iop/apps/edge/internal/openai/responses_handler.go

434 lines
18 KiB
Go

package openai
import (
"bytes"
"encoding/json"
"errors"
"fmt"
"go.uber.org/zap"
"io"
edgeservice "iop/apps/edge/internal/service"
"net/http"
"strconv"
"strings"
)
func (s *Server) handleResponses(w http.ResponseWriter, r *http.Request) {
if r.Method != http.MethodPost {
writeError(w, http.StatusMethodNotAllowed, "method_not_allowed", "method not allowed")
return
}
defer r.Body.Close()
// The raw body is preserved for the provider tunnel passthrough path, which
// forwards the caller's own Responses payload (model rewritten to the served
// target) without strict normalization.
rawBody, err := io.ReadAll(r.Body)
if err != nil {
writeError(w, http.StatusBadRequest, "invalid_request_error", "failed to read request body")
return
}
// The route decision runs before strict normalization: only the
// routing-relevant envelope fields are decoded leniently so provider
// passthrough can preserve Codex/Responses unknown fields (max_output_tokens,
// tools, store, ...). Strict field validation stays on the normalized path.
env, err := decodeResponsesEnvelope(rawBody)
if err != nil {
writeError(w, http.StatusBadRequest, "invalid_request_error", err.Error())
return
}
dispatch, ok := s.resolveRouteDispatch(env.Model)
if !ok {
writeError(w, http.StatusBadRequest, "invalid_request_error", "model is required")
return
}
if dispatch.ProviderPool {
// Provider-pool path: the one-shot SubmitProviderPool selects a candidate
// and dispatches either tunnel passthrough (OpenAI-compatible) or
// normalized RunEvent (Ollama/CLI/native) based on the candidate's
// executionPath. Raw body is preserved for the tunnel branch; the
// normalized branch performs strict decode + prompt build below.
requestCtx, ok := s.newResponsesRequestContext(w, r, dispatch, rawBody, env)
if !ok {
return
}
s.handleResponsesProviderPool(w, requestCtx)
return
}
// Non-provider routes: tunnel or normalized. Tunnel routes relay the raw
// body verbatim; normalized routes require strict decode + prompt build.
if routeUsesProviderTunnel(dispatch) {
// Provider routes always relay pure passthrough. Caller metadata is
// arbitrary context and never selects the route or response shape
// (SDD S01/S04); there is no caller-facing response mode selector.
requestCtx, ok := s.newResponsesRequestContext(w, r, dispatch, rawBody, env)
if !ok {
return
}
s.tunnelResponsesPassthrough(w, requestCtx)
return
}
// Normalized RunEvent path: strict decode, stream/background rejection,
// prompt build, and SubmitRun.
var req responsesRequest
if err := decodeResponsesRequest(json.NewDecoder(bytes.NewReader(rawBody)), &req); err != nil {
writeError(w, http.StatusBadRequest, "invalid_request_error", err.Error())
return
}
// Keep direct normalized validation ahead of identity resolution, matching
// the existing strict Responses contract. The dispatch-context constructor
// repeats these local checks when it is invoked from provider-pool PrepareRun.
if req.Stream {
writeError(w, http.StatusBadRequest, "invalid_request_error", "streaming is not supported for /v1/responses")
return
}
if req.Background {
writeError(w, http.StatusBadRequest, "invalid_request_error", "background is not supported for /v1/responses")
return
}
if _, err := parseResponsesInput(req.Input); err != nil {
writeError(w, http.StatusBadRequest, "invalid_request_error", err.Error())
return
}
requestCtx, ok := s.newResponsesRequestContext(w, r, dispatch, rawBody, env)
if !ok {
return
}
dc, err := s.newResponsesDispatchContext(requestCtx, req)
if err != nil {
writeError(w, http.StatusBadRequest, "invalid_request_error", err.Error())
return
}
s.logger.Info("openai responses input",
zap.String("model", dc.req.Model),
zap.String("target", dc.route.Target),
zap.String("adapter", dc.route.Adapter),
zap.Bool("strict_output", dc.outputPolicy.Strict),
zap.String("xml_completion_tool", dc.outputPolicy.XMLCompletionTool),
zap.Bool("contract_instruction", dc.outputPolicy.ContractInstruction),
zap.Int("prompt_len", len(dc.prompt)),
)
handle, err := s.service.SubmitRun(r.Context(), dc.submitReq)
if err != nil {
writeError(w, http.StatusBadGateway, "node_dispatch_error", err.Error())
return
}
defer handle.Close()
s.logger.Info("openai responses dispatch",
zap.String("run_id", handle.Dispatch().RunID),
zap.String("node_id", handle.Dispatch().NodeID),
zap.String("provider_id", handle.Dispatch().ProviderID),
zap.String("provider_type", handle.Dispatch().ProviderType),
zap.String("execution_path", handle.Dispatch().ExecutionPath),
zap.String("model_group", handle.Dispatch().ModelGroupKey),
zap.String("adapter", handle.Dispatch().Adapter),
zap.String("target", handle.Dispatch().Target),
zap.Int("estimated_input_tokens", handle.Dispatch().EstimatedInputTokens),
zap.String("context_class", handle.Dispatch().ContextClass),
zap.String("queue_reason", handle.Dispatch().QueueReason),
)
s.completeResponse(w, dc, handle)
}
// newResponsesRequestContext resolves the identity, estimate, and long-context
// classification a /v1/responses request carries into either dispatch path. It
// writes the client error and reports false when the caller's identity or
// workspace is rejected.
func (s *Server) newResponsesRequestContext(w http.ResponseWriter, r *http.Request, dispatch routeDispatch, rawBody []byte, env responsesEnvelope) (*responsesRequestContext, bool) {
runMeta, workspace, err := resolveCallerIdentity(r, dispatch, env.Metadata)
if err != nil {
writeError(w, http.StatusBadRequest, "invalid_request_error", err.Error())
return nil, false
}
// The Responses estimate is taken from the raw body so both the tunnel and
// the normalized branch carry the same observability and long-context
// information (SDD S02/S03).
estimate := estimateInputTokens(string(rawBody), runMeta, nil, nil)
return &responsesRequestContext{
openAIRequestContext: openAIRequestContext{
r: r,
route: dispatch,
rawBody: rawBody,
callerMetadata: runMeta,
workspace: workspace,
estimate: estimate,
contextClass: classifyContext(estimate, s.longContextThreshold()),
endpoint: usageEndpointResponses,
},
envelope: env,
}, true
}
// newResponsesDispatchContext freezes the strict-decoded request and the
// normalized Run request after ingress routing. Raw tunnel data stays on the
// parent responsesRequestContext, and the copied metadata below is the only
// mutable execution metadata for this normalized branch.
func (s *Server) newResponsesDispatchContext(requestCtx *responsesRequestContext, req responsesRequest) (*responsesDispatchContext, error) {
if req.Stream {
return nil, fmt.Errorf("streaming is not supported for /v1/responses")
}
if req.Background {
return nil, fmt.Errorf("background is not supported for /v1/responses")
}
inputStr, err := parseResponsesInput(req.Input)
if err != nil {
return nil, err
}
prompt := buildResponsesPrompt(req.Instructions, inputStr)
outputPolicy := s.resolveOutputPolicy(prompt)
if instruction := strictOutputContractInstruction(outputPolicy); instruction != "" {
prompt = instruction + "\n" + prompt
}
defaultThinkingTokenBudget := 0
if catalogEntry := s.findProviderPoolEntry(req.Model); catalogEntry != nil {
applyModelCatalogGenerationPolicyToResponses(&req, *catalogEntry)
defaultThinkingTokenBudget = catalogEntry.DefaultThinkingTokenBudget
}
runMetadata := cloneMetadata(requestCtx.callerMetadata)
runMetadata["openai_model"] = req.Model
runMetadata["openai_stream"] = fmt.Sprintf("%t", req.Stream)
runMetadata["strict_output"] = fmt.Sprintf("%t", outputPolicy.Strict)
input := map[string]any{"prompt": prompt}
if defaultThinkingTokenBudget > 0 {
input["think"] = true
input["thinking_token_budget"] = defaultThinkingTokenBudget
} else if outputPolicy.Strict {
input["think"] = false
}
if options := req.providerOptions(); len(options) > 0 {
input["options"] = options
}
estimate := estimateInputTokens(prompt, runMetadata, nil, nil)
contextClass := classifyContext(estimate, s.longContextThreshold())
runMetadata["estimated_input_tokens"] = strconv.Itoa(estimate)
runMetadata["context_class"] = contextClass
dc := &responsesDispatchContext{
responsesRequestContext: requestCtx,
req: req,
prompt: prompt,
input: input,
outputPolicy: outputPolicy,
runMetadata: runMetadata,
}
dc.submitReq = edgeservice.SubmitRunRequest{
NodeRef: dc.route.NodeRef,
ModelGroupKey: strings.TrimSpace(dc.req.Model),
Adapter: dc.route.Adapter,
Target: dc.route.Target,
SessionID: dc.route.SessionID,
Workspace: dc.workspace,
Prompt: dc.prompt,
Input: dc.input,
TimeoutSec: dc.route.TimeoutSec,
MaxQueue: dc.route.MaxQueue,
QueueTimeoutMS: dc.route.QueueTimeoutMS,
Metadata: dc.runMetadata,
EstimatedInputTokens: estimate,
ContextClass: contextClass,
ProviderPool: dc.route.ProviderPool,
}
return dc, nil
}
// handleResponsesProviderPool dispatches a provider-pool /v1/responses request
// through the one-shot SubmitProviderPool surface. After a single queue
// admission, the selected candidate's executionPath determines whether the
// response is delivered via tunnel passthrough (OpenAI-compatible providers)
// or the normalized RunEvent path (Ollama/CLI/native). Provider auth is
// validated via providerTunnelAuthHeaders for tunnel-path only (SDD D01).
// Strict-output policy only applies to normalized dispatch, not to raw tunnel
// passthrough (SDD D02).
//
// Selection-first flow: the provider is selected by SubmitProviderPool before
// any path-specific decode, validation, or stream gating. The tunnel branch
// only rewrites the model field and relays the raw body (SDD S04). The
// normalized branch performs strict decode + stream/background rejection +
// prompt build, but only after the candidate is known. This prevents
// tunnel-preferring unknown or streaming requests from bypassing the selected
// provider's executionPath rule (SDD S02/S03).
func (s *Server) handleResponsesProviderPool(w http.ResponseWriter, requestCtx *responsesRequestContext) {
// Build a populated base Run for the normalized branch and a Tunnel for
// the tunnel branch. Both are sent to SubmitProviderPool so the one-shot
// admission selects the provider before any path-specific behavior.
r := requestCtx.r
dispatch := requestCtx.route
rawBody := requestCtx.rawBody
workspace := requestCtx.workspace
estimate := requestCtx.estimate
contextClass := requestCtx.contextClass
env := requestCtx.envelope
runMeta := cloneMetadata(requestCtx.callerMetadata)
runMeta["openai_model"] = env.Model
runMeta["openai_stream"] = strconv.FormatBool(env.Stream)
runMeta["estimated_input_tokens"] = strconv.Itoa(estimate)
runMeta["context_class"] = contextClass
// Base Run carries the model-group key that SubmitProviderPool requires
// for candidate resolution. Metadata, estimate, and context class are
// preserved by PrepareRun on the normalized branch; the tunnel branch
// inherits them from the base so tunnel dispatch keeps observability.
baseRun := edgeservice.SubmitRunRequest{
NodeRef: dispatch.NodeRef,
ModelGroupKey: strings.TrimSpace(env.Model),
SessionID: dispatch.SessionID,
Workspace: workspace,
Metadata: runMeta,
EstimatedInputTokens: estimate,
ContextClass: contextClass,
ProviderPool: true,
}
baseTunnel := edgeservice.SubmitProviderTunnelRequest{
Metadata: runMeta,
EstimatedInputTokens: estimate,
ContextClass: contextClass,
ModelGroupKey: strings.TrimSpace(env.Model),
SessionID: dispatch.SessionID,
Method: http.MethodPost,
Path: "/v1/responses",
Stream: env.Stream,
TimeoutSec: dispatch.TimeoutSec,
MaxQueue: dispatch.MaxQueue,
QueueTimeoutMS: dispatch.QueueTimeoutMS,
ProviderPool: true,
}
poolReq := edgeservice.ProviderPoolDispatchRequest{
Run: baseRun,
Tunnel: baseTunnel,
}
// Pre-dispatch provider auth header injection. Runs inside SubmitProviderPool
// BEFORE buildProviderTunnelRequest and the Node Send step, so the auth
// header lands on the actual wire request. On failure the slot is released
// and no tunnel request is sent (SDD S03).
poolReq.PrepareTunnel = func(tunnelReq edgeservice.SubmitProviderTunnelRequest) (edgeservice.SubmitProviderTunnelRequest, error) {
headers, err := s.providerTunnelAuthHeaders(r)
if err != nil {
return tunnelReq, err
}
tunnelReq.Headers = headers
return tunnelReq, nil
}
// Tunnel branch: rewrite only the model field, preserve all other fields
// (tools, max_output_tokens, custom fields). Stream/background gating is
// skipped: the provider itself enforces those constraints (SDD S04).
poolReq.Tunnel.BuildBody = func(target string) ([]byte, error) {
return rewriteResponsesModel(rawBody, target)
}
// The normalized context is created only after candidate selection. That
// preserves raw passthrough for tunnel candidates while giving the
// normalized completion one immutable request/policy/metadata snapshot.
var preparedDispatch *responsesDispatchContext
// Normalized branch: strict decode, stream/background rejection, prompt
// build, and generation policy. Only runs AFTER the candidate is selected
// by SubmitProviderPool (i.e. only when executionPath is normalized).
// Validation errors are wrapped with errProviderRequestValidation so
// SubmitProviderPool can relay them as HTTP 400.
poolReq.PrepareRun = func(runReq edgeservice.SubmitRunRequest) (edgeservice.SubmitRunRequest, error) {
req := responsesRequest{Model: env.Model}
if err := decodeResponsesRequest(json.NewDecoder(bytes.NewReader(rawBody)), &req); err != nil {
return edgeservice.SubmitRunRequest{}, fmt.Errorf("%w: invalid /v1/responses request format: %s", errProviderRequestValidation, err.Error())
}
dc, err := s.newResponsesDispatchContext(requestCtx, req)
if err != nil {
return edgeservice.SubmitRunRequest{}, fmt.Errorf("%w: %s", errProviderRequestValidation, err.Error())
}
// SubmitProviderPool has already filled the common queue-admission
// fields in runReq. Keep those and overlay the normalized context's
// body, metadata, and execution values.
runReq.Prompt = dc.submitReq.Prompt
runReq.Input = dc.submitReq.Input
runReq.Metadata = dc.submitReq.Metadata
runReq.EstimatedInputTokens = dc.submitReq.EstimatedInputTokens
runReq.ContextClass = dc.submitReq.ContextClass
runReq.TimeoutSec = dc.submitReq.TimeoutSec
runReq.MaxQueue = dc.submitReq.MaxQueue
runReq.QueueTimeoutMS = dc.submitReq.QueueTimeoutMS
preparedDispatch = dc
return runReq, nil
}
result, err := s.service.SubmitProviderPool(r.Context(), poolReq)
if err != nil {
metricLabels := s.usageLabelsFor(r.Context(), strings.TrimSpace(env.Model), usageEndpointResponses, responseModePassthrough)
emitUsageMetrics(metricLabels, usageStatusForError(err), usageObservation{})
// Provider auth failure is a client request error (400), not a
// backend dispatch error. The auth check runs inside
// SubmitProviderPool (PrepareTunnel) before any tunnel request is sent.
// Normalized validation failures (e.g. strict decode, stream) are
// also client errors: map to 400.
if errors.Is(err, errProviderAuthRequired) {
writeError(w, http.StatusBadRequest, "invalid_request_error", "provider auth token is required")
return
}
if isValidationError(err) {
writeError(w, http.StatusBadRequest, "invalid_request_error", err.Error())
return
}
writeError(w, http.StatusBadGateway, "node_dispatch_error", err.Error())
return
}
s.logger.Info("openai responses provider-pool dispatch",
zap.String("run_id", result.DispatchInfo.RunID),
zap.String("node_id", result.DispatchInfo.NodeID),
zap.String("provider_id", result.DispatchInfo.ProviderID),
zap.String("provider_type", result.DispatchInfo.ProviderType),
zap.String("execution_path", result.DispatchInfo.ExecutionPath),
zap.String("model_group", result.DispatchInfo.ModelGroupKey),
zap.String("adapter", result.DispatchInfo.Adapter),
zap.String("target", result.DispatchInfo.Target),
zap.String("path", string(result.Path)),
zap.Int("estimated_input_tokens", result.DispatchInfo.EstimatedInputTokens),
zap.String("context_class", result.DispatchInfo.ContextClass),
zap.String("queue_reason", result.DispatchInfo.QueueReason),
)
switch result.Path {
case edgeservice.ProviderPoolPathTunnel:
// Tunnel path: provider auth was already validated and injected via
// PrepareTunnel before dispatch; on failure SubmitProviderPool returns
// an error and no tunnel handle exists. Provider bytes are relayed as
// pure passthrough; caller metadata never selects a sideband surface.
metricLabels := s.usageLabelsFor(r.Context(), strings.TrimSpace(env.Model), usageEndpointResponses, responseModePassthrough)
s.writeProviderTunnelResponse(w, r, result.Tunnel, env.Stream, env.Model, metricLabels)
case edgeservice.ProviderPoolPathNormalized:
// Normalized path: no auth required, collect from RunEvent stream.
handle := result.Run
if handle == nil {
writeError(w, http.StatusInternalServerError, "run_error", "provider-pool selection returned normalized path but no run result")
return
}
// PrepareRun is only invoked on the normalized path. A normalized
// result without a successful PrepareRun is an internal mismatch and
// must surface as run_error (SDD D02).
if preparedDispatch == nil {
writeError(w, http.StatusInternalServerError, "run_error", "provider-pool normalized result received without a successful PrepareRun")
return
}
// Relay the prepared normalized context so strict-output XML wrapping
// and the exact derived metadata survive the provider-pool path.
s.completeResponse(w, preparedDispatch, handle)
}
}