insika 0.0.1 → 0.1.0

data/lib/insika/server/app.rb

@@ -0,0 +1,850 @@
+# frozen_string_literal: true
+
+require "json"
+require "rack"
+require "async"
+require_relative "sse_body"
+require_relative "admin_auth" # Bearer checker shared by the gateway edge (fail-closed)
+require_relative "a2a/app" # A2A edge adapter (pulls protocol/errors/message/projection/card)
+require_relative "responses" # OpenAI Responses adapter (/v1/responses) — drop-in for the OpenClaw gateway
+
+module Insika
+  module Server
+    # Rack app. Transports ONLY
+    # translate requests into Commands — the server holds no business
+    # logic. It parses JSON, builds `Command.build(...)`, dispatches on the
+    # CommandBus and projects the Event Stream to SSE. Reads are NOT Commands: they are
+    # direct reads from the stores.
+    #
+    # AUDITABLE constitutional rule: `server/` does not import the Executor,
+    # store WRITE methods, or RubyLLM. Requires: json, rack, async and the
+    # core types (Command/Event/errors) already loaded by the composition root.
+    class App
+      SSE_HEADERS = {
+        "content-type" => "text/event-stream",
+        "cache-control" => "no-cache",
+        "connection" => "keep-alive"
+      }.freeze
+
+      # Terminal events of a turn (close the task subscription in the
+      # transport). The overflow :error self-closes its own subscription at the
+      # EventStream (enqueues CLOSED), so it needs no entry here to end `each`.
+      TERMINAL_EVENTS = %i[task_completed task_failed task_cancelled].freeze
+      private_constant :TERMINAL_EVENTS
+
+      # RFC-0016 A5: the `/v1` contract, versioned by date. A caller PINS behaviour
+      # with `Insika-Version: YYYY-MM-DD` so a future breaking change does not move
+      # silently underneath it; absent header = today's (only) version. Only one
+      # entry exists so far — the day a second one is added, the routes that
+      # changed branch on this value instead of being served whichever behaviour
+      # happened to be current.
+      KNOWN_VERSIONS = ["2026-08-08"].freeze
+      private_constant :KNOWN_VERSIONS
+
+      # The operator control UI now lives in the Studio (§12 G5); server/ is a
+      # pure transport surface (/v1, /a2a). The constitutional rule holds: server/
+      # only READS stores and never imports the Executor, store writes, or RubyLLM.
+      def initialize(command_bus:, event_stream:, session_store:, task_store:,
+                     config:, pending_action_store: nil, a2a: nil, provisioner: nil,
+                     workflow_registry: nil, onboarding: nil, profiles: nil,
+                     channels: nil)
+        @command_bus = command_bus
+        @event_stream = event_stream
+        @session_store = session_store
+        @task_store = task_store
+        @config = config
+        @pending_action_store = pending_action_store # read for GET /v1/tasks/:id
+        @a2a = a2a # A2A edge. nil = server does not expose A2A (parity).
+        @provisioner = provisioner # PackImporter. nil = provisioning not exposed.
+        # Item 22 / §4.4: READ-ONLY registry, injected only where workflows are
+        # exposed (the minimal wiring). nil = no /v1/workflows routes (parity — the
+        # deployment does not expose workflows). Reading a catalog is a READ, like a
+        # store read: the constitutional rule (no Executor/store-writes/RubyLLM) holds.
+        @workflow_registry = workflow_registry
+        # Item 20 / §5.6: LLM-first onboarding surface (start.md + models.json +
+        # docs). PUBLIC (no auth — the whole point of the "read <base>/start.md" trick
+        # is that the developer's coding agent can fetch it), and READ-ONLY, so the
+        # constitutional rule holds. nil = routes not exposed (parity — the production
+        # deployment opts in). Reading files/masked stores is a READ, like a store read.
+        @onboarding = onboarding
+        # RFC-0014 §3.2: READ-ONLY ProfileSource, so `GET /v1/agents/:id` can answer
+        # what an agent has — the eval is a client and cannot read a store. Same
+        # constitutional footing as the workflow registry: reading a catalog is a
+        # READ. nil = the route 404s (parity).
+        @profiles = profiles
+        # RFC-0011 §4.4: ONE generic route family for every channel, opt-in by
+        # injecting the registry (nil ⇒ the routes do not exist, parity with @a2a).
+        # The channel does the translating; this class keeps doing only transport.
+        @channels = channels
+        @heartbeat = config.fetch(:heartbeat, 15)
+        @sync_timeout = config.fetch(:sync_timeout, 10) # synchronous control
+      end
+
+      # Explicit routing, NO framework: ~10 routes in a `case`. A single
+      # `rescue` centralizes the error->status mapping. Only SYNCHRONOUS
+      # errors (before the fiber) become HTTP status; a task failure travels as
+      # an event on the stream and lands in GET /v1/tasks/:id.
+      def call(env)
+        req = Rack::Request.new(env)
+        route(req)
+      rescue JSON::ParserError => e
+        error_response(400, e) # malformed JSON, before any dispatch
+      rescue Insika::ValidationError => e
+        error_response(422, e)
+      rescue Insika::NotFoundError => e
+        error_response(404, e)
+      rescue Async::TimeoutError => e
+        error_response(504, e) # synchronous control request exceeded the ceiling
+      rescue StandardError => e
+        error_response(500, e)
+      end
+
+      private
+
+      def route(req)
+        # UTF-8, not the ASCII-8BIT Rack hands us. A path segment becomes a STORE KEY
+        # (`/v1/agents/:id`, `/v1/sessions/:id`), and the sqlite3 driver binds a
+        # BINARY string as a BLOB — which never matches a TEXT column. So every such
+        # read answered 404 on a durable deployment while passing every spec, because
+        # the in-memory store is a Ruby Hash and a binary string is `eql?` to its
+        # UTF-8 twin. Found by calling `GET /v1/agents/:id` against a real database.
+        segments = req.path_info.split("/").reject(&:empty?).map { |s| Coercion.utf8(s) }
+        if segments.first == "v1"
+          version_error = version_gate(req)
+          return version_error if version_error
+        end
+
+        gate = public_route?(req.request_method, segments) ? nil : gateway_gate(req)
+        return gate if gate
+
+        case [req.request_method, segments]
+        in ["GET", ["up"]]
+          health # readiness/liveness (Railway/k8s) — no auth, no store access
+        in ["GET", ["start.md"]] if @onboarding
+          markdown_response(200, @onboarding.start_md(base_url: public_base(req)))
+        in ["GET", ["models.json"]] if @onboarding
+          json_response(200, @onboarding.models_json(base_url: public_base(req)))
+        in ["GET", ["docs"]] if @onboarding
+          json_response(200, { docs: @onboarding.docs_index(base_url: public_base(req)) })
+        in ["GET", ["docs", file]] if @onboarding && file.end_with?(".md")
+          handle_doc(file)
+        in ["POST", ["v1", "commands", type]]
+          handle_command(req, type)
+        in ["POST", ["v1", "sessions"]]
+          handle_create_session(req)
+        in ["POST", ["v1", "messages"]]
+          handle_send_message(req)
+        in ["GET", ["v1", "workflows"]] if @workflow_registry
+          handle_list_workflows
+        in ["POST", ["v1", "workflows", name]] if @workflow_registry
+          handle_trigger_workflow(req, name)
+        in ["POST", ["v1", "responses"]]
+          handle_responses(req)
+        in ["POST", ["v1", "tools", "manifest"]]
+          handle_import_tools(req)
+        in ["POST", ["v1", "mcp", name, "import"]]
+          handle_import_mcp_tools(req, name)
+        in ["POST", ["v1", "agents"]] if @provisioner
+          handle_provision(req)
+        in ["DELETE", ["v1", "agents", id]] if @provisioner
+          handle_deprovision(req, id)
+        in ["GET", ["v1", "agents", id]] if @profiles
+          handle_read_agent(id)
+        in ["GET", ["v1", "sessions", id]]
+          handle_read_session(id)
+        in ["GET", ["v1", "tasks", id]]
+          handle_read_task(id)
+        in ["GET", ["v1", "events"]]
+          handle_events(req)
+        in ["POST", ["channels", id, "events"]] if @channels
+          handle_channel_event(req, id)
+        in ["POST", ["channels", id, "sessions"]] if @channels
+          handle_channel_session(req, id)
+        in ["POST", ["channels", id, "messages"]] if @channels
+          handle_channel_message(req, id)
+        in ["GET", ["channels", id, "asset", file]] if @channels
+          handle_channel_asset(req, id, file)
+        in ["OPTIONS", ["channels", id, *]] if @channels
+          handle_channel_preflight(req, id)
+        in ["POST", ["a2a"]] if @a2a
+          handle_a2a(req)
+        in ["GET", [".well-known", "agent-card.json"]] if @a2a
+          json_response(200, @a2a.agent_card)
+        else
+          not_found # wrong method/route (or A2A not exposed -> @a2a nil)
+        end
+      end
+
+      # The ONLY routes that answer without the gateway Bearer. Everything else is gated
+      # in `route`, before the dispatch — an ALLOWLIST, because the previous shape (each
+      # handler calling `gateway_gate` itself) is a rule you have to remember: the generic
+      # `POST /v1/commands/:type` never called it, so every authoring Command
+      # (`write_agent_file`, `upsert_llm_provider`, `delete_agent`…) was reachable by
+      # anyone who knew the URL, as were the session/task/event reads. A route added
+      # tomorrow is closed by default; making it public is now a deliberate edit here.
+      #
+      # `/up` is the health probe (no store access). The onboarding surface is opt-in
+      # (INSIKA_ONBOARDING) and exists to be read by a coding agent before it has any
+      # credential — turning it on is the operator choosing to publish it.
+      PUBLIC_ROUTES = [
+        ["GET", ["up"]],
+        ["GET", ["start.md"]],
+        ["GET", ["models.json"]],
+        ["GET", ["docs"]],
+        ["GET", [".well-known", "agent-card.json"]] # A2A discovery: the card is the ad
+      ].freeze
+
+      def public_route?(method, segments)
+        return true if PUBLIC_ROUTES.include?([method, segments])
+        return true if channel_route?(method, segments)
+
+        method == "GET" && segments.length == 2 && segments.first == "docs"
+      end
+
+      # A channel route skips the GATEWAY bearer because the channel authenticates
+      # it ITSELF — with the platform's own scheme (a relay's shared secret, a Slack
+      # HMAC signature, the widget's origin allowlist plus its mandatory rate limit),
+      # which is the only credential the caller has. Requiring the gateway token here
+      # instead would mean handing every platform — and every anonymous browser — a
+      # second secret it has no way to send.
+      #
+      # This is NOT an ungated route family: every handler below calls `channel_gate`
+      # before parsing anything, and a channel that implements no `authenticate`, or
+      # whose credential is unconfigured, answers `:disabled` rather than open.
+      # ENUMERATED rather than prefix-matched, so a route added to this family
+      # tomorrow is gated by default and publishing it is a deliberate edit here.
+      def channel_route?(method, segments)
+        case [method, segments]
+        in ["POST", ["channels", _, "events" | "sessions" | "messages"]] then true
+        in ["GET", ["channels", _, "asset", _]] then true
+        in ["OPTIONS", ["channels", _, *]] then true # CORS preflight carries no credential, by spec
+        else false
+        end
+      end
+
+      # Bearer-gate error (503 disabled / 401 unauthorized), shared by the
+      # gateway surfaces. JSON body, fail-closed.
+      def auth_error(status, message, extra_headers = {})
+        [status,
+         { "content-type" => "application/json" }.merge(extra_headers),
+         [JSON.generate(error: { class: "Insika::Error", message: message })]]
+      end
+
+      # POST /v1/commands/:type — generic: every new Command is born with a
+      # transport. The control vs turn distinction is BY THE SHAPE of the result (the
+      # transport knows no semantics).
+      def handle_command(req, type)
+        command = Insika::Command.build(type.to_sym, parse_body(req), transport: :http)
+        command_response(dispatch_with_timeout(command))
+      end
+
+      # POST /v1/sessions — sugar for create_session; 201 {session}.
+      def handle_create_session(req)
+        body = parse_body(req)
+        command = Insika::Command.build(:create_session, { vars: body[:vars] || {} },
+                                         transport: :http)
+        session = dispatch_with_timeout(command)
+        json_response(201, { session: session.to_h })
+      end
+
+      # POST /v1/messages — sugar for send_message; ?stream missing/"true" -> SSE,
+      # "false" -> 200 JSON aggregated at the terminal event.
+      def handle_send_message(req)
+        stream = req.GET["stream"] != "false"
+        # RFC-0015 §5.5: only the aggregated-JSON form has room for the `merged`/`steered`
+        # verdict, so only it may join a message to another turn. Once the stream is open
+        # there is no way to tell the caller it does not own the reply.
+        message_flow(parse_body(req), stream: stream, transport: stream ? :http : :"http:json")
+      end
+
+      # GET /docs/:name.md — one public doc as raw markdown (item 20 / §5.6). The
+      # slug is a KEY of the onboarding allowlist, so no filesystem traversal is
+      # possible; an unknown slug -> 404. `file` still carries the ".md" suffix.
+      def handle_doc(file)
+        markdown = @onboarding.doc(file.sub(/\.md\z/, ""))
+        return not_found if markdown.nil?
+
+        markdown_response(200, markdown)
+      end
+
+      # Public base url for the interpolated onboarding links. Prefers an explicit
+      # config[:public_url] (behind a proxy/TLS terminator the request scheme is the
+      # internal http), else the request's own base_url.
+      def public_base(req)
+        Insika::Coercion.presence(@config[:public_url]) || req.base_url
+      end
+
+      # GET /v1/workflows — discovery (item 22 / §4.4). Direct read of the
+      # registry catalog (name + description + the I/O schema contract). Not a
+      # Command; opt-in via the injected registry.
+      def handle_list_workflows
+        json_response(200, { workflows: @workflow_registry.catalog })
+      end
+
+      # POST /v1/workflows/:name — triggers a workflow RUN. The name comes from the
+      # ROUTE; agent/input/session_id from the body. Two shapes:
+      #   · default        -> 202 { run_id, task_id } immediately (async at-most-once
+      #     run; observe via GET /v1/tasks/:run_id or GET /v1/events?task_id=:run_id).
+      #   · ?stream=true    -> SSE of the run's events (incl. :workflow_started /
+      #     :workflow_completed), closing at the terminal event.
+      # A bad input (input_schema) is a synchronous 422 with no run (WorkflowSchemaError
+      # -> ValidationError in #call); an unknown workflow/agent -> 404/422.
+      def handle_trigger_workflow(req, name)
+        body = parse_body(req)
+        payload = { workflow: name, agent: body[:agent],
+                    input: body[:input], session_id: body[:session_id] }.compact
+        workflow_flow(payload, stream: req.GET["stream"] == "true")
+      end
+
+      # POST /v1/responses — OpenAI Responses adapter (drop-in for the OpenClaw
+      # gateway). Bearer via `config[:gateway_token]` (fail-closed). Translates
+      # the request -> :send_message and the turn's
+      # Event Stream -> OpenAI Responses SSE. Always streams (the consumer asks for SSE).
+      def handle_responses(req)
+        gate = gateway_gate(req)
+        return gate if gate
+
+        parsed = Responses.parse_request(parse_body(req), req) # ValidationError -> 422
+        ensure_session(parsed[:user])
+        payload = { agent: parsed[:agent], session_id: parsed[:user], message: parsed[:message] }
+        payload[:origin] = parsed[:origin] if parsed[:origin] # declared, else absent
+        message_flow(payload, stream: true, serialize: Responses.method(:frame_for))
+      end
+
+      # POST /v1/agents — provisions (upserts) an agent from a standardized
+      # PACK (Phase 6/D4/F7). Same Bearer as /v1/responses (gateway_token,
+      # fail-closed). The consumer (GatewayClient/ProvisionStore) sends the pack as
+      # JSON; the PackImporter emits the authoring Commands. -> 200 { summary }.
+      # Raw body (string keys): the pack's file/skill names are data keys,
+      # not symbols.
+      def handle_provision(req)
+        gate = gateway_gate(req)
+        return gate if gate
+
+        pack = Insika::Pack.from_h(parse_raw_body(req))
+        json_response(200, @provisioner.import(pack)) # Validation/NotFound -> 422/404 in #call
+      end
+
+      # POST /v1/tools/manifest — BATCH ingestion of data-tools via manifest
+      # (Phase 7, Step B). Same Bearer as provisioning (gateway_token, fail-
+      # closed): it's an authoring/provisioning surface and resolves the
+      # deployment's secrets. RAW body (string keys): the JSON Schema property names and
+      # the headers are DATA, not symbols. Dispatches :import_tools -> 200 { per-tool
+      # report }. Structural manifest error -> 422 via the #call rescue; per-tool
+      # failure stays isolated in `errors[]` (R4). Dynamic base_url: the egress guard
+      # + host_allowlist block destinations outside the allowlist at CALL time (R5).
+      def handle_import_tools(req)
+        gate = gateway_gate(req)
+        return gate if gate
+
+        command = Insika::Command.build(:import_tools, parse_raw_body(req), transport: :http)
+        json_response(200, dispatch_with_timeout(command))
+      end
+
+      # POST /v1/mcp/:name/import — LIVE MCP ingestion (Phase 7, Step E). Same
+      # Bearer as provisioning (gateway_token, fail-closed): it's an authoring
+      # surface. Discovers the tools of the MCP instance `:name` (via a client injectable
+      # at the composition root) and ingests them as data-tools (reuses :import_tools:
+      # upsert + hot reload). Dispatches :import_mcp_tools -> 200 { per-tool report
+      # + instance }. Missing instance -> 404; disabled/no-url -> 422; per-tool
+      # failure isolated in `errors[]` (R4). The name comes from the ROUTE (data), not the body.
+      def handle_import_mcp_tools(req, name)
+        gate = gateway_gate(req)
+        return gate if gate
+
+        command = Insika::Command.build(:import_mcp_tools, { name: name }, transport: :http)
+        json_response(200, dispatch_with_timeout(command))
+      end
+
+      # DELETE /v1/agents/:id — removes the agent (delete_agent). NotFoundError
+      # (missing) -> 404 via the #call rescue.
+      def handle_deprovision(req, id)
+        gate = gateway_gate(req)
+        return gate if gate
+
+        json_response(200, @provisioner.delete(id))
+      end
+
+      # GET /v1/agents/:id — what this deployment HAS for that agent, so an eval
+      # (a client: it never reads a store) can tell "this case cannot run here" from
+      # "this case failed" (RFC-0014 §3.2). Deliberately NOT the profile: the prompt,
+      # the model and the guardrail config are none of the caller's business. Just the
+      # two facts a case declares `requires` against.
+      #
+      # `tools` is the DECLARED allowlist, and `null` means the agent has an open one
+      # (every registered tool) — the client reads that as "cannot rule anything out"
+      # and runs the case rather than skipping it.
+      def handle_read_agent(id)
+        profile = @profiles.fetch(id)
+        raise Insika::NotFoundError, "agent not found: #{id}" if profile.nil?
+
+        allow = profile.tools_allow
+        deny = Array(profile.tools_deny).map(&:to_s)
+        json_response(200, {
+                        id: profile.id,
+                        tools: allow.nil? ? nil : (Array(allow).map(&:to_s) - deny),
+                        capabilities: Array(profile.capabilities_declared).map(&:to_s)
+                      })
+      end
+
+      # `/v1` only — `/a2a` is versioned by its own JSON-RPC spec and a channel's
+      # shape is the platform's, so neither reads this header. Runs BEFORE the
+      # gateway gate: which version the caller asked for is a contract question,
+      # answerable regardless of whether the request is authorized. Absent header
+      # -> nil (current behaviour); an unknown value -> 400, not a silent fallback.
+      def version_gate(req)
+        version = req.get_header("HTTP_INSIKA_VERSION")
+        return nil if Coercion.blank?(version) || KNOWN_VERSIONS.include?(version)
+
+        error_response(400, Insika::ValidationError.new("unknown Insika-Version: #{version.inspect}"))
+      end
+
+      # Gateway Bearer (fail-closed). -> error response (503/401) OR nil when
+      # ok (the handler proceeds).
+      def gateway_gate(req)
+        case Insika::Server::AdminAuth.check(@config[:gateway_token], req.get_header("HTTP_AUTHORIZATION"))
+        when :disabled then auth_error(503, "gateway disabled")
+        when :unauthorized then auth_error(401, "unauthorized", "www-authenticate" => "Bearer")
+        end
+      end
+
+      # POST /channels/:id/events — the Shape B inbound webhook (RFC-0011 §4.4).
+      # ACK FAST and never the reply: the platform (or the relay consumer) is holding
+      # a connection open with a retry timer on it, so this dispatches the turn and
+      # answers with its id. The answer itself leaves later, out of band, through the
+      # channel's own `deliver` (§6.5).
+      #
+      # The channel does ALL the translating — auth, envelope, session correlation —
+      # and this handler stays what `server/` is allowed to be: a route that turns a
+      # request into a Command. Four answers, and each one is a different fact:
+      #   202 {task_id}            a turn is running; its reply will be delivered
+      #   200 {task_id, duplicate} we already ran this event id (§6.4)
+      #   200 {task_id, merged}    it joined a turn at the door (RFC-0015 §5.5)
+      #   200 {task_id, steered}   it was appended to a turn already running
+      # A consumer that treats the last three as 202 delivers the same answer twice.
+      def handle_channel_event(req, id)
+        channel = @channels.find(id)
+        return not_found if channel.nil?
+
+        gate = channel_gate(channel, req)
+        return gate if gate
+
+        parsed = channel.parse(req, body: parse_raw_body(req)) # ValidationError -> 422
+        session_id = channel.session_id_for(parsed[:external_id])
+        ensure_session(session_id, vars: session_vars(id, parsed))
+
+        payload = { agent: parsed[:agent], session_id: session_id,
+                    message: parsed[:message], event_id: parsed[:event_id] }.compact
+        channel_ack(payload, transport: :"channel:#{id}")
+      end
+
+      # POST /channels/:id/sessions — mint a conversation for a PUBLIC Shape A
+      # channel (RFC-0011 §4.3). The engine issues the id and the client never
+      # proposes one: an endpoint that created a session from a caller-supplied id
+      # would let anyone read someone else's conversation by guessing.
+      #
+      # Only a channel that mints answers here — the relay's id is the consumer's own
+      # key, so `/sessions` does not exist for it (404, the same parity every other
+      # optional surface has).
+      def handle_channel_session(req, id)
+        channel = @channels.find(id)
+        return not_found if channel.nil? || !channel.respond_to?(:mint_session_id)
+
+        gate = channel_gate(channel, req)
+        return cors(channel, req, gate) if gate
+
+        session_id = channel.mint_session_id
+        ensure_session(session_id, vars: { "channel" => id.to_s })
+        cors(channel, req, json_response(201, { session_id: session_id }))
+      end
+
+      # POST /channels/:id/messages — the Shape A turn: the reply comes back on THIS
+      # connection as SSE, so there is no outbox and nothing to deliver later. It is
+      # `handle_responses` with the hardcoded `Responses` module swapped for the
+      # looked-up channel, which is the whole point of naming the seam.
+      #
+      # The session must already exist AND belong to this channel. Both halves
+      # matter: create-on-write would reopen the enumeration hole `/sessions` closes,
+      # and skipping the ownership check would let a widget visitor stream a
+      # relay customer's conversation by pasting its id.
+      def handle_channel_message(req, id)
+        channel = @channels.find(id)
+        return not_found if channel.nil? || !channel.respond_to?(:frame_for)
+
+        gate = channel_gate(channel, req)
+        return cors(channel, req, gate) if gate
+
+        parsed = channel.parse(req, body: parse_raw_body(req))
+        return cors(channel, req, error_response(404, unknown_session)) unless channel_session?(id, parsed[:session_id])
+
+        payload = { agent: parsed[:agent], session_id: parsed[:session_id], message: parsed[:message] }
+        cors(channel, req, message_flow(payload, stream: true, transport: :"channel:#{id}",
+                                                 serialize: channel.method(:frame_for)))
+      rescue Insika::ValidationError => e
+        # Answered here rather than through #call's rescue so the CORS headers ride
+        # along: without them the browser cannot read the 422 and the visitor sees a
+        # generic network failure instead of what was wrong.
+        cors(channel, req, error_response(422, e))
+      end
+
+      # GET /channels/:id/asset/:f — the channel's static file (the widget's JS).
+      # The name is a KEY of the channel's own closed map, never a path, so there is
+      # no traversal to find. Public and unauthenticated by nature: it is a
+      # `<script src>` on someone else's page.
+      # The cache policy is short and the ETag does the rest: the URL carries no
+      # version (the install snippet an adopter pasted has none), so a long max-age
+      # would strand every browser on the widget it already has, while a revalidation
+      # that answers 304 costs one empty round trip and ships an upgrade in minutes.
+      def handle_channel_asset(req, id, file)
+        channel = @channels.find(id)
+        return not_found if channel.nil? || !channel.respond_to?(:asset)
+
+        asset = channel.asset(file)
+        return not_found if asset.nil?
+
+        headers = { "content-type" => asset[:content_type],
+                    "cache-control" => asset[:cache_control] || "no-cache",
+                    "etag" => asset[:etag] }.compact
+        return [304, headers, []] if asset[:etag] && req.get_header("HTTP_IF_NONE_MATCH") == asset[:etag]
+
+        [200, headers, [asset[:body]]]
+      end
+
+      # OPTIONS /channels/:id/* — the CORS preflight. Answered BEFORE the channel's
+      # own check on purpose: a preflight carries no credentials (the browser strips
+      # them, by spec), so gating it would only mean the real request never happens.
+      # It grants nothing — an origin off the allowlist gets no headers back and the
+      # browser refuses the response itself.
+      def handle_channel_preflight(req, id)
+        channel = @channels.find(id)
+        return not_found if channel.nil?
+
+        cors(channel, req, [204, {}, []])
+      end
+
+      # Does this session exist AND belong to this channel? `vars["channel"]` is
+      # written when the session is minted (§4.3).
+      def channel_session?(id, session_id)
+        session = session_id && @session_store.find(session_id)
+        return false if session.nil?
+
+        vars = session.vars || {}
+        (vars["channel"] || vars[:channel]).to_s == id.to_s
+      end
+
+      def unknown_session = Insika::NotFoundError.new("session not found")
+
+      # Merges the channel's CORS headers into a response it is about to return. A
+      # channel with no opinion (the relay: its consumer is a server, not a browser)
+      # changes nothing.
+      def cors(channel, req, response)
+        return response unless channel.respond_to?(:cors_headers)
+
+        headers = channel.cors_headers(req.get_header("HTTP_ORIGIN"))
+        return response if headers.nil? || headers.empty?
+
+        status, existing, body = response
+        [status, existing.merge(headers), body]
+      end
+
+      # The channel's OWN credential check. A channel returns a verdict, not a status
+      # code — HTTP is this file's vocabulary, not lib/'s — and a channel that never
+      # implements one is refused rather than defaulted open: an unauthenticated
+      # public inbound route with an LLM behind it is a money faucet.
+      def channel_gate(channel, req)
+        verdict = channel.respond_to?(:authenticate) ? channel.authenticate(req) : :disabled
+        case verdict
+        when :disabled then auth_error(503, "channel disabled")
+        when :unauthorized then auth_error(401, "unauthorized", "www-authenticate" => "Bearer")
+        end
+      end
+
+      # Dispatch + ack, with NO subscription: nothing about this request waits for the
+      # turn. That is the difference between Shape B and every other surface here.
+      def channel_ack(payload, transport:)
+        result = @command_bus.dispatch(Insika::Command.build(:send_message, payload, transport: transport))
+        verdict = %i[duplicate merged steered].find { |k| result[k] }
+        return json_response(200, { task_id: result[:task_id], verdict => true }) if verdict
+
+        json_response(202, { task_id: result[:task_id] })
+      end
+
+      # RFC-0011 §4.3: `channel` + `external_id` on the session are how a later turn
+      # (and the outbox) know where a reply goes. The consumer's own `vars` ride along
+      # on first contact, but never over those two — a caller must not be able to
+      # rewrite its own conversation's address.
+      def session_vars(channel_id, parsed)
+        (parsed[:vars] || {}).merge("channel" => channel_id.to_s,
+                                    "external_id" => parsed[:external_id].to_s)
+      end
+
+      # Session correlated by an explicit id (`user`=chat.id on /v1/responses, the
+      # namespaced `<channel>:<external_id>` for a channel): creates if new, continues
+      # if it exists (multi-turn). Via Command (server/ does not write to a store).
+      # Benign race (two near-simultaneous turns creating) -> ArgumentError from the
+      # store, treated as "already exists".
+      def ensure_session(id, vars: { channel: "responses" })
+        return if @session_store.find(id)
+
+        @command_bus.dispatch(
+          Insika::Command.build(:create_session, { id: id, vars: vars }, transport: :http)
+        )
+      rescue ArgumentError
+        nil
+      end
+
+      # GET /v1/sessions/:id — direct read (not a Command).
+      def handle_read_session(id)
+        session = @session_store.find(id)
+        raise Insika::NotFoundError, "session not found: #{id}" if session.nil?
+
+        json_response(200, { session: session.to_h })
+      end
+
+      # GET /v1/tasks/:id — direct read. This is where the consumer observes
+      # PolicyDenied/post-202 failures: the terminal state lives in the Task
+      # Store; nothing is lost if the client disconnected.
+      def handle_read_task(id)
+        task = @task_store.find(id)
+        raise Insika::NotFoundError, "task not found: #{id}" if task.nil?
+
+        body = { task: task_to_h(task) }
+        # pending approvals: this is where the consumer/operator sees
+        # what needs approval after an :approval_requested.
+        if @pending_action_store
+          body[:pending_actions] = @pending_action_store.open_for(id).map(&:to_h)
+        end
+        json_response(200, body)
+      end
+
+      # GET /v1/events?task_id=&session_id= — here the filters ARE known.
+      # CONTINUOUS stream (post-crash reconnection route): does not close on a
+      # terminal event — ends on client disconnect or cap.
+      def handle_events(req)
+        subscription = @event_stream.subscribe(task_id: req.GET["task_id"],
+                                                session_id: req.GET["session_id"])
+        sse_response(subscription)
+      end
+
+      # POST /a2a — JSON-RPC 2.0: HTTP 200 ALWAYS (the error travels in the envelope,
+      # not in the status). Malformed JSON -> -32700 (A2A envelope, not the generic
+      # HTTP error of #call). The A2A::App never leaks an exception. Parse with STRING
+      # keys (the A2A wire is generic JSON — does NOT reuse `parse_body`, which
+      # symbolizes for Command payloads).
+      def handle_a2a(req)
+        raw = req.body&.read
+        body =
+          begin
+            raw.nil? || raw.empty? ? {} : JSON.parse(raw)
+          rescue StandardError
+            return json_response(200, A2A::Protocol.error(nil, A2A::Errors::PARSE_ERROR, "parse error"))
+          end
+        json_response(200, @a2a.rpc(body))
+      end
+
+      # --- Turn flow (SSE or aggregated) ------------------------------------
+
+      # Subscribe BEFORE dispatching: under Async the task fiber may
+      # run eagerly and emit :task_started before dispatch returns. The
+      # task_id only exists AFTER dispatch -> subscribe WITHOUT a filter and filter in the
+      # transport (TaskFilter). A SYNCHRONOUS handler error (Validation/NotFound)
+      # happens here, BEFORE the SSE opens -> closes the subscription and propagates to the
+      # #call rescue (becomes an HTTP status).
+      def message_flow(payload, stream:, serialize: nil, transport: :http)
+        command = Insika::Command.build(:send_message, payload, transport: transport)
+        subscription = @event_stream.subscribe
+        result =
+          begin
+            @command_bus.dispatch(command)
+          rescue StandardError
+            subscription.close
+            raise
+          end
+
+        # RFC-0015 §5.5: the message joined another turn — one still waiting at the
+        # door (`merged`) or one already running (`steered`). Either way this call
+        # owns no reply; the one holding `task_id` does. Say exactly that and open no
+        # stream: a caller that delivered this response's (empty) output would
+        # duplicate the answer.
+        if result[:merged] || result[:steered]
+          subscription.close
+          verdict = result[:merged] ? :merged : :steered
+          return json_response(200, { task_id: result[:task_id], verdict => true })
+        end
+
+        task_id = result[:task_id]
+        # Bind the subscription to the task_id now that it exists: the cap now
+        # counts only events for THIS task and the overflow :error goes out with the
+        # right task_id. The already-enqueued events (eager fiber) belong to this task —
+        # none is lost.
+        subscription.bind(task_id: task_id)
+        filtered = TaskFilter.new(subscription, task_id)
+        stream ? sse_response(filtered, serialize: serialize) : aggregate_response(filtered, task_id)
+      end
+
+      # Workflow trigger flow (item 22). Async by default (the honest workflow
+      # contract: fire the run, return the runId); ?stream=true streams the run's
+      # events like a turn. Same subscribe-before-dispatch discipline as
+      # message_flow so no eager event is lost when streaming. A synchronous handler
+      # error (bad input / unknown workflow) closes the subscription and propagates
+      # to #call (HTTP status).
+      def workflow_flow(payload, stream:)
+        command = Insika::Command.build(:trigger_workflow, payload, transport: :http)
+
+        unless stream
+          result = dispatch_with_timeout(command)
+          return json_response(202, { run_id: result[:run_id] || result[:task_id], task_id: result[:task_id] })
+        end
+
+        subscription = @event_stream.subscribe
+        result =
+          begin
+            @command_bus.dispatch(command)
+          rescue StandardError
+            subscription.close
+            raise
+          end
+        task_id = result[:task_id]
+        subscription.bind(task_id: task_id)
+        sse_response(TaskFilter.new(subscription, task_id))
+      end
+
+      # stream=false: aggregates by iterating the filtered subscription in the
+      # request's own fiber. Accumulates :content deltas; responds at the
+      # terminal. The `error:` shape mirrors the :task_failed data (smallest
+      # coherent extension — the state is also in GET /v1/tasks/:id). Non-happy
+      # terminals (:task_cancelled, overflow :error) also become `error:` —
+      # a cancelled/truncated turn is NEVER reported as a 200 success.
+      def aggregate_response(subscription, task_id)
+        content = +""
+        events = []
+        error = nil
+
+        subscription.each do |event|
+          events << event.to_h
+          case event.type
+          when :content        then content << event.data[:delta].to_s
+          when :task_failed    then error = { class: event.data[:error], message: event.data[:message] }
+          when :task_cancelled then error = { class: "Insika::CancelledError", message: "task cancelled" }
+          when :error          then error ||= { class: nil, message: event.data[:message] }
+          end
+        end
+
+        if error
+          json_response(200, { task_id: task_id, events: events, error: error })
+        else
+          json_response(200, { content: content, task_id: task_id, events: events })
+        end
+      end
+
+      def sse_response(subscription, serialize: nil)
+        [200, SSE_HEADERS.dup, SSEBody.new(subscription: subscription, heartbeat: @heartbeat, serialize: serialize)]
+      end
+
+      # --- Dispatch and serialization --------------------------------------
+
+      # Control Commands may exceed 10s -> 504. For
+      # turn Commands the dispatch returns immediately (the turn lives in the fiber) —
+      # the timeout is harmless. NEVER Timeout.timeout from the stdlib. With no current
+      # reactor (pure control test), dispatches directly.
+      def dispatch_with_timeout(command)
+        task = Async::Task.current?
+        return @command_bus.dispatch(command) if task.nil?
+
+        task.with_timeout(@sync_timeout) { @command_bus.dispatch(command) }
+      end
+
+      # Turn -> {task_id:} -> 202. Any other shape (control:
+      # Session/Task, which are Data) -> 200 with serialized to_h.
+      def command_response(result)
+        if turn_result?(result)
+          json_response(202, { task_id: result[:task_id] })
+        else
+          json_response(200, result.to_h)
+        end
+      end
+
+      # Turn = Hash with task_id PRESENT and non-nil. A control that
+      # returned a Hash without a useful task_id is not mistaken for a turn.
+      def turn_result?(result)
+        result.is_a?(Hash) && !result[:task_id].nil?
+      end
+
+      # Empty body or no content-type -> {} (transport validates only
+      # well-formed JSON; the payload belongs to the handler). Does NOT use req.params (it would
+      # consume the body as a form) — reads the raw body.
+      def parse_body(req)
+        raw = req.body&.read
+        return {} if raw.nil? || raw.empty?
+
+        JSON.parse(raw, symbolize_names: true)
+      end
+
+      # Like parse_body, but keeps STRING keys: for payloads with arbitrary
+      # DATA keys (a pack's file/skill names), which must not become
+      # symbols. Malformed JSON -> JSON::ParserError (#call maps it to 400).
+      def parse_raw_body(req)
+        raw = req.body&.read
+        return {} if raw.nil? || raw.empty?
+
+        JSON.parse(raw)
+      end
+
+      # Task#to_h is shallow (Data#to_h doesn't recurse): `executions` stays as an Array of
+      # Execution (Data), which JSON.generate would serialize as an opaque string
+      # (`"#<data ...>"`) — unreadable for the consumer observing failures via
+      # GET /v1/tasks/:id. Recurses the Executions serialization.
+      def task_to_h(task)
+        task.to_h.merge(executions: task.executions.map(&:to_h))
+      end
+
+      def json_response(status, body)
+        [status, { "content-type" => "application/json" }, [JSON.generate(body)]]
+      end
+
+      # Raw markdown (start.md / a public doc). charset is explicit so a coding agent
+      # fetching over HTTP decodes accents correctly.
+      def markdown_response(status, text)
+        [status, { "content-type" => "text/markdown; charset=utf-8" }, [text]]
+      end
+
+      def error_response(status, error)
+        json_response(status, { error: { class: error.class.name, message: error.message } })
+      end
+
+      def not_found
+        [404, { "content-type" => "text/plain" }, ["not found"]]
+      end
+
+      # Liveness/readiness. Fixed 200: if the process accepts the connection and recovery
+      # has already run (Boot only returns the app afterward — doc 07 §4), it's ready. Does NOT
+      # touch a store (health cannot fail on IO nor require auth).
+      def health = json_response(200, { status: "ok" })
+
+      # Thin Subscription decorator: discards
+      # events from OTHER tasks and CLOSES the subscription after forwarding the task's
+      # terminal event. Solves the subscribe-before-task_id gap without touching
+      # the Subscription's signature.
+      class TaskFilter
+        def initialize(subscription, task_id)
+          @subscription = subscription
+          @task_id = task_id
+        end
+
+        def each
+          @subscription.each do |event|
+            next unless (event.meta || {})[:task_id] == @task_id
+
+            yield event
+            break if TERMINAL_EVENTS.include?(event.type)
+          end
+        ensure
+          @subscription.close
+        end
+
+        def close = @subscription.close
+      end
+      private_constant :TaskFilter
+    end
+  end
+end