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

# Conflicts:
#	e2e/agentnetwork/chat_test.go
#	e2e/agentnetwork/guardrail_test.go

.github/workflows/agent-network-e2e.yml

@@ -51,6 +51,9 @@ jobs:
           # token (and URL, for gateways) is unset, so partial coverage is fine.
           OPENAI_TOKEN: ${{ secrets.E2E_OPENAI_TOKEN }}
           ANTHROPIC_TOKEN: ${{ secrets.E2E_ANTHROPIC_TOKEN }}
+          # Moonshot AI platform key (platform.kimi.ai); drives both Kimi wire
+          # shapes (OpenAI /v1 and Anthropic /anthropic) through kimi_api.
+          KIMI_TOKEN: ${{ secrets.E2E_KIMI_TOKEN }}
           VERCEL_URL: ${{ secrets.E2E_VERCEL_URL }}
           VERCEL_TOKEN: ${{ secrets.E2E_VERCEL_TOKEN }}
           OPENROUTER_URL: ${{ secrets.E2E_OPENROUTER_URL }}

e2e/agentnetwork/chat_test.go

@@ -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

@@ -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:

e2e/harness/client.go

@@ -56,6 +56,11 @@ func StartClient(ctx context.Context, c *Combined, setupKey string) (*Client, er
 			// Match the proxy: the combined relay is WebSocket-only, so the
 			// client must use WS transport to keep a stable relay link to it.
 			"NB_RELAY_TRANSPORT": "ws",
+			// Lazy connections defer the peer dial until traffic flows, which
+			// leaves `netbird status` at "0/N Connected" and starves
+			// WaitProxyPeer. Force eager connections so the proxy peer shows
+			// Connected as soon as the tunnel is up.
+			"NB_LAZY_CONN": "false",
 		},
 		HostConfigModifier: func(hc *container.HostConfig) {
 			hc.CapAdd = append(hc.CapAdd, "NET_ADMIN", "SYS_ADMIN", "SYS_RESOURCE")

e2e/harness/proxy.go

@@ -77,6 +77,10 @@ func StartProxy(ctx context.Context, c *Combined, proxyToken string) (*Proxy, er
 			// Management is plain HTTP in-cluster, so allow the proxy token to
 			// ride a non-TLS gRPC connection.
 			"NB_PROXY_ALLOW_INSECURE": "true",
+			// Keep the embedded overlay peer eager (see the client env note):
+			// lazy connections leave peers unconnected until traffic, which
+			// starves the harness's WaitProxyPeer.
+			"NB_LAZY_CONN": "false",
 			// The combined server multiplexes the relay over WebSocket on :8080
 			// (no QUIC listener). The proxy's embedded relay client defaults to
 			// QUIC, which fails here and flaps the relay link, churning the

management/internals/modules/agentnetwork/catalog/catalog.go

@@ -420,6 +420,47 @@ var providers = []Provider{
 			{ID: "mistral-embed", Label: "Mistral Embed", InputPer1k: 0.0001, OutputPer1k: 0, ContextWindow: 8192},
 		},
 	},
+	{
+		ID:                 "kimi_api",
+		Kind:               KindProvider,
+		Name:               "Kimi (Moonshot AI) API",
+		Description:        "Kimi K3 / K2 models via the Moonshot AI platform",
+		DefaultHost:        "api.moonshot.ai",
+		AuthHeaderName:     "Authorization",
+		AuthHeaderTemplate: "Bearer ${API_KEY}",
+		DefaultContentType: "application/json",
+		BrandColor:         "#1A1A2E",
+		// ParserID empty on purpose: Moonshot serves two body shapes on
+		// the same host and key, and the proxy's URL sniffer dispatches
+		// both (same pattern as Bifrost). /v1/chat/completions matches
+		// OpenAIParser; the Anthropic-compatible endpoint the official
+		// Claude Code guide uses (/anthropic/v1/messages) contains
+		// "/v1/messages" and matches AnthropicParser. Pinning "openai"
+		// here would misparse the Claude Code path — the primary way
+		// teams consume Kimi for coding today. Both endpoints accept the
+		// same Moonshot key via Authorization: Bearer (Claude Code's
+		// ANTHROPIC_AUTH_TOKEN rides that header too).
+		//
+		// api.moonshot.ai is the international platform; mainland-China
+		// accounts live on api.moonshot.cn with separate billing —
+		// operators there override the host on the provider record. The
+		// kimi.com subscription coding endpoint (api.kimi.com/coding,
+		// model id "k3") is account-bound seat licensing rather than a
+		// meterable platform key, so it's deliberately not the default.
+		ParserID: "",
+		// Pricing per Moonshot's platform rates at K3 launch (July 2026):
+		// $3/$15 per MTok with $0.30 cached input, flat across the 1M-token
+		// window. kimi-k3 is the ONLY model the platform serves newer
+		// accounts — K2-era ids (kimi-k2-thinking) and even the kimi-latest
+		// alias return resource_not_found_error, verified live 2026-07-21 —
+		// so it's the only catalog entry. Grandfathered accounts with K2
+		// access can still type those ids on the provider's model rows.
+		// The consumer app's "K3 Swarm Max" mode is not an API SKU, so it
+		// doesn't appear here.
+		Models: []Model{
+			{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
+		},
+	},
 	{
 		ID:                 "litellm_proxy",
 		Kind:               KindGateway,

proxy/internal/llm/pricing/defaults_pricing.yaml

@@ -161,6 +161,16 @@ openai:
     input_per_1k: 0.0001
     output_per_1k: 0
 
+  # Kimi / Moonshot AI (kimi_api) — OpenAI-compatible /v1 endpoint. Moonshot
+  # reports cache hits OpenAI-style when present; cached input is 10% of
+  # input ($0.30 vs $3.00 per MTok). kimi-k3 is the only model the platform
+  # serves newer accounts (K2-era ids and kimi-latest 404), matching the
+  # management catalog.
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cached_input_per_1k: 0.0003
+
 anthropic:
   # Claude 4.x family — cache reads ≈10% of input, cache writes ≈125% of input.
   # Pricing source: Anthropic's current published rates per million tokens,
@@ -206,6 +216,20 @@ anthropic:
     cache_read_per_1k: 0.0001
     cache_creation_per_1k: 0.00125
 
+  # Kimi / Moonshot AI (kimi_api) via the Anthropic-compatible endpoint
+  # (/anthropic/v1/messages — the official Claude Code setup). Same rates
+  # as the OpenAI-shape entry above. "kimi-k3[1m]" is the model id some
+  # Claude Code guides set for the 1M-context alias; priced identically so
+  # cost metering doesn't silently skip those requests.
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+  "kimi-k3[1m]":
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+
 bedrock:
   # AWS Bedrock model ids, normalised by the request parser (cross-region
   # inference-profile prefix + version/throughput suffix stripped), e.g.