[management, proxy] add Kimi (Moonshot AI) provider to Agent Network (#6853)

This commit is contained in:
Maycon Santos
2026-07-23 13:32:32 +02:00
committed by GitHub
parent d4a4418969
commit 0936918d24
6 changed files with 121 additions and 15 deletions
+37 -12
View File
@@ -20,14 +20,15 @@ import (
// covers whatever credentials are present (source ~/.llm-keys locally / set the
// Actions secrets in CI).
type providerCase struct {
name string
catalogID string
upstream string
apiKey string
model string // body model (chat/messages) or path model@version (vertex)
kind string // harness.WireChat, harness.WireMessages, or harness.WireVertex
project string // vertex only: GCP project for the rawPredict path
region string // vertex only: GCP region for the rawPredict path
name string
catalogID string
upstream string
apiKey string
model string // body model (chat/messages) or path model@version (vertex)
kind string // harness.WireChat, harness.WireMessages, or harness.WireVertex
project string // vertex only: GCP project for the rawPredict path
region string // vertex only: GCP region for the rawPredict path
pathPrefix string // base-URL path prefix the agent carries (e.g. "/anthropic" for Kimi)
}
// availableProviders builds the matrix from the provider env vars that are set.
@@ -39,6 +40,24 @@ 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 provider keeps the bare
// default upstream and the AGENT carries the /anthropic prefix in
// its base URL — exactly the documented Claude Code / Kimi CLI
// setup (ANTHROPIC_BASE_URL=https://<endpoint>/anthropic) — so one
// provider serves both shapes and the prefix rides through to
// Moonshot. 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). 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).
ps = append(ps, providerCase{name: "kimi", catalogID: "kimi_api", upstream: "https://api.moonshot.ai", apiKey: k, model: "kimi-k3", kind: harness.WireMessages, pathPrefix: "/anthropic"})
}
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})
}
@@ -84,12 +103,18 @@ 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
@@ -246,7 +271,7 @@ func TestProvidersMatrix(t *testing.T) {
case harness.WireBedrock:
c, b, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, pc.model, "Reply with exactly: pong", sessionID)
default:
c, b, cerr = cl.Chat(ctx, settings.Endpoint, proxyIP, pc.kind, pc.model, "Reply with exactly: pong", sessionID)
c, b, cerr = cl.ChatPrefixed(ctx, settings.Endpoint, proxyIP, pc.pathPrefix, pc.kind, pc.model, "Reply with exactly: pong", sessionID)
}
if cerr == nil {
code, body = c, b
+4 -2
View File
@@ -52,7 +52,9 @@ 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:
@@ -72,7 +74,7 @@ func sendModel(ctx context.Context, t *testing.T, cl *harness.Client, endpoint,
case harness.WireVertex:
code, _, err = cl.Vertex(ctx, endpoint, proxyIP, pc.project, pc.region, model, "Reply with exactly: pong", "")
default:
code, _, err = cl.Chat(ctx, endpoint, proxyIP, pc.kind, model, "Reply with exactly: pong", "")
code, _, err = cl.ChatPrefixed(ctx, endpoint, proxyIP, pc.pathPrefix, pc.kind, model, "Reply with exactly: pong", "")
}
require.NoError(t, err, "request must reach the proxy for %s", pc.name)
return code