Give discovered Bedrock models real output limits

Live discovery returns model IDs only. IDs without an exact catalog
entry, such as apac. or us-gov. inference profiles, were merged with
MaxOutput 0, so the Bedrock client fell back to maxTokens 4096 and long
responses stopped with max_tokens.

Discovered profile IDs now inherit limits, reasoning flags, and pricing
from their base model's catalog entry. Claude 4+ models absent from the
catalog get a 64k output / 200k context default (Haiku 4.5's limits,
the smallest among current Claude models). Legacy Claude and other
unknown models keep the existing fallback. Also recognize the jp. geo
prefix, which the catalog already uses.

commit: 8cb588bf0844a8febc1c198da22fe71e6e0841a9
author: Raymond Gasper · 23d ago · 2026-09-29 13:33 UTC
parent: 2ff82fa
signature: unsigned

4 files · +230 -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. Claude models that aren't in the catalog at all get a 64k output / 200k context default; other unknown IDs show up with placeholder pricing and default limits until the catalog is updated.
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. Claude models that aren't in the catalog at all get a 64k output / 200k context default; other unknown IDs show up with placeholder pricing and default limits until the catalog is updated.
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 +125 -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 +81,134 @@ 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	+	// bedrockAnthropicDefaultMaxOutput is the output budget assumed for a
	90	+	// discovered Claude model with no catalog entry. It matches Claude Haiku
	91	+	// 4.5, the smallest output limit among current Claude models, so it is
	92	+	// safe for any Claude 4+ model and far above the client's 4096 fallback.
	93	+	const bedrockAnthropicDefaultMaxOutput = 64000
	94	+
	95	+	// bedrockAnthropicDefaultContextWindow is the context window assumed
	96	+	// for an uncatalogued Claude model, the smallest among current models.
	97	+	const bedrockAnthropicDefaultContextWindow = 200000
	98	+
	99	+	// bedrockDiscoveredModel builds the Model for a discovered ID. The
	100	+	// control plane returns IDs only, so limits and capabilities come from
	101	+	// the static catalog. MergeCatalog already uses an exact catalog match;
	102	+	// this handles IDs without one, most often a geo-prefixed inference
	103	+	// profile (apac., us-gov., ...) whose base model is catalogued. Without
	104	+	// this the model would carry MaxOutput 0 and the Bedrock client would
	105	+	// cap every response at its 4096-token fallback.
	106	+	func bedrockDiscoveredModel(id, region string) Model {
	107	+	m := Model{
	108	+	Provider: "amazon-bedrock",
	109	+	ID: id,
	110	+	Source: "live",
	111	+	BaseURL: "https://bedrock-runtime." + region + ".amazonaws.com",
	112	+	}
	113	+	if bedrockInCatalog(id) {
	114	+	// MergeCatalog keeps the exact catalog entry's metadata.
	115	+	return m
	116	+	}
	117	+	prefix, base := bedrockSplitGeoPrefix(id)
	118	+	if c, ok := bedrockCatalogBase(base); ok {
	119	+	m.ContextWindow = c.ContextWindow
	120	+	m.MaxOutput = c.MaxOutput
	121	+	m.Reasoning = c.Reasoning
	122	+	m.ReasoningLevelMap = maps.Clone(c.ReasoningLevelMap)
	123	+	m.AdaptiveThinking = c.AdaptiveThinking
	124	+	m.PriceInput = c.PriceInput
	125	+	m.PriceOutput = c.PriceOutput
	126	+	m.PriceCacheRead = c.PriceCacheRead
	127	+	m.PriceCacheWrite = c.PriceCacheWrite
	128	+	m.PriceTierInputTokens = c.PriceTierInputTokens
	129	+	m.PriceInputAbove = c.PriceInputAbove
	130	+	m.PriceOutputAbove = c.PriceOutputAbove
	131	+	m.PriceCacheReadAbove = c.PriceCacheReadAbove
	132	+	m.PriceCacheWriteAbove = c.PriceCacheWriteAbove
	133	+	if c.DisplayName != "" {
	134	+	m.DisplayName = c.DisplayName
	135	+	if cp, _ := bedrockSplitGeoPrefix(c.ID); cp != "" {
	136	+	m.DisplayName = strings.TrimSuffix(m.DisplayName, " ("+strings.ToUpper(cp)+")")
	137	+	}
	138	+	if prefix != "" {
	139	+	m.DisplayName += " (" + strings.ToUpper(prefix) + ")"
	140	+	}
	141	+	}
	142	+	return m
	143	+	}
	144	+	if bedrockIsModernClaude(base) {
	145	+	m.ContextWindow = bedrockAnthropicDefaultContextWindow
	146	+	m.MaxOutput = bedrockAnthropicDefaultMaxOutput
	147	+	}
	148	+	return m
	149	+	}
	150	+
	151	+	// bedrockSplitGeoPrefix separates a cross-region inference-profile geo
	152	+	// prefix from the foundation-model ID. prefix is "" when there is none.
	153	+	func bedrockSplitGeoPrefix(id string) (prefix, base string) {
	154	+	for _, p := range bedrockGeoPrefixes {
	155	+	if strings.HasPrefix(id, p+".") {
	156	+	return p, id[len(p)+1:]
	157	+	}
	158	+	}
	159	+	return "", id
	160	+	}
	161	+
	162	+	func bedrockInCatalog(id string) bool {
	163	+	for _, c := range Catalog {
	164	+	if c.Provider == "amazon-bedrock" && c.ID == id {
	165	+	return true
	166	+	}
	167	+	}
	168	+	return false
	169	+	}
	170	+
	171	+	// bedrockCatalogBase finds catalog metadata for a foundation-model ID:
	172	+	// the bare entry if present, otherwise any geo-prefixed variant of it.
	173	+	func bedrockCatalogBase(base string) (Model, bool) {
	174	+	var variant Model
	175	+	found := false
	176	+	for _, c := range Catalog {
	177	+	if c.Provider != "amazon-bedrock" {
	178	+	continue
	179	+	}
	180	+	if c.ID == base {
	181	+	return c, true
	182	+	}
	183	+	if !found {
	184	+	if _, b := bedrockSplitGeoPrefix(c.ID); b == base {
	185	+	variant, found = c, true
	186	+	}
	187	+	}
	188	+	}
	189	+	return variant, found
	190	+	}
	191	+
	192	+	// bedrockIsModernClaude reports whether a foundation-model ID is a Claude
	193	+	// model whose output limit is at least bedrockAnthropicDefaultMaxOutput.
	194	+	// Legacy Claude 2 / Instant / 3 / 3.5 models cap output at 4096-8192 and
	195	+	// would reject a 64000-token request, so they keep the client fallback.
	196	+	func bedrockIsModernClaude(base string) bool {
	197	+	if !strings.HasPrefix(base, "anthropic.claude-") {
	198	+	return false
	199	+	}
	200	+	switch {
	201	+	case strings.HasPrefix(base, "anthropic.claude-v"),
	202	+	strings.HasPrefix(base, "anthropic.claude-instant"),
	203	+	strings.HasPrefix(base, "anthropic.claude-3-5-"),
	204	+	strings.HasPrefix(base, "anthropic.claude-3-haiku"),
	205	+	strings.HasPrefix(base, "anthropic.claude-3-sonnet"),
	206	+	strings.HasPrefix(base, "anthropic.claude-3-opus"):
	207	+	return false
	208	+	}
	209	+	return true
	210	+	}
	211	+
93	212		// bedrockListModelIDs enumerates foundation-model IDs and inference-
94	213		// profile IDs available in the region. Profile IDs are the geo-prefixed
95	214		// 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 +81,134 @@ 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	+	// bedrockAnthropicDefaultMaxOutput is the output budget assumed for a
			90	+	// discovered Claude model with no catalog entry. It matches Claude Haiku
			91	+	// 4.5, the smallest output limit among current Claude models, so it is
			92	+	// safe for any Claude 4+ model and far above the client's 4096 fallback.
			93	+	const bedrockAnthropicDefaultMaxOutput = 64000
			94	+
			95	+	// bedrockAnthropicDefaultContextWindow is the context window assumed
			96	+	// for an uncatalogued Claude model, the smallest among current models.
			97	+	const bedrockAnthropicDefaultContextWindow = 200000
			98	+
			99	+	// bedrockDiscoveredModel builds the Model for a discovered ID. The
			100	+	// control plane returns IDs only, so limits and capabilities come from
			101	+	// the static catalog. MergeCatalog already uses an exact catalog match;
			102	+	// this handles IDs without one, most often a geo-prefixed inference
			103	+	// profile (apac., us-gov., ...) whose base model is catalogued. Without
			104	+	// this the model would carry MaxOutput 0 and the Bedrock client would
			105	+	// cap every response at its 4096-token fallback.
			106	+	func bedrockDiscoveredModel(id, region string) Model {
			107	+	m := Model{
			108	+	Provider: "amazon-bedrock",
			109	+	ID: id,
			110	+	Source: "live",
			111	+	BaseURL: "https://bedrock-runtime." + region + ".amazonaws.com",
			112	+	}
			113	+	if bedrockInCatalog(id) {
			114	+	// MergeCatalog keeps the exact catalog entry's metadata.
			115	+	return m
			116	+	}
			117	+	prefix, base := bedrockSplitGeoPrefix(id)
			118	+	if c, ok := bedrockCatalogBase(base); ok {
			119	+	m.ContextWindow = c.ContextWindow
			120	+	m.MaxOutput = c.MaxOutput
			121	+	m.Reasoning = c.Reasoning
			122	+	m.ReasoningLevelMap = maps.Clone(c.ReasoningLevelMap)
			123	+	m.AdaptiveThinking = c.AdaptiveThinking
			124	+	m.PriceInput = c.PriceInput
			125	+	m.PriceOutput = c.PriceOutput
			126	+	m.PriceCacheRead = c.PriceCacheRead
			127	+	m.PriceCacheWrite = c.PriceCacheWrite
			128	+	m.PriceTierInputTokens = c.PriceTierInputTokens
			129	+	m.PriceInputAbove = c.PriceInputAbove
			130	+	m.PriceOutputAbove = c.PriceOutputAbove
			131	+	m.PriceCacheReadAbove = c.PriceCacheReadAbove
			132	+	m.PriceCacheWriteAbove = c.PriceCacheWriteAbove
			133	+	if c.DisplayName != "" {
			134	+	m.DisplayName = c.DisplayName
			135	+	if cp, _ := bedrockSplitGeoPrefix(c.ID); cp != "" {
			136	+	m.DisplayName = strings.TrimSuffix(m.DisplayName, " ("+strings.ToUpper(cp)+")")
			137	+	}
			138	+	if prefix != "" {
			139	+	m.DisplayName += " (" + strings.ToUpper(prefix) + ")"
			140	+	}
			141	+	}
			142	+	return m
			143	+	}
			144	+	if bedrockIsModernClaude(base) {
			145	+	m.ContextWindow = bedrockAnthropicDefaultContextWindow
			146	+	m.MaxOutput = bedrockAnthropicDefaultMaxOutput
			147	+	}
			148	+	return m
			149	+	}
			150	+
			151	+	// bedrockSplitGeoPrefix separates a cross-region inference-profile geo
			152	+	// prefix from the foundation-model ID. prefix is "" when there is none.
			153	+	func bedrockSplitGeoPrefix(id string) (prefix, base string) {
			154	+	for _, p := range bedrockGeoPrefixes {
			155	+	if strings.HasPrefix(id, p+".") {
			156	+	return p, id[len(p)+1:]
			157	+	}
			158	+	}
			159	+	return "", id
			160	+	}
			161	+
			162	+	func bedrockInCatalog(id string) bool {
			163	+	for _, c := range Catalog {
			164	+	if c.Provider == "amazon-bedrock" && c.ID == id {
			165	+	return true
			166	+	}
			167	+	}
			168	+	return false
			169	+	}
			170	+
			171	+	// bedrockCatalogBase finds catalog metadata for a foundation-model ID:
			172	+	// the bare entry if present, otherwise any geo-prefixed variant of it.
			173	+	func bedrockCatalogBase(base string) (Model, bool) {
			174	+	var variant Model
			175	+	found := false
			176	+	for _, c := range Catalog {
			177	+	if c.Provider != "amazon-bedrock" {
			178	+	continue
			179	+	}
			180	+	if c.ID == base {
			181	+	return c, true
			182	+	}
			183	+	if !found {
			184	+	if _, b := bedrockSplitGeoPrefix(c.ID); b == base {
			185	+	variant, found = c, true
			186	+	}
			187	+	}
			188	+	}
			189	+	return variant, found
			190	+	}
			191	+
			192	+	// bedrockIsModernClaude reports whether a foundation-model ID is a Claude
			193	+	// model whose output limit is at least bedrockAnthropicDefaultMaxOutput.
			194	+	// Legacy Claude 2 / Instant / 3 / 3.5 models cap output at 4096-8192 and
			195	+	// would reject a 64000-token request, so they keep the client fallback.
			196	+	func bedrockIsModernClaude(base string) bool {
			197	+	if !strings.HasPrefix(base, "anthropic.claude-") {
			198	+	return false
			199	+	}
			200	+	switch {
			201	+	case strings.HasPrefix(base, "anthropic.claude-v"),
			202	+	strings.HasPrefix(base, "anthropic.claude-instant"),
			203	+	strings.HasPrefix(base, "anthropic.claude-3-5-"),
			204	+	strings.HasPrefix(base, "anthropic.claude-3-haiku"),
			205	+	strings.HasPrefix(base, "anthropic.claude-3-sonnet"),
			206	+	strings.HasPrefix(base, "anthropic.claude-3-opus"):
			207	+	return false
			208	+	}
			209	+	return true
			210	+	}
			211	+
93		// bedrockListModelIDs enumerates foundation-model IDs and inference-	212		// bedrockListModelIDs enumerates foundation-model IDs and inference-
94		// profile IDs available in the region. Profile IDs are the geo-prefixed	213		// profile IDs available in the region. Profile IDs are the geo-prefixed
95		// forms zot invokes; both are returned so the picker shows every	214		// forms zot invokes; both are returned so the picker shows every

A packages/provider/discover_bedrock_metadata_test.go +103 -0

@@ -0,0 +1,103 @@
	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-5-5"
	55	+	m, wire := discoveredWireMaxTokens(t, id)
	56	+	if wire != 128000 \|\| !m.AdaptiveThinking {
	57	+	t.Errorf("%s: wire=%d adaptive=%v, want 128000 adaptive", id, wire, m.AdaptiveThinking)
	58	+	}
	59	+	if m.DisplayName != "Claude Opus 5.5 (APAC)" {
	60	+	t.Errorf("DisplayName = %q", m.DisplayName)
	61	+	}
	62	+	}
	63	+
	64	+	// Claude models absent from the catalog get Haiku 4.5's 64k output, the
	65	+	// smallest among current Claude models, instead of the 4096 fallback.
	66	+	func TestBedrockDiscoveredUnknownClaudeUses64kFloor(t *testing.T) {
	67	+	for _, id := range []string{"anthropic.claude-future-9", "us.anthropic.claude-future-9-v1:0"} {
	68	+	m, wire := discoveredWireMaxTokens(t, id)
	69	+	if wire != bedrockAnthropicDefaultMaxOutput \|\| m.ContextWindow != bedrockAnthropicDefaultContextWindow {
	70	+	t.Errorf("%s: wire=%d ctx=%d", id, wire, m.ContextWindow)
	71	+	}
	72	+	}
	73	+	}
	74	+
	75	+	// Legacy Claude and non-Anthropic models cap output well below 64k and
	76	+	// would reject a larger request, so they keep the client fallback.
	77	+	func TestBedrockDiscoveredLegacyAndOtherModelsKeepFallback(t *testing.T) {
	78	+	for _, id := range []string{
	79	+	"anthropic.claude-v2:1",
	80	+	"anthropic.claude-instant-v1",
	81	+	"us.anthropic.claude-3-5-sonnet-20241022-v2:0",
	82	+	"anthropic.claude-3-haiku-20240307-v1:0",
	83	+	"mistral.unknown-model-v1:0",
	84	+	} {
	85	+	m, wire := discoveredWireMaxTokens(t, id)
	86	+	if m.MaxOutput != 0 \|\| wire != 4096 {
	87	+	t.Errorf("%s: MaxOutput=%d wire=%d, want 0/4096", id, m.MaxOutput, wire)
	88	+	}
	89	+	}
	90	+	}
	91	+
	92	+	// Exact catalog IDs keep their own catalog metadata through MergeCatalog.
	93	+	func TestBedrockDiscoveredExactCatalogIDUnchanged(t *testing.T) {
	94	+	const id = "au.anthropic.claude-sonnet-4-6"
	95	+	want, err := FindModel("amazon-bedrock", id)
	96	+	if err != nil {
	97	+	t.Fatal(err)
	98	+	}
	99	+	m, wire := discoveredWireMaxTokens(t, id)
	100	+	if m.MaxOutput != want.MaxOutput \|\| wire != want.MaxOutput \|\| m.DisplayName != want.DisplayName {
	101	+	t.Errorf("got MaxOutput=%d wire=%d name=%q, want %d %q", m.MaxOutput, wire, m.DisplayName, want.MaxOutput, want.DisplayName)
	102	+	}
	103	+	}

@@ -0,0 +1,103 @@
	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-5-5"
	55	+	m, wire := discoveredWireMaxTokens(t, id)
	56	+	if wire != 128000 \|\| !m.AdaptiveThinking {
	57	+	t.Errorf("%s: wire=%d adaptive=%v, want 128000 adaptive", id, wire, m.AdaptiveThinking)
	58	+	}
	59	+	if m.DisplayName != "Claude Opus 5.5 (APAC)" {
	60	+	t.Errorf("DisplayName = %q", m.DisplayName)
	61	+	}
	62	+	}
	63	+
	64	+	// Claude models absent from the catalog get Haiku 4.5's 64k output, the
	65	+	// smallest among current Claude models, instead of the 4096 fallback.
	66	+	func TestBedrockDiscoveredUnknownClaudeUses64kFloor(t *testing.T) {
	67	+	for _, id := range []string{"anthropic.claude-future-9", "us.anthropic.claude-future-9-v1:0"} {
	68	+	m, wire := discoveredWireMaxTokens(t, id)
	69	+	if wire != bedrockAnthropicDefaultMaxOutput \|\| m.ContextWindow != bedrockAnthropicDefaultContextWindow {
	70	+	t.Errorf("%s: wire=%d ctx=%d", id, wire, m.ContextWindow)
	71	+	}
	72	+	}
	73	+	}
	74	+
	75	+	// Legacy Claude and non-Anthropic models cap output well below 64k and
	76	+	// would reject a larger request, so they keep the client fallback.
	77	+	func TestBedrockDiscoveredLegacyAndOtherModelsKeepFallback(t *testing.T) {
	78	+	for _, id := range []string{
	79	+	"anthropic.claude-v2:1",
	80	+	"anthropic.claude-instant-v1",
	81	+	"us.anthropic.claude-3-5-sonnet-20241022-v2:0",
	82	+	"anthropic.claude-3-haiku-20240307-v1:0",
	83	+	"mistral.unknown-model-v1:0",
	84	+	} {
	85	+	m, wire := discoveredWireMaxTokens(t, id)
	86	+	if m.MaxOutput != 0 \|\| wire != 4096 {
	87	+	t.Errorf("%s: MaxOutput=%d wire=%d, want 0/4096", id, m.MaxOutput, wire)
	88	+	}
	89	+	}
	90	+	}
	91	+
	92	+	// Exact catalog IDs keep their own catalog metadata through MergeCatalog.
	93	+	func TestBedrockDiscoveredExactCatalogIDUnchanged(t *testing.T) {
	94	+	const id = "au.anthropic.claude-sonnet-4-6"
	95	+	want, err := FindModel("amazon-bedrock", id)
	96	+	if err != nil {
	97	+	t.Fatal(err)
	98	+	}
	99	+	m, wire := discoveredWireMaxTokens(t, id)
	100	+	if m.MaxOutput != want.MaxOutput \|\| wire != want.MaxOutput \|\| m.DisplayName != want.DisplayName {
	101	+	t.Errorf("got MaxOutput=%d wire=%d name=%q, want %d %q", m.MaxOutput, wire, m.DisplayName, want.MaxOutput, want.DisplayName)
	102	+	}
	103	+	}