sendnrw/netbird
2
0
Fork 0
mirror of https://github.com/netbirdio/netbird.git synced 2026-09-16 19:59:07 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/feat/kimi-3-agent-networks' into feature/kimi-3-agent-networks-dns-warmup

Browse Source 
# Conflicts:
#	e2e/agentnetwork/chat_test.go
#	e2e/agentnetwork/guardrail_test.go
This commit is contained in:
Claude
2026-07-22 04:51:16 +00:00
parent 11b7709af2 ea4500e6b2
commit baf60e35b2
7 changed files with 126 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
e2e/agentnetwork/chat_test.go
+42 -6
View File
@@ -39,6 +39,21 @@ func availableProviders() []providerCase {
if k := os.Getenv("ANTHROPIC_TOKEN"); k != "" {
ps = append(ps, providerCase{name: "anthropic", catalogID: "anthropic_api", upstream: "https://api.anthropic.com", apiKey: k, model: "claude-haiku-4-5", kind: harness.WireMessages})
}
if k := os.Getenv("KIMI_TOKEN"); k != "" {
// Kimi (Moonshot AI) serves two body shapes from the same key: OpenAI
// Chat Completions on the bare host (/v1/...) and the Anthropic
// Messages API under the /anthropic path prefix (the endpoint
// Moonshot's Claude Code guide uses). The platform serves this
// account exactly ONE model — kimi-k3 (kimi-k2-thinking and even
// kimi-latest return resource_not_found_error on both surfaces) — so
// two concurrent provider records would claim the same model and
// route ambiguously. Run the Anthropic shape, the flagship Claude
// Code path; the OpenAI wire shape is covered live by the other
// chat-shaped matrix providers, and Kimi-over-chat passed with
// kimi-k3 before the single-model constraint surfaced (run #73 on
// the kimi feature branch).
ps = append(ps, providerCase{name: "kimi", catalogID: "kimi_api", upstream: "https://api.moonshot.ai/anthropic", apiKey: k, model: "kimi-k3", kind: harness.WireMessages})
}
if k, u := os.Getenv("VERCEL_TOKEN"), os.Getenv("VERCEL_URL"); k != "" && u != "" {
ps = append(ps, providerCase{name: "vercel", catalogID: "vercel_ai_gateway", upstream: u, apiKey: k, model: "openai/gpt-4o-mini", kind: harness.WireChat})
}
@@ -84,25 +99,46 @@ func availableProviders() []providerCase {
}
}
// Bedrock: path-routed, bearer auth. Model is a cross-region inference
// profile id (distinct string from the first-party Anthropic case).
// Bedrock: path-routed, bearer auth. Model is the FULL cross-region
// inference-profile id exactly as AWS issues it — region-family prefix
// plus the date/version suffix. A bare or wrong-region id makes Bedrock
// reject the request with "The provided model identifier is invalid"
// before any inference runs. The proxy normalizes this id to the catalog
// key (anthropic.claude-haiku-4-5) for routing/pricing/allowlists.
// Defaults pair eu-central-1 with the eu.* profile; AWS_REGION overrides
// the region and the prefix follows its family.
if k := os.Getenv("AWS_BEARER_TOKEN_BEDROCK"); k != "" {
region := os.Getenv("AWS_REGION")
if region == "" {
region = "us-east-1"
region = "eu-central-1"
}
// A valid Bedrock inference-profile id (region prefix + date + version),
// overridable per account. `global.` profiles can be invoked from any
// region; set AWS_BEDROCK_MODEL to match the enabled profile for the token.
// overridable per account via AWS_BEDROCK_MODEL (e.g. a `global.`
// profile, invokable from any region). The default derives the
// region-family prefix from the configured region.
model := os.Getenv("AWS_BEDROCK_MODEL")
if model == "" {
model = "global.anthropic.claude-haiku-4-5-20251001-v1:0"
model = bedrockProfilePrefix(region) + ".anthropic.claude-haiku-4-5-20251001-v1:0"
}
ps = append(ps, providerCase{name: "bedrock", catalogID: "bedrock_api", upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k, model: model, kind: harness.WireBedrock})
}
return ps
}
// bedrockProfilePrefix maps an AWS region to its cross-region inference
// profile prefix: us-east-1 -> us, eu-central-1 -> eu, ap-southeast-1 -> apac,
// us-gov-west-1 -> us-gov.
func bedrockProfilePrefix(region string) string {
switch {
case strings.HasPrefix(region, "us-gov-"):
return "us-gov"
case strings.HasPrefix(region, "ap-"):
return "apac"
default:
return strings.SplitN(region, "-", 2)[0]
}
}
// providerRequest builds a create request for a matrix provider: enabled, with
// a uniquely-priced model for body-routed providers and none for the
// path-routed Vertex (whose model lives in the request path).
e2e/agentnetwork/guardrail_test.go
+7 -3
View File
@@ -17,8 +17,10 @@ import (
)
// bedrockRegionPrefixes and bedrockVersionSuffix mirror the proxy's Bedrock
// model normalization (region/inference-profile prefix + version suffix) so the
// provider is registered under the same catalog key the router matches against.
// model normalization (llm.NormalizeBedrockModel, not importable from here —
// proxy/internal): region/inference-profile prefix + version suffix are
// stripped so the provider is registered under the same catalog key the
// router and guardrail allowlist compare against.
var (
bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
@@ -52,7 +54,9 @@ func catalogModel(pc providerCase) string {
func disallowedModel(pc providerCase) string {
switch pc.kind {
case harness.WireBedrock:
return "us.anthropic.claude-opus-4-8"
// Same profile prefix as the allowed model so only the model name
// differs; the guardrail must deny it before it reaches AWS.
return strings.SplitN(pc.model, ".", 2)[0] + ".anthropic.claude-opus-4-8"
case harness.WireVertex:
return "claude-opus-4-8@20250101"
default: