Merge remote-tracking branch 'origin/main' into fix/model-switch-token-limits

commit: 418747c7f4034c46a206f54b37071005d7a7b5a1
author: patriceckhart · 23d ago · 2026-09-29 20:13 UTC
merge of: 7cf3d6c + de90baf · diff against the first parent
signature: unsigned

4 files · +206 -8

M README.md +1 -1


@@ -798,7 +798,7 @@ If both are present, the bearer token wins for inference, while live model disco
798	798		Unlike most providers, Bedrock model pricing, context windows, and capability flags are hand-maintained in zot's baked-in catalog. AWS does not expose usable pricing programmatically for current models — its Price List API returns only stale, input-only rates for a few legacy models — so the catalog is authoritative for cost and metadata. This means:
799	799
800	800		- Bearer-token-only users still get the full curated Bedrock model list with correct pricing from the catalog. No AWS API listing is attempted.
801		-	- Users with SigV4 credentials additionally get live model discovery, even if inference uses a bearer token: on startup, zot lists the foundation models and cross-region inference profiles your account can actually use (via `ListFoundationModels` and `ListInferenceProfiles`) and merges them into the catalog. New Bedrock models therefore appear in `/model` without waiting for a zot release. Discovered IDs already known to the catalog keep their curated pricing; genuinely new IDs show up with placeholder pricing until the catalog is updated.
	801	+	- Users with SigV4 credentials additionally get live model discovery, even if inference uses a bearer token: on startup, zot lists the foundation models and cross-region inference profiles your account can actually use (via `ListFoundationModels` and `ListInferenceProfiles`) and merges them into the catalog. New Bedrock models therefore appear in `/model` without waiting for a zot release. Discovered IDs already known to the catalog keep their curated metadata. An inference profile the catalog doesn't list under that exact ID (for example `apac.` or `us-gov.` variants) inherits the limits, reasoning flags, and pricing of its base model. Models with no matching base catalog entry, including new Claude models, retain placeholder pricing and the client's 4096-token output fallback until their limits are verified and added to the catalog.
802	802
803	803		Live discovery requires SigV4 because the listing APIs live on the Bedrock control plane, which rejects the runtime bearer token. It runs in the background at startup and is cached for 6h like other providers, so `/model` reflects the latest list without any manual refresh. If discovery fails (no SigV4 creds, expired SSO login, missing IAM permissions), zot silently falls back to the catalog. Discovery needs `bedrock:ListFoundationModels` and `bedrock:ListInferenceProfiles`.
804	804


@@ -798,7 +798,7 @@ If both are present, the bearer token wins for inference, while live model disco
798		Unlike most providers, Bedrock model pricing, context windows, and capability flags are hand-maintained in zot's baked-in catalog. AWS does not expose usable pricing programmatically for current models — its Price List API returns only stale, input-only rates for a few legacy models — so the catalog is authoritative for cost and metadata. This means:	798		Unlike most providers, Bedrock model pricing, context windows, and capability flags are hand-maintained in zot's baked-in catalog. AWS does not expose usable pricing programmatically for current models — its Price List API returns only stale, input-only rates for a few legacy models — so the catalog is authoritative for cost and metadata. This means:
799			799
800		- Bearer-token-only users still get the full curated Bedrock model list with correct pricing from the catalog. No AWS API listing is attempted.	800		- Bearer-token-only users still get the full curated Bedrock model list with correct pricing from the catalog. No AWS API listing is attempted.
801	-	- Users with SigV4 credentials additionally get live model discovery, even if inference uses a bearer token: on startup, zot lists the foundation models and cross-region inference profiles your account can actually use (via `ListFoundationModels` and `ListInferenceProfiles`) and merges them into the catalog. New Bedrock models therefore appear in `/model` without waiting for a zot release. Discovered IDs already known to the catalog keep their curated pricing; genuinely new IDs show up with placeholder pricing until the catalog is updated.	801	+	- Users with SigV4 credentials additionally get live model discovery, even if inference uses a bearer token: on startup, zot lists the foundation models and cross-region inference profiles your account can actually use (via `ListFoundationModels` and `ListInferenceProfiles`) and merges them into the catalog. New Bedrock models therefore appear in `/model` without waiting for a zot release. Discovered IDs already known to the catalog keep their curated metadata. An inference profile the catalog doesn't list under that exact ID (for example `apac.` or `us-gov.` variants) inherits the limits, reasoning flags, and pricing of its base model. Models with no matching base catalog entry, including new Claude models, retain placeholder pricing and the client's 4096-token output fallback until their limits are verified and added to the catalog.
802			802
803		Live discovery requires SigV4 because the listing APIs live on the Bedrock control plane, which rejects the runtime bearer token. It runs in the background at startup and is cached for 6h like other providers, so `/model` reflects the latest list without any manual refresh. If discovery fails (no SigV4 creds, expired SSO login, missing IAM permissions), zot silently falls back to the catalog. Discovery needs `bedrock:ListFoundationModels` and `bedrock:ListInferenceProfiles`.	803		Live discovery requires SigV4 because the listing APIs live on the Bedrock control plane, which rejects the runtime bearer token. It runs in the background at startup and is cached for 6h like other providers, so `/model` reflects the latest list without any manual refresh. If discovery fails (no SigV4 creds, expired SSO login, missing IAM permissions), zot silently falls back to the catalog. Discovery needs `bedrock:ListFoundationModels` and `bedrock:ListInferenceProfiles`.
804			804

M packages/provider/amazon_bedrock.go +1 -1


@@ -600,7 +600,7 @@ func resolveBedrockInferenceProfileID(modelID, region string) string {
600	600		// bedrockGeoPrefixes are the cross-region inference-profile geo
601	601		// prefixes Bedrock uses. A model ID that starts with one of these
602	602		// (followed by a dot) is already a profile reference.
603		-	var bedrockGeoPrefixes = []string{"us-gov", "us", "eu", "apac", "ap", "global", "au"}
	603	+	var bedrockGeoPrefixes = []string{"us-gov", "us", "eu", "apac", "ap", "global", "au", "jp"}
604	604
605	605		func bedrockHasGeoPrefix(modelID string) bool {
606	606		for _, p := range bedrockGeoPrefixes {


@@ -600,7 +600,7 @@ func resolveBedrockInferenceProfileID(modelID, region string) string {
600		// bedrockGeoPrefixes are the cross-region inference-profile geo	600		// bedrockGeoPrefixes are the cross-region inference-profile geo
601		// prefixes Bedrock uses. A model ID that starts with one of these	601		// prefixes Bedrock uses. A model ID that starts with one of these
602		// (followed by a dot) is already a profile reference.	602		// (followed by a dot) is already a profile reference.
603	-	var bedrockGeoPrefixes = []string{"us-gov", "us", "eu", "apac", "ap", "global", "au"}	603	+	var bedrockGeoPrefixes = []string{"us-gov", "us", "eu", "apac", "ap", "global", "au", "jp"}
604			604
605		func bedrockHasGeoPrefix(modelID string) bool {	605		func bedrockHasGeoPrefix(modelID string) bool {
606		for _, p := range bedrockGeoPrefixes {	606		for _, p := range bedrockGeoPrefixes {

M packages/provider/discover_bedrock.go +97 -6


@@ -34,6 +34,7 @@ import (
34	34		"encoding/json"
35	35		"fmt"
36	36		"io"
	37	+	"maps"
37	38		"net/http"
38	39		"net/url"
39	40		"strings"

@@ -80,16 +81,106 @@ func DiscoverBedrock(ctx context.Context, region string) ([]Model, error) {
80	81
81	82		out := make([]Model, 0, len(ids))
82	83		for _, id := range ids {
83		-	out = append(out, Model{
84		-	Provider: "amazon-bedrock",
85		-	ID: id,
86		-	Source: "live",
87		-	BaseURL: "https://bedrock-runtime." + region + ".amazonaws.com",
88		-	})
	84	+	out = append(out, bedrockDiscoveredModel(id, region))
89	85		}
90	86		return out, nil
91	87		}
92	88
	89	+	// bedrockDiscoveredModel builds the Model for a discovered ID. The
	90	+	// control plane returns IDs only, so limits and capabilities come from
	91	+	// the static catalog. MergeCatalog already uses an exact catalog match;
	92	+	// this handles IDs without one, most often a geo-prefixed inference
	93	+	// profile (apac., us-gov., ...) whose base model is catalogued. Without
	94	+	// this the model would carry MaxOutput 0 and the Bedrock client would
	95	+	// cap every response at its 4096-token fallback.
	96	+	func bedrockDiscoveredModel(id, region string) Model {
	97	+	m := Model{
	98	+	Provider: "amazon-bedrock",
	99	+	ID: id,
	100	+	Source: "live",
	101	+	BaseURL: "https://bedrock-runtime." + region + ".amazonaws.com",
	102	+	}
	103	+	if bedrockInCatalog(id) {
	104	+	// MergeCatalog keeps the exact catalog entry's metadata.
	105	+	return m
	106	+	}
	107	+	prefix, base := bedrockSplitGeoPrefix(id)
	108	+	if c, ok := bedrockCatalogBase(base); ok {
	109	+	m.ContextWindow = c.ContextWindow
	110	+	m.MaxOutput = c.MaxOutput
	111	+	m.Reasoning = c.Reasoning
	112	+	m.ReasoningLevelMap = maps.Clone(c.ReasoningLevelMap)
	113	+	m.AdaptiveThinking = c.AdaptiveThinking
	114	+	m.PriceInput = c.PriceInput
	115	+	m.PriceOutput = c.PriceOutput
	116	+	m.PriceCacheRead = c.PriceCacheRead
	117	+	m.PriceCacheWrite = c.PriceCacheWrite
	118	+	m.PriceTierInputTokens = c.PriceTierInputTokens
	119	+	m.PriceInputAbove = c.PriceInputAbove
	120	+	m.PriceOutputAbove = c.PriceOutputAbove
	121	+	m.PriceCacheReadAbove = c.PriceCacheReadAbove
	122	+	m.PriceCacheWriteAbove = c.PriceCacheWriteAbove
	123	+	if c.DisplayName != "" {
	124	+	m.DisplayName = c.DisplayName
	125	+	if cp, _ := bedrockSplitGeoPrefix(c.ID); cp != "" {
	126	+	suffix := " (" + cp + ")"
	127	+	if strings.HasSuffix(strings.ToUpper(m.DisplayName), strings.ToUpper(suffix)) {
	128	+	m.DisplayName = m.DisplayName[:len(m.DisplayName)-len(suffix)]
	129	+	}
	130	+	}
	131	+	if prefix != "" {
	132	+	m.DisplayName += " (" + strings.ToUpper(prefix) + ")"
	133	+	}
	134	+	}
	135	+	return m
	136	+	}
	137	+	// Unknown models have no verified output or context limits. Keep the
	138	+	// client's 4096-token fallback rather than guessing from the family:
	139	+	// even Claude Opus 4 supports only 32k, not the 64k of newer models.
	140	+	return m
	141	+	}
	142	+
	143	+	// bedrockSplitGeoPrefix separates a cross-region inference-profile geo
	144	+	// prefix from the foundation-model ID. prefix is "" when there is none.
	145	+	func bedrockSplitGeoPrefix(id string) (prefix, base string) {
	146	+	for _, p := range bedrockGeoPrefixes {
	147	+	if strings.HasPrefix(id, p+".") {
	148	+	return p, id[len(p)+1:]
	149	+	}
	150	+	}
	151	+	return "", id
	152	+	}
	153	+
	154	+	func bedrockInCatalog(id string) bool {
	155	+	for _, c := range Catalog {
	156	+	if c.Provider == "amazon-bedrock" && c.ID == id {
	157	+	return true
	158	+	}
	159	+	}
	160	+	return false
	161	+	}
	162	+
	163	+	// bedrockCatalogBase finds catalog metadata for a foundation-model ID:
	164	+	// the bare entry if present, otherwise any geo-prefixed variant of it.
	165	+	func bedrockCatalogBase(base string) (Model, bool) {
	166	+	var variant Model
	167	+	found := false
	168	+	for _, c := range Catalog {
	169	+	if c.Provider != "amazon-bedrock" {
	170	+	continue
	171	+	}
	172	+	if c.ID == base {
	173	+	return c, true
	174	+	}
	175	+	if !found {
	176	+	if _, b := bedrockSplitGeoPrefix(c.ID); b == base {
	177	+	variant, found = c, true
	178	+	}
	179	+	}
	180	+	}
	181	+	return variant, found
	182	+	}
	183	+
93	184		// bedrockListModelIDs enumerates foundation-model IDs and inference-
94	185		// profile IDs available in the region. Profile IDs are the geo-prefixed
95	186		// forms zot invokes; both are returned so the picker shows every


@@ -34,6 +34,7 @@ import (
34		"encoding/json"	34		"encoding/json"
35		"fmt"	35		"fmt"
36		"io"	36		"io"
			37	+	"maps"
37		"net/http"	38		"net/http"
38		"net/url"	39		"net/url"
39		"strings"	40		"strings"

@@ -80,16 +81,106 @@ func DiscoverBedrock(ctx context.Context, region string) ([]Model, error) {
80			81
81		out := make([]Model, 0, len(ids))	82		out := make([]Model, 0, len(ids))
82		for _, id := range ids {	83		for _, id := range ids {
83	-	out = append(out, Model{	84	+	out = append(out, bedrockDiscoveredModel(id, region))
84	-	Provider: "amazon-bedrock",
85	-	ID: id,
86	-	Source: "live",
87	-	BaseURL: "https://bedrock-runtime." + region + ".amazonaws.com",
88	-	})
89		}	85		}
90		return out, nil	86		return out, nil
91		}	87		}
92			88
			89	+	// bedrockDiscoveredModel builds the Model for a discovered ID. The
			90	+	// control plane returns IDs only, so limits and capabilities come from
			91	+	// the static catalog. MergeCatalog already uses an exact catalog match;
			92	+	// this handles IDs without one, most often a geo-prefixed inference
			93	+	// profile (apac., us-gov., ...) whose base model is catalogued. Without
			94	+	// this the model would carry MaxOutput 0 and the Bedrock client would
			95	+	// cap every response at its 4096-token fallback.
			96	+	func bedrockDiscoveredModel(id, region string) Model {
			97	+	m := Model{
			98	+	Provider: "amazon-bedrock",
			99	+	ID: id,
			100	+	Source: "live",
			101	+	BaseURL: "https://bedrock-runtime." + region + ".amazonaws.com",
			102	+	}
			103	+	if bedrockInCatalog(id) {
			104	+	// MergeCatalog keeps the exact catalog entry's metadata.
			105	+	return m
			106	+	}
			107	+	prefix, base := bedrockSplitGeoPrefix(id)
			108	+	if c, ok := bedrockCatalogBase(base); ok {
			109	+	m.ContextWindow = c.ContextWindow
			110	+	m.MaxOutput = c.MaxOutput
			111	+	m.Reasoning = c.Reasoning
			112	+	m.ReasoningLevelMap = maps.Clone(c.ReasoningLevelMap)
			113	+	m.AdaptiveThinking = c.AdaptiveThinking
			114	+	m.PriceInput = c.PriceInput
			115	+	m.PriceOutput = c.PriceOutput
			116	+	m.PriceCacheRead = c.PriceCacheRead
			117	+	m.PriceCacheWrite = c.PriceCacheWrite
			118	+	m.PriceTierInputTokens = c.PriceTierInputTokens
			119	+	m.PriceInputAbove = c.PriceInputAbove
			120	+	m.PriceOutputAbove = c.PriceOutputAbove
			121	+	m.PriceCacheReadAbove = c.PriceCacheReadAbove
			122	+	m.PriceCacheWriteAbove = c.PriceCacheWriteAbove
			123	+	if c.DisplayName != "" {
			124	+	m.DisplayName = c.DisplayName
			125	+	if cp, _ := bedrockSplitGeoPrefix(c.ID); cp != "" {
			126	+	suffix := " (" + cp + ")"
			127	+	if strings.HasSuffix(strings.ToUpper(m.DisplayName), strings.ToUpper(suffix)) {
			128	+	m.DisplayName = m.DisplayName[:len(m.DisplayName)-len(suffix)]
			129	+	}
			130	+	}
			131	+	if prefix != "" {
			132	+	m.DisplayName += " (" + strings.ToUpper(prefix) + ")"
			133	+	}
			134	+	}
			135	+	return m
			136	+	}
			137	+	// Unknown models have no verified output or context limits. Keep the
			138	+	// client's 4096-token fallback rather than guessing from the family:
			139	+	// even Claude Opus 4 supports only 32k, not the 64k of newer models.
			140	+	return m
			141	+	}
			142	+
			143	+	// bedrockSplitGeoPrefix separates a cross-region inference-profile geo
			144	+	// prefix from the foundation-model ID. prefix is "" when there is none.
			145	+	func bedrockSplitGeoPrefix(id string) (prefix, base string) {
			146	+	for _, p := range bedrockGeoPrefixes {
			147	+	if strings.HasPrefix(id, p+".") {
			148	+	return p, id[len(p)+1:]
			149	+	}
			150	+	}
			151	+	return "", id
			152	+	}
			153	+
			154	+	func bedrockInCatalog(id string) bool {
			155	+	for _, c := range Catalog {
			156	+	if c.Provider == "amazon-bedrock" && c.ID == id {
			157	+	return true
			158	+	}
			159	+	}
			160	+	return false
			161	+	}
			162	+
			163	+	// bedrockCatalogBase finds catalog metadata for a foundation-model ID:
			164	+	// the bare entry if present, otherwise any geo-prefixed variant of it.
			165	+	func bedrockCatalogBase(base string) (Model, bool) {
			166	+	var variant Model
			167	+	found := false
			168	+	for _, c := range Catalog {
			169	+	if c.Provider != "amazon-bedrock" {
			170	+	continue
			171	+	}
			172	+	if c.ID == base {
			173	+	return c, true
			174	+	}
			175	+	if !found {
			176	+	if _, b := bedrockSplitGeoPrefix(c.ID); b == base {
			177	+	variant, found = c, true
			178	+	}
			179	+	}
			180	+	}
			181	+	return variant, found
			182	+	}
			183	+
93		// bedrockListModelIDs enumerates foundation-model IDs and inference-	184		// bedrockListModelIDs enumerates foundation-model IDs and inference-
94		// profile IDs available in the region. Profile IDs are the geo-prefixed	185		// profile IDs available in the region. Profile IDs are the geo-prefixed
95		// forms zot invokes; both are returned so the picker shows every	186		// forms zot invokes; both are returned so the picker shows every

A packages/provider/discover_bedrock_metadata_test.go +107 -0

@@ -0,0 +1,107 @@
	1	+	package provider
	2	+
	3	+	import "testing"
	4	+
	5	+	// discoveredWireMaxTokens runs a discovered ID through the same path as
	6	+	// production (discovery metadata -> MergeCatalog -> FindModel -> the
	7	+	// agent's MaxTokens -> Bedrock buildRequest) and returns the merged model
	8	+	// plus the maxTokens that would be sent on the wire.
	9	+	func discoveredWireMaxTokens(t *testing.T, id string) (Model, int) {
	10	+	t.Helper()
	11	+	withLiveModels(t, []Model{bedrockDiscoveredModel(id, "us-east-1")})
	12	+	m, err := FindModel("amazon-bedrock", id)
	13	+	if err != nil {
	14	+	t.Fatalf("FindModel(%q): %v", id, err)
	15	+	}
	16	+	c := &bedrockClient{region: "us-east-1"}
	17	+	req, err := c.buildRequest(Request{Model: id, MaxTokens: m.MaxOutput})
	18	+	if err != nil {
	19	+	t.Fatalf("buildRequest(%q): %v", id, err)
	20	+	}
	21	+	return m, req.InferenceConfig.MaxTokens
	22	+	}
	23	+
	24	+	// Discovered inference-profile IDs with no exact catalog entry used to
	25	+	// carry MaxOutput 0, so every request fell back to maxTokens 4096 and
	26	+	// long responses stopped with max_tokens.
	27	+	func TestBedrockDiscoveredProfileInheritsBaseMetadata(t *testing.T) {
	28	+	base, err := FindModel("amazon-bedrock", "anthropic.claude-sonnet-4-6")
	29	+	if err != nil {
	30	+	t.Fatal(err)
	31	+	}
	32	+	for _, id := range []string{"apac.anthropic.claude-sonnet-4-6", "us-gov.anthropic.claude-sonnet-4-6"} {
	33	+	m, wire := discoveredWireMaxTokens(t, id)
	34	+	if m.MaxOutput != base.MaxOutput \|\| wire != base.MaxOutput {
	35	+	t.Errorf("%s: MaxOutput=%d wire=%d, want %d", id, m.MaxOutput, wire, base.MaxOutput)
	36	+	}
	37	+	if m.ContextWindow != base.ContextWindow \|\| m.Reasoning != base.Reasoning {
	38	+	t.Errorf("%s: metadata not inherited: %+v", id, m)
	39	+	}
	40	+	if m.PriceOutput != base.PriceOutput {
	41	+	t.Errorf("%s: PriceOutput=%v, want %v", id, m.PriceOutput, base.PriceOutput)
	42	+	}
	43	+	}
	44	+
	45	+	m, _ := discoveredWireMaxTokens(t, "apac.anthropic.claude-sonnet-4-6")
	46	+	if m.DisplayName != "Claude Sonnet 4.6 (APAC)" {
	47	+	t.Errorf("DisplayName = %q", m.DisplayName)
	48	+	}
	49	+	}
	50	+
	51	+	// A base model catalogued only under geo-prefixed IDs still donates its
	52	+	// metadata, and the donor's region label is not carried over.
	53	+	func TestBedrockDiscoveredProfileUsesPrefixedCatalogVariant(t *testing.T) {
	54	+	const id = "apac.anthropic.claude-opus-5"
	55	+	m, wire := discoveredWireMaxTokens(t, id)
	56	+	if wire != 128000 \|\| !m.AdaptiveThinking \|\| m.PriceOutput != 25 {
	57	+	t.Errorf("%s: wire=%d adaptive=%v price=%v, want 128000 adaptive and $25", id, wire, m.AdaptiveThinking, m.PriceOutput)
	58	+	}
	59	+	if m.DisplayName != "Claude Opus 5 (APAC)" {
	60	+	t.Errorf("DisplayName = %q", m.DisplayName)
	61	+	}
	62	+	}
	63	+
	64	+	// An uncatalogued Claude ID must not assume a 64k output limit. Opus 4
	65	+	// supports only 32k, and an unknown future model has no verified limit.
	66	+	func TestBedrockDiscoveredUnknownClaudeKeepsFallback(t *testing.T) {
	67	+	for _, id := range []string{
	68	+	"anthropic.claude-opus-4-20250514-v1:0",
	69	+	"us.anthropic.claude-opus-4-20250514-v1:0",
	70	+	"anthropic.claude-future-9",
	71	+	} {
	72	+	m, wire := discoveredWireMaxTokens(t, id)
	73	+	if m.MaxOutput != 0 \|\| m.ContextWindow != 0 \|\| wire != 4096 {
	74	+	t.Errorf("%s: MaxOutput=%d ctx=%d wire=%d, want 0/0/4096", id, m.MaxOutput, m.ContextWindow, wire)
	75	+	}
	76	+	}
	77	+	}
	78	+
	79	+	// Legacy Claude and non-Anthropic models cap output well below 64k and
	80	+	// would reject a larger request, so they keep the client fallback.
	81	+	func TestBedrockDiscoveredLegacyAndOtherModelsKeepFallback(t *testing.T) {
	82	+	for _, id := range []string{
	83	+	"anthropic.claude-v2:1",
	84	+	"anthropic.claude-instant-v1",
	85	+	"us.anthropic.claude-3-5-sonnet-20241022-v2:0",
	86	+	"anthropic.claude-3-haiku-20240307-v1:0",
	87	+	"mistral.unknown-model-v1:0",
	88	+	} {
	89	+	m, wire := discoveredWireMaxTokens(t, id)
	90	+	if m.MaxOutput != 0 \|\| wire != 4096 {
	91	+	t.Errorf("%s: MaxOutput=%d wire=%d, want 0/4096", id, m.MaxOutput, wire)
	92	+	}
	93	+	}
	94	+	}
	95	+
	96	+	// Exact catalog IDs keep their own catalog metadata through MergeCatalog.
	97	+	func TestBedrockDiscoveredExactCatalogIDUnchanged(t *testing.T) {
	98	+	const id = "au.anthropic.claude-sonnet-4-6"
	99	+	want, err := FindModel("amazon-bedrock", id)
	100	+	if err != nil {
	101	+	t.Fatal(err)
	102	+	}
	103	+	m, wire := discoveredWireMaxTokens(t, id)
	104	+	if m.MaxOutput != want.MaxOutput \|\| wire != want.MaxOutput \|\| m.DisplayName != want.DisplayName {
	105	+	t.Errorf("got MaxOutput=%d wire=%d name=%q, want %d %q", m.MaxOutput, wire, m.DisplayName, want.MaxOutput, want.DisplayName)
	106	+	}
	107	+	}

@@ -0,0 +1,107 @@
	1	+	package provider
	2	+
	3	+	import "testing"
	4	+
	5	+	// discoveredWireMaxTokens runs a discovered ID through the same path as
	6	+	// production (discovery metadata -> MergeCatalog -> FindModel -> the
	7	+	// agent's MaxTokens -> Bedrock buildRequest) and returns the merged model
	8	+	// plus the maxTokens that would be sent on the wire.
	9	+	func discoveredWireMaxTokens(t *testing.T, id string) (Model, int) {
	10	+	t.Helper()
	11	+	withLiveModels(t, []Model{bedrockDiscoveredModel(id, "us-east-1")})
	12	+	m, err := FindModel("amazon-bedrock", id)
	13	+	if err != nil {
	14	+	t.Fatalf("FindModel(%q): %v", id, err)
	15	+	}
	16	+	c := &bedrockClient{region: "us-east-1"}
	17	+	req, err := c.buildRequest(Request{Model: id, MaxTokens: m.MaxOutput})
	18	+	if err != nil {
	19	+	t.Fatalf("buildRequest(%q): %v", id, err)
	20	+	}
	21	+	return m, req.InferenceConfig.MaxTokens
	22	+	}
	23	+
	24	+	// Discovered inference-profile IDs with no exact catalog entry used to
	25	+	// carry MaxOutput 0, so every request fell back to maxTokens 4096 and
	26	+	// long responses stopped with max_tokens.
	27	+	func TestBedrockDiscoveredProfileInheritsBaseMetadata(t *testing.T) {
	28	+	base, err := FindModel("amazon-bedrock", "anthropic.claude-sonnet-4-6")
	29	+	if err != nil {
	30	+	t.Fatal(err)
	31	+	}
	32	+	for _, id := range []string{"apac.anthropic.claude-sonnet-4-6", "us-gov.anthropic.claude-sonnet-4-6"} {
	33	+	m, wire := discoveredWireMaxTokens(t, id)
	34	+	if m.MaxOutput != base.MaxOutput \|\| wire != base.MaxOutput {
	35	+	t.Errorf("%s: MaxOutput=%d wire=%d, want %d", id, m.MaxOutput, wire, base.MaxOutput)
	36	+	}
	37	+	if m.ContextWindow != base.ContextWindow \|\| m.Reasoning != base.Reasoning {
	38	+	t.Errorf("%s: metadata not inherited: %+v", id, m)
	39	+	}
	40	+	if m.PriceOutput != base.PriceOutput {
	41	+	t.Errorf("%s: PriceOutput=%v, want %v", id, m.PriceOutput, base.PriceOutput)
	42	+	}
	43	+	}
	44	+
	45	+	m, _ := discoveredWireMaxTokens(t, "apac.anthropic.claude-sonnet-4-6")
	46	+	if m.DisplayName != "Claude Sonnet 4.6 (APAC)" {
	47	+	t.Errorf("DisplayName = %q", m.DisplayName)
	48	+	}
	49	+	}
	50	+
	51	+	// A base model catalogued only under geo-prefixed IDs still donates its
	52	+	// metadata, and the donor's region label is not carried over.
	53	+	func TestBedrockDiscoveredProfileUsesPrefixedCatalogVariant(t *testing.T) {
	54	+	const id = "apac.anthropic.claude-opus-5"
	55	+	m, wire := discoveredWireMaxTokens(t, id)
	56	+	if wire != 128000 \|\| !m.AdaptiveThinking \|\| m.PriceOutput != 25 {
	57	+	t.Errorf("%s: wire=%d adaptive=%v price=%v, want 128000 adaptive and $25", id, wire, m.AdaptiveThinking, m.PriceOutput)
	58	+	}
	59	+	if m.DisplayName != "Claude Opus 5 (APAC)" {
	60	+	t.Errorf("DisplayName = %q", m.DisplayName)
	61	+	}
	62	+	}
	63	+
	64	+	// An uncatalogued Claude ID must not assume a 64k output limit. Opus 4
	65	+	// supports only 32k, and an unknown future model has no verified limit.
	66	+	func TestBedrockDiscoveredUnknownClaudeKeepsFallback(t *testing.T) {
	67	+	for _, id := range []string{
	68	+	"anthropic.claude-opus-4-20250514-v1:0",
	69	+	"us.anthropic.claude-opus-4-20250514-v1:0",
	70	+	"anthropic.claude-future-9",
	71	+	} {
	72	+	m, wire := discoveredWireMaxTokens(t, id)
	73	+	if m.MaxOutput != 0 \|\| m.ContextWindow != 0 \|\| wire != 4096 {
	74	+	t.Errorf("%s: MaxOutput=%d ctx=%d wire=%d, want 0/0/4096", id, m.MaxOutput, m.ContextWindow, wire)
	75	+	}
	76	+	}
	77	+	}
	78	+
	79	+	// Legacy Claude and non-Anthropic models cap output well below 64k and
	80	+	// would reject a larger request, so they keep the client fallback.
	81	+	func TestBedrockDiscoveredLegacyAndOtherModelsKeepFallback(t *testing.T) {
	82	+	for _, id := range []string{
	83	+	"anthropic.claude-v2:1",
	84	+	"anthropic.claude-instant-v1",
	85	+	"us.anthropic.claude-3-5-sonnet-20241022-v2:0",
	86	+	"anthropic.claude-3-haiku-20240307-v1:0",
	87	+	"mistral.unknown-model-v1:0",
	88	+	} {
	89	+	m, wire := discoveredWireMaxTokens(t, id)
	90	+	if m.MaxOutput != 0 \|\| wire != 4096 {
	91	+	t.Errorf("%s: MaxOutput=%d wire=%d, want 0/4096", id, m.MaxOutput, wire)
	92	+	}
	93	+	}
	94	+	}
	95	+
	96	+	// Exact catalog IDs keep their own catalog metadata through MergeCatalog.
	97	+	func TestBedrockDiscoveredExactCatalogIDUnchanged(t *testing.T) {
	98	+	const id = "au.anthropic.claude-sonnet-4-6"
	99	+	want, err := FindModel("amazon-bedrock", id)
	100	+	if err != nil {
	101	+	t.Fatal(err)
	102	+	}
	103	+	m, wire := discoveredWireMaxTokens(t, id)
	104	+	if m.MaxOutput != want.MaxOutput \|\| wire != want.MaxOutput \|\| m.DisplayName != want.DisplayName {
	105	+	t.Errorf("got MaxOutput=%d wire=%d name=%q, want %d %q", m.MaxOutput, wire, m.DisplayName, want.MaxOutput, want.DisplayName)
	106	+	}
	107	+	}