
are fuzz-safe (no exception escapes; malformed input yields partial metadata + a warning).
ExifWriter (prototype) is kept for proxies/exports: writes a minimal, valid Exif block.
10. Analysis advisor
Two passes; both deterministic, both explainable (every finding and every recommendation carries a resource key + parameters that render as a localised reason in tooltips).
10.1 Light analysis (at import/scan, metadata only, seconds per 10k frames)
Cadence (median interval, jitter, gaps → session boundaries; SessionGapMinutes default
45, MinimumSessionFrames 8), total capture span, exposure series (EV =
log2(N²/t) − log2(ISO/100)), exposure step detection (ISO/shutter/aperture changes and
their EV amplitude), camera/lens identity, orientation consistency, resolution
consistency, timestamp sanity (clock resets, duplicates, non-monotonic), format tiers,
GPS → sunrise/sunset/civil twilight via an in-house solar-position calculation (NOAA
algorithm) → day/night/transition labelling.
10.2 Deep analysis (background job after import; "speed is not an issue")
Every frame decoded at analysis width (default half native, capped 2048): per-frame
trimmed log-luminance mean, histogram, clip fractions, per-channel means (WB drift), colour
temperature where applicable (McCamy with Planckian-locus validity check), sharpness
(Laplacian variance), noise estimate, phase-correlation shift vs previous frame (shake
magnitude, drift → panning), region mask (horizon by strongest row-luminance step, refined
per column; fallback by luminance), flicker spectrum (autocorrelation of luminance
residuals → deflicker window), exposure steps as actually received by the pixels,
star-field detection (small bright point density on night frames), cloud coverage estimate
(sky-region texture), scene-change detection (histogram distance), duplicate/black/white
frames, motion magnitude histogram (for motion-blur recommendation), suggested hero frame.
Results → analysis.json + compact summary in DB; charts in the Edit section (§12.3; amended 2026-10-16, R7).
10.3 Recommendations (rules, each with reason and confidence)
- Frame timing mode: Constant when jitter < tolerance; Adaptive when the intervalometer paused deliberately; Interpolated when gaps are sparse and motion is smooth.
- Frame rate and time compression: choose 24/25/30/60 fps and duration targets from cadence and motion magnitude; offer "fit to N seconds".
- Motion blur target shutter angle: from measured shutter angle (360×exposure/interval) and motion magnitude; skip flow when exposure already exceeds the target.
- Deflicker window (4× flicker period), region deflicker when horizon confidence is high, holy-grail transition length (half the typical distance between exposure steps), WB smoothing, stabilization on/off (measured shake > 0.3 px) with crop cost, crop/aspect suggestions from subject placement, output resolution cap from source/preview size and codec level limits, quality profile.
- Auto-generated sequence presets ("Night sky — 4K 10-bit HEVC 24 fps, deflicker 41,
stabilise", "Holy grail sunset — 4K HEVC 30 fps, transitions 18", "Construction cam —
1080p H.264 30 fps, adaptive timing", "Social vertical 1080×1920") = full project settings
- export preset; one click to apply; saveable as user presets.
10.4 Auto vs manual
Every deducible parameter shows an Auto indicator with the chosen value and a reason tooltip; touching the control takes it manual; a small reset returns it to Auto. Per section Optimise applies heavier decisions (module on/off) and reports what changed and why. Where data is missing, the advisor says so and changes nothing.
11. Rendering engine and export
11.1 Render plan (port)
The engine walks a list of output frames, each with a (possibly fractional) source position
and a duration; constant, adaptive, interpolated and remapped timings all produce this one
shape. Fractional positions synthesise frames from the motion field; motion blur scales
with √(target² − real²) shutter angle composition (prototype rule).
11.2 Modules (port, keep the validated behaviour and its tests)
Colour science (linear light, sRGB transfer), orientation detection (declared vs decoded size), deflicker (weighted local linear regression with Tukey robustness blended by surviving-window fraction), region deflicker (horizon/luminance masks), holy-grail exposure step redistribution (only steps the pixels actually received), WB smoothing (channel ratios), optical flow (pyramidal differential, median filtered), motion blur (gather filter), stabilization (phase correlation, nine patches, similarity fit, texture gating, smoothed path), virtual camera (Bézier eased pan/zoom keyframes, composed with stabilization into one affine), non-linear time remap (Fritsch–Carlson monotone spline), temporal stacking (median, progressive max with decay), geometry stage, codec level limits, disk spill, cancellation semantics.
11.3 Encoders and containers
| Output | Client (Windows) | Server (Linux/Windows) |
|---|---|---|
| MP4 H.264 (Baseline/Main/High, 8-bit) | Media Foundation: hardware MFT (NVENC/QSV/AMF) else Microsoft software encoder | render node; P2 in-house intra-only encoder |
| MP4 HEVC Main / Main10 (10-bit, P010) | hardware MFT only | render node |
| MP4 AV1 | hardware MFT only (Windows 11 + AV1 encoder) | render node |
| AVI MJPEG | managed | managed |
| Image sequence JPEG / PNG / TIFF-16 / linear DNG | managed | managed |
| Animated GIF, APNG | managed | managed |
Encoder session design as prototype: encoder MFT driven directly (not SinkWriter), NV12/P010
conversion in-house, Annex-B → MP4 samples, streaming muxer; rate control (VBR/CBR/quality),
GOP, B-frames, profile/level via ICodecAPI. Capability discovery (MFTEnumEx hardware
flag, friendly names, DXGI adapter) is shown in Settings › Rendering and used to hide
unavailable codecs.
11.4 Presets
Presets bind resolution, aspect, codec, profile, bit depth, fps, bitrate (≈0.10 bpp/frame H.264, 0.07 HEVC, 0.05 AV1; night-sky content is the most demanding), GOP and container. Editing any value flips the name to "Custom"; matching values restore the preset name.
Built-in presets: 8K UHD 10-bit HEVC archive · 4K UHD 10-bit HEVC archive · 4K UHD HEVC · 4K UHD H.264 compatible · DCI 4K HEVC 24 fps · Cinema 21:9 4K HEVC 24 fps · 4K 60 fps · 1440p H.264 · 1080p H.264 share · 1080p 60 · 1080×1920 vertical social · 1080×1350 4:5 · 1440×1440 square · 720p quick preview (GOP 1 s) · AV1 4K · MJPEG AVI intermediate · PNG sequence · JPEG sequence · TIFF-16 sequence · GIF preview (480 px) · APNG. Resolutions: 720p/1080p/1440p/4K UHD/DCI 4K/5K/6K/8K/custom; aspects 16:9, 21:9, 4:3, 3:2, 1:1, 9:16, 4:5, source, custom; fps 23.976/24/25/29.97/30/48/50/59.94/60/120.
11.5 Export UI behaviour
Preview stays visible during encoding (optional live preview); single global progress bar
(frames, fps, elapsed, ETA) mirrored to the taskbar; cancel is always available and leaves
no partial files behind; output name from {title} - {preset} template; "Reveal in
Explorer" and "Render another preset" follow-ups; batch render of several presets in one
job; per-sequence render history.
11.6 P2: in-house H.264 intra-only encoder
Baseline profile, I-frames only, I_16x16/I_4x4 prediction, 4×4 integer transform, CAVLC, fixed QP or simple rate control. Large files but universally playable; enables MP4 output on a server without render nodes. Validate with the in-house MP4 inspector and a system player smoke test on Windows CI.
12. Chronos desktop client — UI specification
12.1 Shell
┌────┬─────────────────────────────────────────────────────────────────────────────┐
│ ≡ │ ▮▮▮▮▮▮▮▮▮▮▮▮▮░░░░░░░ Building proxies · Harbour sunset · 1,240/3,600 · 12 fps · 3 min │ ⏹ │ ← global job bar, only while a job runs
│ ├─────────────────────────────────────────────────────────────────────────────┤
│ ▤ │ │
│Lib │ │
│ ⬇ │ section content (edge to edge) │
│Imp │ │
│ ✎ │ commands of the section live INSIDE the section, in one row above the │
│Edit│ content they act on; every command has a tooltip │
│ ▣ │ │
│Arch│ │
│ ☰ │ │
│Jobs│ │
│ ☁ │ │
│Srv │ │
│ │ │
│ ⚙ │ │
│Set │ │
│ ├─────────────────────────────────────────────────────────────────────────────┤
│ │ 412 sequences · 618,204 photos · 9.4 TB · ☁ Connected to "Studio NAS" · 3 jobs │ ← status strip
└────┴─────────────────────────────────────────────────────────────────────────────┘
- Left navigation rail (prototype pattern): Library, Import, Edit, Archive, Jobs and Server at the top, Settings alone at the bottom (About is the last page of Settings, §12.9; amended 2026-10-26, R5); full names in 200 px, collapsible to 48 px of icons with ≡ or Ctrl+B (state saved at once; a collapsed icon's tooltip names its section before describing it; amended 2026-10-16, R1), badge counts (warnings, running jobs), hover tooltip with section description. The Settings icon is a toothed gear (8 teeth, centre hole), the same in Chronos, the Uploader and the admin panel; the circle with rays is the Sun icon only (amended 2026-10-26, R4).
- Fluent theme, system/dark/light, system accent; density "comfortable"; 8 px grid; one typeface (Segoe UI Variable, system fallback for CJK/Arabic/etc.).
- Window chrome custom (dark title bar), min size 1200×760; per-monitor DPI.
- Keyboard: standard shortcuts, listed in tooltips ("Render · Ctrl+R").
- Notifications: unobtrusive toasts inside the window (bottom-right) for job completion, card detected, server events; optional OS toast when minimised.
- Modals only for destructive confirmations and login.
12.2 Library (home)
│ [🔍 Search title, camera, tag] [Year ▾] [Camera ▾] [Status ▾] [Tags ▾] [Sort ▾] [⊞][☰][▦] [⟳] │
├──────────────────────────────────────────────────────────────────────────────────────────────┤
│ ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐ │
│ │ hero thumb │ │ │ │ │ │ │ │ │ │ │ │
│ │ │ │ │ │ │ │ │ │ │ │ │ │
│ ├────────────┤ ├────────────┤ ├────────────┤ ├────────────┤ ├────────────┤ ├────────────┤ │
│ │Harbour sun…│ │2026-09-27 │ │… │ │ │ │ │ │ │ │
│ │3,600 · R5 │ │1,020 · GoPr│ │ │ │ │ │ │ │ │ │
│ │✓ ▣ ▶2 ☁ │ │✓ ▣ · ☁ │ │ │ │ │ │ │ │ │ │
│ └────────────┘ └────────────┘ └────────────┘ └────────────┘ └────────────┘ └────────────┘ │
- Unit of the library = sequence (never a flat list of 600k photos). Views: grid (default), list (sortable columns), timeline (year/month bands with density).
- Card: hero thumbnail, title (user title or
{date} {camera}), count · camera · span, status chips:✓analysed,▣proxies,▶nrenders,☁synced/origin,!warnings. Hover reveals quick actions: Open, Analyse, Render preset, Sync. - Multi-select → mass actions bar (Analyse, Build proxies, Verify, Tag, Rate, Reorganize…, Upload, Download, Delete, Export list as CSV). Reorganize… (amended 2026-10-26, R7; formerly Move) takes the selected sequences of one archive, never a server-only one, into the dialog of §12.5.
- Filters are chips; saved views (e.g. "Night, 4K, not rendered"); search is instant
(indexed columns); everything virtualized (in-house
VirtualizingWrapPanel). - Photos deleted from a sequence (amended 2026-10-26, R6; on the admin panel, §14.8, or by another program sharing the archive) leave it at its next index sync: its count follows, a cover whose photo went is cleared, the project's positions on those photos (exclusions, hero, keyframes) wait as pending frames until a restore brings the photos back, and the others follow their photos without a settings edit. Delete… still moves whole sequences into the archive's trash, which lists photo entries too (§12.5).
12.3 Edit (amended 2026-10-16, R7; formerly "Sequence editor")
│ ← Harbour sunset ▾ · 3,600 frames · Canon EOS R5 · 2 s interval · 2 h 00 m [Analyse] [Optimise all] [Render ▾] │
│ ✓ Analysis ── ▲ Frames ── ✓ Exposure ── ✓ Motion ── ✓ Camera ── ✓ Time ── ⑦ Stack ── ⑧ Export │ ← the path
├──────────────────────────────────────────────────────────────────┬──────────────────────────────────┤
│ ┌──────────────────────────────────────────────────────────────┐ │ Exposure │
│ │ PREVIEW │ │ Deflicker [●] Auto 41 ⓘ │
│ │ [◧ compare] [⊹ vectors] [▭ regions] [▦ crop mask] [1:1] [⛶] │ │ Strength ───●──── 0.8 │
│ └──────────────────────────────────────────────────────────────┘ │ Transitions [●] Auto 18 ⓘ │
│ ▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏▏ filmstrip + markers │ [Optimise section] [Reset] │
│ ┌──────────────────────────────────────────────────────────────┐ │ │
│ │ the step's measuring tool (charts, frame table, curves) │ │ [‹ Previous] 3 / 8 [Next ›] │
│ └──────────────────────────────────────────────────────────────┘ │ │
- One sequence at a time. Without an open sequence, or after Change sequence (the title's ▾, Ctrl+O, or Next on the last step), Edit shows a picker: this PC's sequences and the signed-in server's, newest first, searched by title, camera or tag, the open sequence (else the one edited last) selected and in view; Enter (on what was typed) or a double-click opens one, Esc clears the search, then goes back. The next sequence is read first: the open one stays open behind the picker until another one opens (a server sequence whose proxies are still coming, or one that cannot open, leaves it open), its last edit saved first, or discarded on the user's word and then never written. The Library, Jobs, the Server screen, project files and the command line still open sequences in Edit; ← goes back to the Library.
- The steps form a guided path, freely navigable: Analysis (every finding of the analyses with the step that deals with it; light, flicker, exposure steps received, shake and flagged frames; Analyse), Frames (virtualized table: index, time, exposure, ISO, f, size, tier, flags; exclude/include; set hero; the findings about single frames), Exposure (deflicker, regions, holy-grail transitions, WB) with luminance/waveform charts, Motion (blur, flow quality, stabilization) with shake path chart, Camera (crop, aspect, virtual camera keyframes with a spline editor), Time (fps, timing mode, remap curve, fit-to-duration, plan chart), Stack (median/max, trail length; optional), Export (preset, values, output name, batch list, render history, sponsor slot — Phase 9 only).
- Each step shows where it stands, by its own glyph as well as its colour: to do (its number), done ✓ (nothing is missing: Auto, or Edited once a setting or value of the step was set by hand), attention ▲ (a saved analysis of other frames or one that cannot be read now, frames still flagged, data an Auto setting that applies lacks, a render blocked or none made from the current settings), running ◐ (its analysis or render), waiting ○ (it needs the deep analysis). Statuses guide; they never block a step. Previous and Next (Ctrl+PgUp/PgDn), Alt+1…8 and the step bar (←/→) move between steps, the keyboard following; each step's tooltip starts with its name and ends with its shortcut, and screen readers hear the step and its state. The editor opens on the step last shown for that sequence in this run; else on Analysis while the sequence has no current deep analysis or one waits, runs or is paused; else on Frames.
- Preview always on top with tools: zoom/pan (1:1 shows raw pixels), wipe compare original/corrected, motion-vector overlay, region boundary, crop mask ("what I lose" vs "how it will look"), full-screen native-resolution window with collapsible shot data.
- Every deducible control shows Auto/manual state (§10.4); every control has a tooltip with the reason and the effect of moving it the wrong way.
- While a job runs on this sequence, settings that would change the rules mid-render are disabled; viewing tools stay live.
12.4 Import (cards)
│ Devices │ Profile "EOS R5 · card A" (SD 128 GB · serial 3F2A-9C11 · Canon EOS R5 #0431) [Edit ▾] [New] │
│ ● SD 128 GB EOS R5 card A ▲ │ Destination (•) Archive ( ) Folder [D:\Shoots …] ( ) Server ☁ "Studio NAS" │
│ ● microSD 64 GB GoPro 12 │ Scan [✓] Subfolders Sessions [✓] gap [45] min · min [8] frames [✓] Verify hashes │
│ ○ E:\ no DCIM │ Names {year}/{date} {camera} / {date}_{time}_{n:0000} → 2026/2026-09-28 Canon EOS R5/2026-09-28_18-02-11_0001.CR2 │
│ ○ Folder… │ On existing (•) Skip ( ) Rename ( ) Overwrite Convert (•) Copy ( ) TIFF-16 ( ) Linear DNG ( ) JPEG │
│ │ After import (•) Keep card ( ) Delete imported ( ) Format card [✓] Eject [✓] Deep analysis │
│ [✓] Import automatically when │ Plan 3 sessions · 4,812 photos · 96.2 GB · 12 min [Show plan] [ Import ▶ ] │
│ a known card is inserted │ │
- Devices list: removable volumes and any volume with
DCIM; identity = volume serial + label + capacity, camera identity from the first file; known cards show their nickname and profile; unknown cards prompt for a nickname (one field). - Plan before touching disk: sessions, folders to be created, first three file names, duplicates already in the archive (by hash), disk space check.
- A picked folder (
Folder…, and Archive › Adopt folder › Copy into archive…) adds a Group row: Sessions (as for a card) or Whole folder (amended 2026-10-26, R9). Whole folder makes the folder (its subfolders too while Subfolders is on) one sequence per file format: camera bodies and pauses are ignored, formats are not (RAW+JPEG of the same shots stays two sequences, so no frame is doubled). Sessions, gap and minimum are then disabled, their tooltips saying why. The title defaults to the folder's name (a typed title wins; it also fills{title}); a repeated import gives the folder's name only to a sequence without a title, so a rename made since survives. Each plan row shows the camera, or "n cameras", and the format, and both dates when its photos span several days; Split in Show plan cuts a part off, Merge joins it back. A folder starts on Sessions every time (folders have no profile); cards, drives and the admin panel have no Group choice. - Copy with hash while streaming; verify by re-reading destination; only then delete or
format (OS dialog); eject via
IOCTL_STORAGE_EJECT_MEDIAafterFSCTL_LOCK_VOLUME. - Destination "Server" streams frames through the client to the server (§14.4); the server confirms each hash before the client counts it as imported.
- Auto-import on insertion (per card), with the OS toast "Card X detected — Import".
12.5 Archive manager
│ [Archive: D:\ChronosArchive ▾] [Scan] [Rebuild index] [Verify] [Duplicates] [Orphans] [Adopt folder…] [Maintenance ▾] │
├──────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ 618,204 photos · 412 sequences · 9.4 TB · 14 cameras · 2019–2026 │
│ ▆▇█▅▂▁▃▆ photos per month (heatmap) ██████░░ storage by year ▉▉▉▍ by camera ▉▉▎ by format │
│ Index health: ✓ sidecars 412/412 · ⚠ 3 sequences without proxies · ⚠ 1 unverified since 90 days │
│ Recent maintenance: 2026-09-20 verify ✓ · 2026-09-01 vacuum ✓ │
├──────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Mass operations on filtered set (412): [Analyse] [Proxies] [Thumbnails] [Verify] [Reorganize…] [Retag] [Export CSV] [Delete] │
Operations: Scan (incremental, mtime/size), Rebuild index (from sidecars; falls back to full metadata scan; never deletes originals), Verify (hashes; bit-rot report), Duplicates (by hash across sequences), Orphans (files not indexed / rows without files, with fix actions), Adopt folder (index an existing tree in place or copy into the layout), Merge/split sequences, Move archive (relink), Maintenance (SQLite vacuum/backup/restore, cache cleanup, proxy purge/rebuild, thumbnail pack compaction, log rotation). Statistics drawn with in-house chart elements. Every operation is a Job (resumable, cancellable, logged). Recurring scans, integrity checks, thumbnails, proxies, analyses, maintenance, backups, clean-ups and reports are schedules in Jobs › Scheduled (§12.6, §14.6); the buttons here run them once (amended 2026-10-26, R10).
Maintenance › Trash… (amended 2026-10-26, R6) lists the archive's trash (§7), newest
first: deleted sequences and photos deleted from sequences, each with its kind, photos,
size, folder and when it went, and Restore. A sequence goes back as its folder is (into
its old place, or "folder (2)" when that is taken; its photos are not re-hashed, Verify does
that). Photos are re-hashed against their record and go back into their sequence with their
frame records (a name taken since gets _2, its companion following), or into a new
sequence when theirs is gone, every entry of that sequence joining the same new one. A photo
that stays in the trash says why (its files no longer match their record, they could not be
moved back, the sequence holds that content already, or its file was missing when it was
deleted): one sentence on the dialog's line, or how many stayed when the reasons differ, and
the tooltip with each reason and the first names. When no file of an entry matches, nothing
moves and the line says so. Empty… measures every entry (dry run), lists them, asks for the
number of entries to be typed, then queues the Empty trash job. The job erases exactly
those entries and nothing when the trash changed since; an entry that changes while it runs
is kept, and the job then fails after erasing the others. It is the only way the product
erases trash entries; no schedule runs it (§14.6).
Reorganize… (amended 2026-10-26, R7; it replaces Rename… and Move…) applies new templates to sequences already filed: the set (online), the Library's selection (§12.2), or every sequence of the archive from the toast of Settings › Archive (§12.8). The admin panel offers the same (§14.8), through the same engine and job.
- An in-window dialog: Folders (
{title}is the new title), File names (as filing names them, §7.1) and Titles ({title}is the current title), each with its switch and field, checked as it is typed against the rules of §7.1; a template switched off never blocks the others. Folders and File names start on with the archive's templates, read again when the dialog opens; Titles starts off, with{title}. - The dry run (nothing touched) follows the typing (400 ms; Enter, or leaving a field, at
once) and lists, in the order the changes apply, each sequence with its folder, title and
first three file names before → after, Follows its folder for a sequence nested in a folder
that moves, how many files get a new name, and the photos that keep their names with why:
kept in place outside
originals/(a tree someone else keeps), never verified (verify the sequence first), missing; or that the sequence stays as it is, with every folder around it, because its sequence file cannot be read (damaged, another sequence's, or from a newer Chronos). The summary reads "12 of 36 change · 3,600 files renamed · 20 names kept". - Apply (enabled only for the dry run of the templates on screen, when something changes) queues the Reorganize job of exactly that plan at user priority; the dialog closes and a toast links to Jobs. Cancel is the default button.
- The job (desktop and server) plans again and applies the plan only while it is the one confirmed (its fingerprint, §14.3); otherwise it fails and nothing changes, a sequence deleted since named in its log. It changes one step at a time. Pause or cancel stops between two steps and holds the rest for the job's resume or retry; a failed job holds it too (also when the archive's volume went away), and its retry completes it while it still fits. A retry while the archive is still offline fails as offline and keeps the plan: a plan journal it cannot reach is never taken for one completed. An archive open completes a plan an interruption left (a crash, a yield to another job, the program closing), not one its job holds.
- What changed after the dry run: a sequence moved or gone, one indexed inside a folder that moves, a new folder taken, a title edited that the plan changes or names a folder or files after, a sidecar of a newer Chronos it would rewrite: the rest of the plan is refused, the steps done stay, and a new dry run is needed. Photos changed since (a sequence still receiving them) refuse nothing: that sequence keeps its file names, and its folder and title change while the templates still give them from its first photo as it is now, otherwise it stays as it was, and so do the sequences its folder was to carry. A sequence left as it was inside a folder that a later step moves goes with that folder, keeping its title and file names, so the folder around it still moves. A photo in use or a new name taken keeps that sequence's file names (every file back). These sequences are listed in the job's log, and the job fails at the end, after the others changed, saying how many; its retry plans again.
- Deleting sequences of the archive is refused while a Reorganize job runs or is paused. The Archive and Library screens read the archive again when it ends; an editor open on a sequence it changed reloads where the sequence is now after that sequence's step.
12.6 Jobs
Two tabs, a Queue | Scheduled segmented control first in the toolbar (amended 2026-10-26, R10; a tab, not a rail entry, so a schedule and the jobs it queued are side by side):
- Queue: queue + history table: kind, target, progress, fps, ETA, node, state; pause/resume/cancel/retry; log tail on expand; filters. A job a schedule queued carries a Scheduled pill. The history line gives the job-history days of the Housekeeping schedule (Full history kept while it is off or at 0).
- Scheduled: every schedule of this PC, the system ones first: Archive scan (every 10 minutes, or the interval the former Settings › Archive › Rescan every had, no catch-up; a scan follows Settings › Analysis › After a scan at every run; it also runs once after the archive is reopened at start and when an unplugged archive is back), Index backup (every day, catch-up, 3 copies kept) and Housekeeping (daily at 04:00, catch-up: job history older than 90 days, the cache trimmed to Settings › Storage › Cache limit, oldest files first, never files in use, synced proxies, the render node's cache or import staging). Each row: an on/off switch, name (a system schedule by its key, in the UI language), action, when (in the UI language), next run (Today, Tomorrow or a date, and a time), last result (the state of the job it queued, else the outcome), Run now, Edit, and Reset (system schedules) or Delete (the user's), both confirmed. Selecting a schedule lists its last runs; Show in queue opens the job in the Queue tab with its log. Which actions this PC runs is read again when the tab is shown and after a Run now that finds one unavailable; without an index the tab says the schedules do not run.
- New schedule (the toolbar's end) and Edit open an in-window editor: name (a system schedule keeps its name and action); action (those this PC cannot run disabled, the reason in the tooltip); Repeat Every / Daily / Weekly / Monthly with their fields (interval 5 minutes to a day, 1–168 hours or 1–366 days; a time of this PC's clock; days of the week in the UI language's order; day 1–28 of the month); the action's parameters (full scan, rebuild all, deep analysis, unverified for or older than n days, at most a share of the sequences, report only, copies kept, clean-up steps); priority; catch-up; on. Every field is checked as it is typed.
- No work recurs out of sight: every timer that queues work is a schedule here (§14.6). A schedule's background run raises a toast only when it fails, unless it is upkeep the next run repeats (scan, thumbnails, proxies, light analysis, clean-up); Run now raises one only when it queued nothing.
Global job bar mirrors the active job.
12.7 Server (client side)
│ Server [https://nas.local:8443 ……] [Connect] Status ● Connected · v1.3 · TLS fingerprint 9A:… (trusted 2026-09-01) │
│ Signed in as alberto · Editor · 2.1 TB of 5 TB used [Sign out] │
│ [✓] Offer this PC as render node (NVENC HEVC/H.264, AV1) [✓] Keep proxies of synced sequences offline │
├───────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Sequence Local Server State Action │
│ Harbour sunset ✓ originals ✓ in sync — │
│ Comet 2026-08-13 ✓ originals ✗ local only [Upload] │
│ Construction cam wk 38 proxies ✓ remote [Download originals] [Open with proxies]│
First connection shows the certificate fingerprint (trust-on-first-use); login modal; per-sequence sync actions and states; conflict resolution only for settings JSON (keep mine / take server / save both).
12.8 Settings
Sections (list on the left, controls on the right, no descriptions — tooltips only): General (language, theme, startup, notifications), Archive (root, templates, proxy policy, thumbnail size; the rescans are the Archive scan schedule of Jobs › Scheduled, §12.6, so Rescan every is gone; amended 2026-10-26, R10; a new folder or file template names only what is filed next, so once the archive took it, and holds sequences, a toast offers Reorganize… for every sequence with the changed templates switched on, a change made while that archive's toast is up joining it; on the Archive screen it never replaces an open dialog, and opens once that one closes: amended 2026-10-26, R7), Import (defaults, session rules, verify, post-actions), Analysis (widths, thresholds, deep analysis auto-run), Rendering (quality profile, memory budget, spill folder, parallelism = photos the running job processes at once, priority queue with a level per job kind, encoders detected, hardware preference; amended 2026-10-16, R2), Formats (per format tier on this PC, codec install links), Server (last servers, trusted fingerprints, render node), Updates (check on start; latest version: check now, download, releases page), Advanced (portable mode, data folder; logs: the logs folder, Log level (Trace…Error, default Info), Keep logs (1–365 days, default 14), File size (MB, default 50, or No limit), Logs total (MB, default 1,024, or No limit; never below the file size), Rotation (Daily or Hourly, UTC), Diagnostics logs (MB of the end of each log in the bundle, default 4), applied to the log files from their first line and at once when changed; diagnostics bundle, database backup/restore (Back up now keeps as many copies as the backup schedules, §6.2, R10), reset; Reset restores the log rows; amended 2026-10-26, R11), About last (§12.9). Each action has one place in Settings (amended 2026-10-26, R5). Import/export settings as JSON.
12.9 About
The last page of Settings (§12.8), no longer an entry of the rail (amended 2026-10-26, R5): version, build, runtime, system, licence/EULA link, third-party notice (none) and, in every language but English, the machine-translation notice with its review link. Nothing is offered twice: the update check, its notice and the releases page are on Updates, the data folder and the diagnostics bundle button (zips logs + environment report) on Advanced. The update check on start (§18) runs whatever screen is open.
13. Card import — behaviour details
- Detection:
WM_DEVICECHANGE(DBT_DEVICEARRIVAL/REMOVECOMPLETE) +DriveInfopoll fallback; card identity{volumeSerial}-{label}-{capacity}; camera identity from EXIF Make/Model/BodySerial of the first image found. - Card pool (
media_cards): nickname, profile, history; profiles copyable; a card can inherit a "camera profile" (rules shared by all cards of that camera). - Session detection: gap >
SessionGapMinutes; groups with <MinimumSessionFramesmerge into the previous session (test shots); each session becomes a sequence; the user can merge/split in the plan. A folder imported as Whole folder (§12.4) skips detection: one session per file format whatever the bodies and pauses, broken only by the user's splits. Its session key isfolder-+ the first 16 hex digits of SHA-256(PC name | full path without trailing separator, upper-cased on Windows | format), plus-yyyyMMdd-HHmmssof its first photo for a part split off; so an interrupted import resumes into the same sequence and a later import of the folder adds to it. The PC name is in the key because the server files uploads by session key for every user alike (amended 2026-10-26, R9). - Duplicate handling: hash lookup against the archive index (skip/rename/overwrite); "already imported" state remembered per card + file (path, size, mtime) for fast re-inserts.
- Rename during import: templates (§7.1) with live preview; original name kept in the sidecar. Every photo of a session is named with its first photo's camera and lens, the ones its sequence carries, as Reorganize names it (amended 2026-10-26, R7).
- Post-actions: keep / delete imported files (only those verified) / format card (OS dialog) / eject; run deep analysis automatically (default on, background, low priority).
- Failure handling: card removed mid-copy → partial files deleted, plan resumable on re-insert; errors listed per file, never a modal per error.
14. Chronos Server
14.1 Stack
ASP.NET Core minimal API on Kestrel (Microsoft.AspNetCore.App shared framework, no
packages), System.Text.Json source-generated contexts from Chronos.Protocol, WebSockets
for events, in-house structured file logging (Microsoft.Extensions.Logging abstractions
are part of the shared framework; the file provider is in-house), in-house rate limiting for
login and uploads, health endpoint, graceful shutdown (jobs checkpoint).
Logging (amended 2026-10-26, R11): files /data/logs/server-yyyyMMdd.log, or
server-yyyyMMdd-HH.log with hourly rotation (UTC). A file that reaches the size limit (50
MB by default, MB of 2²⁰ bytes; 0 = no limit) continues in .1.log, .2.log…, and a
restart continues the period's newest part. At every new file, and when the settings
change, files older than the retention (14 days) are deleted, then the oldest while all
the files exceed the total limit (1,024 MB; 0 = no limit; never below one file), never
the file being written and never a name the sink did not write. The console and the files
have their own levels (Information by default), which the logging filters read at every
line, so a change in the admin panel applies at once; ASP.NET Core's own categories never
write below Warning, the host's lifetime never below Information; the engine's level is
the lower of the console's and the files'. Logging never fails what logs: a line that
cannot be written (the folder removed, a full disk) is dropped and the next one tries the
file again (a removed folder is not created again), so a job still ends Failed and the job
engine and the scheduler go on. The desktop programs use the same sink (chronos-,
uploader-, setup- files, with the default limits; Chronos takes its own from Settings ›
Advanced, §12.8).
14.2 Security
- TLS:
auto(self-signed cert generated at first start withCertificateRequest, stored in/data/tls, fingerprint printed in logs and shown in admin),pem(cert + key paths),proxy(HTTP behind a reverse proxy,ForwardedHeadersmiddleware, admin warns if not local). - Passwords: PBKDF2-SHA512, ≥ 210,000 iterations, 16-byte salt, versioned hash format;
minimum length 10; admin-created users get
mustChangePassword. - Tokens: 32 random bytes, stored as SHA-256,
Authorization: Bearer; sliding expiry (default 30 days), per-client name, revocation list; web admin uses anHttpOnly; Secure; SameSite=Strictcookie session + double-submit CSRF token. - Authorization: roles admin/editor/uploader/viewer + permission bits; sequence ownership; admin-only management endpoints; path traversal impossible (ids → paths via DB only).
- Limits: max upload size, chunk size (default 8 MiB), per-user quota, request timeouts, login rate limit (5/min/IP), audit log.
14.3 API v1 (JSON; errors as { code, message, details }; X-Chronos-Api: 1)
POST /api/v1/auth/login {username,password,clientName} → {token,user,mustChangePassword}
POST /api/v1/auth/logout · POST /auth/password · GET /auth/me
GET /api/v1/server/info name, version, apiVersion, capabilities, tlsFingerprint, formats, limits, extensions
GET /health
GET /api/v1/users · POST · GET/PATCH/DELETE /users/{id} · POST /users/{id}/reset-password (admin)
GET /api/v1/sequences?year&camera&status&rendered&tag&q&cursor&limit
GET /api/v1/sequences/facets years, camera models (with their label) and tags of the archive, each with its count
POST /api/v1/sequences create (title, camera hint) → id
GET /api/v1/sequences/{id} · PATCH (title, tags, rating, settings) · DELETE
GET /api/v1/sequences/{id}/manifest frames with hashes/sizes/versions, tombstones (sync)
GET /api/v1/sequences/{id}/analysis · GET /stats · GET /renders
GET /api/v1/sequences/{id}/frames?cursor&limit
GET /api/v1/sequences/{id}/thumb the hero frame's thumbnail, else the first frame's (ETag, 304)
POST /api/v1/sequences/{id}/frames/delete {frameIds, expectedFramesVersion, dryRun} → {photos, bytes, framesVersion, frameCount, trashEntry}
POST /api/v1/sequences/bulk/delete {ids, dryRun} · /bulk/retag {ids, add, remove} · /bulk/jobs {ids, kind} → an item per id
GET /api/v1/frames/{id}/original · /proxy · /thumb (streamed, Range supported)
POST /api/v1/uploads {sequenceId|null, fileName, size, sha256, capturedAt?, sessionKey?, restore?} → {uploadId, chunkSize, received[]} | 200 {alreadyExists, frameId?}
PUT /api/v1/uploads/{id}/chunks/{n} body bytes (idempotent) → {received}
POST /api/v1/uploads/{id}/complete[?restore=true] → frame id (server verifies sha256, commits into layout, updates manifest)
GET /api/v1/uploads/{id} resume state
DELETE /api/v1/uploads/{id}
GET /api/v1/jobs?state · POST /jobs {kind, target, params} · GET /jobs/{id} · POST /jobs/{id}/cancel · /retry
GET /api/v1/schedules schedules (system ones first), actions with availability, time zone, clean-up steps (admin)
POST /api/v1/schedules · GET/PUT/PATCH/DELETE /schedules/{id} · POST /schedules/{id}/run · /reset · GET /schedules/{id}/runs?cursor&limit
WS /api/v1/events {type: jobProgress|jobState|sequenceChanged|cardEvent|nodeState|scheduleChanged, …}
POST /api/v1/nodes/register · POST /nodes/{id}/heartbeat · POST /nodes/{id}/claim → job · POST /nodes/{id}/jobs/{jobId}/result (chunked upload of output)
GET/PUT /api/v1/archive/settings · POST /archive/scan · /rebuild · /verify · GET /archive/stats · POST /archive/maintenance (admin)
POST /api/v1/archive/reorganize {sequenceIds|all, folderTemplate?, fileTemplate?, titleTemplate?, dryRun, fingerprint?} → plan | 202 move job (admin)
GET /api/v1/archive/trash · POST /archive/trash/{name}/restore · POST /archive/trash/empty {names?, dryRun, fingerprint} (admin)
GET /api/v1/i18n/{culture} strings for the web admin
Library and trash (amended 2026-10-26, R6):
frames/deleteneeds the Delete permission and the sequence's owner or an admin, like deleting the sequence. It takes 1–100,000 ids, under a body limit of its own (100,000 × 48 bytes + 4 KiB, about 4.8 MB; other JSON bodies stay at 1 MiB); a dry run answers the photos, bytes and the frames version now, a deletion the frames version after it (the next request'sexpectedFramesVersion), the frame count and the trash entry's name. Errors: 400bad_request(frameIds); 404not_found(sequence); 409frames_conflict(the frames changed sinceexpectedFramesVersion, or an id is not the sequence's; details: the frames version now),sequence_busy(a job works on it, as for deleting it),would_empty(every photo would go: delete the sequence),photo_unreadable(a photo never hashed whose original is on disk but cannot be read, so it could leave no tombstone; details: how many, the message naming the first),sidecar_mismatch,folder_in_use; 413too_large; 503archive_offline. Auditedframes_deleted; publishessequenceChangedframes.- The bulk routes take 1–1,000 distinct ids and answer
{items, ok, skipped, dryRun, photos, bytes, job}, one item per id in the request's order:ok, orskippedwith the reason's code and sentence. delete (Delete) moves each sequence asDELETE /sequences/{id}does, whole or not at all, and skipsnot_found,forbidden(another user's, unless an admin),sequence_busy,archive_root,contains_sequences,sidecar_mismatch,folder_missing,archive_offline,folder_in_use. retag (Edit) is the PATCH tag edit (trimmed, at most 64 characters, at most 50 a sequence, case-insensitive, up to 3 attempts on a settings conflict) and skipsnot_found,bad_request,settings_conflict. jobs (Edit; kindanalyse,deep,proxies,thumbnailsorverify) queues one job at user priority over the sequences found (202; 200 when none), a verify targeting their archive with the sequences in its parameters, asPOST /jobswithverifynow does too; 400invalid_job_kind, 409job_unavailable. 400bad_request(ids,tags). GET /archive/trash(ManageArchive, like the rest of the trash) answers the entries newest first,{name, kind: sequence|photos|replaced, sequenceId, title, folder, photos, files, bytes, deletedBy, deletedAt}, and their total bytes. A name is<32 hex>[-<14 digits>[-n]](a sequence) orphotos-<32 hex>-<14 digits>[-n](photos), checked against that shape and never a path.- Restore answers
{name, kind, sequenceId, photos, kept, newSequence, keptReasons, message}: the photos that stayed and, in the same order, why (mismatch,not_moved,in_sequence,no_file), with one sentence per reason in the request's language. Errors: 404not_found(trash_entry); 409sequence_in_library,sequence_busy,sidecar_mismatch(the sidecar of the sequence the photos return to is damaged or another's),folder_in_use; 422trash_mismatch(no file matches its record any more, or a deleted sequence's own sidecar no longer names it); 503archive_offline. Auditedtrash_restored; publishessequenceChangedcreated(a sequence, or a new one) orframes. - Empty: a dry run answers the entries measured on disk, their photos, files and bytes, the
plan's
fingerprintand the names asked for that the trash does not hold (unknown); the erase sends the fingerprint back and answers 202 with theempty-trashjob. Errors: 400bad_request(names,fingerprint); 409trash_changed(details: the fingerprint now),job_unavailable; 503archive_offline(offline, and for the erase also read-only; the listing answers it too, offline). The request is auditedtrash_empty_requested; what the job erased is in its log. - The manifest gains
removed: [{sha256, replacement, reason: deleted|replaced, at}](null when there are none; covered by its ETag, since a deletion moves the frames version).
Reorganize (amended 2026-10-26, R7; ManageArchive; §7, §12.5, §14.8):
- The request names exactly one of
sequenceIds(1–1,000) andall; a null template keeps that part, at least one is given, each usable and within the rules of §7.1. A dry run answers{steps, sequences, files, kept, fingerprint}and changes nothing: a step for each sequence that changes, is carried or keeps photo names, in the order they apply,{sequenceId, fromFolder, toFolder, fromTitle, toTitle, carried, framesVersion, fileChanges, fileSamples: [{from, to}] (at most 3, relative to the sequence folder), blocked: [{reason: in_place|unhashed|missing|unreadable, photos, message}]}(unreadable: the whole sequence stays, 0 photos; the message in the request's language). The change sends the dry run's fingerprint back and answers 202 with themovejob. Errors: 400bad_request(sequenceIds,templates,fingerprint),invalid_template(below); 404not_found(sequence); 409nested_target(details: the folder that would lie inside a sequence's folder),plan_changed(details: the fingerprint now),job_unavailable; 503archive_offline,plan_dropped. Auditedarchive_reorganized(sequences, files renamed, templates). The job publishessequenceChangedwith the changereorganizedfor each sequence once its step is journaled (at once for a step that failed after its folder moved and its index rows followed); a plan the server completes as it starts tells no one, as no client is connected yet. - The fingerprint is a SHA-256 of the steps as the dry run shows them (folders, titles, carried, file counts, samples, what stays and why), every file rename by its digest, and the file template; not the frames version, so photos a sequence receives under the names the plan leaves them keep it.
- The server keeps a dry run's plan, and a refusal (
nested_target,not_found), for 5 minutes, keyed by the archive, the selection (all, or the ids in their order, which decides which folder gets " (2)"), the templates, and a stamp of what the plan reads from the index: every sequence's id and folder; the planned sequences' title, first capture, camera, lens, photo count, frames version and how many of their photos are hashed; the title of the others inside their folders. The same dry run again, and its confirmation, take the kept plan. A plan is made apart from its request (one a proxy cuts off leaves it to be made, and asking again waits for it), at most 2 at once and 8 kept or being made; one still being made 15 minutes after it was asked, let go for newer ones, or when the server stops is stopped, and its requests answer 503plan_dropped. A change made on disk alone is not in the stamp: the confirmation then answers 202, and the job, which plans again, fails without changing anything. invalid_template, here and fromPUT /archive/settings: the details name the field and, after a colon, the rule a template that parses breaks (folderTemplate:reserved_folder,fileTemplate:file_separator,fileTemplate:extension,fileTemplate:not_unique), the message saying why in the request's language (Errors.Template.<Rule>); a blank, invalid or unusable template keeps the bare field and the generic message.PUT /archive/settingsholds a template to the rules only when the save changes it; every template must still parse and expand.- Writers that file into a sequence's folder take it from the index row read under the
sequence's lease, so they write where a Reorganize put it. Completing a render upload or a
render node's result that cannot be filed for now answers 503 and keeps the upload
receiving, to be completed again:
sequence_busy(the sequence held past the lease wait),archive_offline(the archive offline or read-only) orfolder_moved(the sequence's folder not where the index has it, or holding another sequence, after another program moved them; the server's next scan finds it). The node keeps its job while it retries. A proxy upload answers 503sequence_busywhile the sequence is held, 404framewhen the sequence or its folder is gone, or the folder is another sequence's.
Uploads (amended 2026-10-26, R3, R6): server/info lists extensions, the photo file
extensions the archive files (lower case, no dot), which the admin panel's dialog keeps to. A
frame upload whose file name starts with . is 400 invalid_file_name. A start whose
SHA-256 the target sequence tombstoned (and none of its frames holds) answers 200
alreadyExists without a frameId, unless restore is true; the completion checks again
unless ?restore=true and answers 409 photo_removed, the upload cancelled (§14.4).
Schedules (amended 2026-10-26, R10; ManageArchive, every change audited): PUT and
PATCH change the fields given; an invalid field is 400 invalid_schedule with the field
as details (name, recurrence, parameters.percent…); a system schedule is never
deleted (409 schedule_system) and keeps its name and action; run now is refused while the
schedule's last job is queued, running or paused (409 schedule_pending, nothing recorded);
only system schedules reset (409 schedule_no_defaults); changing an age limit an
environment variable fixes, or turning that schedule off while the limit is above 0, is 409
setting_locked.
JobDto.scheduleId names the schedule that queued a job; POST /jobs with kind
duplicates or orphans queues a read-only report job. scheduleChanged {scheduleId, deleted} follows every create, edit, reset, delete and run (also one that queued nothing)
and every move of a next run; older clients ignore it. GET/PUT /server/settings carries
the log group logs (R11; a PUT without it keeps the log settings) and lockedBy, the
environment variable that fixes each locked setting (ignored when saving).
14.4 Upload and ingest
Client computes SHA-256 while reading; POST /uploads short-circuits if the hash exists;
chunks are written to /data/uploads/{id}.part with a received-bitmap; complete verifies
the hash, moves the file into originals/ under the layout templates (temp → fsync →
rename), extracts metadata, appends to sequence.json, inserts the DB row, enqueues
thumbs/proxies/light analysis, emits sequenceChanged. Frames without a target sequence
are grouped by the same session rules (§13) into auto-created sequences. Card imports from
a connected client use exactly this path. POST /uploads answers 507 insufficient_storage
when /data lacks the upload's size plus a 64 MiB reserve; before its staging copy,
complete checks the archive's own volume (never the space path, §14.9) the same way and
answers 507 with the upload still receiving, so the client completes it again once there is
room; render uploads and render-node results are checked the same way at completion, and a
volume that cannot be measured refuses nothing (amended 2026-10-26, R2).
Deleted photos (amended 2026-10-26, R6): a photo deleted from a sequence is not filed into it
again by an upload. POST /uploads answers a SHA-256 the target sequence tombstoned, and none
of its frames still holds, with alreadyExists and no frame id, so an old client syncing a
copy that still has it, or a watch folder sending it again, neither sends it nor counts a
failure. complete checks again once the upload stops receiving, and the ingest checks the
sidecar's tombstones a third time under the sequence's lease, which a deletion takes: a photo
deleted while its bytes were on the way is not filed, its upload is cancelled and the answer
is 409 photo_removed (the protocol client reports it as not sent). Only an explicit user
action brings such a photo back by upload: restore: true on the start and ?restore=true
on the completion, sent by the admin panel's Upload again alone (never by sync, the
Uploader or a card import); it is then filed into its own sequence whatever other sequences
hold, and filing it drops the tombstone, also after the trash was emptied. A frame upload
whose file name starts with . (the AppleDouble ._ files a Mac leaves on cards, hidden
files) is refused for every client (400 invalid_file_name), as the archive scanner leaves
such files out.
The admin panel uploads through the same path (amended 2026-10-26, R3; the dialog, §14.8):
- It hashes each photo in the browser with an in-house incremental SHA-256 (FIPS 180-4,
sha256.js) in a same-origin module worker (hash-worker.js, under the panel'sscript-src 'self') that reads the file in 8 MiB blocks. Before its first hash the worker runs the FIPS 180-4 known answers (the empty message, "abc", the 448- and 896-bit messages, one million "a"), whole and fed in uneven pieces, and, wherecrypto.subtleexists, compares a 1 MiB random buffer with it; on any mismatch it refuses every hash and the dialog uploads nothing (fail closed). The server verifies every upload again. - Each start carries the photo's capture time: the Exif DateTimeOriginal of a JPEG or a TIFF-based file (TIFF, DNG, most RAW formats) read from its first 256 KiB, else the file's modification time as local wall time without an offset; the ingest uses it only for a photo that records no time of its own (a PNG, an export stripped of Exif).
- At most two photos and two chunks per photo travel at once (four requests, half the
server's eight per account). A call that got no answer, 429 or 503 is retried after 1, 2
and 4 s; a completion refused with 409
upload_state(another completion is filing that upload) is read again after 1, 2, 4 and then every 8 s, about a minute in all. Two copies of one photo in a run go one after the other, so the second finds the first. alreadyExistswith a frame is Already there; without one, Deleted before, whose row offers Upload again between runs (therestoreabove). A quota or permission refusal (401, 403), no space (507), the target sequence gone (404sequence) or a failed self-test stops the remaining photos (Not sent) with the server's message. A new target sequence is created once (POST /sequences, never retried) and stays the dialog's target.- Cancel aborts, stops the worker and deletes the uploads the run started and did not complete; a photo that failed otherwise keeps its upload receiving, so the next run resumes it, and abandoned uploads expire after the upload retention. While photos are sent, leaving the page asks first and Esc is held, so only Cancel stops the run.
- Only files whose extension
server/infolists, not empty, not starting with.and no larger thanlimits.maxUploadBytesare taken. Folders dropped on the dialog are counted (Folders skipped) and left out, and it has no folder picker: importing a whole folder stays a desktop feature (R9).
14.5 Sync
Manifest diff per sequence: originals are content-addressed (never conflict); settings and tags carry version counters (last-writer-wins with an explicit conflict prompt in the client when both changed); proxies/thumbs are regenerable and never synced upward; renders are uploaded on request. "Edit with proxies, render where the originals are": a client can open a remote sequence with proxies only and submit a render job to the server, which runs it itself (managed formats) or hands it to a render node.
Removed photos (amended 2026-10-26, R6). A photo deleted on one side leaves a tombstone there
(sidecar removed, index frame_removals, manifest removed), and sync never offers it back:
no upload of a photo the server deleted, no download of one this PC deleted; sync never
deletes local files either way. A tombstone leaves a photo out only while its own side holds
no matching photo (by hash, or by size and capture time while unhashed), so a duplicate that
stays or a file put back by hand still syncs. A photo replaced by a converted file (R7)
counts as its replacement on both sides, an alias followed up to 8 steps, for the
comparison, the transfers and the frame positions. Content known only once hashed is checked
then: a download skips an unhashed server frame this PC deleted (or holds as its
replacement), and an upload skips an unhashed local frame the server deleted and then runs
the archive verify, so the index gets the hashes and the next comparison leaves it out.
Known limitations: until the server verifies such a frame, the sync table still counts it as
missing here and Download fetches and drops it each time; a photo never hashed whose
original was already missing when it was deleted leaves no tombstone, so a copy elsewhere
(another PC, a card) is not known to be deleted and can come back.
Folders moved by another program (amended 2026-10-26, R7). When a desktop and a server share an archive, a Reorganize by one index (or a folder moved by hand) is found by the other's next scan, which moves the sequence's row, its frames and the paths of the exports that moved with the folder (a file or, for an image sequence, a folder; an export written at the old place after the move keeps its path there, and another sequence's export saved inside follows only when it moved with the folder). The other program may also have moved another sequence into the old folder (its sidecar names that sequence): the moving sequence's exports then follow to its new folder, and an export of the sequence that took the old folder never follows, so no row names another sequence's file. Until that scan nothing writes into the old place, which would create it again, nor into another sequence that took it: Save both and Take server answer that the folder is not where the index has it (try again after the next scan), and so do a render job as it starts, exports fetched by sync, a proxy pass, and a thumbnail pass that finds another sequence's sidecar there; the editor's save says the sequence is no longer in the archive, and a deep analysis that loses its folder records nothing. A render of the other program whose folder moves while it runs is not stopped. Known limitation: an export a sequence saved into another sequence's folder (the Export folder picker), before the other program moved that sequence away and the first one into its folder, keeps naming the old path; the file is intact, moved with the other sequence.
14.6 Job engine
Persistent queue in the DB; jobs run one at a time, and the parallelism (configurable, default = cores/2) is how many photos the running job processes at once; job kinds in §5; checkpoints every 5 s (progress + resumable cursor); crash-safe restart; priorities (user-initiated > background analysis); cancellation cooperative; logs per job; finished jobs kept 90 days by the housekeeping schedule, below (amended 2026-10-26, R10). By default a new job waits its turn, and pausing the running job lets the next one start. An optional priority queue (Settings › Rendering on the desktop, the admin panel's Settings on the server) gives each job kind a level (high, normal, low): higher kinds are claimed first, and a queued job of a higher level makes the running job of a lower level yield without asking; that job returns to the queue with its checkpoint and resumes by itself afterwards. Jobs of the same level never interrupt each other, and a job's own follow-ups never interrupt it. (Amended 2026-10-16, R2.)
Scheduled jobs (amended 2026-10-26, R10). Recurring work is queued only by schedules (§5),
shown in Jobs › Scheduled on the desktop (§12.6) and the admin panel (§14.8): no job is
pre-scheduled out of sight. Each host runs one scheduler beside its job engine. Every timer
that used to queue work is a system schedule: seeded at start by its key when missing
(INSERT … ON CONFLICT (system_key) DO NOTHING), so the user's edits survive restarts and
upgrades; it can be edited, turned off and reset to the host's defaults (the options as
they are then), never deleted, and it keeps its name and action. On the server:
| Key | Action | When | Catch-up | Notes |
|---|---|---|---|---|
archive-scan |
incremental scan, background | every 10 min, and once at start | no | |
upload-cleanup |
uploads step: uploads idle for UploadRetentionHours (48) expire, orphan part files and day-old staging files go, records of uploads finished over 90 days ago and expired tokens are purged |
every 6 h | yes | an idle upload expires up to 6 h after its 48 h |
index-backup |
backup, 3 copies kept | daily 03:00 | yes | SQLite only (PostgreSQL uses its own tools) |
housekeeping |
job-history step, JobHistoryDays (90) |
daily 04:00 | yes | |
audit-cleanup |
audit step, AuditRetentionDays (0) |
daily 04:30 | yes | off at 0: the audit log is kept |
The server's options are what its system schedules are seeded and reset with. The audit
retention has one source, the audit-cleanup schedule: Settings › Logs › Keep audit
shows its age limit (0 while it is off) wherever it was last set, and saving another value
sets it (on with that limit, off at 0, no restart) and stores it as AuditRetentionDays,
the reset default. CHRONOS_AUDIT_RETENTION_DAYS is put into the schedule at every start
before the first pass, so a run caught up at start already uses it, and locks it (§14.3).
JobHistoryDays is likewise only the housekeeping schedule's reset default.
On the desktop: archive-scan (every 10 minutes, seeded from the former Rescan every; no
catch-up; also once after the archive is reopened at start and when an unplugged archive is
back), index-backup (every day, catch-up, 3 copies) and housekeeping (daily 04:00,
catch-up: job-history 90 days and cache-trim). A system schedule whose work was last
done a whole period ago (the newest index copy, for index-backup) is seeded due at once,
so the first start after an upgrade copies an index whose newest copy is a day old;
otherwise a new schedule first runs at its recurrence's first occurrence (an interval one
interval after it is created). Times are wall-clock times of the host's zone: the PC's, or
the container's TZ (UTC, with a warning, when the image does not know it), which the
admin panel's Scheduled tab names.
The scheduler looks every 30 s (the desktop's also at once when its archive is back or its
job settings are saved). A run is claimed by moving its next run with a compare-and-set on
next_run, so processes sharing a database fire it once (a crash between the claim and the
queueing loses that one run). A run queues the action's jobs at the schedule's priority
with jobs.schedule_id, at most 500 sequences per job; it queues nothing, and records why, while
the schedule's previous job, or a job doing the same work on the archive, is queued, running
or paused (skipped-pending), while the archive is missing or offline (or read-only, for an
action that writes: skipped-offline), when there is nothing to do (skipped-nothing), or
when the host cannot run the action (unavailable: no handler, PostgreSQL or an index
without a file for backups, proxies turned off). A run more than 5 minutes late (the host
was off, the PC slept) was missed: with catch-up it fires once, otherwise the schedule moves
to its next occurrence; intervals stay on the grid of their due times. A time a
spring-forward gap skips moves forward by the gap; a time the clocks go back over runs once,
at its first occurrence; a next run more than two periods ahead means the clock went back
and is computed again. Run now queues at once, also for a disabled schedule, without moving
the next run.
Actions: scan (force: full; deep: deep analysis of what it indexes, unset: as the host
says at the run, Settings › Analysis › After a scan on the desktop, off on the server),
verify (percent of the sequences per run, never-verified then least recently verified
first, those unverified for olderThanDays), thumbnails and proxies (missing ones, or
force all; proxies only when the archive makes them), analysis (light, or deep),
maintenance (dryRun), backup (keep), cleanup (steps, olderThanDays) and the
read-only duplicates and orphans reports in the job log. Clean-up steps: job-history
(finished jobs; the engine no longer purges them at start), audit (0 days keeps
everything), trash-report (lists trash entries older than the limit with their sizes and
never deletes: emptying the trash stays a user action, §0.2 rule 8), and the host's own
(uploads on the server, cache-trim on the desktop). A step that fails fails the job after
the others ran.
Live services are not schedules: they queue no recurring work and change no user data, they keep a connection, a device or the archive's settings current from moment to moment, and they show in their own screens. On the server: the read of archive.json every 10 minutes, whatever the scan schedule does (it only takes the settings a desktop sharing the archive wrote, so uploads follow them while scans are pending, rare or off), the render-node monitor and the LAN beacon. On the desktop: the sync agent (the transfers it decides are jobs in the queue), the render node, card and drive detection, the look for an unplugged archive every minute and the update check at start. The log files prune themselves (§14.1).
14.7 Render nodes
A Chronos client with "Offer this PC as render node" registers its encoder capabilities and polls/claims render jobs matching them, downloads what it needs (proxies or originals per job settings), renders with its GPU, uploads the result as a render of the sequence. Node state shown in admin › Jobs; the server falls back to managed formats when no node is available and says so in the job log.
14.8 Admin web panel (/admin, vanilla HTML/CSS/JS, ES modules, no build step, same API)
┌─────────────┬────────────────────────────────────────────────────────────────────────────┐
│ Chronos │ Dashboard │
│ ● Dashboard │ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ Library │ │ 9.4 TB │ │ 412 seq │ │ 3 jobs │ │ 2 nodes │ │
│ Users │ │ of 16 TB │ │ 618k pho │ │ running │ │ online │ │
│ Archive │ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │
│ Jobs │ Storage by year ▁▂▃▅▇█ Uploads last 30 days ▂▁▃▆▇▅▂ │
│ Logs │ Recent: alberto uploaded 1,020 frames · node "studio-pc" rendered 4K HEVC │
│ │ │
│ Settings │ │
│ (A) admin ⏻ │ │
└─────────────┴────────────────────────────────────────────────────────────────────────────┘
Pages: Login · Dashboard · Library (amended 2026-10-26, R6; below) · Users (table;
create/edit/disable/reset/roles/quota) · Archive (path, templates, proxy policy,
scan/rebuild/verify, stats, and Trash, which opens the Library's trash view; R6; a save
that changes a template offers Reorganize, R7) · Jobs
(two tabs in the header, amended 2026-10-26, R10: Queue, the queue/history/nodes with
cancel/retry, where
#/jobs?job=<id> opens a job's log and a job a schedule queued carries a Scheduled pill
linking to it; Scheduled (#/jobs?tab=schedules), every schedule of the server, its
system schedules first and marked System, with when it runs (in the panel's language), its
next run and last result (every time on the tab in the server's time zone, named there), an
on/off switch, Run now, Edit, and Delete (user schedules) or Reset (system schedules), both
confirmed, its runs paged under it, each linking to its job; New schedule and Edit open a
side panel: name, action (those the server cannot run disabled, the reason in the
tooltip), recurrence (interval, daily, weekly, monthly), the action's parameters, priority,
catch-up, on; a field the server refuses is marked; the tab follows scheduleChanged and
jobState events, so skipped runs and other administrators' edits appear at once) · Logs
(tail, download; schedule changes and runs in words, a system schedule by its name in the
panel's language) and, docked at the foot of the sidebar above the account, Settings
(server name, ports, TLS mode + fingerprint, discovery, jobs: parallelism and priority
queue, limits, language, retention (tokens, uploads; job history is the Job history
clean-up schedule's age limit, R10), Logs (amended 2026-10-26, R11: file level and console
level (Trace…None, default Information), keep logs (days), file size (MB), logs total (MB),
rotation (daily or hourly, UTC), keep audit (days; the Audit log clean-up schedule's age
limit, §14.6), diagnostics logs (MB), applied without a restart)). A field an environment
variable fixes is read-only, its tooltip naming that variable (lockedBy, §14.3; R11). The
last group of Settings, About, is read-only: version, API version, database and uptime (the
server name and the fingerprint are in the groups above); an old #/about link opens Settings on it (amended 2026-10-26,
R4, R5). Same UI rules: tooltips only, light/dark, localised via /api/v1/i18n, keyboard
accessible. The sidebar is 256 px and collapses to its icons with ≡ or Ctrl+B like the
desktop rail, the choice kept in the browser and the links' tooltips then naming their page;
on narrow screens it stays the top bar, Settings at its end (amended 2026-10-16, R1;
2026-10-26, R4). Settings has the desktop's gear icon (§12.1).
Library (amended 2026-10-26, R6), the second page of the sidebar. The panel manages the
library and never edits a video: it has no project settings and no render. It browses,
views and deletes sequences and photos, restores them from the archive's trash and runs mass
operations; it queues no deep analysis (the API's bulk/jobs takes deep).
#/librarylists the sequences as cards (DESIGN §4: 16:9 cover, count badge, status glyphs, check box) or a list, the choice kept in the browser; search over title, camera and tag; Year, Camera, Status and Tag filters whose menus come fromGET /sequences/facets, each value with its count (a camera named as on the desktop, the brand once), Status offering analysed or not, proxies or none, rendered or not, verified or not, warnings and errors; sort newest, oldest, title, most photos, largest; Load more by cursor. The filters live in the address, so Back from a sequence finds them again. A cover is the sequence's thumbnail at an address that changes with its cover and its frames version, cached an hour.- Selection by check box, Shift range or the keyboard (one tab stop for the grid, arrows, Space, Shift+Space, Enter opens, Ctrl+A, Esc), and Select all N of the filter. The selection's bar: Analyse, Proxies, Thumbnails and Verify (one job per 1,000 sequences, the toast saying how many were queued), Tags… (add and remove), Export CSV (title, camera, photos, bytes, first and last capture, tags; built in the browser, a cell a spreadsheet would run as a formula starting with an apostrophe) and Delete… (a dry run, then a dialog listing the sequences that cannot go with why, then those that go, at most 500 names; no typed confirmation, since the trash keeps them). Bulk requests are sent in chunks of 1,000.
#/library?id=<sequence>: title (edited in place), tags, rating and notes, saved with the same conflict-checked PATCH as the desktop; its facts; its photos in pages of 500 with lazy thumbnails; a photo viewer (the proxy, else the thumbnail; ← and →, Esc; time, exposure, aperture, ISO, focal length, size in pixels and bytes, format, flags, SHA-256); Set cover; Download the original; Delete… of the selected photos (a dry run, then a confirmation that asks for their number to be typed when more than 20 go, the sequence's frames version sent so a changed sequence refuses); Delete… of the sequence; its renders, to download; Add photos (the upload dialog, R3).#/library?view=trash, also opened from Archive: the trash's entries (kind, sequence, photos, size, deleted by and when), each with Restore, whose answer says which photos stayed and why; Empty… measures the trash (dry run), lists what goes, asks for the word "empty" in the panel's language (any case), sends back the plan's fingerprint and follows the empty-trash job to its end, reading it again after a reconnect of the events or when it ended before the request answered.- The pages follow
sequenceChangedandjobState: the list reads every loaded row again (1,000 a request) with its filters; a sequence reads itself and its renders again after a job ends (at most every 1.5 s), redraws only what changed, keeps the keyboard focus, and asks again for thumbnails not built yet (from 2 s up to 60 s apart, at most 8 times).jobStatecarries no kind or target, so a page reacts to any job that ends. - The audit log and the Dashboard name photo deletions, trash restores and emptyings asked for.
Reorganize (amended 2026-10-26, R7), the panel's way to the Reorganize of §12.5 through
POST /archive/reorganize (§14.3):
- The Library's selection bar has Reorganize…: the selected sequences, up to 1,000 (one plan, one job); above that, every sequence of the archive, only while the selection is exactly Select all without a filter and the archive still holds as many; any other larger selection is refused with a message, never split.
- On Archive, a save that changes the folder or file template, when the archive has sequences, shows Saved with Reorganize…: the dialog for every sequence, with the saved templates and only the changed parts switched on. Saving alone never renames anything.
- The dialog: Folders (
{title}the new title), Files (named as filing names them) and Titles ({title}the current title), each with its switch and field, from the archive's settings; Titles starts off and empty. Each change asks the server for a dry run 400 ms after typing stops (Enter at once), one at a time, a change made meanwhile planned right after. The table shows each sequence in the order the changes apply, 200 at a time: title, folder and renamed files before and after (up to three samples), Carried for a sequence that moves only with a folder it is inside, and the photos that keep their names with why, or why the sequence stays as it is; totals of sequences changed, files renamed and names kept. A refused template is marked on its field with the server's reason; a nested target on Folders. - Apply… (enabled only while the table is the plan of the fields and changes something)
asks for a confirmation that counts the sequences, folders and photos (nothing is deleted,
so nothing is typed), then sends that plan's fingerprint; the toast offers Open job. When
the archive changed meanwhile (
plan_changed), or a new plan answered during the confirmation, the dialog shows the new plan and Apply waits. - A toast with an action stays 15 s and pauses under the pointer or the keyboard focus. The audit log and the Dashboard read "… reorganized the archive · N sequences · M files renamed · templates".
Upload (amended 2026-10-26, R3): Library › Upload (no sequence) and a sequence's Add photos open the upload dialog (§14.4), a module the pages import only on demand.
- Target: the given sequence, or a choice between an existing sequence (a search box, then a list) and a new one (title optional), fixed once the first run starts.
- Photos: the file picker (several files, the photo extensions only) or a drop on the dialog. Folders, files that are not photos (hidden dot files among them) and files above the size limit are counted in chips and left out.
- Each photo shows its state: waiting, hashing %, starting, uploading %, verifying, uploaded, already there, deleted before (with Upload again), failed (the reason in the tooltip), cancelled, not sent; a total bar shows photos and bytes.
- Buttons: Upload, which sends whatever is not on the server yet; Cancel while sending; Close otherwise. Esc closes only while nothing is being sent.
The panel's files (amended 2026-10-26, R6) are served with Cache-Control: no-cache and an
ETag of their content: the browser asks again before each use and an unchanged file costs a
304 (If-None-Match lists, weak W/ tags and * count), so after an upgrade it never runs
new modules next to old cached ones.
Storage (Dashboard and Archive) shows the space volume: the folder CHRONOS_SPACE_PATH
names (§14.9) when it can be measured, else the archive's own volume. When a space path is
in use and the archive's own volume (the write volume, where uploads land; on Unraid the
share's primary storage, often the cache) has other figures (another size, or free space
more than 64 MiB apart), a second line Write volume shows its free space. The figures'
tooltip names the volume and its file system (on Linux the mount's type from
/proc/self/mountinfo, so an Unraid share reads fuse.shfs); on such a share without a
space path it says the figures are the share's primary storage and to map the share's array
view (/mnt/user0/<share>) to /space and set CHRONOS_SPACE_PATH=/space. The server's
diagnostics report lists the archive, the space path and /data, each with free space,
file system and mount (amended 2026-10-26, R2).
14.9 Configuration and Docker
Environment variables (also readable from /data/server.json):
CHRONOS_ADMIN_USER, CHRONOS_ADMIN_PASSWORD (used only when no admin exists; forced change
at first login), CHRONOS_ARCHIVE_PATH=/archive, CHRONOS_DATA_PATH=/data,
CHRONOS_SPACE_PATH (SpacePath in server.json; optional: a folder, e.g. /space, whose
volume's free space the admin panel shows instead of the archive's; one that cannot be
measured falls back to the archive's with one warning, logged at start and again only once
it has worked in between; uploads always check the archive's own volume, §14.4),
CHRONOS_DB=postgres|sqlite (Linux: postgres only), CHRONOS_DB_HOST, CHRONOS_DB_PORT=5432,
CHRONOS_DB_NAME, CHRONOS_DB_USER, CHRONOS_DB_PASSWORD, CHRONOS_DB_SSLMODE=prefer,
CHRONOS_HTTP_PORT=8080, CHRONOS_HTTPS_PORT=8443, CHRONOS_TLS_MODE=auto|pem|proxy,
CHRONOS_TLS_CERT, CHRONOS_TLS_KEY, CHRONOS_SERVER_NAME, CHRONOS_LANG=en,
CHRONOS_PARALLELISM (photos the running job processes at once; CHRONOS_WORKERS, its former
name, still sets it), CHRONOS_PRIORITY_QUEUE=false, CHRONOS_DISCOVERY=true (LAN UDP beacon),
CHRONOS_LOG_LEVEL (Trace|Debug|Information|Warning|Error|Critical|None, also info,
warn, off; default Information: sets the level of the log files and of the console and
locks both in the admin panel), CHRONOS_LOG_CONSOLE_LEVEL (the console alone, read after
it, so it wins for the console), CHRONOS_LOG_RETENTION_DAYS=14 (1–3650),
CHRONOS_LOG_MAX_FILE_MB=50 (0–10240, 0 = no limit), CHRONOS_LOG_MAX_TOTAL_MB=1024
(0–1048576, 0 = no limit, not below the file size), CHRONOS_LOG_ROTATION=daily|hourly,
CHRONOS_AUDIT_RETENTION_DAYS=0 (0–3650 days of audit log the Audit log clean-up
schedule keeps, 0 = forever; set, it is applied to that schedule at every start and locked
there), CHRONOS_DIAGNOSTICS_LOG_MB=4 (1–1024 MB of the end of each log file in the
diagnostics zip); in server.json LogLevel sets the files only and ConsoleLogLevel the
console; a value out of range stops the start naming it (R11). PUID, PGID, TZ (also
the time zone of the schedules, R10). A variable fixes its setting: the admin panel shows
it read-only and names the variable. (Amended 2026-10-16, R2; 2026-10-26, R2, R10, R11.)
Dockerfile: multi-stage on mcr.microsoft.com/dotnet/sdk:10.0 → aspnet:10.0, non-root
user via PUID/PGID entrypoint, volumes /archive and /data (the optional read-only
/space bind mount is not a declared volume; amended 2026-10-26, R2), HEALTHCHECK on
/health, linux/amd64 + linux/arm64. docker-compose.yml with a PostgreSQL 17 service.
Windows: Chronos.Server.exe runs as console or Windows Service (sc create), SQLite or
PostgreSQL.
15. Chronos Uploader
Single small window (tray-capable), zero analysis: connect + sign in (token saved with
DPAPI), drop zone (files/folders/cards), queue with per-file state, pause/resume, bandwidth
limit, retry with backoff, resumable across restarts (queue persisted in a JSON file next to
the app data), hash verification, optional "delete source after server confirms" (off by
default, confirmation), watch folders (poll + FileSystemWatcher where reliable; upload
new files after they stop growing), run at startup, minimise to tray, per-session target
(auto-group / choose sequence / new sequence with title). P2: "make proxies here" toggle
for HEIC/RAW so the server can render without a node.
│ Chronos Uploader ☁ https://nas.local:8443 · alberto · ● Connected [⚙] │
│ ┌────────────────────────────────────────────────────────────────────────────────┐ │
│ │ Drop photos, folders or a card here │ │
│ └────────────────────────────────────────────────────────────────────────────────┘ │
│ Target (•) Group into sessions automatically ( ) Sequence [Harbour sunset ▾] ( ) New │
│ Queue 4,812 files · 96.2 GB · 38 MB/s · 41 min left [⏸] [⏹] Limit [50] MB/s │
│ ▮▮▮▮▮▮▮▮▮▮▮▮▮▮░░░░░░░░░░░░░░░░ 2026-09-28_18-02-11_0413.CR2 │
│ Watch folders D:\Cam\Roof (every 60 s) ● [+] │
│ [✓] Delete source after verified upload [✓] Start with Windows │
16. Localization
- Source strings in
i18n/en/*.resx(one file per area:Shell,Library,Editor,Import,Archive,Server,Settings,Tooltips,Errors,Advisor,Admin). Keys are stable identifiers (Library.Filter.Year), never English sentences. - Runtime:
ResourceManager+ aLocXAML markup extension andL("key", args)helper inChronos.Core/Localization; culture switch at runtime without restart (rebinding via aLocalizationChangedevent); web admin receives/api/v1/i18n/{culture}generated from the same resx byChronos.Tools i18n export. - Plurals: an in-house minimal CLDR plural-rule table for the shipped languages
(
one/few/many/other), keys with suffixes (Photos.Count.one,.other). - Dates/numbers/units via
CultureInfo; file sizes localised; RTL viaFlowDirectionforar,he,fa,ur(mirrored rail and icons where directional). - Languages (Phase 8, all machine-translated by Claude Code with a glossary, marked as such
in About and open to community review via the releases repository):
en(source),it,de,fr,es,pt-BR,pt-PT,nl,pl,cs,sk,hu,ro,sv,da,nb,fi,ru,uk,tr,el,bg,hr,sl,lt,lv,et,ja,zh-Hans,zh-Hant,ko,ar,he,hi,id,ms,vi,th. Add more if effort allows. (Amended 2026-10-23, R2.) Chronos.Tools i18n checkfails CI when a key is missing inenorit, when a placeholder count differs between languages, or when a string exceeds the layout budget (tooltips ≤ 200 chars, labels ≤ 30 chars unless flagged).- UI is written to survive +40 % text expansion (German, Finnish): no fixed-width labels.
17. Branding and icon
- Mark: a time-lapse aperture-clock — a circle of 12 short arc segments increasing in opacity clockwise (the sequence of frames) around a solid sun disc slightly off-centre, reading as camera aperture, clock and sun path at once; single accent colour on a rounded square (dark) tile; no text. Works at 16 px (segments merge into a ring; the disc remains).
- Generated programmatically (
Chronos.Tools make-icon): in-house PNG encoder + ICO container (16/20/24/32/48/64/128/256), SVG source kept inbranding/; variants: Server (mark + three horizontal bars badge), Uploader (mark + up-arrow badge). Also favicon for the admin panel and a 1024 px PNG for listings. PrototypeIconGeneratorshows the ICO writer approach. - App name in title bar only; no splash screen; first-run screen is the empty Library with one sentence and two buttons (Import a card · Open an archive).
18. Installer, distribution and updates
- Builds:
dotnet publish -c Release -r win-x64 --self-contained -p:PublishSingleFile=true -p:PublishReadyToRun=truefor App, Uploader, Setup; server as framework-dependent Docker image and a self-contained Windows zip/installer. (No trimming: WPF does not support it.) The Windows programs are cross-compiled on the Linux runner (EnableWindowsTargeting), andChronos.Tools pe-checkstops a release whose executables lost the GUI subsystem, the icon, the version information or the PerMonitorV2 manifest (amended 2026-10-23, R3). Chronos.Setup(in-house, WPF, ~one screen): choose components (Chronos, Uploader, Server for Windows), folder, shortcuts, file associations (.chronosproject bundles), installs to%LOCALAPPDATA%\Programs\Chronosper user (no admin) orProgram Files(elevated), registers Add/Remove Programs entry with uninstaller, silent mode (/S /DIR=), upgrade in place, keeps user data. Payload appended to the setup exe as a compressed archive (in-house format onDeflateStream) with a manifest and per-file hashes. If the human later permits a build-time tool, Inno Setup/MSIX are acceptable alternatives; default is in-house.- Code signing hook in the release workflow (Authenticode with
osslsigncode, the Linux counterpart ofsigntool; amended 2026-10-23, R3) using secrets when present; unsigned builds are still produced. - Updates: on start (opt-out) the client calls the Gitea releases API of the releases
repository (§19.4),
<server>/api/v1/repos/<owner>/<name>/releases?draft=false&pre-release=false, takes the highest version (a pre-release build is offered the final release of its own version; a 404, a repository the client cannot see, is a failed check, not "up to date"), builds the release page and download links fromChronosReleasesUrl(Gitea writes its own with its ROOT_URL, which clients may not reach), and shows a one-line notice with "Download" (P1: download installer, verify SHA-256 fromSHA256SUMS, run it). The repository is the build propertyChronosReleasesUrl(defaulthttps://gitea.encke-hake.ts.net/BalboSoft/Chronos), which also gives the licence link (<repository>/src/branch/main/EULA.md), the issues link (<repository>/issues) and the Add/Remove ProgramsURLInfoAbout. (Amended 2026-10-23, R2.) - Distribution: installers, zips and
SHA256SUMSon the Gitea release of the releases repository; Docker images in the Gitea container registry (gitea.encke-hake.ts.net/balbosoft/chronos-server); Unraid template in the repository (§19.6); who can reach them today: §19.4. (Amended 2026-10-23, R2.)
19. CI/CD, Docker, Unraid
19.1 ci.yml (push to main, pull requests, manual; amended 2026-10-23, R3)
Gitea Actions in .gitea/workflows. The runner is act_runner in Docker mode on the Linux
server that also hosts Gitea, label ubuntu-latest, one job at a time. The .NET jobs run
in the container mcr.microsoft.com/dotnet/sdk:10.0 and check out with plain git (the
image has no Node.js, so no JavaScript actions such as actions/checkout); no artifacts
and no caches are used; builds use 4 cores (BUILD_CPUS). Gitea shows no job summary and
no annotations: ::warning/::error lines appear in the log only. Jobs:
build:dotnet --info, build with warnings as errors, zero-package guard,Chronos.Tools i18n check, XAML lint, format check, CC0 RAW fixtures fetched byChronos.Tools fixtures(a missing sample is a warning in the log and its tests skip),Chronos.Testingfor the Core, Data (SQLite), Server and Tools tests (JUnit XML), and the Windows programs (App, Uploader, Setup) cross-compiled for win-x64, packed withpack-installerand checked withChronos.Tools pe-check.postgres: PostgreSQL 16 and 17 service containers (host namepostgres) forChronos.Data.TestsandChronos.Server.Tests;Chronos.Tools postgres-ciprepares the MD5, cleartext and trust roles and TLS through SQL only, because Gitea gives no id of the service container.docker: builds the image with the host's daemon (no push) and smoke-tests/healthwith PostgreSQL 17 on a network of the run.yaml: inruby:3.3-alpine, with the sources from the Gitea archive API; checks the workflows, the.gitea/ISSUE_TEMPLATEforms and the compose file.windows(optional): runs only when the repository variableWINDOWS_RUNNERnames the label of a Windows act_runner in host mode with the .NET SDK, Git for Windows (inC:/Program Files/Git, whose bash the job names by path) and PowerShell 7: the Core.Windows and App tests, the--capturescreenshots and their diff, and the installer test. Without a Windows runner the Windows-only tests are run locally on a Windows PC.
perf.yml runs the performance gates weekly on SQLite and on PostgreSQL 17 in a service
container; the timings are printed at the end of the log.
19.2 Zero-package guard
Directory.Build.targets target AssertNoPackageReferences fails the build if any
PackageReference exists; CI also greps **/*.csproj and **/*.props and checks that
obj/project.assets.json lists no packages. The workflows use only shell (run) steps: no
third-party actions at all (amended 2026-10-23, R3).
19.3 release.yml (tag v*, or manual re-publish of an existing tag; amended 2026-10-23, R3; amended 2026-10-26, R1)
Job release, in the SDK container: the version from the tag
(vMAJOR.MINOR.PATCH[-PRERELEASE], SemVer 2.0 as release-channels reads it, checked
before anything is built) passed as -p:ChronosVersion, with
-p:ChronosReleasesUrl (the releases repository's address, §18); build and the Linux
tests; self-contained single-file win-x64 publish of App and Uploader (ReadyToRun), Server
and Setup (ReadyToRun), cross-compiled on Linux; Chronos.Tools pe-check (GUI subsystem,
icon, version information, PerMonitorV2 manifest); Authenticode signing with
osslsigncode when the secrets SIGN_PFX_BASE64 and SIGN_PASSWORD exist; the
installers Chronos-Setup-<ver>.exe (components Chronos, Uploader, Server) and
Chronos-Uploader-Setup-<ver>.exe, Chronos-Server-win-x64-<ver>.zip and
Chronos-win-x64-portable-<ver>.zip (with portable.txt); release notes
(Chronos.Tools changelog from Conventional Commits) and SHA256SUMS; the Gitea release
made by Chronos.Tools gitea-release (a draft first, every file uploaded and its size
verified, then published). Then Chronos.Tools release-channels decides what the tag moves
(amended 2026-10-26, R1): it reads every page of the target repository's published releases
(releases?draft=false&limit=50&page=<n> up to the first empty page, at most 100), stable
and pre-releases alike, ignores drafts and tags that are not v + SemVer 2.0, always counts
the tag itself, takes a release Gitea flags as a pre-release for one, and compares by
SemVer 2.0 precedence (not Gitea's releases/latest, which is the most recently created
release; equal precedence counts as highest). latest is true for the highest stable
release and moves the releases repository's content (§19.4); image_latest is true for
the highest stable release or, while no stable release exists, the highest pre-release,
and moves :latest. Re-publishing an older tag, or a pre-release once a stable release
exists, moves neither. When the channels cannot be decided (a page unreadable or not a
list, or a tag older than release-channels, whose own Chronos.Tools lacks it) the tag
moves neither, with a warning. Then job image, in the runner's default image with the
docker CLI and the host's daemon: it builds linux/amd64 only (arm64 would need QEMU on
the host) and pushes
gitea.encke-hake.ts.net/balbosoft/chronos-server:<ver>, plus :latest when image_latest
is true. Pushing needs the repository secret REGISTRY_TOKEN, an access token of a user who
can write the owner's packages (scope "package: read and write"): the workflow's own token
(user gitea-actions) is refused by the organisation's registry. Pre-releases (tags with a
-) are published as Gitea pre-releases; the in-app update check ignores them.
Variables, all optional: PUBLIC_URL (default https://gitea.encke-hake.ts.net),
RELEASES_REPO (§19.4), REGISTRY (default the host of PUBLIC_URL), REGISTRY_USER
(default the user who pushed the tag), WINDOWS_RUNNER (§19.1). Secrets:
REGISTRY_TOKEN (needed for images), RELEASES_TOKEN (only with RELEASES_REPO),
SIGN_PFX_BASE64 + SIGN_PASSWORD (optional signing).
19.4 Releases repository (amended 2026-10-23, R2)
By default releases are published in BalboSoft/Chronos itself: on the owner's own Gitea
the owner decides who sees the organisation, its releases and its packages, and one
repository needs no second token and no copy step. A separate public repository stays
optional, for when the sources must stay hidden while the releases are public: with the
repository variable RELEASES_REPO (owner/name of another repository on the same Gitea)
and the secret RELEASES_TOKEN, the release goes there, and for the newest stable release
release.yml copies into it deploy/releases-repo/README.md (what it is, screenshots,
requirements), EULA.md, SECURITY.md, .gitea/ISSUE_TEMPLATE,
deploy/unraid/chronos.xml and branding/chronos-server.svg (as unraid/), the compose
file and .env.example (as docker/), i18n/ translations for community review and
SHA256SUMS, rewriting the links to that repository and its image; nothing is developed
there. EULA.md and SECURITY.md live at the repository root, the issue forms in
.gitea/ISSUE_TEMPLATE (Gitea issue-form YAML: name, about, title, labels, body).
Today the BalboSoft organisation is not visible to anonymous visitors, so the update
check, release downloads and image pulls work for other people only once the owner makes
the organisation (and its packages) public, or publishes to a public RELEASES_REPO (the
image stays in the organisation's registry either way); while it is private,
docker login gitea.encke-hake.ts.net is needed before pulling. Gitea has no private
vulnerability reporting: security problems are reported privately to the maintainers (the
contact on the owner's Gitea profile, or an e-mail to them), never in a public issue.
19.5 docker.yml
Nightly build of main (02:17 daily) to
gitea.encke-hake.ts.net/balbosoft/chronos-server:nightly in the Gitea container registry,
with the host's daemon and the REGISTRY_TOKEN secret as in §19.3 (amended 2026-10-23, R3);
deploy/docker/docker-compose.yml for users (server + PostgreSQL + volumes), downloaded
from https://gitea.encke-hake.ts.net/BalboSoft/Chronos/raw/branch/main/deploy/docker/docker-compose.yml
together with .../deploy/docker/.env.example. :nightly is never a release: it carries
main as it was that night (version 0.0.0-nightly.<run>), for testing only; users follow
:latest or pin :<ver> (amended 2026-10-26, R1).
19.6 Unraid template (deploy/unraid/chronos.xml)
Community Applications XML: repository gitea.encke-hake.ts.net/balbosoft/chronos-server:latest
(the newest release: the highest stable release, or the highest pre-release until the first
stable one, §19.3; never :nightly, so the template finds an image from the first
pre-release on; amended 2026-10-26, R1), WebUI http://[IP]:[PORT:8080]/admin, ports
8080/8443, paths /archive → /mnt/user/Timelapse, /space → /mnt/user0/Timelapse
(Array view, read-only, optional: the same share seen on the array only, so the admin
panel shows the array's free space, §14.8), /data → /mnt/user/appdata/chronos,
variables from §14.9 (admin user/password, DB host/port/name/user/password, TLS mode,
Space path CHRONOS_SPACE_PATH=/space (advanced), PUID/PGID, TZ; amended 2026-10-26, R2;
the log variables, Log level CHRONOS_LOG_LEVEL included, and Audit retention
CHRONOS_AUDIT_RETENTION_DAYS are advanced and empty by default (each description gives
the default), so the admin panel's Settings › Logs chooses them; a value given in the
template locks its setting; TZ also sets the time zone of the schedules; amended
2026-10-26, R10, R11), description ≤ 2 sentences.
Registry and Project https://gitea.encke-hake.ts.net/BalboSoft/Chronos,
Support https://gitea.encke-hake.ts.net/BalboSoft/Chronos/issues, TemplateURL
https://gitea.encke-hake.ts.net/BalboSoft/Chronos/raw/branch/main/deploy/unraid/chronos.xml,
Icon https://gitea.encke-hake.ts.net/BalboSoft/Chronos/raw/branch/main/branding/chronos-server.svg;
release.yml rewrites these exact prefixes for a separate releases repository (§19.4), so
they stay verbatim. (Amended 2026-10-23, R2.) Document (in the README) how to install the
template as a user template (Unraid's template repositories read GitHub and GitLab
addresses, not Gitea's) and to pair it with the official PostgreSQL template.
20. Testing and quality
20.1 Chronos.Testing (in-house)
Attributes [Test], [Theory] + [Case(...)], [Skip(reason)], [Timeout(ms)],
[Category]; assertions with expressive messages (Assert.Equal, Approximately,
Throws, SequenceEqual, Monotonic, Within(px)); fixtures (temp dirs, synthetic
sequences, in-process server, database per test); parallel execution; filters by name/
category; console output + JUnit XML; exit codes; dotnet run --project tests/X -- --junit out.xml --filter core.
20.2 Test layers
- Unit: parsers (fuzz corpus: truncated/garbage files must not throw), codecs (round-trip PSNR thresholds), analysis rules (synthetic sequences with known answers — port the prototype's 63-check self-test as tests), naming templates, SQL repositories (SQLite in-memory), PostgreSQL wire client (against service container), protocol (chunk resume, hash mismatch, idempotency), advisor recommendations (golden JSON).
- Integration: import a synthetic card folder → archive layout + sidecars + DB; rebuild index from sidecars equals original DB; verify detects a flipped byte; server ingest via API; sync round trip; render node claim/complete; job resume after simulated crash.
- Windows: WIC decode of fixtures, Media Foundation encode smoke (software H.264) +
in-house MP4 inspector +
MediaPlayerplayback check;--capturescreenshots of every screen for visual regression (pixel-diff tolerance) and for the phase report. - Fixtures: tiny synthetic files checked in; CC0 RAW samples (one per major format, from a
public CC0 sample set) fetched by
Chronos.Tools fixturesintotests/fixtures/external(git-ignored; CI fetches them on every run, without a cache; amended 2026-10-23, R3).
20.3 Static rules
Warnings as errors; nullable enabled; dotnet format; XAML lint: every Button,
ToggleButton, Slider, ComboBox, CheckBox, MenuItem has a ToolTip; no literal
Content="…"/Text="…" outside resources; no PackageReference; no Windows APIs in
Chronos.Core/Chronos.Data/Chronos.Protocol/Chronos.Server (checked by building them on
Linux); public API of Chronos.Protocol versioned.
20.4 Quality gates carried over from the prototype (regression thresholds)
Deflicker: flicker reduction ≥ 60 % on the synthetic ramp sequence with the ramp preserved within 0.05 EV; region deflicker: landscape residual ≤ 25 % of the global-curve residual; phase correlation: sub-pixel error ≤ 0.10 px; stabilization on a textureless frame: shake ≤ 0.1 px (texture gating); holy grail: no step > 0.3 EV in the target curve; time remap monotonic; star-trail luminance non-decreasing on ≥ 99 % of steps; memory: peak RSS independent of sequence length for a fixed window; cancellation: no partial output, no unobserved exceptions, spill file removed.
21. Monetization — absolute final step (Phase 9; do not start before every gate passed)
- In-house ad delivery, no SDK, no tracking by default: the client fetches a signed
campaign manifest (JSON: id, image URL, link, weight, languages, countries, start/end,
minimum version) from
ads/manifest.json, a raw file of the releases repository on Gitea (by defaulthttps://gitea.encke-hake.ts.net/BalboSoft/Chronos/raw/branch/main/ads/manifest.json), or from a URL set by the human (amended 2026-10-23, R2), caches creatives, renders them natively in one slot — the Export section while a render runs (prototype rationale: the only moment an ad takes nothing from anyone) — and optionally a small footer slot in Library. Clicks open the browser. Manifest signature: ECDSA P-256 with an embedded public key. - Consent: first-run one-line choice for anonymous impression counting (off by default); when on, a daily aggregated ping (version, language, country from OS locale, counts) — no identifiers. Privacy policy link in About and EULA.
- Ad-free: "Chronos Pro" licence key (ECDSA-signed, offline validation, tied to a display name and edition, no phone-home) entered in Settings › Licence; a server licence makes all its connected clients ad-free (team scenario). Fallback content when the manifest is empty: internal tips about the program.
- Everything above is a toggle in Settings for the human's own builds; the feature ships
disabled by default until the human flips a build property
ChronosAdsEnabled=true.
22. Phased delivery plan
Each phase ends with the Definition of Done (§0.2 rule 9) and a phase report. Phases may be split into sub-phases; order inside a phase is Claude Code's choice.
Phase 0 — Bootstrap (repo, rules, shells)
Solution, Directory.Build.*, global.json, CLAUDE.md, DECISIONS.md, CHANGELOG.md;
Chronos.Testing; zero-package guard; ci.yml; i18n infrastructure with en + it;
WPF shell (rail, theme, job bar, status strip, settings skeleton, tooltip lint); server
skeleton (Kestrel, /health, /server/info, login page, admin shell); Dockerfile builds;
branding mark + ICO generator.
Gate: both apps start, server container answers /health, CI green on both OSes.
Phase 1 — Metadata and imaging core
Parsers (§9), managed codecs (§8.3), WIC provider, decode tiers, thumbnails/proxies,
ImageBuffer/FrameWindow port, fixtures tool, fuzz tests, DNG full pipeline.
Gate: every fixture decodes at the expected tier on Windows and Linux; PSNR/round-trip
tests green; Chronos.Tools diagnose <folder> reports per-file tier and metadata.
Phase 2 — Data layer and archive
SQLite binding, PostgreSQL wire client, migrations, repositories, archive layout, sidecars, scanner, rebuild, verify, duplicates, orphans, adopt folder, naming templates. Gate: import synthetic 100k-frame archive; rebuild equals original; verify detects corruption; PG tests green against 16 and 17.
Phase 3 — Client Library and Archive manager
Virtualized grid/list/timeline, filters, saved views, mass operations, jobs screen, stats
charts, maintenance, settings screens, --capture screenshots.
Gate: 500k-photo synthetic library meets §1.1 timings.
Phase 4 — Card import
Device detection, card pool and profiles, plan preview, sessions, hash-verified copy, post-actions (delete/format/eject), auto-import, direct-to-server destination stub. Gate: end-to-end import from a simulated card folder; destructive actions gated by verification; screenshots.
Phase 5 — Analysis advisor
Light + deep analysis, reports, charts, recommendations with reasons, auto presets, Auto/manual controls, Optimise. Gate: golden recommendations for synthetic sequences; every recommendation has a localised reason; deep analysis of 10k frames completes unattended.
Phase 6 — Engine, encoding and Sequence editor
6a modules port with regression thresholds (§20.4); 6b Media Foundation encoder, MP4 muxer (H.264/HEVC/AV1), managed encoders (AVI/GIF/APNG/sequences), presets; 6c preview tools, section charts, spline editor, export UI, batch render. Gate: self-test suite green; renders play in Windows Media Player and the in-house inspector validates containers; hardware MFT used when present (logged).
Phase 7 — Server, sync, uploader
Full API, auth, uploads/ingest, sync, jobs, render nodes, admin panel, Docker + compose, client server mode, Uploader app, LAN discovery. Gate: two clients + uploader against a Docker server: upload, sync, render via node and via server; crash-restart resumes jobs; audit log complete.
Phase 8 — Worldwide release readiness
All languages (§16), RTL pass, installer, release.yml, releases repository content
(README, EULA draft, template; a separate public repository optional since 2026-10-23, R2),
Unraid template, update check, diagnostics bundle,
accessibility pass, performance pass, docs.
Gate: tagged pre-release published end-to-end from CI; install/uninstall/upgrade tested;
Unraid template validated.
Phase 9 — Monetization (§21)
Gate: human explicitly approves start; feature behind build property; privacy text done.
23. Assumptions and open decisions (for the human; do not block on these)
- Product name Chronos, given by the owner on 2026-10-23 in place of the name kept from the prototype (amended 2026-10-23, R1); settled.
- Linux server MP4 output relies on render nodes (and P2 intra-only encoder). If an
FFmpeg subprocess were ever allowed as an optional, opt-in backend, it would be a
single
IVideoEncoderProvider; not planned unless the human approves. - HEIC pixels are only decodable on Windows (WIC); server-side rendering of HEIC-only sequences requires proxies from a client/node.
- Native RAW decoders beyond DNG are phased (CR2, NEF, ARW, PEF, ORF, RW2, RAF uncompressed first); until then those formats render from full-size embedded previews on the server and via WIC + Raw Image Extension on Windows.
- Windows 10 21H2+/Windows 11, x64 only (ARM64 later);
winsqlite3.dllpresence assumed. - Tests use the in-house runner to keep the zero-package rule absolute; xUnit would be acceptable only if the human relaxes the rule for test projects.
- Installer is in-house; Inno Setup/MSIX acceptable if the human allows build-time tools.
- Ads and licence keys are designed but disabled by build property until Phase 9.
24. Amendments
The product owner's change requests amend this brief in place; each amended paragraph says
"(amended , R)". The request files in docs/requests/ keep the owner's words, the
questions asked and the answers; DECISIONS rows give the reasons.
| Date | Request | Sections | What changed |
|---|---|---|---|
| 2026-10-16 | Owner review 1, R4 | §8.1, §8.2 | DNG decodes through the managed decoder (WIC only through the Raw Image Extension, never DNG); a GoPro DNG is GPR only when VC-5; analyses and previews take an adequate embedded preview by choice, renders decode natively. |
| 2026-10-16 | Owner review 1, R2 | §12.8, §14.6, §14.8, §14.9 | Jobs run one at a time; the parallelism is photos the running job processes at once; an optional priority queue by job kind pauses lower jobs and resumes them by itself; CHRONOS_PARALLELISM (former CHRONOS_WORKERS) and CHRONOS_PRIORITY_QUEUE. |
| 2026-10-16 | Owner review 1, R1 | §12.1, §14.8 | The rail is 200 px and collapses to 48 px of icons with ≡ or Ctrl+B, their tooltips naming the sections, saved at once; the admin panel's sidebar is 256 px and collapses the same way, kept in the browser. |
| 2026-10-16 | Owner review 1, R7 | §4.1, §10.2, §12.1, §12.3 | The Sequence section is now Edit: one sequence at a time, chosen in a picker inside the section (Change sequence, Ctrl+O); the editor's tabs are the steps of a guided, freely navigable path that starts with a new Analysis step, each step showing where it stands, with Previous and Next. |
| 2026-10-23 | Chronos on Gitea, R1 | whole brief, §23 | The product and its programs are named Chronos (Chronos, Chronos Server, Chronos Uploader) everywhere: projects and namespaces, files, settings, CHRONOS_* variables, %LOCALAPPDATA%\Chronos, .chronos project bundles and folders, the image chronos-server; the product name is no longer an open decision. |
| 2026-10-23 | Chronos on Gitea, R2 | §3, §16, §18, §19.4, §19.6, §21, §22 | The repository is BalboSoft/Chronos on the self-hosted Gitea (https://gitea.encke-hake.ts.net): the update check reads the Gitea releases API of ChronosReleasesUrl; licence, issue and support links, the Unraid template and the ads manifest point to that repository; images go to the Gitea container registry; releases are published in the repository itself, a separate public releases repository is optional (RELEASES_REPO); EULA.md and SECURITY.md at the root, issue forms in .gitea/ISSUE_TEMPLATE. |
| 2026-10-23 | Chronos on Gitea, R3 | §0.2, §18, §19.1, §19.2, §19.3, §19.5, §20.2, Appendix A | Gitea Actions on a Linux act_runner in Docker mode: CI jobs build, postgres, docker and yaml, plus a windows job once a Windows runner is registered; shell steps only, no artifacts, caches, job summaries or annotations; the release cross-compiles the Windows programs on Linux, checks them with pe-check, signs them with osslsigncode when secrets exist, publishes them with gitea-release and pushes a linux/amd64 image with REGISTRY_TOKEN; Definition of Done is green on Gitea Actions, the Windows-only tests on a Windows runner or a Windows PC. |
| 2026-10-26 | Owner review 2, R1 | §19.3, §19.5, §19.6 | :latest follows the highest stable release, or the highest pre-release while none is stable, as decided by Chronos.Tools release-channels by SemVer 2.0 precedence; the releases repository's content and the update check still follow stable releases only; :nightly is the nightly build of main, never a release. |
| 2026-10-26 | Owner review 2, R2 | §14.4, §14.8, §14.9, §19.6 | CHRONOS_SPACE_PATH (SpacePath) names the volume whose free space the admin panel shows, the Unraid template mapping the share's array view /mnt/user0/<share> read-only to /space; the archive's own volume shows apart as Write volume when it differs; the figures' tooltip names the file system (fuse.shfs from mountinfo); completing an upload checks the archive's own volume first (507, resumable). |
| 2026-10-26 | Owner review 2, R4 | §12.1, §14.8 | The Settings icon is a toothed gear in Chronos, the Uploader and the admin panel; the admin panel docks Settings at the foot of its sidebar (the end of the top bar on narrow screens). |
| 2026-10-26 | Owner review 2, R5 | §12.1, §12.8, §12.9, §14.8 | About is the last page of the desktop Settings and the last group of the admin panel's Settings, no longer a rail or sidebar entry; nothing is offered twice (update check and releases on Updates, data folder and diagnostics on Advanced); #/about opens the admin Settings on About. |
| 2026-10-26 | Owner review 2, R9 | §12.4, §13 | Import › Folder… can group the folder as Whole folder: one sequence per file format whatever the cameras and pauses, titled after the folder, with a session key per PC, folder and format so a repeated import adds to the same sequence; desktop only. |
| 2026-10-26 | Owner review 2, R10 | §5, §6.2, §7, §12.5, §12.6, §12.8, §14.3, §14.6, §14.8, §14.9, §19.6 | Scheduled jobs: a Schedule entity (migration 0007, jobs.schedule_id) with an in-house recurrence form, a scheduler per host that claims runs by compare-and-set, with catch-up and DST rules; every timer that queued work is a visible system schedule (server: archive scan, upload clean-up every 6 h, index backup at 03:00, job history and audit log clean-ups; desktop: archive scan, daily index backup, housekeeping with cache trim), seeded once and kept as the user edits it; new Backup and Cleanup job kinds and duplicates/orphans report jobs; Jobs has Queue and Scheduled tabs on the desktop and the admin panel; /api/v1/schedules and the scheduleChanged event; one backup retention; the audit retention is the audit clean-up schedule's; Settings › Archive › Rescan every and the admin Settings' job history are gone; the trash is only reported, never emptied. |
| 2026-10-26 | Owner review 2, R11 | §12.8, §14.1, §14.3, §14.8, §14.9, §19.6 | Log settings: file and console levels apart on the server (CHRONOS_LOG_LEVEL locks both, CHRONOS_LOG_CONSOLE_LEVEL the console), retention, a file size limit with numbered parts, a total size limit, daily or hourly rotation, audit retention and the diagnostics share, in the admin panel's Settings › Logs (applied at once, lockedBy naming the variable) and the desktop's Settings › Advanced (from the first log line); the server's files take Debug and Trace when asked; logging never fails a job. |
| 2026-10-26 | Owner review 2, R6 | §5, §7, §12.2, §12.5, §14.3, §14.4, §14.5, §14.8 | The admin panel manages, never edits: a Library page (filters from facets, cards or list, keyboard selection, mass Analyse, Proxies, Thumbnails, Verify, Tags, Export CSV and Delete), a sequence's page with a photo viewer, cover, downloads and photo deletion, and the trash view with Restore and Empty; photos deleted into trash/photos under journals, a photo never hashed hashed as it goes, tombstones in the sidecar (removed) and the index (frame_removals, migration 0008) that keep sync and uploads from bringing a deleted photo back unless the user restores it or uploads it again on purpose; trash kinds, restore reasons, and Empty as the empty-trash job behind a dry run, a typed confirmation and a fingerprint, never scheduled; the library, trash and bulk endpoints; the panel's files revalidated (no-cache, ETag). |
| 2026-10-26 | Owner review 2, R3 | §14.3, §14.4, §14.8 | Upload from the admin panel into an existing or a new sequence (file picker, drag and drop, no folders), hashed in the browser by a self-tested in-house SHA-256 in a module worker, the capture time read from Exif, sent in resumable chunks through the upload API and verified by the server; server/info lists the photo extensions; hidden dot files are refused for every client. |
| 2026-10-26 | Owner review 2, R7 (Reorganize) | §5, §7, §12.2, §12.5, §12.8, §14.3, §14.5, §14.8 | Reorganize on the desktop (Archive manager, Library, the toast of Settings › Archive) and the admin panel (Library, the toast of Archive): folder, file and title templates applied to sequences already filed, behind a dry run that lists each sequence's changes, the photos that keep their names and why (in place outside originals/, never hashed, missing) and the sequences that stay whole (unreadable sidecar); applied by the move job of exactly the confirmed plan (fingerprint), which holds the rest when paused, cancelled or failed and lets a sequence still receiving photos keep its file names; crash-safe plan and file journals (two-phase renames through hidden temporary names, recovery by the next writer); a sequence of the plan left as it was inside a moving folder goes with it; the index, frames and exports follow a moved folder in one transaction, on this index and at another index's next scan; writers take the folder under the lease, never create a left one again and never write into a folder another sequence took; POST /api/v1/archive/reorganize with a kept, bounded plan cache and sequenceChanged reorganized; Rename… and Move… replaced. |
| 2026-10-26 | Owner review 2, R7 (naming) | §7.1, §13, §14.3 | One naming rule for import, upload and Reorganize: the sequence's camera and lens, {n} its place, a frame without a time at the sequence's start; template rules everywhere a template is typed or sent (no folder level scans skip, one a value such as a title gives escaped with _; a file template keeps the extension, tells photos apart, holds no slash; long names shortened before .{ext}; titles as text), held to a template only when it is given anew; invalid_template names the rule. |
Appendix A — CLAUDE.md content (create verbatim in Phase 0, keep updated)
# Chronos — working rules
Read docs/BRIEF.md when in doubt; it wins over these notes.
- Zero <PackageReference>. Anything not in the .NET BCL / Microsoft shared frameworks / OS APIs is written here.
- No external processes at runtime (no ffmpeg, exiftool, powershell). Only OS dialogs listed in BRIEF §8.6.
- Chronos.Core / Data / Protocol / Server target net10.0 and build on Linux; Windows-only code lives in *.Windows or the desktop apps.
- English identifiers, comments, commits (Conventional Commits). Prototype in docs/reference is Italian: port algorithms, not language.
- Every UI string is a resource (en source, it translation). Labels ≤ 3 words; explanations only in tooltips; every actionable control has a tooltip; no helper paragraphs.
- Destructive operations: dry-run plan first, hash verification, explicit user action.
- When the brief is silent: choose the simplest safe option, log it in docs/DECISIONS.md, continue. Ask only at phase gates or before irreversible choices.
- Definition of Done per phase: CI green on Gitea Actions (the Linux jobs; the windows job once a Windows runner is registered, until then the Windows-only tests run on a Windows PC), tests green, no new warnings, DECISIONS + CHANGELOG updated, i18n check green, phase report with "how to try it".
- Build: dotnet build -c Release; tests: dotnet run --project tests/<X> -- --junit out.xml; tools: dotnet run --project src/Chronos.Tools -- <command>.
- Current phase and next steps: see docs/STATUS.md (update at every gate).
Appendix B — Prototype module map (grep targets in docs/reference/chronos-prototype.md)
| Prototype (Italian) | New home | Notes |
|---|---|---|
Metadata/TiffDirectory, TiffPrimitives, MetadataReader, XmpScanner, FrameMetadata |
Core/Metadata |
extend with BMFF/RIFF/RAF walkers, MakerNotes |
Imaging/WicInterop, ImageDecoder |
Core.Windows/Wic |
keep pull-based scaling chain |
Imaging/ImageBuffer, ColorSpace, GeometryStage, RasterWriter, OrientationDetector |
Core/Imaging |
RasterWriter → Codecs/TiffWriter, LinearDngWriter |
Analysis/* (LuminanceAnalyzer, DeflickerEngine, RegionSegmenter, HolyGrailEngine, ExposureProcessor, ColorScience) |
Core/Analysis |
keep validated behaviour + tests |
Motion/* (Fourier, PhaseCorrelator, GrayPyramid, OpticalFlowEngine, MotionField, MotionBlurRenderer, Stabilizer, VirtualCamera, TemporalStacker, FrameInterpolator) |
Core/Motion |
|
Core/Spline, LocalRegression, TimelapseSequence, FrameRecord |
Core/Math, Core/Sequences |
|
Pipeline/RenderPlan, RenderPipeline, FrameWindow, AutoDirector, AutoOptimizer, SequenceWarnings |
Core/Render, Core/Advisor |
AutoDirector/Optimizer → advisor rules with reasons |
Pipeline/MediaImporter, PathTemplate |
Core/Import, Core/Naming |
tokens translated to English |
Video/MediaFoundationInterop, VideoEncoderSession, Nv12Converter, CodecLimits |
Core.Windows/MediaFoundation |
add P010, AV1, capability enumeration |
Video/Mp4Muxer, BoxWriter, AnnexBParser, ExportPresets, ExportSettings |
Core/Video, Core/Export |
add av1C, new presets |
Diagnostics/SelfTest, SyntheticSequence, SequenceDiagnostics, Mp4Inspector, Mp4Playback, AdvancedModuleTests, ExifWriter, IconGenerator, CrashGuard |
tests/*, Chronos.Tools, Core/Diagnostics |
63 checks → tests |
UI/* (WinForms) |
not ported | reuse only the interaction ideas (rail, preview tools, Auto indicator, presets) |
UI/SponsorPanel |
Phase 9 only |
Appendix C — Glossary
Sequence (a shoot: one folder of frames = one project) · Frame (one photo) · Hero frame
(thumbnail of the sequence) · Light/Deep analysis · Advisor (rule engine producing
recommendations with reasons) · Tier (how a frame's pixels are obtained) · Proxy (JPEG
stand-in) · Sidecar (sequence.json manifest) · Render (an exported video) · Preset
(bundle of export values) · Sequence preset (project settings + export preset produced by
the advisor) · Render node (client lending its GPU to the server) · Job (any long
operation, resumable) · Card profile (import rules for a memory card).