Skip to content

v0.5.0

Latest

Choose a tag to compare

@github-actions github-actions released this 07 Oct 23:25
· 184 commits to main since this release
66e4ec3

0.5.0 (2026-10-06)

forgec build now emits ForgeGraph's contract IR (contract.json), and
forgec import-contract and forgec contract-diff take it back in and compare
it. Projections' latest aggregate is breaking: it is now
latest <field> by <orderField>. References accept "$self", floor and
datetime arithmetic evaluate at runtime, and the console gains Forgejo Git
projects and an Integrations tab backed by ForgeGraph's contract registry.

Added

  • Console Forgejo Git projects: a project may set "provider": "forgejo" and
    commit through FORGEJO_URL and FORGEJO_TOKEN (minted by
    forge forgejo-token) instead of a GitHub token. /healthz reports a
    configured provider's missing credential, and a Forgejo-only instance no
    longer needs GITHUB_TOKEN.

  • Console integrations: with FORGEGRAPH_URL and a read-scope FORGEGRAPH_TOKEN, every
    ForgeGraph app is listed as an integration described by its published contract (contract IR v1,
    converted to OpenAPI 3.1), and its operations can be called through the console. Base URLs,
    credentials (bearer, header or Cloudflare Access service token, INTEGRATION_* secrets only) and
    write access are set per app in INTEGRATIONS_JSON; reads are the default, writes need an opt-in
    and a per-call confirmation and are audited. The playground's new Integrations tab browses
    operations by group, builds requests from their schemas, confirms writes, keeps drafts per
    operation, and can hand an operation to the Graph tab's OpenAPI importer.

  • forgec build writes contract.json, ForgeGraph's contract IR v1, next to
    openapi.json and derived from it: one operation per @crud route, @http
    function and workflow route, ids <serviceId>.<group>.<endpoint>
    (--service-id sets the app slug; the default is the unscoped package name),
    authentication from the security requirement, and each operation's SLO as
    policy.slo. Fingerprints are hashed over the bytes JSON.stringify would
    produce, so ForgeGraph's validateContract accepts the file as written; the
    conformance suite runs that validator, vendored, over a curated set of contracts
    and CI sweeps every package in the repository with scripts/validate-contracts.sh.
    fg contract publish --file generated/contract.json publishes it.

  • forgec import-contract <contract.json> --package <name> --out <dir> turns a
    contract IR document into a Forge package through OpenAPI 3.1 and the OpenAPI
    importer, reporting RPC procedures, unbindable methods, anonymous or public
    operations, SLA/SLO policy, middleware, alternate media and non-Problem error
    bodies instead of approximating them. A contract whose fingerprints do not
    match is refused.

  • forgec contract-diff <a> <b> [--ignore <kind>] compares two contracts per
    route (request and response schemas after $ref inlining, statuses, policy)
    and exits 1 on any difference not declared expected. See docs/contract-ir.md
    for the round trip and which differences it is expected to show.

  • @forgegraph/foundation/configuration (experimental): typed parameterization
    against a pinned specification revision, layered defaults and overrides,
    deterministic resolution, and resolutions identified by a digest over their
    values. Configuration in @forgegraph/runtime resolves a chain and records it.

    Configuration is not arbitrary JSON: every value assigns a parameter the pinned
    revision declared, so an unknown key cannot be set and a key retyped in a later
    revision cannot be silently inherited. Layers carry an ordinal and a child's is
    strictly greater than its parent's, so precedence is a total order and resolution
    gives the same answer whichever end of the chain you start from. Every resolved
    value records which configuration supplied it, so a surprising value is
    attributable without re-deriving the chain by hand.

    Secret parameters carry a reference and never a literal, may not carry a default,
    and resolve to that reference. This package cannot dereference a binding, and a
    test scans the compiled contract for any field that could hold secret material.

    Seven criteria have local evidence; scalar type validation remains tracked in #94.

  • @forgegraph/foundation/reachability (experimental): contact points and service
    endpoints as durable facts, with purpose labels, preference ranking, half-open
    validity, verification evidence and explicit supersession. Reachability in
    @forgegraph/runtime declares, verifies, supersedes and resolves them.

    Contact points and endpoints are separate resources rather than one polymorphic
    locator, because a data classification is a property of a field. A single value
    column has to be classified once: as data.contact, which sweeps webhook URLs into
    subject-rights erasure that does not apply to them, or as structural, which
    under-classifies an email address. LocatorKind.personal decides which profile a
    kind belongs to and a rule on each profile enforces it.

    No transport mechanics are foundation semantics here: a locator says where something
    can be reached, never whether a message was sent, how many attempts it took, or what
    a provider answered. A test asserts that structurally against the compiled contract.
    Notifications now consumes typed contact points and endpoints and checks their
    validity at notification time. All eight criteria have local evidence.

Added

  • @forgegraph/foundation/consent (experimental): consent grants, withdrawal,
    supersession and evidence as durable facts, with purpose, processing activity and
    data scope as three independent axes. Consent in @forgegraph/runtime grants,
    evidences, withdraws, supersedes and answers effectiveAt.

    Valid time and knowledge time are recorded separately, which is the point of the
    package. A withdrawal recorded on 1 March but backdated to 1 February gives three
    different correct answers: asked on 20 February about 15 February, consent held,
    because the withdrawal was not yet known; asked on 5 March about the same moment,
    it did not; asked on 5 March about 15 January, it held, because a backdated
    withdrawal reaches back only as far as it says. A decision made last Tuesday has to
    stay explainable with what was known last Tuesday.

    Consent is not authorization: these facts are one input to a decision, and a null
    result means no permission was recorded rather than a denial. Notification
    preference, agreement and entitlement are different facts with different lifecycles,
    and a test asserts structurally that none of them has appeared in this contract.

    This is a bounded valid/knowledge interpretation built from ordinary fields, not an
    L0 temporal facet; ConceptIR #75 is what would make those axes language-level.
    All eight criteria on #93 have executable local evidence.

Changed

  • Breaking: a projection's latest aggregate now names its ordering field:
    latest <field> by <orderField>. latest previously returned the greatest value
    of the field, so it was max; it now returns the value from the contribution with
    the greatest ordering field (a required integer, decimal, money, date or datetime
    field), with the record id breaking ties. Write order is never inferred. Bare
    latest x is E-PROJ-006 with the fix-it latest x by <field>. Bundles from
    older compilers keep their previous behaviour. Migrate with, for example,
    latest remainingFraction by observedAt. (#198, #207)

Added

  • A reference to a record's own resource accepts "$self" on create and update,
    resolved to the id of the record being written, so a required self-reference
    (root : Thread @immutable) can be created. Admin import orders a
    self-referencing resource's records so in-snapshot targets come first. (#200, #206)
  • floor and datetime arithmetic in resource expressions evaluate at runtime
    (hour := floor(at / 3600) * 3600), and the compiler accepts exactly the
    expression functions the runtime implements (E-EXPR-004); a parity test keeps
    the two lists equal. Previously any call compiled and then failed every create.
    (#196, #205)

Fixed

  • An optional json field's contract schema stays {} (which already admits
    null) instead of becoming {"type": ["string", "null"]}, which rejected every
    JSON value but strings.
  • D1Storage.list omits WHERE when there are no predicates; a resource with
    neither @tenant nor @softDelete listed without filters failed with a SQLite
    syntax error. (#197, #203)
  • OpenAPI contracts include null in the enum of a nullable enum field, so
    generated validators accept the null the API returns. (#199, #204)
  • /healthz can now fail. Both hosts return 503 with the names of the absent
    settings when the instance's own configuration implies something it does not
    have — an R2 bucket, REGISTRY_TOKEN_SECRET, INSTANCE_AUTHORITY, the Access
    team domain and audience tag. It previously returned a constant ok, so a
    Worker deployed with no secrets reported healthy on every probe while
    answering 503 to every authenticated request, which is exactly what happened
    to forge.gmac.io. .forgegraph.yaml points a deployment health check here,
    so a probe that could not fail could not detect the only outage this
    deployment has had. Only names are reported, never values, and only absence is
    checked — a present but invalid signing key is a different failure and
    claiming to have checked it here would be the same kind of lie.
  • /v2 refuses outright when REGISTRY_TOKEN_SECRET is unset, instead of
    challenging the client with a WWW-Authenticate realm naming a /v2/token
    endpoint that does not exist. docker login could not exit that loop.
  • src/node.ts passes a credential store to createApi and applies
    migrations/0003_registry_credentials.sql. Registry credentials were
    Workers-only by accident, so the panel answered 503 on every Docker
    deployment — and on the Playwright suite, which runs this host rather than the
    Worker, which is why no test saw it.
  • README: corrected the release status. It claimed all five required profiles
    were certified against this build; RELEASE_MANIFEST.json reports
    allRequiredProfilesCertified: false, with two profiles' evidence bound to a
    superseded build and three carrying none. The manifest's auto-demotion rule
    worked; the prose had not followed it down. It also still described 0.3.0 as
    the current release.

Added

  • packages/console/test/worker.test.ts — the first test of the Workers
    entrypoint, which had none despite being the only place /v2, R2, D1, the
    credential store and the token endpoint are wired together. D1 is real
    node:sqlite running the real migrations, which no test had executed. It
    covers the missing-secret and missing-binding cases, and carries a credential
    from HTTP Basic through token exchange to an authorized request — a seam every
    component of which was tested and none of which was tested together.
  • packages/console/e2e/credentials.spec.ts — issues a credential through the
    real UI against the real server. Verified to fail when the wiring above is
    removed.

Install the compiler:

brew install gmackorg/tap/forgec   # macOS and Linux
cargo install forgegraph-cli       # from source

Runtime packages are on npm under @forgegraph/* with build provenance.
Certified profiles and their evidence: RELEASE_MANIFEST.json.