mirror of
https://github.com/netbirdio/netbird.git
synced 2026-10-05 04:59:06 +02:00
[management,proxy] Agent network: per-account LLM gateway (policy, metering, multi-provider) (#6555)
* [agent-network] Shared proto, OpenAPI schema, and generated types * [agent-network] Management: store, manager, synthesizer, policy engine, provider catalog, HTTP/gRPC API Adds the account-scoped agent-network module: provider/policy/budget CRUD and store, the reverse-proxy service synthesizer, policy selection + limit enforcement, the provider catalog (incl. Vertex AI and AWS Bedrock entries), and the management HTTP + proxy gRPC surfaces. * [management] Fix agent-network proxy-peer fan-out on affected-peer recompute The affected-peers resolver loaded only persisted reverse-proxy services, but agent-network services are synthesized on demand and never persisted. As a result the embedded proxy peer was never folded into the affected set when a client's group changed, so the proxy received no network-map update for a newly authorised client and rejected its handshake until a full resync (restart). loadProxyServices now merges the synthesized agent-network services (injected via a registration hook to avoid an import cycle), so proxy peers learn newly authorised clients immediately. * [proxy] Reverse-proxy middleware framework, chain, and request plumbing The per-target middleware chain (slots, dispatcher, mutation gate, metadata merger), body capture, access-log terminal sink, and the proxy wiring that builds + runs chains for synthesized agent-network services. * [proxy] LLM parsers, pricing, and builtin middlewares (OpenAI, Anthropic, Vertex AI, AWS Bedrock) Request/response parsers and SSE/event-stream metering, the embedded pricing table, and the builtin middleware set: request parser, router, policy limit-check/record, cost meter, guardrail, identity inject, response parser. Includes the path-routed providers — Google Vertex AI (keyfile:: service-account OAuth minting) and AWS Bedrock (bearer auth, invoke/converse/streaming, optional /bedrock prefix) — plus the Models allowlist and unmeterable-publisher deny. * [proxy] IPv6 in-place apply and TCP accept-loop hardening on netstack listeners * [agent-network] End-to-end test suite, module docs, and deployment preset * [agent-network] Fix codespell typos and exclude false positives - labelgen word pool: vermillion -> vermilion, racoon -> raccoon. - codespell ignore list: add flate (Go compress/flate package), recordin (a test-local identifier), and unparseable (a valid alternative spelling used consistently across identifiers + a metadata-value constant). * [management] Set LastSeen on injected proxy peer in realstack test (MySQL strict-mode) The injected embedded proxy peer had a PeerStatus with a zero LastSeen, which serializes to '0000-00-00' and is rejected by MySQL in strict mode (SQLite tolerates it). Set LastSeen to a valid time so SaveAccount succeeds on both engines. * [agent-network] Remove e2e shell-script suite from this branch The end-to-end shell scripts under scripts/e2e/ are maintained in a separate testing suite and are not part of this change set. * [agent-network] Polish module docs: remove internal review scaffolding, fix links, verify diagrams Strip PR-review framing, commit references, absolute paths, and stale internal references from the agent-network module docs; fix broken relative links; verify all diagrams against the current architecture. Remove the internal AI-reviewer prompt file. * [management] Refine session expiration handling to support 3-state encoding for SSO deadlines * [agent-network] Relocate agentnetwork package to internals/modules Move management/server/agentnetwork (and its catalog/, labelgen/, types/ subpackages) to management/internals/modules/agentnetwork, alongside the reverse-proxy module, and rewrite all importers. Pure relocation: package names, the synthesizer + affectedpeers registration hook, and store access (shared store.Store) are unchanged, so no import cycle is introduced (affectedpeers still depends only on the agentnetwork/types leaf). * [agent-network] Co-locate HTTP handlers in the module (RegisterEndpoints) Move the agent-network HTTP handlers from server/http/handlers/agentnetwork into the module at internals/modules/agentnetwork/handlers (package handlers) and rename the entrypoint AddEndpoints -> RegisterEndpoints, matching the reverse-proxy module convention. Wiring in http/handler.go updated accordingly.
This commit is contained in:
Maycon Santos
@@ -0,0 +1,269 @@
|
||||
package llm_response_parser
|
||||
|
||||
import (
|
||||
"bytes"
|
||||
"encoding/json"
|
||||
"errors"
|
||||
"io"
|
||||
"strings"
|
||||
|
||||
"github.com/netbirdio/netbird/proxy/internal/llm"
|
||||
)
|
||||
|
||||
// openAIDoneSentinel is the OpenAI end-of-stream marker. The scanner
|
||||
// stops once this data frame is observed.
|
||||
const openAIDoneSentinel = "[DONE]"
|
||||
|
||||
// accumulateStream walks the SSE byte slice, dispatches per provider,
|
||||
// and returns the running token-usage and concatenated completion text.
|
||||
// Errors from the scanner short-circuit accumulation but never panic
|
||||
// — partial results are returned for truncated bodies.
|
||||
func accumulateStream(provider string, body []byte) (llm.Usage, string) {
|
||||
switch provider {
|
||||
case "openai":
|
||||
return accumulateOpenAIStream(body)
|
||||
case "anthropic":
|
||||
return accumulateAnthropicStream(body)
|
||||
case llm.ProviderNameBedrock:
|
||||
return accumulateBedrockStream(body)
|
||||
default:
|
||||
return llm.Usage{}, ""
|
||||
}
|
||||
}
|
||||
|
||||
// openAIStreamUsage is the usage block shared by both OpenAI streaming
|
||||
// envelopes. Pointer fields tell "absent" from zero; the chat.completions
|
||||
// (prompt_/completion_) and Responses-API (input_/output_) names are both
|
||||
// accepted so a single decode covers either endpoint.
|
||||
type openAIStreamUsage struct {
|
||||
PromptTokens *int64 `json:"prompt_tokens"`
|
||||
CompletionTokens *int64 `json:"completion_tokens"`
|
||||
InputTokens *int64 `json:"input_tokens"`
|
||||
OutputTokens *int64 `json:"output_tokens"`
|
||||
TotalTokens *int64 `json:"total_tokens"`
|
||||
PromptTokensDetails *struct {
|
||||
CachedTokens *int64 `json:"cached_tokens"`
|
||||
} `json:"prompt_tokens_details"`
|
||||
InputTokensDetails *struct {
|
||||
CachedTokens *int64 `json:"cached_tokens"`
|
||||
} `json:"input_tokens_details"`
|
||||
}
|
||||
|
||||
// openAIStreamChunk matches both OpenAI streaming envelopes. The
|
||||
// chat.completions chunk carries text in choices[].delta.content and a
|
||||
// trailing top-level usage block. The Responses API (/v1/responses) emits
|
||||
// typed events instead: completion text rides response.output_text.delta
|
||||
// (top-level "delta" string) and the final usage rides response.completed
|
||||
// under response.usage. Only fields used for accumulation are declared.
|
||||
type openAIStreamChunk struct {
|
||||
Choices []struct {
|
||||
Delta struct {
|
||||
Content string `json:"content"`
|
||||
} `json:"delta"`
|
||||
} `json:"choices"`
|
||||
Usage *openAIStreamUsage `json:"usage"`
|
||||
|
||||
Type string `json:"type"`
|
||||
Delta json.RawMessage `json:"delta"`
|
||||
Response *struct {
|
||||
Usage *openAIStreamUsage `json:"usage"`
|
||||
} `json:"response"`
|
||||
}
|
||||
|
||||
// accumulateOpenAIStream sums per-chunk content deltas and lifts the usage
|
||||
// block off the final frame, handling both the chat.completions and the
|
||||
// Responses-API event shapes. Clients without stream_options.include_usage
|
||||
// (chat.completions) and any provider that omits the final usage simply
|
||||
// leave tokens at zero; the caller chooses what to emit.
|
||||
func accumulateOpenAIStream(body []byte) (llm.Usage, string) {
|
||||
var (
|
||||
usage llm.Usage
|
||||
completion strings.Builder
|
||||
)
|
||||
scanner := llm.NewScanner(bytes.NewReader(body))
|
||||
for {
|
||||
ev, err := scanner.Next()
|
||||
if err != nil {
|
||||
if errors.Is(err, io.EOF) {
|
||||
break
|
||||
}
|
||||
break
|
||||
}
|
||||
if ev.Data == "" || ev.Data == openAIDoneSentinel {
|
||||
if ev.Data == openAIDoneSentinel {
|
||||
break
|
||||
}
|
||||
continue
|
||||
}
|
||||
|
||||
var chunk openAIStreamChunk
|
||||
if err := json.Unmarshal([]byte(ev.Data), &chunk); err != nil {
|
||||
continue
|
||||
}
|
||||
for _, c := range chunk.Choices {
|
||||
completion.WriteString(c.Delta.Content)
|
||||
}
|
||||
if chunk.Type == "response.output_text.delta" {
|
||||
if s, ok := decodeJSONString(chunk.Delta); ok {
|
||||
completion.WriteString(s)
|
||||
}
|
||||
}
|
||||
|
||||
u := chunk.Usage
|
||||
if u == nil && chunk.Response != nil {
|
||||
u = chunk.Response.Usage
|
||||
}
|
||||
if u != nil {
|
||||
usage.InputTokens = pickInt64(u.InputTokens, u.PromptTokens)
|
||||
usage.OutputTokens = pickInt64(u.OutputTokens, u.CompletionTokens)
|
||||
usage.TotalTokens = derefInt64(u.TotalTokens)
|
||||
if u.InputTokensDetails != nil {
|
||||
if v := derefInt64(u.InputTokensDetails.CachedTokens); v > 0 {
|
||||
usage.CachedInputTokens = v
|
||||
}
|
||||
}
|
||||
if usage.CachedInputTokens == 0 && u.PromptTokensDetails != nil {
|
||||
usage.CachedInputTokens = derefInt64(u.PromptTokensDetails.CachedTokens)
|
||||
}
|
||||
if usage.TotalTokens == 0 && (usage.InputTokens > 0 || usage.OutputTokens > 0) {
|
||||
usage.TotalTokens = usage.InputTokens + usage.OutputTokens
|
||||
}
|
||||
}
|
||||
}
|
||||
return usage, completion.String()
|
||||
}
|
||||
|
||||
// decodeJSONString unmarshals a JSON-encoded string value, returning
|
||||
// ok=false when the raw message is empty or not a string.
|
||||
func decodeJSONString(raw json.RawMessage) (string, bool) {
|
||||
if len(raw) == 0 {
|
||||
return "", false
|
||||
}
|
||||
var s string
|
||||
if err := json.Unmarshal(raw, &s); err != nil {
|
||||
return "", false
|
||||
}
|
||||
return s, true
|
||||
}
|
||||
|
||||
// anthropicStreamEvent captures the union of Messages-API stream event
|
||||
// payloads we care about. Each named event on the wire fills only its
|
||||
// shape's fields; unknown keys are ignored.
|
||||
type anthropicStreamEvent struct {
|
||||
Type string `json:"type"`
|
||||
Message *struct {
|
||||
Usage *struct {
|
||||
InputTokens *int64 `json:"input_tokens"`
|
||||
OutputTokens *int64 `json:"output_tokens"`
|
||||
CacheReadInputTokens *int64 `json:"cache_read_input_tokens"`
|
||||
CacheCreationInputTokens *int64 `json:"cache_creation_input_tokens"`
|
||||
} `json:"usage"`
|
||||
} `json:"message"`
|
||||
Delta *struct {
|
||||
Type string `json:"type"`
|
||||
Text string `json:"text"`
|
||||
} `json:"delta"`
|
||||
Usage *struct {
|
||||
InputTokens *int64 `json:"input_tokens"`
|
||||
OutputTokens *int64 `json:"output_tokens"`
|
||||
CacheReadInputTokens *int64 `json:"cache_read_input_tokens"`
|
||||
CacheCreationInputTokens *int64 `json:"cache_creation_input_tokens"`
|
||||
} `json:"usage"`
|
||||
}
|
||||
|
||||
// accumulateAnthropicStream tracks input_tokens from message_start,
|
||||
// output_tokens from message_delta, and concatenates text_delta payloads
|
||||
// from content_block_delta events. Final usage prefers message_delta
|
||||
// values which carry the post-completion totals.
|
||||
func accumulateAnthropicStream(body []byte) (llm.Usage, string) {
|
||||
var (
|
||||
usage llm.Usage
|
||||
completion strings.Builder
|
||||
)
|
||||
scanner := llm.NewScanner(bytes.NewReader(body))
|
||||
for {
|
||||
ev, err := scanner.Next()
|
||||
if err != nil {
|
||||
if errors.Is(err, io.EOF) {
|
||||
break
|
||||
}
|
||||
break
|
||||
}
|
||||
if ev.Data == "" {
|
||||
continue
|
||||
}
|
||||
|
||||
var payload anthropicStreamEvent
|
||||
if err := json.Unmarshal([]byte(ev.Data), &payload); err != nil {
|
||||
continue
|
||||
}
|
||||
|
||||
eventType := ev.Type
|
||||
if eventType == "" {
|
||||
eventType = payload.Type
|
||||
}
|
||||
applyAnthropicStreamEvent(eventType, payload, &usage, &completion)
|
||||
}
|
||||
if usage.InputTokens > 0 || usage.OutputTokens > 0 {
|
||||
usage.TotalTokens = usage.InputTokens + usage.OutputTokens + usage.CachedInputTokens + usage.CacheCreationTokens
|
||||
}
|
||||
return usage, completion.String()
|
||||
}
|
||||
|
||||
// applyAnthropicStreamEvent folds one parsed Anthropic Messages stream event
|
||||
// into the running usage/completion. Shared by the SSE accumulator and the
|
||||
// Bedrock InvokeModel event-stream, whose chunks wrap the same event JSON.
|
||||
func applyAnthropicStreamEvent(eventType string, payload anthropicStreamEvent, usage *llm.Usage, completion *strings.Builder) {
|
||||
switch eventType {
|
||||
case "message_start":
|
||||
if payload.Message != nil && payload.Message.Usage != nil {
|
||||
if v := derefInt64(payload.Message.Usage.InputTokens); v > 0 {
|
||||
usage.InputTokens = v
|
||||
}
|
||||
if v := derefInt64(payload.Message.Usage.OutputTokens); v > 0 {
|
||||
usage.OutputTokens = v
|
||||
}
|
||||
if v := derefInt64(payload.Message.Usage.CacheReadInputTokens); v > 0 {
|
||||
usage.CachedInputTokens = v
|
||||
}
|
||||
if v := derefInt64(payload.Message.Usage.CacheCreationInputTokens); v > 0 {
|
||||
usage.CacheCreationTokens = v
|
||||
}
|
||||
}
|
||||
case "content_block_delta":
|
||||
if payload.Delta != nil && payload.Delta.Type == "text_delta" {
|
||||
completion.WriteString(payload.Delta.Text)
|
||||
}
|
||||
case "message_delta":
|
||||
if payload.Usage != nil {
|
||||
if v := derefInt64(payload.Usage.InputTokens); v > 0 {
|
||||
usage.InputTokens = v
|
||||
}
|
||||
if v := derefInt64(payload.Usage.OutputTokens); v > 0 {
|
||||
usage.OutputTokens = v
|
||||
}
|
||||
if v := derefInt64(payload.Usage.CacheReadInputTokens); v > 0 {
|
||||
usage.CachedInputTokens = v
|
||||
}
|
||||
if v := derefInt64(payload.Usage.CacheCreationInputTokens); v > 0 {
|
||||
usage.CacheCreationTokens = v
|
||||
}
|
||||
}
|
||||
case "message_stop":
|
||||
// No-op; Anthropic does not emit usage here.
|
||||
}
|
||||
}
|
||||
|
||||
func pickInt64(preferred, fallback *int64) int64 {
|
||||
if preferred != nil {
|
||||
return *preferred
|
||||
}
|
||||
return derefInt64(fallback)
|
||||
}
|
||||
|
||||
func derefInt64(v *int64) int64 {
|
||||
if v == nil {
|
||||
return 0
|
||||
}
|
||||
return *v
|
||||
}
|
||||
Reference in New Issue
Block a user