Skip to main content

unitysvc-core

Shared data models, JSON schemas, and validation helpers for the UnitySVC ecosystem. This package is audience-neutral: it contains only the foundational types and helpers that are consumed by the UnitySVC backend, the customer SDK, the admin CLI, and the seller SDK.

It intentionally does not include any CLI, HTTP client, or audience-specific helpers. Those live in the corresponding packages:

  • unitysvc-sellers — seller SDK + usvc_seller CLI, catalog builders (populate_from_iterator, render_template_file, convert_convenience_fields_to_documents, etc.)
  • unitysvc — customer SDK + usvc CLI
  • unitysvc-admin — admin CLI

What's inside

src/unitysvc_core/
├── models/
│   ├── base.py            # All enums (status, type, currency, rate-limit, etc.)
│   ├── pricing.py         # PriceStr, UsageData, all *PriceData classes,
│   │                      # Pricing union, validate_pricing, cost calculation
│   ├── channel.py         # UpstreamAccessConfig, ChannelConfig, per-protocol
│   │                      # servers, ChannelParams, RETIRED_CHANNEL_KEYS
│   ├── documents.py       # DocumentData
│   ├── service.py         # RateLimit, ServiceConstraints, AccessInterfaceData,
│   │                      # UpstreamAccessConfigData
│   ├── validators.py      # validate_name, validate_service_options,
│   │                      # suggest_valid_name, SUPPORTED_SERVICE_OPTIONS
│   ├── provider_v1.py     # Provider file schema + ProviderData
│   ├── offering_v1.py     # Offering file schema + ServiceOfferingData
│   ├── listing_v1.py      # Listing file schema + ServiceListingData
│   ├── promotion_v1.py    # Promotion file schema + PromotionData
│   └── service_group_v1.py
├── schema/                # Generated JSON schemas (run scripts/generate_schemas.py; CI enforces sync)
├── validator.py           # DataValidator — per-file schema validation,
│                          # union-field reference checks, API-key secret
│                          # scanning, Jinja2 syntax validation, etc.
└── utils.py               # Content hashing, mime/extension helpers,
                           # data file loaders (JSON/TOML + overrides),
                           # schema-based file discovery

What's deliberately NOT here

  • CLI code — typer, rich, no entry points
  • HTTP clients — no httpx, no API wrappers
  • Seller catalog builders — populate_from_iterator, render_template_file, convert_convenience_fields_to_documents, execute_script_content live in unitysvc-sellers
  • Seller-catalog-layout validation — validate_provider_status, validate_all_service_directories, resolve_provider_name, resolve_service_name_for_listing live in unitysvc-sellers

Installation

pip install unitysvc-core

Dependencies: pydantic, email-validator, jsonschema, jinja2, tomli-w. No typer, no rich, no httpx.

Usage

from pathlib import Path

import unitysvc_core
from unitysvc_core.models import ProviderV1, OfferingV1, ListingV1
from unitysvc_core.models.pricing import UsageData, validate_pricing
from unitysvc_core.validator import DataValidator

# Validate a single catalog directory
validator = DataValidator(
    data_dir=Path("./catalog"),
    schema_dir=Path(unitysvc_core.__file__).parent / "schema",
)
results = validator.validate_all()

# Typed pricing + cost calculation
pricing = validate_pricing({
    "type": "one_million_tokens",
    "input": "0.50",
    "output": "1.50",
})
cost = pricing.calculate_cost(UsageData(input_tokens=1_000, output_tokens=500))

History

This package was split out of unitysvc-services (see unitysvc/unitysvc-services#99). The seller CLI, catalog builders, and seller-catalog-layout validators moved to unitysvc-sellers.

License

MIT

Required enrollment parameter usage

Each channel's required parameters must be read by that channel or by document content (a .j2 document's template source, which the seller validator reads from disk as the SDK uploads it); another channel's use does not count. Metadata such as display_name or description is never rendered with params, and a user access interface never reads them. File validation checks a listing against its paired offering when present; backend ingest calls the same validate_required_parameter_usage helper with the complete listing and offering. Schema declarations alone are not consumers. Jinja attribute access, string-key lookups, and params.get("name") are inspected; dynamic access and whole-map iteration conservatively consume all parameters. Deprecated content can still be retired.

Canonical user access interfaces

A user access interface is the service's canonical address, the same for every caller. validate_listing_interfaces_are_canonical renders {{ service_name }} (the one value the platform fills in when it stores an interface) and then requires:

  • a base_url that is one of the four gateways, ${API_GATEWAY_BASE_URL}, ${S3_GATEWAY_BASE_URL}, ${SMTP_GATEWAY_BASE_URL} or ${MCP_GATEWAY_BASE_URL}, optionally followed by / and a path (/a/<alias> included); never an external URL, another variable or a secret reference;
  • no template left over. {{ routing_vars.X }} would stay literal in the stored interface and never route, and a per-enrollment value (enrollment.*, params.*) is never canonical: an enrollment is reached through /e/<CODE> on the API gateway, which a listing never declares.

The one exception is {{ enrollment.code }} as the whole SMTP routing_key.username or the whole S3 bucket segment, a reminder of the per-enrollment address on protocols without a path for it. The shared validator runs it on every listing_v1; backend ingest calls the same function (unitysvc/unitysvc#2726).

Listing audience

A listing may say who the service is for and what a buyer must know before enrolling, as audience: a short list of plain-text phrases (unitysvc/unitysvc#2717):

"audience": [
  "Rice University students (enrollment at Rice required)",
  "US customers only / export restrictions may apply"
]

At most 8 phrases, each 1-120 characters after trimming, no URLs or markup (details belong in the description or documents), and no phrase twice. The phrases are shown before enrollment and read by assistants matching services to customers; the platform never enforces them. capabilities says what a service does; audience says who it is for.

Exports and customer access requests

Three stages (unitysvc/unitysvc#2720): a target service's exports (what it exposes), a seller's customer_access_requests (what the seller asks the enrolling customer to approve: using other services through that customer's account and authority), and the issued key's policy (what the customer actually approved). There is no central list of services. unitysvc_core.grants holds the formats and the checks; it does no I/O.

Invariant. Calls made under an approved request run under the enrolling customer's delegated identity, within the approved permissions, options and limits. Looking up a target's exports validates a request's capabilities; customer consent and key issuance establish its authority. A request grants nothing by itself.

1. The exports format (the offering, top level). How other services may use this one:

"exports": {
  "permissions": {"read": ["GET /records/*"], "write": ["POST /records", "PUT /records/*"]},
  "options": {
    "type": "object",
    "properties": {
      "log_level": {"type": "string", "enum": ["debug", "info", "warn", "error"], "default": "info",
                    "title": "Lowest level recorded"},
      "max_size": {"type": "integer", "minimum": 1, "maximum": 10000}
    }
  },
  "options_ui_schema": {"log_level": {"ui:widget": "radio"}}
}
  • Without permissions (most services, including one that declares only options), the service exports one permission, invoke, on every route ("* /*"); service:<name> and the service:* wildcard grant it.
  • With permissions, the service exports exactly those, and invoke only if it declares it too. A data service (me, log, needs) declares read / write and no invoke: it holds the customer's own data, so only a grant that names its permissions reaches it, never invoke or a wildcard.
  • permissions: a name (lowercase letters, digits and _) → the routes it covers, each "<METHOD> /<path>" (* for any method, * globs in the path). Permissions are never inferred from HTTP methods.
  • options: a JSON Schema (draft 7) object of the customizations a grant may set on the service, like a channel's params.schema, with an optional react-jsonschema-form options_ui_schema; the frontend renders it with the same form component. It covers limits only the service can apply (a message size). Restrictions so a grant can write every value in query-string syntax: property names are lowercase letters, digits and _; each property is a string, integer, number, boolean or an array of strings, serialized as an OpenAPI form-style query parameter with arrays comma-joined (style: form, explode: false); options are optional (no required) and only declared properties are accepted; the schema is self-contained (no $ref), so validating against it never fetches anything. A property's default is what the service does when the policy does not set the option.

2. A customer access request (upstream_access_config.<channel>.customer_access_requests), and the policy entry it becomes. Any channel may declare its own list: a fixed list of items the customer works on, and nothing else (unitysvc/unitysvc#2831). Each approved request becomes an entry (a grant) in the issued key's policy, in the same shape:

"customer_access_requests": [{"object": "service:log", "permissions": ["invoke"], "required": true,
                "justification": "Records each check you run",
                "options": "log_level=warn",
                "limits": {"count_per_day": 1000}}]
  • object names one service (service:<name>, no wildcards), a resource the platform provisions for the enrollment (resource:room, below), or apikey for the leased key's lifespan (below). permissions are plain names (lowercase letters, digits and _; no wildcards), checked against the target's exports with a catalog.
  • required (true or false, on every entry, apikey included) says whether the customer may decline the request. A stored declaration from before this rule reads as optional when a policy is built.
  • options is written in query-string syntax (parse_options; a notation, not how options are delivered): each name once, each value non-empty; with a catalog, each value is converted to its declared type and validated against the target's options schema.
  • A request's options are fixed. Every option it declares is in the policy of an approved request, with the declared value; the customer may repeat it but not change or drop it. An option the request does not declare is the customer's to set at enrollment (a form rendered from the target's options schema). So {"object": "service:ai", ...} lets the customer pick the AI service, and "options": "service=labs/llm-balanced" pins it.
  • Several requests may name one object (two kinds of room) only if each fixes some option to a different value (kind=chat, kind=feed), so an approval says which one it is.
  • limits (validate_limits): count_per_hour / count_per_day / count_per_month / count_total (positive integer) and the same for amount_ (finite positive decimal, as a number or a string). Limits are universal: the platform meters every call, so any service can be limited without exporting anything. The seller's limits only pre-fill the form; the customer sets them at enrollment, and an enrollment's own limits (unitysvc/unitysvc#2687) use the same shape. They are counted per policy entry (the enrollment's calls to that service) from the usage stream (counters keyed by scope, metric and window bucket), and the gateway refuses calls once one is exceeded.
  • A room the platform provisions (unitysvc/unitysvc#2830) is asked for with resource:room: {"object": "resource:room", "permissions": ["create"], "required": true, "justification": "Creates your chat room", "options": "kind=chat&max_members=100"}. Core knows the resource, so no catalog is consulted. The request must fix kind, the room's identity; each kind bounds the permissions (chat: create, a room the customer owns; support: join, read, post, the seller's private room with the customer; announcement: join, read; discussion: join, read, post). max_members (a positive integer) caps only a room the customer creates; max_retention (seconds, 0 or more) caps how long any provisioned room keeps messages, within the platform's own ceiling for the kind, and the room's owner may keep less. A channel requests each kind at most once, and a resource takes no limits.
  • A longer-lived leased key is asked for with one apikey entry: {"object": "apikey", "required": false, "limits": {"api_key_lifespan": 14400}, "justification": "Batch runs take up to four hours"}. It names the key itself, so it takes only limits (exactly api_key_lifespan, whole seconds from 1 to 86400), required and a justification, needs no catalog lookup, and grants nothing. Without it, a key the seller leases for the enrollment lives DEFAULT_API_KEY_LIFESPAN (3600 s). api_key_lifespan is valid only on this entry.

3. How options reach the service. Not through the upstream URL: the gateway never changes a target's base_url or query to carry policy. How an option is applied is up to the service's implementation:

  • A service implemented in the gateway (me, log, ai, notify, email, bill) reads the calling key's policy from the routing context and applies the option itself (me gives values only for the grant's profile paths; log would drop entries below a grant's log_level).
  • A proxied service can only be influenced through its own upstream's interface, so it applies an option only where its configuration maps it into the upstream request in a form that upstream understands, for example a request transformer setting a body field. Until such a mapping exists, options are enforced by gateway-implemented services only.
  • The policy itself never reaches a seller: a seller's upstream sees at most the effect of an option, in its own vocabulary.
  • An option the policy does not set takes the service's default. A plain service:* key sets none.
  • Enrollment parameters are unchanged and keep working alongside: {{ params.X }} still renders into the channel config.

4. Where exports are published. In the public catalog bundle: every service row carries its exports exactly as declared (null when none). GET /v1/catalog/bundle returns the bundle's URL; a service:<name> is the row whose name is <name>.

5. Validating customer_access_requests. validate_channel_access_requests(upstream_access_config, resolve) checks every channel's requests. Limits need no lookup. Every claim's permissions and options, invoke included (a service that declares permissions may not export it), are looked up: resolve(name) returns the target's exports from the catalog ({} when it declares none) or None when no such service is published; core parses it with parse_exports and checks the claimed permissions and the options (names, then values against its options schema) against it, refusing an unpublished target. Without a resolver only the shape is checked.

  • DataValidator(data_dir, schema_dir, exports_resolver=...) runs both checks on every offering_v1 (exports and each channel's customer_access_requests). The seller SDK passes a resolver that downloads the catalog bundle once and reads the target row's exports; offline it passes none and gets the shape checks.
  • Backend ingest calls the same function with a resolver over the published services, so an unknown target, permission or option fails at upload.

6. From approvals to a policy (build_policy). When the customer enrolls, build_policy(requests, approvals, resolve) turns their approvals of the channel's requests into the policy the enrollment keeps and every key leased for it copies (unitysvc/unitysvc#2766):

[{"object": "service:log", "permissions": ["invoke"], "options": "log_level=error", "limits": {"count_per_day": 200}},
 {"object": "apikey", "limits": {"api_key_lifespan": 7200}}]
  • Each approval matches one request (match_request): the request of its object whose fixed options it does not contradict. One that fits none is refused, as is one that fits several (it names the option that tells them apart). A request is approved at most once; every required one must be; optional ones may be left out. Nothing else may be added, apart from the enrollment entry below.
  • An approved request keeps exactly its permissions (in any order) and its fixed options; the approval may leave the options out. Values compare as the target's schema types them (max_size=02000 repeats 2000).
  • The options the request leaves open, and the limits, are the customer's, checked against the target's exports (with resolve) and validate_limits.
  • The apikey lifespan may be shortened, never lengthened beyond the request.
  • The result drops required / justification and empty options / limits, and writes options as one canonical query string (options_text: names sorted, only what must be escaped escaped), so equal policies are equal. match_request finds a stored entry's request again, for a readout or a diff. api_key_lifespan(policy) reads the lifespan back (the default when there is none).

A policy has three kinds of entry (unitysvc/unitysvc#2770):

Entry Means
{"object": "enrollment", "limits": {"amount_per_day": "1.00"}} This enrollment's own limits (total spending, say). The customer adds it without a request; it takes only limits, at least one, at most one entry. It grants nothing.
{"object": "service:ai", ...} A service the app calls for the customer. A seller may name a specific one (service:labs/llm-premium instead of service:ai with options="service=labs/llm-premium"), at the cost of the customer's choice of provider, which a customer may decline.
{"object": "service:<name>", "enrollment": "<CODE>", ...} One of the customer's enrollments of a service that requires enrolling (a chain enrollment), called through /e/<CODE>. The customer picks the enrollment when approving the request for service:<name>, so a request never carries a code; a request with fixed options can't be approved with one, since an enrollment carries no options. A canonical code, no options (the enrollment's params are fixed), permissions and limits as for any service entry. Core checks the code's form; the caller checks the enrollment is the customer's.

Upstream access channels

upstream_access_config maps a channel name to a channel (UpstreamAccessConfig, ChannelConfig; unitysvc/unitysvc#2730). A channel describes one way the platform reaches the upstream, as a whole, and its servers hold the connection, typed by the channel's protocol:

Channel key What it is
access_method http (default), smtp, s3 or mcp; selects the server type
type normally derived by the platform; declare enrollable to state it
description, is_active, sort_order display text; inactive channels are skipped; rank, lowest first
request_formats, applicable_interfaces the formats the channel serves; the interfaces that reach it (display only)
rate_limit_refs the provider rate limits it consumes
customer_access_requests what the enrolling customer is asked to approve (see above)
params the enrollment parameters (below)
request_rules http channels only: rules the gateway answers itself before dispatch (below)
servers the connection: HttpServer, SmtpServer, S3Server or McpServer

The common one-server case writes the connection fields directly on the channel; they are read as servers[0]. When a channel has both, each server inherits the direct fields it does not set. Any string may be a template ({{ params.X }}, ${ secrets.X }, ${ customer_secrets.X ?? default }), so numeric and boolean connection fields also accept strings.

Unknown keys are refused. Retired keys are refused with what to do instead (RETIRED_CHANNEL_KEYS): raw, is_primary, the listing-style parameter trio, grants, rate_limits (nothing reads it; use rate_limit_refs), storage_type (use access_method: s3), response_rules, timeout_ms, transport, and the platform-derived customer_secrets_required / customer_secrets_optional.

Request rules

An http channel may list up to 32 request_rules (unitysvc/unitysvc#2947, unitysvc/apisix-gateways#94). The gateway checks them in order after authentication, channel selection and rate limiting and before anything is dispatched. The first rule whose when holds answers with its static respond, and the request goes no further; when none holds, the request continues unchanged. A rule without when answers every request.

"request_rules": [
  {"name": "github-ping", "when": "header[\"x-github-event\"] == \"ping\"", "respond": {"status": 200}},
  {"name": "own-echo", "when": "body.user_id == \"${ customer_secrets.DISCORD_WEBHOOK_ID ?? }\"", "respond": {"status": 204}},
  {"name": "filter", "when": "body.count >= {{ params.MIN_COUNT }}", "respond": {"status": 204}}
]

when is a Lua expression that the gateway compiles and evaluates against the request. Its environment (the request's body, header, query and method, the approved functions), its template references (${ customer_secrets.NAME ?? }, ${ secrets.NAME }, {{ params.NAME }}, {{ enrollment.code }}, bound as values and never spliced into the source; a secret may only be compared whole with == / ~=) and its limits are the gateway's, documented in apisix-gateways' README. Core does not parse it: it checks only that it is non-empty and at most 1024 bytes. Lua syntax errors and invalid references are reported by the gateway (the channel answers 503) and by a gateway-run test (services run-tests), not by specs validate.

respond takes a status (200–599), an optional body (an object or array is sent as JSON; a string needs content_type), and optional headers (not the gateway-managed or hop-by-hop ones).

Versions: 0.3.7 accepted a two-operator grammar; 0.3.8 refused template references and added customer_secrets; 0.3.9 follows the gateway as merged (template references in when, no customer_secrets). No published service used 0.3.7 or 0.3.8 rules.

An offering that declares no channel (upstream_access_config omitted or {}) is an external app: customers reach it at its own site, not through the gateway (unitysvc/unitysvc#2873). external is not a channel type. An external app:

  • gives its homepage in details.links.homepage (required; the provider's homepage is not substituted);
  • has no user_access_interfaces and no list_price (signup and payment happen at its site);
  • needs no connectivity-test document.

details.links holds an offering's product links, for any service: up to 10 free display labels (at most 40 characters each) mapped to https:// URLs without embedded credentials. When present it must include homepage, the primary product link. The rest of details stays free-form. validate_details_links checks it on every offering_v1.

{
  "details": {
    "links": {
      "homepage": "https://app.example",
      "documentation": "https://app.example/docs"
    }
  }
}

Enrollment parameters on channels

A channel declares the enrollment parameters it reads, so each plan of a service asks for what it needs (unitysvc/unitysvc#2726). A listing declares none: user_parameters_schema, user_parameters_ui_schema and service_options.ops_testing_parameters are refused.

"upstream_access_config": {
  "plus": {
    "base_url": "https://u/{{ params.folder }}",
    "params": {
      "schema": {"type": "object", "properties": {"folder": {"type": "string"}}, "required": ["folder"]},
      "ui_schema": {},
      "testing_defaults": {"folder": "ops"}
    }
  }
}

params.schema is a JSON Schema of the parameters (a property's own default pre-fills the customer's form), params.ui_schema holds form hints, and params.testing_defaults holds the values the platform tests the channel with. Rules:

  • a channel reads only the parameters its params.schema declares, and a channel that reads {{ params.* }} must declare params (validate_channel_parameters, on every offering_v1);
  • params.schema is a valid, self-contained JSON Schema (draft 7, no $ref), and params.testing_defaults validates against it with format checking (an email property needs an email address), so a malformed schema is refused at ingest rather than when a customer enrolls;
  • each required parameter has a value in that channel's params.testing_defaults, which is also what ${ customer_secrets.{{ params.X }} } renders with when secret references are checked;
  • each required parameter is read by its own channel or by document content (validate_required_parameter_usage, which backend ingest also calls).

Metadata

Release files for unitysvc-core 0.3.9

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for unitysvc-core 0.3.9
File Size Uploaded Provenance Actions
unitysvc_core-0.3.9.tar.gz 203.6 kB With provenance Details for unitysvc_core-0.3.9.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for unitysvc-core 0.3.9
File Interpreter ABI Platform Actions
unitysvc_core-0.3.9-py3-none-any.whl Python 3 none any Details for unitysvc_core-0.3.9-py3-none-any.whl

Total release size: 354.2 kB

Release files / unitysvc_core-0.3.9.tar.gz

Download URL unitysvc_core-0.3.9.tar.gz
Size 203.6 kB
Tags Source
SHA-256 checksum
How to use checksums
45929369464935b056954a94bcb5b970d6c6bd6866bcab6685f3a49ced5345d0
BLAKE2b-256 checksum
How to use checksums
a79f4323b190a6e7454277b20151fb95ececd59c5e58c0fd99e7307c9dc5fa02
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 9, 2026.

Transparency log

Release files / unitysvc_core-0.3.9-py3-none-any.whl

Download URL unitysvc_core-0.3.9-py3-none-any.whl
Size 150.7 kB
Tags Python 3
SHA-256 checksum
How to use checksums
60a0f933069a520b066e8a91ef05b6329fa8dd37a56f25a3a835b2613e6099e2
BLAKE2b-256 checksum
How to use checksums
4e037c295bed5815d7ddf80c4191e519eaac0005dddaf4c15c7688cd8d99a82d
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 9, 2026.

Transparency log

Security for unitysvc-core 0.3.9

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for unitysvc_core-0.3.9.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 9, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.3.9 This release

2 release files Full provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page