Files
alknet/docs/architecture/decisions/042-openapi-gateway-pattern.md
glm-5.2 125cb49cc4 docs(http): defer h3/WebTransport (ADR-044); browsers use WebSocket for v1
Working through the WebTransport implementation path surfaced a scope
question distinct from the hedging-as-deferral anti-pattern ADR-038 was
written to correct. Three findings drove the re-evaluation:

1. The browser bidirectional call-protocol path doesn't require
   WebTransport — WebSocket is full-duplex, EventEnvelope fits a WS
   binary message boundary cleanly, and the Dispatcher is stream-
   agnostic (ADR-012). What WebTransport gives over WebSocket (native
   multi-stream multiplexing, the ALPN-as-stream substrate) benefits the
   proxy use case, not the call protocol.
2. WebTransport is a draft standard (-07, not RFC) on an experimental
   Rust dependency stack (wtransport/h3 both self-describe as not
   production-ready). Either choice puts a draft protocol on the
   security surface of the first release.
3. The ALPN-stream-proxy (ADR-040) is speculative — its WASM parser
   consumers (browser SSH/SFTP/git clients) don't exist yet, and the
   downstream crates WebTransport deferral blocks (SSH, git, SFTP)
   expose their ALPNs natively over QUIC regardless.

This is a scope decision (per ADR-009: a decision that 'genuinely
doesn't need to be made yet because the use case isn't concrete'), not
hedging. The reversal trigger is concrete: a real deployment needing
the ALPN-stream-proxy.

ADR-038 is superseded (its anti-pattern correction stands; its specific
'h3 in scope now' decision is reversed). ADR-040 and ADR-043 are
parked, not superseded — their designs revive unchanged when WebTransport
revives, with §2 (bidirectionality) and §3 (no-PeerId overlay) of ADR-043
transferring to WebSocket for v1.

ADR-044 §5 also states the 'browser is not a peer' rationale that
ADR-034 §4 closed without arguing: peer = addressable node in the
call-protocol peer graph (stable PeerId, PeerRef::Specific-reachable,
identity stable across reconnects), not 'any endpoint that exchanges
calls during a live session.' A browser is the second but not the first
(no stable crypto identity of its own, ephemeral, not addressable from
other nodes). ADR-034 §4 and Assumption 2 are amended by reference.

The wtransport-vs-hyperium dependency question is recorded (not
resolved — WebTransport is deferred) in ADR-044 §'Research note' and
webtransport.md so the revival doesn't re-derive it: wtransport probably
isn't the right choice (axum-bridge friction — it owns its own HTTP
serving path); the hyperium stack (h3 + h3-quinn + h3-webtransport) fits
the axum integration better but its server-side WebTransport API needs
verification before commitment.

Reviewed by architecture-review subagent; all critical cross-reference
issues (ADR-034 §5 stale 'in scope' assertion, ADR-036 Context listing
h3 as implemented, webtransport.md Design Decisions table) resolved.
2026-06-30 05:55:55 +00:00

13 KiB

ADR-042: OpenAPI Gateway Pattern for to_openapi

Status

Proposed

Context

The current to_openapi spec (crates/http/http-adapters.md) describes to_openapi as generating a traditional OpenAPI document with one path entry per alknet External operation — POST /fs/readFile, POST /agent/chat, etc., each with parameters, request body, and responses built from the operation's input_schema/output_schema/ error_schemas. This is the "inverse of from_openapi" framing: since from_openapi merges OpenAPI path params / query params / request body into a single flat JSON input schema, to_openapi should split them back out.

The flat→structured problem

The inverse is genuinely messy. The call protocol's input is a flat JSON object (e.g., { path, content, encoding } for fs/writeFile). To generate a traditional OpenAPI path entry (POST /fs/{path} with path param path, body content), to_openapi would need to know which fields are path params, which are query params, and which is the body. That information isn't in the flat schema — it's metadata the call protocol doesn't carry because it doesn't care about HTTP parameter structure. to_openapi would need either:

  1. HTTP-specific metadata on OperationSpec (which fields are path params, etc.) — a leaky abstraction that puts HTTP concerns in the protocol-foundation crate (alknet-call), or
  2. Heuristics (guess that fields named id are path params?) — fragile and wrong, or
  3. Manual annotation per operation — boilerplate that defeats the "pure projection" promise.

All three are messy. The flat→structured split is the hard direction, and it's the one to_openapi has to do.

The per-caller API surface problem

A traditional OpenAPI document is static — it describes the full API surface regardless of who's reading it. Real APIs have per-caller authorization: an admin sees admin operations, a regular user sees a subset. OpenAPI has no standard mechanism for "show me only what I have access to." The Gitea API is a concrete failure case: its OpenAPI spec dumps the full API (including admin operations) to every caller, regardless of privilege. A user reading the spec can't tell which endpoints they can actually call without trial-and-error 403s.

The call protocol already has the per-caller filtering primitive: services/list is AccessControl::check(identity)-filtered — the caller sees only the operations they are authorized to call. A to_openapi that generates a static full-surface doc loses this property. A to_openapi that uses the gateway pattern preserves it.

The pattern that works

The same tool-gateway pattern ADR-041 applies to to_mcp applies here: to_openapi exposes a small fixed set of endpoints that gate access to the full operation registry. The external client (a code generator, a human developer, a fetch-based client) calls search to discover operations, schema to learn an operation's input shape, call to invoke. The input is always a flat JSON body — no path/query/body split to reverse-engineer. JSON Schema for the input/output is already in the OperationSpec — no conversion beyond wrapping it in OpenAPI's schema format.

The OpenAPI gateway has one endpoint the MCP gateway doesn't: subscribe (SSE). OpenAPI/SSE supports streaming; MCP tool calls don't. So the OpenAPI gateway is 5 endpoints; the MCP gateway is 4.

Decision

1. to_openapi exposes a fixed gateway endpoint set, not one path per operation

to_openapi generates an OpenAPI document with a small fixed set of endpoints that gate access to the full operation registry. The external client discovers and invokes operations through the gateway.

The gateway endpoint set (initial, two-way-door extensible):

OpenAPI path Call protocol operation HTTP method Purpose
/search services/list GET List/search available operations (filtered by the caller's AccessControl). Returns names + descriptions.
/schema services/schema GET Get an operation's full OperationSpec (input/output JSON Schemas, error schemas).
/call call.requested (Query/Mutation) POST Invoke an operation by name with a JSON input. Returns the output or a typed error (ADR-023).
/batch multiple call.requested POST Invoke multiple operations in one request (correlated request IDs, OQ-14). Returns an array of results.
/subscribe call.requested (Subscription) GET (SSE) Invoke a streaming operation. Returns text/event-stream — each call.responded is an SSE frame, call.completed closes the stream.

Five endpoints. The client calls /search to find operations, /schema to learn the input shape, /call (or /subscribe for streaming) to invoke. The input is always a flat JSON body ({ operation: "/fs/readFile", input: { ... } }); the output is the operation's result as JSON. No path/query/body split to reverse-engineer.

2. subscribe is the OpenAPI gateway's streaming endpoint (SSE)

The OpenAPI gateway includes subscribe (which the MCP gateway excludes — ADR-041, MCP tool calls are request/response). The subscribe endpoint maps Subscription operations onto SSE: GET /subscribe with Accept: text/event-stream, each call.responded event is an SSE data: frame, call.completed closes the stream, call.aborted closes with an error frame. This is the same SSE projection ADR-036 describes for h2/http/1.1 clients — the gateway's subscribe endpoint is the single SSE entry point instead of per-operation SSE streams.

3. The generated OpenAPI doc is per-caller (AccessControl-filtered)

The /search endpoint's results are filtered by the caller's AccessControl::check(identity) — the client sees only the operations it is authorized to call. The /call and /subscribe endpoints run the same AccessControl check on dispatch. The generated OpenAPI doc describes the gateway endpoints (5 fixed paths); the per-caller operation surface is discovered through /search, not preloaded into the doc.

This is the key advantage over a traditional per-operation-paths OpenAPI doc: the per-caller API surface is the default, not an afterthought. A client reading the gateway OpenAPI doc learns the gateway's shape (5 endpoints, stable); a client calling /search learns what it can call (per-caller, AccessControl-filtered). The Gitea failure mode (dumping admin ops to every caller) is structurally impossible — /search doesn't return operations the caller can't call.

4. The gateway OpenAPI doc is a compatibility contract

Once published, the gateway endpoint set (5 endpoints) and the request/response shapes are a compatibility contract (ADR-017 Consequences). Adding endpoints is additive (non-breaking); removing or renaming is a one-way door. The initial 5-endpoint set is the published contract. The versioning strategy for the generated doc is tracked as OQ-39 (same as the per-operation-paths versioning question — the gateway pattern doesn't eliminate the versioning concern, it simplifies it to 5 stable endpoints instead of a per-operation surface).

5. A traditional per-operation-paths projection is additive, not replacement

A deployment that wants a traditional REST OpenAPI doc (per-operation paths with split parameters) can build it as a separate projection with the HTTP-specific metadata (which fields are path params, etc.). The gateway pattern is the default to_openapi projection; the traditional projection is an additive alternative for deployments that need it. The gateway does not foreclose the traditional projection — it just doesn't require it for the common case.

Consequences

Positive:

  • No flat→structured split. The gateway's input is always a flat JSON body ({ operation, input }); the operation's input/output schemas are already JSON Schemas in the OperationSpec. No reverse- engineering of path/query/body semantics. The messy direction of the from_openapi inverse is sidestepped.
  • Per-caller API surface by default. /search is AccessControl-filtered; the client sees only what it can call. The Gitea failure mode (dumping admin ops to every caller) is structurally impossible. This is a property the traditional per-operation-paths OpenAPI doc cannot provide (OpenAPI has no per-caller filtering concept).
  • Easy to build clients for. Any language's fetch + JSON Schema libraries can call the gateway: POST /call with a JSON body, get a JSON result. No code generator needed for the common case; a code generator produces a CallClient (call/search/schema/batch/ subscribe) instead of typed per-operation methods.
  • 5 stable endpoints instead of a per-operation surface. The versioning concern (OQ-39) is simpler — 5 endpoints that rarely change vs. a per-operation surface that changes on every operation addition/modification.
  • subscribe maps cleanly onto SSE — the same projection ADR-036 describes, just as a single gateway entry point instead of per- operation SSE streams.
  • A deployment that wants the traditional REST surface can build it additively. The gateway doesn't foreclose it.

Negative:

  • The generated OpenAPI doc is not a "nice UI" by default. A Swagger UI rendering shows 5 generic endpoints instead of a REST tree. This is the tradeoff for avoiding the flat→structured split and gaining per- caller filtering. A deployment that wants the nice UI builds the traditional projection (additive, with metadata).
  • A code generator reading the gateway OpenAPI doc produces a CallClient (generic call/search/schema methods), not typed per- operation methods. Typed methods require the traditional projection (with metadata) or a client that reads /search + /schema and generates typed wrappers at build time. The gateway is optimized for the fetch-and-JSON-Schema use case, not the code-generation use case.
  • The gateway doc is less "traditional" — a developer expecting a REST OpenAPI doc sees a small RPC-style surface instead. This is honest (the call protocol is a flat JSON RPC, not a REST API), but it's a departure from OpenAPI conventions.

Assumptions

  1. The gateway endpoint set is stable. Once external clients build against the 5-endpoint gateway, changing the endpoint set (renaming, removing) breaks them. Adding endpoints is additive (non-breaking). The initial 5-endpoint set is the published contract.

  2. AccessControl filtering is the right per-caller mechanism. The client sees the operations it's authorized to call. If an operation's existence is itself sensitive, Visibility::Internal (ADR-015) is the mechanism — Internal ops are excluded from services/list and therefore from /search results. The gateway does not add a separate visibility layer.

  3. The common case is fetch + JSON Schema, not code generation. The gateway is optimized for the developer who calls POST /call with a JSON body and parses the result. The code-generation case (typed per-operation methods) is served by the traditional projection (additive) or a client that generates wrappers from /search + /schema at build time.

  4. subscribe (SSE) is the streaming projection for the gateway. Over h2/http/1.1, subscriptions are SSE. Over WebSocket (the v1 browser bidirectional path, ADR-044), subscriptions project onto the WS connection directly as binary messages — the gateway's /subscribe is the h2/http/1.1 SSE path; the WebSocket path is the native call-protocol session (http-server.md §"WebSocket browser path"). WebTransport (h3, deferred per ADR-044) would project onto WebTransport streams; the deferred design is at webtransport.md.

References

  • ADR-015 — External/Internal visibility (Internal ops excluded from services/list, therefore from /search)
  • ADR-017to_* adapters are projections; published-spec compatibility contract
  • ADR-023 — typed error details mapped to OpenAPI error responses
  • ADR-036 — the SSE projection for subscriptions over h2/http/1.1 (the gateway's /subscribe endpoint uses the same SSE framing)
  • ADR-044 — WebSocket is the v1 browser bidirectional path; h3/WebTransport deferred (the gateway's /subscribe is the h2/http/1.1 SSE path; the WS path is the native call-protocol session). ADR-038 is superseded by ADR-044.
  • ADR-041 — the sibling gateway pattern for to_mcp (4 tools; subscribe excluded because MCP tool calls are request/response)
  • OQ-39 — to_openapi published-spec versioning (simplified by the gateway pattern to 5 stable endpoints)
  • crates/http/http-adapters.md — the spec that implements the gateway