Data Model

Data Model

All tables use UUID primary keys (never auto-increment integers — see invariants.md). Timestamps are timezone-aware (DateTime(timezone=True)), stored in UTC.

Entity relationships

erDiagram
    USER ||--o{ COLLECTION : owns
    USER ||--o{ PROFILE : owns
    USER ||--o{ API_TOKEN : owns
    USER ||--o{ COLLECTION_RATING : rates
    USER ||--o{ COLLECTION_COMMENT : writes
    COLLECTION ||--o{ ARTIFACT : contains
    COLLECTION ||--o{ COLLECTION_RATING : "rated by"
    COLLECTION ||--o{ COLLECTION_COMMENT : "commented on"
    PROFILE }o--|| COLLECTION : "base_collection_id"
    PROFILE }o--o{ COLLECTION : "additional_collection_ids (JSON list, not FK)"

    USER {
        uuid id PK
        string email UK
        string display_name
        string password_hash "nullable — SSO-only users have none"
        string oidc_sub UK "nullable"
        string oidc_provider "nullable"
        bool is_active
        bool is_admin
        string role "user | moderator | admin"
        bool notify_on_download "daily digest opt-in"
        bool notify_on_comment "immediate email opt-in"
        bool mfa_enabled
        string totp_secret "nullable — set once MFA is enrolled"
    }
    COLLECTION {
        uuid id PK
        uuid owner_id FK
        string name
        string git_url "real URL, or imported://name / seed://type/slug for non-git sources"
        string collection_type "base | additional"
        string visibility "private | public"
        bool is_active "soft-delete flag"
        int artifact_count "denormalized cache"
        int download_count "tracked for community collections"
        bool published "true only once moderation_status = approved"
        string category "nullable — browse category"
        bool is_starter_pack "true for the seeded starter collections"
        float avg_rating "denormalized cache of collection_ratings"
        int rating_count "denormalized cache of collection_ratings"
        string moderation_status "draft | submitted | approved | denied"
        string moderation_reason "nullable — set on deny"
        datetime submitted_at "nullable"
        datetime moderated_at "nullable"
        uuid moderated_by FK "nullable — the reviewing moderator/admin"
        int last_digest_download_count "digest-script watermark"
        datetime last_digest_sent_at "nullable"
    }
    ARTIFACT {
        uuid id PK
        uuid collection_id FK
        string artifact_type "rule|skill|agent|workflow|model_config"
        string name
        string target_compatibility "JSON-encoded list[str]"
        string tags "JSON-encoded list[str]"
        int priority
        text body "markdown"
        bool is_enabled
        datetime deleted_at "nullable — soft-delete timestamp"
    }
    PROFILE {
        uuid id PK
        uuid owner_id FK
        uuid base_collection_id FK
        string additional_collection_ids "JSON-encoded list[uuid]"
        string disabled_artifact_ids "JSON-encoded list[uuid]"
        string target_framework "nullable — preferred compile target, informational only"
        bool is_public
        datetime deleted_at "nullable — soft-delete timestamp"
        int version "incremented on every update"
    }
    API_TOKEN {
        uuid id PK
        uuid user_id FK
        string token_prefix "first 8 chars, for lookup"
        string token_hash "bcrypt"
        datetime expires_at
        bool is_active
    }
    COLLECTION_RATING {
        uuid id PK
        uuid collection_id FK
        uuid user_id FK
        int stars "1-5, CHECK constraint"
        datetime created_at
        datetime updated_at
    }
    COLLECTION_COMMENT {
        uuid id PK
        uuid collection_id FK
        uuid user_id FK
        text body "max 2000 chars"
        datetime created_at
        datetime deleted_at "nullable — soft-delete timestamp"
    }

Loading

DOC_CACHE (framework documentation cache for adapter compatibility rules) has no relationships to the rest of the schema — it's a standalone TTL cache, keyed by framework + content_hash.

SYSTEM_SETTINGS also has no relationships — it's a singleton row (id always 1) holding global, admin-editable config: which auth providers are enabled, whether registration/MFA are allowed or forced, and the doc cache TTL. See the system_settings section below.

Tables

users

The only table every other owned table has a foreign key into. A user can authenticate via password (password_hash, nullable — SSO-only accounts never set one), OIDC/GitHub/Google (oidc_sub + oidc_provider, unique together), or both. is_admin bypasses ownership checks everywhere; see invariants.md. mfa_enabled + totp_secret back TOTP-based MFA (pyotp) — totp_secret is only set once enrollment completes via POST /auth/me/mfa/totp/setup + .../verify. reset_token_hash

• reset_token_expires_at back password-reset-by-email (POST /auth/forgot-password / /auth/reset-password) — only the SHA-256 hash of the emailed token is stored, mirroring the API-token-hash pattern; the token is single-use (both fields are cleared on a successful reset) and expires after 1 hour.

role (user/moderator/admin, default user) is additive alongside is_admin — see ADR-0007 for why there are two fields instead of one. is_admin still gates every pre-existing admin check; role is read only by the newer require_moderator_or_admin dependency that gates the community moderation routes. The two are kept in sync by PATCH /auth/users/{id}/role (admin-only), never independently. notify_on_download/notify_on_comment (both default false) are per-user, self-service opt-ins for community-collection notification emails — see the "Community collections" section below.

collections

A named bag of artifacts. git_url is a real repository URL for GitHub-imported collections, or a synthetic URI for anything that didn't come from a real git remote — imported://<name> for a local scan or bulk-exported artifacts, imported://community/<name> for a community import, seed://<collection_type>/<slug> for the built-in starter packs (seed_collections.py) — don't assume it's always dereferenceable. visibility (private/public) is the only access-control dimension beyond ownership; there is no per-collaborator sharing. is_active is a soft-delete flag — deleted collections are never physically removed. artifact_count is a denormalized cache updated by every route that adds/removes artifacts (import, bulk delete, bulk export) — if you add a new artifact-mutating route, update it there too or the count will drift. is_starter_pack marks the collections seeded on every backend startup from collections/base/ and collections/additional/, owned by a dedicated passwordless system account — see the "Starter packs" section in CLAUDE.md. avg_rating/rating_count are a denormalized cache of collection_ratings, recomputed transactionally on every rating write/delete — see the collection_ratings section below. moderation_status (draft/submitted/approved/denied) is the single source of truth for the community-publishing lifecycle — see ADR-0008 and the "Community collections" section below; published/visibility are only ever flipped to public by the approve action, never by submission itself. last_digest_download_count/last_digest_sent_at are watermark fields for the daily download-digest script (app/scripts/send_download_digests.py), not exposed via the API.

artifacts

Belongs to exactly one collection (collection_id, ON DELETE not configured — see invariants.md for what that means in practice). Has no owner_id of its own — authorization is always done via the parent collection. target_compatibility and tags are JSON-encoded into Text columns rather than being normalized into their own tables; every route that returns artifacts must decode them before responding (see debugging.md). is_enabled is per-artifact and independent of the collection's is_active — a disabled artifact is skipped during compilation unless a profile explicitly asks to include_disabled. deleted_at supports soft-delete — artifacts are never physically removed from the database.

collection_ratings

One row per (collection, user) — enforced by a DB-level unique constraint (uq_collection_ratings_collection_user), not just upsert logic in the route handler. stars has a CHECK (stars >= 1 AND stars <= 5) constraint in addition to the 422 the API returns for an out-of-range value. PUT /collections/{id}/rating upserts the caller's own row and recomputes Collection.avg_rating/rating_count transactionally; DELETE removes it and recomputes again. Rating requires Collection.moderation_status == "approved" and blocks self-rating (owner_id == current_user.id → 400).

collection_comments

Soft-deleted (deleted_at), never hard-deleted, per the soft-delete rule below. body is capped at 2000 characters, enforced at the Pydantic schema layer (CollectionCommentCreate), not just informally. Commenting requires Collection.moderation_status == "approved", same gate as ratings. Deletion is allowed for the comment's own author, the collection's owner, or a moderator/admin — checked in the route handler, not by a DB constraint. GET /collections/{id}/comments still requires authentication like every other route in this API (see invariants.md) — the plan that introduced comments described this endpoint as public, but that would have been the first unauthenticated data route in the app, so it was kept consistent with the existing all-routes-require-auth convention instead.

profiles

base_collection_id is a real foreign key; additional_collection_ids and disabled_artifact_ids are not — they're JSON-encoded UUID lists, resolved and validated at request time rather than enforced by the database. This means the database will happily store a profile referencing a collection that's since been deleted or made private; the compiler (compile_profile()) silently skips any collection ID it can't resolve rather than erroring. version increments on every PUT — there's no history table, just the counter. target_framework is an optional, free-string "preferred target" hint shown in the UI — it's never validated against the adapter registry and has no effect on compilation; the actual target is chosen per-compile-request. is_public is the profile's own visibility flag, independent of the visibility of the collections it references (see the documented gap in invariants.md). deleted_at supports soft-delete — profiles are never physically removed from the database.

Compile-time validation warnings (response-only — no table)

POST /profiles/compile's response carries a warnings field (list[ValidationIssue], each {level: "warning", code: str, message: str} — app/models/profile.py), additive to the pre-existing {profile_id, profile_name, target, artifact_count, files} shape. compile_profile() (app/services/compiler.py) computes these fresh on every call, in the same step that deduplicates artifacts by name — nothing here is persisted to the database, there's no validation_issues table, and no migration was needed to add it. POST /profiles/compile/zip can't embed this JSON alongside the zip's binary response, so it instead appends a _myace_warnings.txt file inside the archive when there are any warnings to report. See AGENTS.md rule 32 for the name-collision rule this currently powers. The {level, code, message} shape is deliberately generic rather than name-collision-specific, so a future compile-time check (e.g. a dangling agent-handoff reference) can append its own code to the same warnings list without a schema change or a second parallel warnings mechanism.

The same response also now carries compiled_hash: str — a sha256 over a deterministic serialization of files (compute_compiled_hash(), same codebase). This, too, is response-only: nothing is persisted, and it's recomputed fresh on every compile. GET /profiles/{id}/compile-status returns just {compiled_hash, updated_at} for the same profile+target without the full files payload — see ADR-0009 for why this exists (cheap polling for the CLI's check/watch commands) and its honestly-documented cost trade-off (it still resolves artifacts and runs translate(); it only saves the client the transfer cost of file content, not the server's compute cost).

sync_statuses

Opt-in only — a row here exists exclusively because a user ran myace check --report or myace watch --report from some machine; nothing is written here automatically by pull, check, or watch in their default form (see ADR-0009). user_id and profile_id are real foreign keys; target and machine_label are free strings (the latter defaults to the reporting machine's hostname but is otherwise arbitrary user-supplied text, not validated against anything). locally_modified_files is a JSON-encoded Text column, same pattern as Artifact.tags (AGENTS.md rule 11) — always converted through SyncStatusRead before leaving a route, never returned as a raw row. A unique constraint on (user_id, profile_id, target, machine_label) makes POST /sync/report an upsert: a second report for the same triple updates in_sync/locally_modified_files/ last_checked_at on the existing row rather than inserting a duplicate. GET /sync/status only ever returns the caller's own rows — there is no cross-user visibility here, not even for admins (this is personal local- machine state, not a community feature); see invariants.md.

api_tokens

CLI authentication. The raw key is generated once, split into an 8-char token_prefix (stored in plaintext, used to narrow the lookup) and the full key (bcrypt-hashed into token_hash). The raw key is returned to the caller exactly once, at creation — the database never stores it in recoverable form. expires_at is enforced in get_current_user; there's no background job that revokes expired tokens, they just stop authenticating.

doc_cache

Unrelated to the auth/ownership model above — a TTL-based cache of fetched framework documentation, used by adapter compatibility checks. Not owner-scoped; it's shared, read-only reference data. deleted_at supports soft-delete — cache entries are never physically removed from the database.

system_settings

A singleton row (id is always 1, enforced by convention rather than a DB constraint) holding global, admin-only config: which auth providers (oidc_enabled/github_enabled/google_enabled) and registration (allow_registration) are turned on, whether MFA is available (mfa_enabled) or mandatory (mfa_forced), and doc_cache_ttl_days. Read via GET /admin/settings, written via PATCH /admin/settings (SystemSettings.tsx) — both admin-gated. Not owner-scoped; there is exactly one row.

Also holds the SMTP config used for password-reset emails: smtp_enabled, smtp_host, smtp_port, smtp_username, smtp_from_email, smtp_from_name, smtp_use_tls, and smtp_password_encrypted (the admin-entered password, Fernet-encrypted — see ADR-0006). Any of these left unset falls back to the matching SMTP_* env var at runtime (app/services/effective_settings.py); a non-empty DB value always wins. SystemSettingsRead never returns smtp_password_encrypted itself — only a computed smtp_password_set: bool — and SystemSettingsUpdate accepts a plaintext, write-only smtp_password field that the PATCH /admin/settings handler encrypts before persisting.

Also holds per-provider OAuth credentials — oidc_client_id, oidc_client_secret_encrypted, oidc_issuer_url, oidc_scopes, github_client_id, github_client_secret_encrypted, google_client_id, google_client_secret_encrypted — entered via System Settings' expandable Authentication Providers rows instead of only .env. Same precedence and encryption contract as SMTP above (get_effective_oauth_config() in effective_settings.py; {provider}_client_secret_set computed booleans, never the encrypted value, on SystemSettingsRead). security.py's get_oauth_client() rebuilds a provider's Authlib client whenever its effective config changes (tracked by a fingerprint), so a credential saved here takes effect on the next login/callback request without a restart.

Also holds disabled_adapters — a JSON-encoded list[str] of adapter names (BaseAdapter.adapter_name()) an admin has disabled system-wide, same manual-JSON-in-Text-column convention as Profile.disabled_artifact_ids. Toggled via PATCH /admin/adapters/{name}?enabled=<bool>. Enforced in two places: the frontend's compile target picker (TargetExporter.tsx) filters disabled adapters out of the dropdown, and compile_profile() (app/services/compiler.py) — the single choke point both /profiles/compile and /profiles/compile/zip funnel through — raises AdapterDisabledError (→ HTTP 400) if the requested target is disabled, so the restriction can't be bypassed by calling the API directly.

Why JSON-as-text instead of proper junction tables

Collection.tags/target_compatibility and Profile.additional_collection_ids/disabled_artifact_ids could be normalized into join tables. They're JSON-in-text instead because:

• These lists are always read/written as a whole (never queried element-wise — nothing does "find all artifacts tagged X" at the SQL level today).
• It keeps the schema small for what is, so far, a single-user-per-resource data model with no need for relational queries across these lists.

Community collections

Publishing goes through a moderation queue, not a self-serve DB write — see ADR-0008 for the full state machine. POST /collections/{id}/publish moves a collection from draft/denied to submitted; only the moderator/admin-only POST /moderation/{id}/approve action flips published = True and visibility = "public". GET /collections/community lists anything with published = True AND is_active = True (unchanged query, but now that predicate is only ever true for moderation_status = "approved" rows) and accepts sort=rating|downloads|alpha (default downloads). GET /moderation/queue (moderator/admin only) lists submitted rows, oldest-first by default, accepting the same three sort values as an override. download_count tracks how many times the collection has been imported; category is a free-text browse string.

Denying (POST /moderation/{id}/deny, body {reason}) sets moderation_status = "denied" and moderation_reason, leaves published = False, and lets the owner edit and resubmit (back to submitted). A moderator/admin can also fix name/description/ category directly via PATCH /moderation/{id}/meta, regardless of moderation_status — never git_url/git_branch/artifact content, and never available to the collection's own owner (they use their existing PATCH /collections/{id} instead).

This is entirely separate from the seeded starter-pack set (is_starter_pack = True, owned by the system account — see the collections table above): starter packs are grandfathered straight to moderation_status = "approved" at seed time, never routed through the queue. The migration that introduced moderation_status did the same one-time grandfathering for every collection that was already published = True AND is_active = True, matching this exact query's filter — nothing already visible in the community disappeared or needed re-review.

Importing a community collection creates a new user-owned Collection with copied Artifact rows and increments download_count on the source. If the source owner has notify_on_download = True, they get a daily digest email (not a per-download email) — see last_digest_download_count/ last_digest_sent_at in the collections table above and docs/deployment.md's cron entry for app/scripts/send_download_digests.py. Comment creation on an approved collection emails the owner immediately instead (comments are much lower-volume than downloads) if notify_on_comment is on and they have an email on file.

If a future feature needs to query inside these lists (e.g. "find all public collections tagged python"), that's the point to normalize — don't work around it with LIKE '%...%' queries on the JSON text.