[management, proxy] Management-owned LLM pricing: file-backed defaults + (#6965)

management/internals/modules/agentnetwork/pricing/exampleyaml.go

@@ -0,0 +1,109 @@
+package pricing
+
+import (
+	"bytes"
+	"fmt"
+	"sort"
+	"strconv"
+)
+
+// MarshalDefaultsYAML renders the built-in default pricing table (catalog
+// + supplementals, WITHOUT any operator override) as the YAML schema
+// LoadOverrideFile consumes. It backs the generated
+// defaults_llm_pricing.example.yaml so operators start from a file that
+// matches the compiled-in rates exactly; a golden test keeps the two in
+// sync. Output is deterministic (sorted surfaces and models).
+func MarshalDefaultsYAML() []byte {
+	var b bytes.Buffer
+	b.WriteString(exampleHeader)
+
+	table := buildDefaultTable()
+	surfaces := make([]string, 0, len(table))
+	for s := range table {
+		surfaces = append(surfaces, s)
+	}
+	sort.Strings(surfaces)
+
+	for _, surface := range surfaces {
+		fmt.Fprintf(&b, "\n%s:\n", surface)
+		models := table[surface]
+		ids := make([]string, 0, len(models))
+		for id := range models {
+			ids = append(ids, id)
+		}
+		sort.Strings(ids)
+		for _, id := range ids {
+			e := models[id]
+			fmt.Fprintf(&b, "  %s:\n", yamlKey(id))
+			fmt.Fprintf(&b, "    input_per_1k: %s\n", rate(e.InputPer1k))
+			fmt.Fprintf(&b, "    output_per_1k: %s\n", rate(e.OutputPer1k))
+			if e.CachedInputPer1k > 0 {
+				fmt.Fprintf(&b, "    cached_input_per_1k: %s\n", rate(e.CachedInputPer1k))
+			}
+			if e.CacheReadPer1k > 0 {
+				fmt.Fprintf(&b, "    cache_read_per_1k: %s\n", rate(e.CacheReadPer1k))
+			}
+			if e.CacheCreationPer1k > 0 {
+				fmt.Fprintf(&b, "    cache_creation_per_1k: %s\n", rate(e.CacheCreationPer1k))
+			}
+		}
+	}
+	return b.Bytes()
+}
+
+// rate renders a USD-per-1k rate without float noise ("0.00015", not
+// "0.000150000000...").
+func rate(v float64) string {
+	return strconv.FormatFloat(v, 'f', -1, 64)
+}
+
+// yamlKey quotes model ids that YAML would otherwise misparse (e.g.
+// "kimi-k3[1m]" starts a flow sequence unquoted).
+func yamlKey(id string) string {
+	for _, r := range id {
+		switch r {
+		case '[', ']', '{', '}', ':', '#', ',', '&', '*', '!', '|', '>', '\'', '"', '%', '@', '`':
+			return strconv.Quote(id)
+		}
+	}
+	return id
+}
+
+const exampleHeader = `# Default LLM pricing used by NetBird's Agent Network cost metering.
+# GENERATED from the management catalog — do not edit this file in the
+# repository; regenerate with:
+#
+#   go generate ./management/internals/modules/agentnetwork/pricing
+#
+# Operators: copy this file to <datadir>/defaults_llm_pricing.yaml (or
+# any path configured via management.json:
+#
+#   { "AgentNetwork": { "PricingDefaultsFile": "/path/defaults_llm_pricing.yaml" } }
+#
+# ) and adjust the entries you want to change. Management re-reads the
+# file periodically (mtime poll, every minute): the live table feeds the
+# proxies' cost metering and the dashboard's model-price prefill, so
+# edits apply without a restart. Your file only needs the entries you
+# want to change — but each entry REPLACES the built-in entry for that
+# surface+model whole, so repeat the cache rates you want to keep.
+# Unknown fields and negative or non-finite rates are rejected: at
+# startup that fails boot (for an explicitly configured path); at
+# runtime the previous table is kept and a warning is logged. Deleting
+# the file reverts to the built-in defaults below.
+#
+# Top-level keys are pricing surfaces — the parser shape requests are
+# metered under: "openai" (also Azure, Mistral, and OpenAI-compatible
+# gateways), "anthropic" (also Anthropic-on-Vertex), "bedrock"
+# (normalized ids, e.g. anthropic.claude-sonnet-4-5). Model keys must be
+# the normalized id the proxy meters (version/region suffixes stripped).
+#
+# Values are USD per 1_000 tokens. Optional cache fields:
+#   cached_input_per_1k    OpenAI shape: rate for cached prompt tokens
+#                          (a SUBSET of input tokens). Absent -> cached
+#                          portion bills at input_per_1k.
+#   cache_read_per_1k      Anthropic shape: rate for cache_read tokens
+#                          (ADDITIVE to input). Absent -> input rate.
+#   cache_creation_per_1k  Anthropic shape: rate for cache_creation
+#                          tokens (ADDITIVE to input). Absent -> input
+#                          rate.
+`