OpenNest/docs/nest-storage.md

Nest storage: File mode vs. Database mode

The desktop app can save nests two ways:

• File mode (default, unchanged behavior): Save/Save As write a .nest ZIP archive to disk via the normal file dialog. See nest-file-format.md.
• Database mode: Open browses server records 100 at a time. The Find box searches name, customer, material, made by, comments and status on the server after a short typing pause (Enter in the box searches at once); clicking a column header sorts every match on the server (click again to reverse); the arrow buttons move between pages and the title and find bar show the range and total. Dates and numbers are not searched as text; sort their column instead. The selected sort column and direction are remembered for the current Windows user when the browser closes, including across app restarts. Until a selection is saved, nests appear newest-saved first. Below the list, the Details tabs show the highlighted nest's plates (duplicates, size, parts, drawings, utilization), drawings (required, nested, remaining, area), and Versions (bounded current and earlier snapshots). Plates and drawings come from the archive downloaded shortly after the highlight stops moving; Versions loads on selection and follows the selected nest without opening another window. Beside the tabs, a preview draws one plate of the highlighted nest; the arrow buttons below it step through the plates, and selecting a row on the Plates tab shows that plate. Enter on the nest list or a double-click opens the highlighted nest; Esc closes. Save creates a record on first save and updates that same record afterward. Save As creates a new record (a copy). A failed save keeps the previous record association; switching server URLs creates a record on the new server rather than updating an id from the previous one. On the Versions tab, Open read-only... previews one snapshot without binding an editable document, and Download selected version... saves its exact archived bytes. Select a nest and click Nest Information... above the list (or use the row's right-click menu) to edit the same information as Nest > Nest Info, including the name, without opening the nest in the main editor. OK saves immediately to the same database ID as a new version, then refreshes the list, preview and history; Cancel makes no change. The dialog starts with the latest saved metadata. Plate defaults follow the same rules as the open editor: populated sheets are not resized or re-nested. This action requires a version-aware server and refuses to overwrite a newer save from another PC. After a conflict or network error, refresh and reopen the information dialog before trying again; a lost reply can mean the save already reached the server. An already-open editor is not modified; reload it before saving further changes to that same nest. A separate confirmed Restore selected version... creates a new current revision of the same ID and keeps all intervening versions. Delete... hides the record from ordinary browsing but retains every archive; Deleted nests... opens an explicit recovery view. Neither action purges data. An old desktop without a revision precondition cannot update or delete on an upgraded server. A separate File > Export .nest... command is always available (in both modes) for producing a local file to share or back up. Export does not change the document's file save path or database record association.

Dragging one or more .nest files onto the main window (or onto an open plate) opens them from disk directly, in either mode — this is the only way to open a local file while in Database mode, since Open there browses server records instead of showing a file dialog. A document opened this way is not bound to a server record, so a subsequent Save in Database mode creates a new record on first save, same as any other unbound document.

The mode and server address are stored per-PC at %APPDATA%\OpenNest\storage.json (OpenNest.Data.NestStorageSettings), defaulting to File mode so existing installs are unaffected until an operator opts in via File > Storage Mode....

Wire contract

OpenNest.Data.NestRecord is the metadata DTO shared by the client (RemoteNestRepository) and the server (OpenNest.Server). JSON is camelCase with enums serialized as strings (JsonSerializerDefaults.Web + JsonStringEnumConverter(JsonNamingPolicy.CamelCase)), matching the rest of OpenNest.Data's local JSON stores. Uploaded metadata (POST and both PUTs) must carry string text fields (null is 400) and a defined status: a name, or for existing callers its integer value (0 quote, 1 toBeCut, 2 hasBeenCut, 3 draft); an undefined value is 400 and stores nothing. A server older than draft refuses it, so upgrade the server before desktops that can set it.

Field	Type	Notes
id	guid	Client-generated on first save; kept on updates.
name	string
customer	string
dateCreated / dateModified	datetime	Set by the client from the nest's own metadata.
material	string
thickness	number
status	"draft" | "quote" | "toBeCut" | "hasBeenCut"	Default quote.
plateCount / partCount	integer	Computed client-side at save time (non-cutoff parts only).
comments	string
madeBy	string
fileSize	integer	Server-computed; ignored on upload.
savedAt	datetime	Server-computed; ignored on upload.
revision	integer	Current immutable revision; zero on a legacy server without history.
deleted	boolean	Hidden from ordinary reads when true; retained in recovery/backup.
restoredFromRevision	nullable integer	Historical source promoted by Restore; not the preceding revision.

Endpoints

Method	Path	Body	Response
GET	/healthz	—	200 { "status": "ok" } after a live database query; 503 { "status": "unavailable" } on storage failure (no internal details)
GET	/api/nests	—	NestRecord[], newest savedAt first. Full, unbounded enumeration for backup manifests, restore checks and the container smoke; browsing uses /api/nests/query
GET	/api/nests/query?search=&sort=&order=&offset=&limit=	—	{ "items": NestRecord[], "total", "offset", "limit" }: one bounded page filtered in SQL (see below), or 400 for an invalid parameter
GET	/api/nests/{id}	—	NestRecord or 404
GET	/api/nests/{id}/file	—	.nest archive bytes (application/zip) or 404
POST	/api/nests	multipart: metadata (JSON NestRecord) + file (.nest bytes)	NestRecord with server-assigned id when the client sends an empty guid; 409 with the current record (active or deleted) when the id already exists, storing nothing
PUT	/api/nests/{id}/file	multipart: metadata + file	Updated NestRecord, or 404 if id is unknown
PUT	/api/nests/{id}/metadata	JSON NestRecord	Updated NestRecord (archive untouched), or 404
DELETE	/api/nests/{id}	—	Updated hidden NestRecord, or 404
GET	/api/nests/{id}/revisions?offset=0&limit=100	—	Metadata page, newest first; limit 1–200
GET	/api/nests/{id}/revisions/{revision}/file	—	Exact immutable archive bytes
POST	/api/nests/{id}/revisions/{revision}/restore	—	New current revision of the same ID, with source metadata/bytes
GET	/api/nests/deleted/query	—	Bounded deleted-record page; same query parameters as ordinary browsing
GET	/api/nests/deleted/{id} / /{id}/file / /{id}/revisions / /{id}/revisions/{revision}/file	—	Recovery equivalents of current/history reads
POST	/api/nests/deleted/{id}/recover	—	New nondeleted revision of the same ID
POST	/api/nest-names/next?year=	—	{ "year", "number", "name" }: the year's next generated nest number (see below); 400 outside 2000–2099

All existing-record mutations require If-Match: "<current revision>". Missing precondition returns 428 (upgrade/reload), malformed precondition 400, stale precondition 409 with current metadata; none mutate data. Ordinary reads of hidden records return 404. Changed file/metadata saves and visibility changes append snapshots atomically; an identical save at the current revision is a no-op. Restore appends a new snapshot, never rewrites its source or intervening history.

Generated nest names

File > New names a nest N<yy>-<seq> (OpenNest.Data.NestNames): the two-digit year and the year's sequence number in a base-30 alphabet without look-alike characters, padded to three digits (27,000 a year before a fourth appears). The sequence runs for the whole year. In Database mode the desktop asks POST /api/nest-names/next for the number, so every PC shares one counter per year (nest_sequences, schema version 6). The server also keeps the number above every stored generated name for that year, active or deleted, so names created offline or before the server counted are not reissued; renamed nests and other years' names do not move it.

When the server is unreachable within 3 seconds, has no such endpoint, or the app is in File mode, the PC's own counter (NestNumber user setting) names the nest. It resets when the year changes and is raised past each number the server hands out. Two PCs naming nests offline at the same time can still pick the same name; a nest's identity on the server is its id, never its name.

A nest's name may already be on posted CNC programs or labels, so it is never renamed silently. When Save (not Save As, and not the reconciliation of a pending upload) creates a nest's first record on the current server, it first searches active nests for one with exactly the same name (case-sensitive; deleted nests do not count, NestNameConflicts). Only on a match does it allocate the year's next name and show it in an editable dialog: OK uploads under the chosen name, Cancel saves nothing. If the check or the allocation fails or takes longer than 3 seconds, the nest uploads under its current name.

In Database mode the name can be edited in Nest > Nest Info (it uploads with the next Save) and in the saved-nest browser's Nest Information... (saved immediately). In File mode it stays read-only, since the file name is the nest name. If another active nest already has the exact new name, a warning asks before using it; No reopens the dialog with the edits kept. A name accepted this way is not asked about again by the nest's first Save. A rename that the server cannot check within 3 seconds goes ahead without a warning.

There is no authentication; this is a LAN-only service. Do not expose it outside the shop network without adding one.

Browsing query

GET /api/nests/query filters, orders and pages in SQLite, so a client receives only the requested page of metadata (never archive bytes):

• search (optional): trimmed, at most 200 characters, no NUL characters; blank means no filter. Case-insensitive substring of the whole text in name, customer, material, madeBy, comments or the status (ToBeCut or the display name To Be Cut). %, _ and \ are literal. Dates and numbers are not matched as text. Case folding is ASCII-only (SQLite LIKE).
• offset (default 0, at least 0) and limit (default 100, 1 to 500).
• sort (default savedAt): one of savedAt, name, customer, status, material, dateCreated, dateModified, thickness, plateCount, partCount, madeBy, comments, fileSize (names, case-insensitive; no numbers). Text columns sort case-insensitively (ASCII); dates sort as their stored ISO text. order is asc or desc (default desc).
• Ties are broken by id in the same direction, so pages partition the matches deterministically. total counts every match and is read with the page in one database hold. A save between two page requests can move a row across a page boundary.
• Unknown, repeated, non-integer or out-of-range parameters return 400 instead of being ignored or clamped. RemoteNestRepository.QueryAsync checks the same bounds before sending, and reports an HTTP 404 from an older server as a server version that must be updated.

Drawing library endpoints

The server also keeps a shared library of drawings (schema v5). The desktop adds library drawings to a nest and saves nest drawings to the library (below). A drawing's geometry travels as an opaque drawing package: a .nest archive holding exactly one drawing and no plates, written and read on the client by OpenNest.IO.DrawingPackage. The server never parses it.

OpenNest.Data.DrawingLibraryRecord fields: id (client-allocated before the first upload), revision, name, customer, materialName, materialGrade, thickness (stored rounded to 4 decimals), units (inches | millimeters), cutType (laser | waterjet | plasma), notes, settings, archived, savedBy, and the server-computed packageSha256 (lower-case hex), packageSize and savedAt.

settings holds the drawing's nesting settings, applied when it is added to a nest: cutFirst (sequence its parts first, for example test squares), priority, stepAngle, startAngle, endAngle (radians), allow180Equivalent and colorArgb (32-bit ARGB; alpha 0 means no stored color, so the nest assigns one). They are record fields, not package content (see drawing packages), so changing them adds a revision without new geometry. Default quantity is not stored: quantity belongs to each nest.

Method	Path	Body	Response
POST	/api/drawings	multipart: metadata (JSON record with a nonempty id) + package	Revision 1; 409 with the current record when the id exists
GET	/api/drawings/query?search=&customer=&materialName=&materialGrade=&thickness=&cutType=&cutFirst=&includeArchived=&sort=&order=&offset=&limit=	—	{ items, total, offset, limit }
GET	/api/drawings/{id} / /{id}/file	—	Current record / package bytes (archived drawings included), or 404
GET	/api/drawings/{id}/revisions?offset=0&limit=100	—	Revision page, newest first; limit 1–200
GET	/api/drawings/{id}/revisions/{revision}/file	—	Exact immutable package bytes
PUT	/api/drawings/{id}/file	multipart: metadata + package	New revision
PUT	/api/drawings/{id}/metadata	JSON record (all details and settings; an omitted settings means defaults)	New revision with the same package
POST	/api/drawings/{id}/archive / /restore	—	New revision with archived changed

• Existing-drawing writes require If-Match exactly as for nests: 428 missing, 400 malformed, 409 stale with the current record. Writing to an archived drawing, archiving an archived one or restoring an active one is also 409. None of these change data. There is no delete; archive hides a drawing from the query unless includeArchived=true.
• Every change appends an immutable revision in the same transaction that updates the current row, including metadata-only edits and archive/restore. An identical save (details, settings and package hash) at the current revision is a no-op.
• Packages are content-addressed: each distinct package is stored once, keyed by its SHA-256, and revisions point at it. Only a new package adds a blob; details, settings, archive and restore revisions add none.
• Validation (400, no state change): cutType and units must be single enum names (no numbers or combinations); thickness finite and at least 0; name nonblank; name, customer, material, grade and savedBy at most 200 characters, notes at most 4000; no NUL characters; settings not null, angles finite and stepAngle at least 0; package nonempty, at most 16 MiB and starting with a ZIP signature. Metadata id on a PUT must be empty or match the route.
• Query: search is a literal, case-insensitive (ASCII) substring of name, customer or notes (%, _, \ literal). customer, materialName and materialGrade match exactly after SharedListNames.Key normalization (trimmed, collapsed spaces, case-insensitive); blank means any. thickness matches exactly after rounding to 4 decimals. sort is one of savedAt (default), name, customer, materialName, materialGrade, thickness, cutType, revision; ties break by id. cutFirst=true|false keeps only drawings with that setting. Paging, order and the refusal of unknown, repeated or out-of-range parameters follow the nest query.
• An older server answers 404 on these routes; the client reports that the server must be updated and never falls back to another store.

Schema v5 keeps three drawing tables in the same file: drawing_packages (sha256, size, package), drawing_revisions (every immutable revision's details, settings, archived flag and package hash) and drawings (each drawing's current revision, indexed for browsing). Upgrading from v4 runs in one transaction: each distinct v4 blob is stored once by its hash, every revision and current row is copied as it was with default nesting settings (v4 had none), and the v4 drawing tables are dropped. A pre-v4 database gets the v5 tables directly. The nest tables are untouched. A server older than v5 refuses to open a v5 database, so rolling back after the upgrade needs the pre-upgrade backup (see Backup, restore, and upgrade). The stopped-volume tar backup includes the drawing tables, but nest_manifest.py and restore_check do not verify drawings yet.

Add from Drawing Library (desktop)

Nest > Add from Drawing Library... and the Library button above the drawing list open a paged browser of active library drawings (Database mode only; in File mode both are disabled with "Database mode only"). It starts filtered to the nest's material, grade and thickness; Show all lifts that filter. Selecting a drawing shows its notes. Add to Nest (or double-click / Enter) downloads that exact revision, checks its SHA-256 against the record and plans the insert (OpenNest.Data.LibraryDrawingInsert) before changing anything:

• Material, grade or thickness differing from the nest: a warning names each difference and the operator confirms. The drawing is added, but its cut pins will not apply; the nest's cutting settings do.
• Different units: refused (no conversion).
• The same revision from the same server already in the nest: only the quantity is raised. A different revision of that drawing: refused; change revisions through Update Drawings.
• Another drawing with the same name: refused, naming both.

The added drawing is an independent copy carrying a library link (server key, id, revision and a snapshot of the library details) that is saved with the nest (see the nest file format). Editing it in the nest never changes the library. The copy takes its name, customer, material and nesting settings from the library record, not from the package.

The browser's Cut first column shows the setting. Edit Settings... (OpenNest.Data.DrawingLibraryEditSession) edits the selected active drawing's Cut first flag, priority, rotation step/start/end (degrees), 180° equivalent, color and notes, and saves them through the metadata endpoint guarded by the listed revision: a new revision with the same package. A 409 is shown and the operator reloads. Nests that already hold the drawing keep their copy and revision.

Save to Drawing Library (desktop)

Right-click a drawing in the drawing list > Save to Drawing Library... (Database mode only; disabled with "Database mode only" in File mode). The dialog (OpenNest.Data.DrawingLibrarySaveSession) prefills name and customer from the drawing (the nest's customer when the drawing has none), material, grade, thickness and units from the nest (the drawing's material when the nest has none), and the cut type and notes of the linked revision when the drawing came from this server's library, otherwise the cut type last saved on this PC. Units follow the nest and cannot be changed. The drawing's own nesting settings (Cut first, priority, rotation, color) are saved in the record and summarized read-only in the dialog; change them later with Edit Settings.

Whenever the details change, the dialog looks up active library drawings with the same normalized customer, name, material and grade and the exact thickness, units and cut type, and offers Save as:

• No match: a separate (new) library drawing.
• One match, or the drawing's own linked library drawing among several: the next revision of it, guarded by If-Match on the revision listed.
• Several matches: no default; the operator chooses one or a separate drawing.
• A linked drawing whose details no longer match is offered but never the default, so changing thickness cannot silently rewrite it.

Save stays disabled until the lookup for the current details has finished. A 409 (someone else saved first) is shown and the choices are refreshed. A save that fails without a definite answer stays pending: the next Save first looks for the revision it would have produced and adopts it when the package hash and details match, so it is never stored twice and a retried new drawing keeps its client-allocated id. Saving identical content at the current revision is a no-op. On success only the nest drawing's library link changes, to the saved revision; save the nest to keep it.

Drawing set endpoints

A drawing set (schema v7) is a named, versioned list of library drawings with a quantity each: a kit, an assembly or a recurring order. Members name drawings by id and always mean each drawing's current revision; there is no per-member revision pin. Sets travel as JSON (application/json, at most 1 MiB). The desktop does not use these routes yet.

OpenNest.Data.DrawingSetRecord fields: id (client-allocated before the first save), revision, name, customer, notes, archived, savedBy, the server-computed savedAt and memberCount, and members, in order, each { drawingId, quantity, drawing }. drawing is the member drawing's current library record (archived or not), joined by the server on every read so a client can group members by material without one request per drawing; it is ignored on save, so a record read from the server may be sent back as it is.

Method	Path	Body	Response
POST	/api/sets	JSON set with a nonempty id and members	Revision 1; 409 with the current record when the id exists
GET	/api/sets/query?search=&customer=&includeArchived=&sort=&order=&offset=&limit=	—	{ items, total, offset, limit }; items have memberCount and no members
GET	/api/sets/{id}	—	Current record with joined members (archived sets included), or 404
GET	/api/sets/{id}/revisions?offset=0&limit=100	—	Revision page without members, newest first; limit 1–200
GET	/api/sets/{id}/revisions/{revision}	—	That revision's details and members, each joined to the drawing's current record
PUT	/api/sets/{id}	JSON set (id empty or matching the route)	New revision
POST	/api/sets/{id}/archive / /restore	—	New revision with archived changed and the same members
GET	/api/drawings/{id}/sets?offset=0&limit=100	—	Sets, archived included, whose current revision contains the drawing, by name; 404 for an unknown drawing

• Existing-set writes follow the drawing contract: If-Match 428 missing, 400 malformed, 409 stale with the current record; writing to an archived set, archiving an archived one or restoring an active one is also 409. There is no delete.
• Every change appends an immutable revision, with its own copy of the members, in the transaction that updates the current row. An identical save (name, customer, notes and members in order) at the current revision is a no-op; reordering members is a change.
• Validation (400, no state change): name nonblank; name, customer and savedBy at most 200 characters, notes at most 4000, no NUL characters; members present and not null, at most 500, each with a nonempty drawingId that is in the library (archived drawings allowed), an integer quantity from 1 to 100000, and no drawing listed twice (one member carries the whole quantity). An empty set is allowed.
• Query: search matches name, customer or notes and customer matches by SharedListNames.Key, exactly as for drawings. sort is one of savedAt (default), name, customer, revision; ties break by id. Paging, order and the refusal of unknown, repeated or out-of-range parameters follow the drawing query.

Set releases

A set is never added to a nest, because a nest is one material. It is released: the client makes one nest per material group with every quantity multiplied, uploads each nest, and only then posts a release record naming the set revision and the nests. A failure part-way leaves ordinary nests and no release, and nothing is rolled back silently.

OpenNest.Data.SetReleaseRecord fields: id (client-allocated, so an ambiguous post can be retried), setId (from the route; a body value must match it), setRevision, multiplier, notes, releasedBy, the server-computed releasedAt, setName (the set's name at that revision) and nests, in order, each { nestId, materialName, materialGrade, thickness, units, cutType, nest }. The material fields are the group the nest was made for; nest is the nest's current metadata (deleted or not), joined on every read so status and deletion are always current, and ignored on post.

Method	Path	Body	Response
POST	/api/sets/{id}/releases	JSON release with a nonempty id	The stored release; 404 for an unknown set; 409 with the release that already holds the id or one of the nests
GET	/api/sets/{id}/releases?offset=0&limit=100	—	Release page with nests, newest first; limit 1–200
GET	/api/nests/{id}/release	—	The release the nest (deleted or not) came from, or 404

• Releases are append-only: there is no edit or delete. A mistaken release is undone by deleting its nests, which then read as deleted.
• setRevision must be a revision of the set but need not be the current one, and the set may be archived: the nests already exist by the time the release is posted, so it is never refused for a state the operator cannot fix.
• Each nest belongs to at most one release (a unique index enforces it).
• Validation (400, no state change): multiplier 1–1000; notes at most 4000 and releasedBy at most 200 characters; nests present, 1–500 entries, no nest twice, each an existing, not deleted nest with material name and grade at most 200 characters, thickness finite and at least 0 (stored rounded to 4 decimals), and units and cutType single enum names.

Schema v7 adds drawing_set_revisions (every revision's details and archived flag), drawing_set_members (each revision's members by position, unique per drawing), drawing_sets (each set's current revision), set_releases and set_release_nests. Upgrading from v6 only creates these tables; every other table is untouched. A server older than v7 refuses to open a v7 database, so rolling back needs the pre-upgrade backup. The stopped-volume tar backup includes the set and release tables; nest_manifest.py and restore_check do not verify them yet.

Storage

OpenNest.Server.NestDatabase uses SQLite (Microsoft.Data.Sqlite) with journal_mode=WAL. Each row holds the metadata columns plus the .nest archive as a BLOB. The database file path comes from --database=<path> or --database <path>, then OPENNEST_DB, defaulting to ./data/nests.db; a --database without a path, or given twice, fails startup. The server opens the database at startup, so an unusable path fails startup rather than the first request.

Run one service instance with its database on a local filesystem. Multiple PCs may use that instance, but do not share the SQLite file between replicas or place it on SMB/NFS storage. A private reentrant lock serializes complete database operations (including readers, write readbacks, health checks, and disposal); HTTP upload/body reading happens outside that lock. /healthz executes SELECT 1 on the live connection; it is a connection check, not a backup, integrity scan, or guarantee of future disk capacity.

Revision preconditions refuse stale edits; the database monitor is not a document edit lock. After a conflict, explicitly reload the newer revision or Save As a separate nest. After an ambiguous update failure, reload the current version before another save. For an ambiguous first upload (including Save As or a server switch), OpenNest keeps the client-generated pending ID and blocks another upload. The next Save checks that exact ID on the original server without writing: if it exists with the exact attempted archive and metadata, the document binds to its revision; if it is absent, another explicit Save (or Save As for a pending copy) may retry with the same ID. If it differs, reopen the server record to review it; do not discard the pending identity or blindly create another record. Older servers remain read-only for existing nests until upgraded.

Server integration tests

dotnet test OpenNest.Server.Tests/OpenNest.Server.Tests.csproj

Cross-platform; runs in both CI workflows. Each test hosts the production routes in memory (WebApplicationFactory) against its own temporary SQLite file and drives them through the real RemoteNestRepository and NestSaveSession: create, list, download, same-record update, copy, metadata-only update (archive and fileSize unchanged), and delete. Missing/invalid upload parts must return 400 and unknown ids 404, both leaving every stored record and archive hash unchanged. Tests never open data/nests.db or OPENNEST_DB; the temporary directory is removed when the host is disposed. Concurrent HTTP clients and database-level readers/writers must retain exact record membership, metadata, and archive bytes; a monitor-ownership test checks every operation without relying on stress timing. Health tests check the unchanged success response and a detail-free 503 after the database is disposed. The container smoke remains the image-level check.

Running

Locally:

dotnet run --project OpenNest.Server/OpenNest.Server.csproj

Listens on ASPNETCORE_URLS (default Kestrel ports) unless overridden.

Docker image

Docker must be built from the repository root: the server references OpenNest.Data, which references OpenNest.Core. The Dockerfile copies all three project files before restore and their source/resources before publish, including Data's embedded Defaults/CL-980.json. The allowlist .dockerignore excludes other projects, host bin/obj, local databases, environment files, and launch settings. The server image uses the .NET 8 SDK to build and the .NET 8 ASP.NET runtime to run; IO/Engine are used only by the host-side smoke tool, not the image.

docker build --pull -f OpenNest.Server/Dockerfile -t opennest-server:local .
# A release build stamps its version and exact source commit:
docker build --pull -f OpenNest.Server/Dockerfile --build-arg VERSION=X.Y.Z \
  --build-arg SOURCE_REVISION="$(git rev-parse HEAD)" -t opennest-server:X.Y.Z .

Without build arguments the image is labeled version 0.0.0-dev, revision unknown: a development build, never a release. The OCI labels are org.opencontainers.image.source, version, revision and base.name. SDK_IMAGE/RUNTIME_IMAGE build arguments accept digest-pinned base images; for a release, record the base digests the build log resolved.

Runtime contract:

• Container port 8090 (ASPNETCORE_URLS=http://+:8090). Change the host port mapping, not the container port: the healthcheck probes 8090.
• Runs as the .NET base image's non-root app user (UID/GID 1654). Application files are root-owned and read-only to it; within /app only /app/data is writable, and a fresh named volume inherits that directory's ownership. The database stays OPENNEST_DB=/app/data/nests.db. The server also needs the container's writable /tmp: multipart uploads over 64 KiB are buffered there.
• HEALTHCHECK runs curl --fail http://127.0.0.1:8090/healthz (interval 30 s, timeout 5 s, start period 10 s, retries 3, and a 2 s start interval during the start period on Docker 25+). curl is in the image only for it.
• A read-only or unwritable data location fails startup: the process exits rather than serving requests it cannot store.
• Uploads: Kestrel's default 30,000,000-byte request limit covers the whole multipart request, so an archive must be slightly smaller (the smoke stores a 29,000,000-byte one). A larger request gets 413 and nothing is stored. The desktop client streams without Expect: 100-continue, so it typically reports a closed connection instead of the 413 status. The save still fails.

Data written by a root-run image or a root-owned bind mount is not writable by UID 1654, so startup fails. With the operator's approval, change ownership of that data location only, once, for example docker run --rm --user 0 --entrypoint chown -v <volume>:/app/data <image> -R 1654:1654 /app/data (or chown -R 1654:1654 on the bind-mounted directory itself, never its parents). The image has no root entrypoint that changes ownership.

Deploying with Compose

compose.server.yaml runs a published image (no local build) with no-new-privileges, all capabilities dropped, and a named data volume. Copy it and OpenNest.Server/server.env.example to a deployment directory, save the env file under a local name such as server.env, and set:

• OPENNEST_SERVER_IMAGE: ghcr.io/ajisaacs/opennest-server:X.Y.Z or, preferably, ghcr.io/ajisaacs/opennest-server@sha256:<verified-manifest-digest> from the successful clean pulled-image verification job. Do not substitute Docker's store-dependent image Id or the config blob digest for this verified registry manifest digest; see server image releases.
• OPENNEST_BIND_ADDRESS: 127.0.0.1 for testing on the host; this host's trusted LAN address for shop PCs. Do not use 0.0.0.0. The bind address is only one layer: verify that the network/firewall admits only trusted clients.
• OPENNEST_HOST_PORT (default 8090) and OPENNEST_DATA_VOLUME (default opennest-data).

Publication requires an existing verified private package; first-package initialization needs separate owner authorization outside the release workflow. An authorized deployment operator must obtain a read:packages token out of band and log in on the deployment host using docker login ghcr.io -u <github-user> --password-stdin, piping the token from a secret manager or a protected prompt (never a token literal in shell history). Keep Docker's credentials protected on that host; do not put credentials in server.env, Compose, the image, or source control. Login/pull success does not authorize making the package public. Until an approved image has passed the published-digest smoke, the reference below is only a template, not an available verified image.

compose="docker compose --env-file server.env -f compose.server.yaml"
$compose pull || { printf 'STOP: image pull failed\n' >&2; exit 1; }
$compose up -d
$compose ps                                       # STATUS shows (healthy)
curl --fail http://<bind-address>:<port>/healthz  # {"status":"ok"}

In the desktop app choose File > Storage Mode..., Database, and enter the base URL http://<bind-address>:<port>, without /healthz or /api/nests. To set or change just the saved server address without changing File/Database mode, use Tools > Options > Nest server URL > Set / Change Server. This saves the URL immediately, independently of the Options Save button; it does not migrate nests or enable Database mode. Canceling Options afterward does not undo that address.

Backup, restore, and upgrade

SQLite runs in WAL mode, so copying nests.db from a running service is not a backup. Back up during a maintenance period with ALL client writes disabled, including metadata edits, delete, recover and restore, until capture finishes. The HTTP manifest is not a transaction across requests; writes during capture can invalidate it. Keep the entire stopped data directory, including any WAL/SHM files, original image digest and deployment configuration. Copy scripts/Server.Tests/nest_manifest.py from this source version to the deployment directory as nest_manifest.py (Python 3 standard library only). It enumerates bounded active AND deleted pages plus ALL revision pages, then streams every current/historical archive through SHA-256. Canonical JSON includes complete metadata, IDs, current pointers, visibility, timestamps, byte sizes and restore source relationships. Missing pages, legacy endpoints, failed downloads, inconsistent totals and invalid revision chains fail closed, not as empty history. Protect manifests like the database: they contain names and comments. Each step below is a Bash function that checks every command and stops at the first failure, however it is called; define them in the shell where compose is set:

url=http://<bind-address>:<port>

# Canonical complete manifest; stdout is emitted only after all reads succeed.
nest_manifest() (
  python3 ./nest_manifest.py "$1" || exit 1
)

# Records the expected manifest and metadata, then archives the stopped data directory.
# The service is restarted even if the archive fails.
backup_nests() (  # usage: backup_nests opennest-data-YYYY-MM-DD
  set -o pipefail
  fail() { echo "STOP: $*" >&2; exit 1; }
  for suffix in tar manifest metadata.json; do
    [[ ! -e "$1.$suffix" ]] || fail "$1.$suffix already exists"
  done
  nest_manifest "$url" > "$1.manifest" || fail "could not record the manifest"
  curl -fsS "$url/api/nests" > "$1.metadata.json" || fail "could not record the metadata"
  $compose stop || fail "could not stop the service"
  $compose run --rm --no-deps -T --entrypoint tar opennest-server -C /app/data -cf - . > "$1.tar"
  status=$?
  $compose start || fail "could not restart the service"
  curl -fs --retry 30 --retry-all-errors --retry-delay 1 "$url/healthz" > /dev/null \
    || fail "the restarted service is not healthy"
  [[ $status == 0 ]] || fail "the archive failed; $1.tar is incomplete"
  echo "Backed up all records and revisions to $1.tar"
)

backup_nests opennest-data-YYYY-MM-DD

Restore only into a volume created for that restore, never into an existing one (least of all the active volume). restore_check refuses an existing volume, extracts the backup into a new one, starts a temporary loopback-only container on it, and requires complete active/deleted revision metadata and archive hashes to match. It removes only the container it started; the new volume is kept either way:

restore_check() (  # usage: restore_check opennest-data-YYYY-MM-DD <image> [new-volume]
  set -o pipefail
  fail() { echo "STOP: $*" >&2; exit 1; }
  [[ -r "$1.tar" && -r "$1.manifest" && -r "$1.metadata.json" ]] || fail "backup is incomplete"
  restore=${3:-opennest-data-restore-$(date +%Y%m%d-%H%M%S)}
  existing=$(docker volume ls --format '{{.Name}}') || fail "could not inspect Docker volumes"
  while IFS= read -r volume; do
    [[ "$volume" != "$restore" ]] || fail "volume $restore already exists"
  done <<< "$existing"
  owner=$(python3 -c 'import secrets; print(secrets.token_hex(16))') || fail "could not create ownership token"
  docker volume create --label "com.opennest.restore.owner=$owner" "$restore" > /dev/null \
    || fail "could not create volume $restore"
  [[ $(docker volume inspect --format '{{index .Labels "com.opennest.restore.owner"}}' "$restore") == "$owner" ]] \
    || fail "restore volume ownership differs"
  echo "Restoring into new volume $restore"
  docker run --rm -i --network none --entrypoint tar -v "$restore:/app/data" "$2" \
    -C /app/data -xf - < "$1.tar" || fail "extraction into $restore failed"
  check=$(docker create --cap-drop ALL --security-opt no-new-privileges:true \
    -p 127.0.0.1::8090 -v "$restore:/app/data" "$2") \
    || fail "could not create the check container"
  trap 'docker rm -f "$check" > /dev/null' EXIT
  docker start "$check" > /dev/null || fail "the check container did not start"
  port=$(docker port "$check" 8090/tcp | cut -d: -f2) && [[ -n $port ]] || fail "no check port"
  curl -fs --retry 30 --retry-all-errors --retry-delay 1 "http://127.0.0.1:$port/healthz" > /dev/null \
    || fail "the restored service is not healthy"
  curl -fsS "http://127.0.0.1:$port/api/nests" | cmp - "$1.metadata.json" || fail "metadata differs"
  nest_manifest "http://127.0.0.1:$port" | diff - "$1.manifest" || fail "archives differ"
  echo "Verified $restore. Keep it isolated until the operator explicitly approves switching volumes."
)

restore_check opennest-data-YYYY-MM-DD <image>

Switch only after restore_check prints Verified. If it stops, inspect or remove the new volume it named; the active volume is untouched.

To upgrade, take a stopped-service backup and keep a copy of the current server.env. Prefer a digest in OPENNEST_SERVER_IMAGE so that file records exactly what ran; for a tag, record docker image inspect --format '{{join .RepoDigests " "}}' <image> first. Change only OPENNEST_SERVER_IMAGE after migration/restore rehearsal on a disposable copy, then run $compose up -d only with separate deployment authorization. Never run an old image against a migrated database, even if its current-table GETs still work: old writers can overwrite snapshots or physically delete rows without updating history. Rollback requires a NEW volume restored from the verified PRE-migration backup and the exact old image/configuration. Test its old current metadata/byte manifest on that isolated copy before switching; the revision-aware manifest deliberately refuses old endpoints. Archive the migrated volume unchanged for recovery; reverting loses post-backup changes from the old server's view, so reconcile them deliberately, never replay them automatically. A pre-version server backup needs its own current-record metadata/byte oracle before migration; verify those records become byte-identical revision 1 on the disposable upgraded copy. Never run docker compose down -v: it deletes the data volume.

Capacity and retention

Every revision currently stores a full archive BLOB, including metadata-only and visibility revisions. Budget approximately the sum of ALL revision archive sizes plus current-table archives, SQLite indexes/pages, WAL and backups; do not estimate capacity from active current records alone. Measure volume/database and WAL growth during the disposable rehearsal, reserve space for migration and a separate verified restore copy, and monitor free space before rollout. There is no automatic retention or purge; no revision deletion is authorized. Deduplication requires a separately reviewed migration after measuring actual growth. An old backup is not a substitute for retaining the migrated volume during rollback.

Isolated container smoke

Prerequisites: a local Linux Docker daemon on the default Unix-socket context, Bash, Python 3, curl, GNU timeout, mktemp, tar, cmp, and a .NET SDK able to build net8.0 projects (.NET 8 or newer). Restore needs NuGet access or cached packages. Run from the repository root after building the image:

scripts/Test-ServerContainer.sh --image opennest-server:local
# Optional durable logs outside the default ignored .hermes/progress directory:
scripts/Test-ServerContainer.sh --image opennest-server:local --results /path/to/results

Manifest refusal tests run without Docker. The optional image variables also exercise the Bash backup/restore blocks extracted verbatim from this page in a unique disposable Compose project, including existing-file/volume refusal and tampered-history rejection. The old image test restores ONLY a synthetic pre-migration backup to another fresh volume and compares every old record's metadata/bytes; it never mounts a migrated database into the old image.

python3 -m unittest discover -s scripts/Server.Tests -p test_revision_backup.py -v
OPENNEST_TEST_IMAGE=opennest-server:local \
OPENNEST_TEST_OLD_IMAGE=<local-pre-version-image> \
  python3 -m unittest discover -s scripts/Server.Tests -p test_revision_backup.py -v

Local smoke evidence does not imply registry publication or production deployment.

The wrapper accepts an image, not a URL or an existing volume/container. It uses only the local default Docker context, creates uniquely labeled disposable resources, and publishes to 127.0.0.1 on a Docker-allocated port. Every container runs with the Compose example's --cap-drop ALL and no-new-privileges. It builds the smoke tool once into a private temporary directory, honoring TMPDIR; readiness, HTTP requests, and child commands have watchdogs.

Before starting a container it checks the image's numeric non-root USER, the documented HEALTHCHECK command and timings (including the start interval), and the OCI source/version/revision labels. A service is ready only when Docker reports the image's own healthcheck healthy and the host receives {"status":"ok"}. Each service then checks that the server process (PID 1) runs as the image user with no effective capabilities, owns nests.db, and can write /app/data but not /app or the server DLL.

The test-only scripts/Server.Tests/Server.Tests.csproj console exercises the existing .NET storage client; it is not included in the server image.

Before seeding, the wrapper restores two synthetic legacy-schema rows into its fresh volume, verifies byte-identical revision-1 migration against a precomputed complete metadata/hash oracle, then checks repeated startup preserves history. The synthetic fixture is generated in code: one plate and one rectangular part, with no customer files. seed uses the actual RemoteNestRepository and NestSaveSession to create, update the same ID, and save a copy with a new ID. It also creates metadata-only, historical-restore, delete/recover/delete revisions, leaving a six-revision hidden record with a restore-source relationship. It checks read-back metadata, session binding, exact list membership, raw camelCase/string-enum JSON, identical downloaded bytes, and NestReader geometry and counts. reject snapshots every record's metadata and archive hash, then checks 400 responses for missing/invalid/null metadata and missing/empty files on both upload and file-update routes; unknown metadata/file IDs must return 404. Every rejected operation must leave the snapshot unchanged. limits stores, downloads byte-for-byte, and soft-deletes a 29,000,000-byte synthetic archive (buffered to disk by the non-root runtime), then requires 413 for a request over 30,000,000 bytes sent with Expect: 100-continue, and a reported failure from the real client's oversized upload and file update. The snapshot must be unchanged after each.

The persistence stage stops and removes the first container while retaining its fresh named volume, starts a replacement on another allocated loopback port, and runs verify against saved IDs, all metadata (including server-assigned savedAt), archive hashes, and exact list membership. It then follows the documented stopped-service backup: tar of the whole data directory through the image, restored into a second owned volume whose service must pass the same verify. The Python manifest is captured after all mutations and compared exactly after restart and volume restore, with page size 1 to exercise every pagination boundary. It covers active and deleted current records and ALL revisions, not just the ordinary /api/nests enumeration; the large soft-deleted archive remains included. The original volume mounted read-only, and a root-owned tmpfs, must each make the server exit nonzero at startup with a SQLite error and never report healthy.

A missing/invalid state file or failed assertion exits nonzero. Direct tool use is test-only: it accepts plain loopback URLs, and seed refuses a nonempty server unless --test-allow-nonempty is explicitly supplied for an owned disposable target. Prefer the wrapper; loopback alone is not proof that an existing server is disposable.

Each invocation writes a fresh run.* results directory containing build, smoke, container, image, runtime, ownership, and cleanup logs. On failure those logs remain for diagnosis. The exit trap removes only invocation-owned containers and named volumes; transient state, the backup archive, synthetic databases, and build artifacts are removed on success or failure. It never prunes Docker or alters preexisting services/volumes. Cleanup requires successful empty Docker inventory readback, including for an already-removed container. Unresolved ownership, lookup, removal, or readback errors make the wrapper fail without deleting unproven resources; inspect cleanup.log and outcome.log for diagnostics. An original smoke failure is preserved even if cleanup also fails.

The focused cleanup regression needs only Python 3 and Bash:

python3 -m unittest discover -s scripts -p test_server_container_cleanup.py -v

It executes the wrapper's actual cleanup functions with explicitly mocked Docker transport, including timeouts and missing cidfiles; it does not replace the real container smoke or contact a daemon.