@linxiraos/pi-catalog 1.4.4 → 1.4.6

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (79) hide show
  1. package/CHANGELOG.md +8 -0
  2. package/dist/types/compat/anthropic.d.ts +20 -0
  3. package/dist/types/compat/behavior.d.ts +2 -0
  4. package/dist/types/compat/catalog-policy.d.ts +2 -2
  5. package/dist/types/compat/collapse.d.ts +10 -0
  6. package/dist/types/compat/factory-droid.d.ts +2 -0
  7. package/dist/types/compat/resolve.d.ts +2 -2
  8. package/dist/types/compat/types.d.ts +2 -0
  9. package/dist/types/discovery/antigravity.d.ts +7 -2
  10. package/dist/types/discovery/devin.d.ts +1 -1
  11. package/dist/types/provider-models/google.d.ts +2 -0
  12. package/dist/types/provider-models/openai-compat.d.ts +16 -3
  13. package/dist/types/types.d.ts +58 -3
  14. package/dist/types/wire/devin.d.ts +2 -2
  15. package/dist/types/wire/gemini-headers.d.ts +41 -22
  16. package/dist/types/wire/xiaomi-token-plan.d.ts +2 -0
  17. package/package.json +4 -4
  18. package/src/build.ts +47 -1
  19. package/src/compat/anthropic.d.ts +20 -0
  20. package/src/compat/anthropic.ts +47 -0
  21. package/src/compat/axes.ts +10 -3
  22. package/src/compat/behavior.d.ts +2 -0
  23. package/src/compat/behavior.ts +5 -0
  24. package/src/compat/catalog-policy.ts +10 -8
  25. package/src/compat/collapse.d.ts +10 -0
  26. package/src/compat/collapse.ts +63 -2
  27. package/src/compat/factory-droid.d.ts +2 -0
  28. package/src/compat/factory-droid.ts +3 -0
  29. package/src/compat/resolve.d.ts +2 -2
  30. package/src/compat/resolve.ts +7 -5
  31. package/src/compat/rules/README.md +13 -5
  32. package/src/compat/rules/auth/google-antigravity.kdl +2 -1
  33. package/src/compat/rules/auth/llama.cpp.kdl +1 -1
  34. package/src/compat/rules/auth/xiaomi-token-plan-ams.kdl +1 -1
  35. package/src/compat/rules/auth/xiaomi-token-plan-cn.kdl +1 -1
  36. package/src/compat/rules/auth/xiaomi-token-plan-sgp.kdl +1 -1
  37. package/src/compat/rules/classes/anthropic.kdl +6 -1
  38. package/src/compat/rules/classes/minimax.kdl +10 -0
  39. package/src/compat/rules/providers/anthropic.kdl +15 -1
  40. package/src/compat/rules/providers/deepseek.kdl +13 -10
  41. package/src/compat/rules/providers/factory-droid.kdl +48 -0
  42. package/src/compat/rules/providers/google-antigravity.kdl +7 -0
  43. package/src/compat/rules/providers/kimi-code.kdl +20 -0
  44. package/src/compat/rules/providers/moonshot.kdl +25 -0
  45. package/src/compat/rules/providers/muse-code.kdl +1 -1
  46. package/src/compat/rules/providers/openai-codex.kdl +17 -4
  47. package/src/compat/rules/providers/openai.kdl +23 -2
  48. package/src/compat/rules/providers/opencode-go.kdl +38 -3
  49. package/src/compat/rules/providers/opencode-zen.kdl +7 -0
  50. package/src/compat/rules/providers/xiaomi-token-plan-ams.kdl +1 -1
  51. package/src/compat/rules/providers/xiaomi-token-plan-cn.kdl +1 -1
  52. package/src/compat/rules/providers/xiaomi-token-plan-sgp.kdl +1 -1
  53. package/src/compat/rules/providers/xiaomi.kdl +1 -1
  54. package/src/compat/rules/runtime/behavior.kdl +30 -14
  55. package/src/compat/rules/taxonomy/kimi.kdl +7 -0
  56. package/src/compat/rules.json +1 -1
  57. package/src/compat/types.d.ts +2 -0
  58. package/src/compat/types.ts +2 -0
  59. package/src/discovery/antigravity.d.ts +7 -2
  60. package/src/discovery/antigravity.ts +41 -6
  61. package/src/discovery/cursor.ts +0 -6
  62. package/src/discovery/devin.d.ts +1 -1
  63. package/src/discovery/devin.ts +9 -3
  64. package/src/discovery/factory-droid.ts +38 -6
  65. package/src/model-cache.ts +4 -2
  66. package/src/models.json +1 -1
  67. package/src/provider-models/cache-provider-id.ts +5 -2
  68. package/src/provider-models/google.d.ts +2 -0
  69. package/src/provider-models/google.ts +13 -1
  70. package/src/provider-models/openai-compat.d.ts +16 -3
  71. package/src/provider-models/openai-compat.ts +127 -38
  72. package/src/types.d.ts +58 -3
  73. package/src/types.ts +63 -1
  74. package/src/wire/devin.d.ts +2 -2
  75. package/src/wire/devin.ts +2 -2
  76. package/src/wire/gemini-headers.d.ts +41 -22
  77. package/src/wire/gemini-headers.ts +68 -24
  78. package/src/wire/xiaomi-token-plan.d.ts +2 -0
  79. package/src/wire/xiaomi-token-plan.ts +4 -0
@@ -248,6 +248,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
248
248
  ]),
249
249
  "effort-beta": wire("effortBeta", ["anthropic"]),
250
250
  "escape-builtin-tool-names": wire("escapeBuiltinToolNames", ["anthropic"]),
251
+ "extended-cache-ttl-beta": wire("extendedCacheTtlBeta", ["anthropic"]),
251
252
  "fast-mode": wire("fastMode", ["anthropic"]),
252
253
  "first-party-provider": wire("firstPartyProvider", ["anthropic"]),
253
254
  "honors-claude-code-beta-kill-switch": wire("honorsClaudeCodeBetaKillSwitch", ["anthropic"]),
@@ -257,6 +258,7 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
257
258
  "official-endpoint": wire("officialEndpoint", ["anthropic", "openai-responses"]),
258
259
  "replay-unsigned-thinking": wire("replayUnsignedThinking", ["anthropic"]),
259
260
  "requires-thinking-enabled": wire("requiresThinkingEnabled", ["anthropic"]),
261
+ "requires-thinking-presence": wire("requiresThinkingPresence", ["anthropic"]),
260
262
  "requires-tool-result-id": wire("requiresToolResultId", ["anthropic"]),
261
263
  "signing-endpoint": wire("signingEndpoint", ["anthropic"]),
262
264
  "strip-thinking-history": wire("stripThinkingHistory", ["openai", "anthropic", "ollama"]),
@@ -544,12 +546,17 @@ export const AXES: Readonly<Record<string, AxisDef>> = {
544
546
  "limits-patch": { key: "limitsPatch", set: "catalog", shape: "object" },
545
547
  /** Seeds `maxTokens` only when the row reports none; applied after `limits-patch`. */
546
548
  "max-tokens-fallback": { key: "maxTokensFallback", set: "catalog", shape: "scalar" },
547
- /** Most output tokens the host serves: a larger reported `maxTokens` is lowered to it; an unknown one stays unknown. */
549
+ /** Most output tokens the host serves: a larger reported `maxTokens` (and Antigravity per-SKU `maxOutputTokens`) is lowered to it; an unknown one stays unknown. */
548
550
  "max-tokens-ceiling": { key: "maxTokensCeiling", set: "catalog", shape: "scalar" },
549
551
  "long-context-cost": { key: "longContext", set: "catalog", shape: "object" },
550
552
  "prompt-cache": { key: "promptCache", set: "catalog", shape: "object" },
551
- "task-sibling-prompt-cache": {
552
- key: "taskSiblingPromptCache",
553
+ /**
554
+ * The lane rewards routing related requests — Task siblings, forks, explicit
555
+ * `--prompt-cache-key` callers — under one shared prompt-cache key instead of
556
+ * each conversation's own identity. Unassigned: per-conversation routing.
557
+ */
558
+ "shared-prompt-cache": {
559
+ key: "sharedPromptCache",
553
560
  set: "catalog",
554
561
  shape: "scalar",
555
562
  values: [true, false],
@@ -71,6 +71,8 @@ export declare function isRetiredProvider(provider: string): boolean;
71
71
  * model that happens to advertise the same bare id.
72
72
  */
73
73
  export declare function isBareIdReferenceProvider(provider: string): boolean;
74
+ /** Whether a host's prefix cache should use append-only conversation context by default. */
75
+ export declare function prefersAppendOnlyContext(provider: string): boolean;
74
76
  /** The declared subscription tier required to use a provider model id, if any. */
75
77
  export declare function planRequirementFor(provider: string, model: string): string | undefined;
76
78
  /** Cross-provider pricing-peer resolution for one provider model id. */
@@ -204,6 +204,11 @@ export function isBareIdReferenceProvider(provider: string): boolean {
204
204
  return !behavior.referenceIsolatedProviders.includes(provider);
205
205
  }
206
206
 
207
+ /** Whether a host's prefix cache should use append-only conversation context by default. */
208
+ export function prefersAppendOnlyContext(provider: string): boolean {
209
+ return behavior.appendOnlyContextProviders.includes(provider);
210
+ }
211
+
207
212
  /** The declared subscription tier required to use a provider model id, if any. */
208
213
  export function planRequirementFor(provider: string, model: string): string | undefined {
209
214
  const rule = behavior.planRequirements.find(candidate => candidate.provider === provider);
@@ -1,13 +1,15 @@
1
- import { toModelSpec } from "../provider-models/bundled-references";
2
- import type { Model } from "../types";
1
+ import { withoutUnverifiedThinking } from "../model-thinking";
2
+ import type { CatalogPolicyInputs } from "../types";
3
3
  import { resolveCatalogAxes } from "./resolve";
4
4
 
5
5
  /**
6
6
  * Rule-owned catalog-axis policy by model. Resolve once per process rather
7
7
  * than walking the static policy cascade on every accessor call. The key
8
- * mirrors the cascade target inputs — `providerType ?? provider`, api, model
9
- * id, and the reasoning flag; identity-derived class/family/revision facts are
10
- * a pure function of provider + id. Bounded: one entry per distinct model.
8
+ * mirrors the cascade target inputs — `providerType ?? provider`, api, the
9
+ * logical id plus `requestModelId` (the wire id `policyModelId` resolves
10
+ * unknown logical ids under), and the reasoning flag; identity-derived
11
+ * class/family/revision facts are a pure function of provider + id. Bounded:
12
+ * one entry per distinct model.
11
13
  */
12
14
  const catalogPolicyCache = new Map<string, Readonly<Record<string, unknown>>>();
13
15
  const CATALOG_POLICY_CACHE_MAX = 8192;
@@ -17,11 +19,11 @@ const CATALOG_POLICY_CACHE_MAX = 8192;
17
19
  * memoized per distinct model. Frozen: the bag is shared across callers, so it
18
20
  * must stay read-only — extract the value you need rather than handing it on.
19
21
  */
20
- export function resolveCatalogPolicy(model: Model): Readonly<Record<string, unknown>> {
21
- const key = `${model.provider}\u0000${model.providerType ?? ""}\u0000${model.api}\u0000${model.id}\u0000${model.reasoning ? 1 : 0}`;
22
+ export function resolveCatalogPolicy(model: CatalogPolicyInputs): Readonly<Record<string, unknown>> {
23
+ const key = `${model.provider}\u0000${model.providerType ?? ""}\u0000${model.api}\u0000${model.id}\u0000${model.requestModelId ?? ""}\u0000${model.reasoning ? 1 : 0}`;
22
24
  const cached = catalogPolicyCache.get(key);
23
25
  if (cached !== undefined) return cached;
24
- const policy = Object.freeze({ ...resolveCatalogAxes(toModelSpec(model)) });
26
+ const policy = Object.freeze({ ...resolveCatalogAxes(withoutUnverifiedThinking(model)) });
25
27
  if (catalogPolicyCache.size >= CATALOG_POLICY_CACHE_MAX) catalogPolicyCache.clear();
26
28
  catalogPolicyCache.set(key, policy);
27
29
  return policy;
@@ -139,6 +139,16 @@ export declare function isCollapsedVariantSpec(spec: VariantSpecLike): boolean;
139
139
  * and for routes live discovery never advertised.
140
140
  */
141
141
  export declare function isCursorMaxModeWireId(wireModelId: string): boolean;
142
+ /**
143
+ * Whether `wireModelId` is one tier of a reviewed `provider` family that serves
144
+ * different efforts from distinct sibling SKUs (`-low`/`-medium`/`-high`), so
145
+ * the SKU, not the effort, fixes the thinking strength. Decided from the
146
+ * declared family routes, not a collapsed row's routing: an account roster
147
+ * listing one tier collapses to a row routing every effort to it, yet that SKU
148
+ * still advertises its own fixed thinking. Families routing every effort to
149
+ * one `-thinking` SKU keep the effort's own budget or level.
150
+ */
151
+ export declare function isReviewedTierVariantWireId(provider: string, wireModelId: string): boolean;
142
152
  /**
143
153
  * Collapse a full mixed-provider list: per provider, the compiled reviewed table, Cursor's
144
154
  * conservative live effort-sibling rule, and the automatic `X`/`X-thinking`
@@ -36,7 +36,7 @@
36
36
  */
37
37
  import { buildModel } from "../build";
38
38
  import { Effort, THINKING_EFFORTS } from "../effort";
39
- import type { Api, Model, ModelSpec, Provider, ThinkingConfig } from "../types";
39
+ import type { AntigravityModelWireProfile, Api, Model, ModelSpec, Provider, ThinkingConfig } from "../types";
40
40
  import { resolveModelPolicy } from "./resolve";
41
41
  import { parseRevision, parseRevisionConstraint, type RevisionTerm, revisionSatisfies } from "./revision";
42
42
  import { collapseVariantId, collapseVocabulary, stripThinkingVariantSuffix } from "./taxonomy";
@@ -759,6 +759,31 @@ export function isCursorMaxModeWireId(wireModelId: string): boolean {
759
759
  return effort === Effort.XHigh || effort === Effort.Max;
760
760
  }
761
761
 
762
+ /**
763
+ * Whether `wireModelId` is one tier of a reviewed `provider` family that serves
764
+ * different efforts from distinct sibling SKUs (`-low`/`-medium`/`-high`), so
765
+ * the SKU, not the effort, fixes the thinking strength. Decided from the
766
+ * declared family routes, not a collapsed row's routing: an account roster
767
+ * listing one tier collapses to a row routing every effort to it, yet that SKU
768
+ * still advertises its own fixed thinking. Families routing every effort to
769
+ * one `-thinking` SKU keep the effort's own budget or level.
770
+ */
771
+ export function isReviewedTierVariantWireId(provider: string, wireModelId: string): boolean {
772
+ const table = reviewedCollapseTable(provider);
773
+ if (table === undefined) return false;
774
+ const id = wireModelId.trim();
775
+ const family =
776
+ table.families.find(candidate => candidate.members.includes(id)) ??
777
+ (table.templates === undefined ? undefined : templateFamilyFor(table.templates, id));
778
+ if (family === undefined) return false;
779
+ const tiers = new Set<string>();
780
+ for (const key of VARIANT_ROUTING_KEYS) {
781
+ const target = family.routing[key];
782
+ if (key !== "off" && target !== undefined) tiers.add(target);
783
+ }
784
+ return tiers.size > 1 && tiers.has(id);
785
+ }
786
+
762
787
  /**
763
788
  * Recover Cursor's max-mode marker for a bundled collapsed row. The bundled
764
789
  * snapshot may contain only the logical row, but its effort routing still
@@ -821,6 +846,37 @@ function reconcileCursorMaxMode<TSpec extends VariantSpecLike>(spec: TSpec, memb
821
846
  };
822
847
  }
823
848
 
849
+ /**
850
+ * Union the members' discovered Antigravity wire profiles, keyed by each
851
+ * member's own wire id, so the collapsed row keeps the request constants of
852
+ * every SKU it routes to (`...firstMember` alone keeps only the first tier's).
853
+ * Returns `undefined` when no member carries a profile.
854
+ */
855
+ function antigravityWireProfilesOf<TSpec extends VariantSpecLike>(
856
+ memberSpecs: readonly TSpec[],
857
+ ): Record<string, AntigravityModelWireProfile> | undefined {
858
+ let profiles: Record<string, AntigravityModelWireProfile> | undefined;
859
+ for (const member of memberSpecs) {
860
+ if (member.antigravityWireProfiles === undefined) continue;
861
+ profiles = { ...profiles, ...member.antigravityWireProfiles };
862
+ }
863
+ return profiles;
864
+ }
865
+
866
+ /**
867
+ * Merge live members' wire profiles over an already-collapsed snapshot, which
868
+ * froze them (or none) when it was written. Returns `spec` by reference when no
869
+ * member carries a profile.
870
+ */
871
+ function reconcileAntigravityWireProfiles<TSpec extends VariantSpecLike>(
872
+ spec: TSpec,
873
+ memberSpecs: readonly TSpec[],
874
+ ): TSpec {
875
+ const profiles = antigravityWireProfilesOf(memberSpecs);
876
+ if (profiles === undefined) return spec;
877
+ return { ...spec, antigravityWireProfiles: { ...spec.antigravityWireProfiles, ...profiles } };
878
+ }
879
+
824
880
  /**
825
881
  * Collapse every family in `table` found in `specs`. Non-member specs pass
826
882
  * through verbatim (by reference), order preserved; the collapsed spec
@@ -892,7 +948,10 @@ function collapseWithTable<TSpec extends VariantSpecLike>(
892
948
  // lowest tier.
893
949
  replacement.set(
894
950
  family.id,
895
- reconcileCursorMaxMode(reconcileDefaultMember(reconciled, family, new Set(rawPresent)), memberSpecs),
951
+ reconcileAntigravityWireProfiles(
952
+ reconcileCursorMaxMode(reconcileDefaultMember(reconciled, family, new Set(rawPresent)), memberSpecs),
953
+ memberSpecs,
954
+ ),
896
955
  );
897
956
  continue;
898
957
  }
@@ -939,6 +998,7 @@ function collapseWithTable<TSpec extends VariantSpecLike>(
939
998
  // which tier needed it, so the members' own markers are kept per wire id.
940
999
  const cursorMaxMode = memberSpecs.some(spec => spec.cursorMaxMode === true) ? true : undefined;
941
1000
  const cursorMaxModeRoutes = cursorMaxModeRoutesOf(firstMember.provider, memberSpecs);
1001
+ const antigravityWireProfiles = antigravityWireProfilesOf(memberSpecs);
942
1002
  const collapsed: TSpec = {
943
1003
  ...firstMember,
944
1004
  id: family.id,
@@ -949,6 +1009,7 @@ function collapseWithTable<TSpec extends VariantSpecLike>(
949
1009
  maxTokens: maxOrNull(memberSpecs.map(spec => spec.maxTokens)),
950
1010
  ...(cursorMaxMode === undefined ? {} : { cursorMaxMode }),
951
1011
  ...(cursorMaxModeRoutes === undefined ? {} : { cursorMaxModeRoutes }),
1012
+ ...(antigravityWireProfiles === undefined ? {} : { antigravityWireProfiles }),
952
1013
  };
953
1014
  if (memberSpecs.some(spec => spec.isProviderDefault === true)) {
954
1015
  collapsed.isProviderDefault = true;
@@ -45,6 +45,8 @@ export interface FactoryDroidModelPolicy {
45
45
  entitlement: FactoryDroidEntitlement;
46
46
  /** Native default with no caller effort is thinking off. */
47
47
  defaultReasoningOff: boolean;
48
+ /** A `generate_image` model: gated by the org's image-generation settings, served through Factory's image routes. */
49
+ image: boolean;
48
50
  }
49
51
  /** One registry model: its roster row and resolved policy. */
50
52
  export interface FactoryDroidRegistryModel {
@@ -59,6 +59,8 @@ export interface FactoryDroidModelPolicy {
59
59
  entitlement: FactoryDroidEntitlement;
60
60
  /** Native default with no caller effort is thinking off. */
61
61
  defaultReasoningOff: boolean;
62
+ /** A `generate_image` model: gated by the org's image-generation settings, served through Factory's image routes. */
63
+ image: boolean;
62
64
  }
63
65
 
64
66
  /** One registry model: its roster row and resolved policy. */
@@ -128,6 +130,7 @@ function readPolicy(spec: ModelSpec<"factory-droid-agent">): FactoryDroidModelPo
128
130
  ...(baseVariant !== undefined && { baseVariant }),
129
131
  },
130
132
  defaultReasoningOff: catalog.defaultReasoningOff === true,
133
+ image: catalog.kind === "image",
131
134
  };
132
135
  }
133
136
 
@@ -1,4 +1,4 @@
1
- import type { Api, CompatOf, ModelSpec, ThinkingConfig } from "../types";
1
+ import type { Api, CatalogPolicyInputs, CompatOf, ModelSpec, ThinkingConfig } from "../types";
2
2
  import type { ModelIdentity, RequestPolicy } from "./types";
3
3
  /** Result of resolving one model spec through the compat engine. */
4
4
  export interface ResolvedModelPolicy<TApi extends Api = Api> {
@@ -21,7 +21,7 @@ export declare function resolveDiscoveryApi(spec: ModelSpec<Api>, providerType:
21
21
  * {@link resolveModelPolicy} without resolving the compat and thinking
22
22
  * policies, which dominate its cost on catalog-wide scans.
23
23
  */
24
- export declare function resolveCatalogAxes(spec: ModelSpec<Api>): Record<string, unknown>;
24
+ export declare function resolveCatalogAxes(spec: CatalogPolicyInputs): Record<string, unknown>;
25
25
  /**
26
26
  * Resolves the full policy surface for one model spec: structured identity,
27
27
  * complete compat record, thinking metadata, and catalog-data corrections.
@@ -20,6 +20,7 @@ import { hostMatchesUrl, modelMatchesHost } from "../hosts";
20
20
  import { withoutUnverifiedThinking } from "../model-thinking";
21
21
  import type {
22
22
  Api,
23
+ CatalogPolicyInputs,
23
24
  CompatOf,
24
25
  ModelSpec,
25
26
  OpenAICompat,
@@ -126,14 +127,14 @@ class IdentityFacts {
126
127
  * pairing runs only its lead, so the lead's reviewed rules (cache-write rate,
127
128
  * long-context tier) apply to it at build time.
128
129
  */
129
- function policyModelId<TApi extends Api>(spec: ModelSpec<TApi>): string {
130
+ function policyModelId(spec: CatalogPolicyInputs): string {
130
131
  const wireId = spec.requestModelId;
131
132
  if (wireId === undefined || wireId === spec.id) return spec.id;
132
133
  if (classifyModel(spec.provider, spec.id).class !== "unknown") return spec.id;
133
134
  return classifyModel(spec.provider, wireId).class === "unknown" ? spec.id : wireId;
134
135
  }
135
136
 
136
- function resolveIdentity<TApi extends Api>(spec: ModelSpec<TApi>): ModelIdentity {
137
+ function resolveIdentity(spec: CatalogPolicyInputs): ModelIdentity {
137
138
  // Strict on purpose: ambiguous identity is a rule-authoring defect surfaced
138
139
  // at build/CI time, never silently degraded to `unknown`. Discovery
139
140
  // normalization opts into leniency through `classifyModel` directly.
@@ -924,6 +925,7 @@ function resolveAnthropicPolicy(
924
925
  supportsSamplingParams: true,
925
926
  requiresToolResultId: false,
926
927
  requiresThinkingEnabled,
928
+ requiresThinkingPresence: false,
927
929
  replayUnsignedThinking:
928
930
  !signingEndpoint && (compatReasoning(spec, axes) || modelMatchesHost(spec, "deepseekFamily")),
929
931
  escapeBuiltinToolNames: false,
@@ -1305,8 +1307,8 @@ function fillExplicitThinking<TApi extends Api>(
1305
1307
  // Entry
1306
1308
  // ---------------------------------------------------------------------------
1307
1309
 
1308
- function buildResolveTarget<TApi extends Api>(
1309
- spec: ModelSpec<TApi>,
1310
+ function buildResolveTarget(
1311
+ spec: CatalogPolicyInputs,
1310
1312
  identity: ModelIdentity,
1311
1313
  providerType = spec.providerType ?? spec.provider,
1312
1314
  ): ResolveTarget {
@@ -1338,7 +1340,7 @@ export function resolveDiscoveryApi(spec: ModelSpec<Api>, providerType: string):
1338
1340
  * {@link resolveModelPolicy} without resolving the compat and thinking
1339
1341
  * policies, which dominate its cost on catalog-wide scans.
1340
1342
  */
1341
- export function resolveCatalogAxes(spec: ModelSpec<Api>): Record<string, unknown> {
1343
+ export function resolveCatalogAxes(spec: CatalogPolicyInputs): Record<string, unknown> {
1342
1344
  return resolveCascade(buildResolveTarget(spec, resolveIdentity(spec))).catalog;
1343
1345
  }
1344
1346
 
@@ -223,15 +223,15 @@ One object axis carries a computed form: `long-context-cost` accepts either the
223
223
 
224
224
  `long-context-window-uncapped #true` keeps a model's `long-context-cost` tier from capping its runtime context window while extended context is off: the coding agent normally lowers the window to the tier's input threshold so compaction fires before a request crosses into premium pricing. Use it where the tier is an API-equivalent estimate on a subscription host (`xai-oauth`) or where the model is served at its full window regardless (GPT-5.4, except on GitHub Copilot). `buildModel` materializes it as `Model.longContextWindowUncapped`; explicit user context overrides still apply.
225
225
 
226
- `max-tokens-fallback` seeds `maxTokens` only when the row reports none, and `max-tokens-ceiling` lowers a larger reported `maxTokens` to the host's output cap (an unknown budget stays unknown unless a fallback seeded it). Both apply in `buildModel` after `limits-patch`, so discovery mappers pass upstream output budgets through verbatim. See the Kimi rules in [`providers/fireworks.kdl`](providers/fireworks.kdl).
226
+ `max-tokens-fallback` seeds `maxTokens` only when the row reports none, and `max-tokens-ceiling` lowers a larger reported `maxTokens` to the host's output cap (an unknown budget stays unknown unless a fallback seeded it); the ceiling also lowers discovered Antigravity per-SKU `maxOutputTokens` (`antigravityWireProfiles`, row and per-account), which the transport sends verbatim. Both apply in `buildModel` after `limits-patch`, so discovery mappers pass upstream output budgets through verbatim. See the Kimi rules in [`providers/fireworks.kdl`](providers/fireworks.kdl).
227
227
 
228
- `inline-image-byte-budget` caps retained base64 image characters for a deployment's request body. It leaves room for tool schemas and text beneath the provider's body-size limit; the coding agent reads it through `resolveInlineImageByteBudget` before sending live history, and snapcompact gives its archive frames half of it (`frameDataBytesBudget`). The budget applies only while the request reaches the host the rules describe (the Anthropic API, first-party OpenAI, the Codex backend at its default URL, or Factory Droid's own global or EU API host), which callers decide with pi-ai's `reachesBudgetedHost`: `compat.officialEndpoint` for the configured `baseUrl`, for first-party Anthropic models also the `ANTHROPIC_BASE_URL` or Foundry `FOUNDRY_BASE_URL` reroute applied at request time, and for Factory Droid models the `baseUrl` origin. A custom or rerouted endpoint imposes its own body limit.
228
+ `inline-image-byte-budget` caps retained base64 image characters for a deployment's request body. It leaves room for tool schemas and text beneath the provider's body-size limit; the coding agent reads it through `resolveInlineImageByteBudget` before sending live history, and snapcompact gives its archive frames half of it (`frameDataBytesBudget`). The budget applies only while the request reaches the host the rules describe (the Anthropic API, first-party OpenAI, the Codex backend at its default URL, DeepSeek's API, or Factory Droid's own global or EU API host), which callers decide with pi-ai's `reachesBudgetedHost`: `compat.officialEndpoint` for the configured `baseUrl`, for first-party Anthropic models also the `ANTHROPIC_BASE_URL` or Foundry `FOUNDRY_BASE_URL` reroute applied at request time, and for Factory Droid models the `baseUrl` origin. A custom or rerouted endpoint imposes its own body limit.
229
229
 
230
230
  `request-body-byte-limit` is the largest request body the host accepts (Factory Droid's hosted routes: 4,500,000 bytes), under the same gate. Where it is set, the byte clamp keeps the inline images of the newest user message and then of the newest image-bearing tool result even past `inline-image-byte-budget`, each only while the kept images fit the limit less the request's other serialized bytes, and evicts older images instead; read through `resolveRequestBodyByteLimit`.
231
231
 
232
- `catalog-endpoint-budgets #true` makes these host budgets apply whenever a request reaches the provider's own catalog endpoint (an unset `baseUrl`, or an `endpoint-hosts` host; see `isCatalogEndpointUrl`), for a hosted gateway such as Factory Droid whose compat names no official endpoint. Unassigned, they apply only on the official endpoint.
232
+ `catalog-endpoint-budgets #true` makes these host budgets apply whenever a request reaches the provider's own catalog endpoint (an unset `baseUrl`, or an `endpoint-hosts` host; see `isCatalogEndpointUrl`), for a hosted gateway such as Factory Droid whose compat names no official endpoint, or a first-party provider such as DeepSeek whose `openai-completions` compat resolves no `officialEndpoint`. Unassigned, they apply only on the official endpoint.
233
233
 
234
- `image-count-budget` is the most images one request may carry, archive frames, screenshots and imaged tool results combined. The coding agent reads it through `resolveImageCountBudget`, drops the oldest images past it, and caps snapcompact frames at it; unassigned hosts get a floor of 5. `official-image-count-budget` replaces it while the request reaches the host the rules describe (same gate as `inline-image-byte-budget`), for limits established on the host's own endpoint (Anthropic: 600, 100 on 200k-window models; Codex: 2,000); a custom or rerouted endpoint keeps `image-count-budget` (Anthropic 90, Codex 200).
234
+ `image-count-budget` is the most images one request may carry, archive frames, screenshots and imaged tool results combined. The coding agent reads it through `resolveImageCountBudget`, drops the oldest images past it, and caps snapcompact frames at it; unassigned hosts get a floor of 5. `official-image-count-budget` replaces it while the request reaches the host the rules describe (same gate as `inline-image-byte-budget`), for limits established on the host's own endpoint (Anthropic: 600, 100 on 200k-window models; Codex: 2,000; DeepSeek: 600); a custom or rerouted endpoint keeps `image-count-budget` (Anthropic 90, Codex 200, DeepSeek 5).
235
235
 
236
236
  `endpoint-hosts` lists the HTTPS hostnames a provider serves besides its bundled base URL's host, such as Copilot's per-plan `api.<plan>.githubcopilot.com` hosts. That host plus these define the provider's own endpoint (`isCatalogEndpointUrl`), which separates first-party traffic from a custom model or provider override that keeps a catalog provider id but points `baseUrl` at a gateway, proxy, or local server. Connection prewarming and request compression act only on the provider's own endpoint.
237
237
 
@@ -245,7 +245,12 @@ One object axis carries a computed form: `long-context-cost` accepts either the
245
245
 
246
246
  `prompt-cache-persistent #true` marks a provider whose own endpoint keeps prompt-cache entries indefinitely (logged-in Codex accounts). The coding agent reads it through `isPromptCachePersistent` and skips idle compaction, which would only discard a prefix that stays warm; a custom `baseUrl` reusing the provider id does not qualify. Unassigned means entries expire.
247
247
 
248
- `task-sibling-prompt-cache #true` lets Task children use an automatically derived sibling cache key on the matching request model. Missing or `#false` keeps the provider's per-child default; explicit cache keys still win. The session resolves this catalog policy against the actual model for primary turns, handoff, and ephemeral requests that reuse the live context, including model changes, without sharing conversation IDs or response state. The automatic candidate is persisted separately from explicit keys and survives full forks, `/tan`, and session reopening; new and partial sessions discard it. A fork carrying an automatic candidate uses its own conversation ID on unsupported models, and re-enables the candidate when returning to an eligible model. Ordinary forks without an automatic candidate keep their existing inherited cache routing. Conventional compaction keeps its existing cache routing because its serialized summary prompt differs from the primary context. The opt-in covers Muse Code's `muse-spark-1.3`, `muse-spark-1.3-contributor`, `muse-wonder-1.0`, `muse-wonder-1.0-contributor`, `muse-wonder-full-1.0`, and `muse-wonder-full-1.0-contributor` over `openai-responses`. Meta documents shared-prefix `prompt_cache_key` routing for Responses. Warmed sibling cache reuse was measured on Spark 1.3 Contributor; the standard Spark variant shares this scope following maintainer guidance. A bounded synthetic probe observed cache reuse on all four Wonder variants; it does not establish production or latency benefits. Auxiliary-request latency has not been measured. Provider selectors also match custom providers that declare the same `providerType`; no benefit is established for custom endpoints.
248
+ `shared-prompt-cache #true` marks a lane that rewards routing related requests under one shared prompt-cache key instead of each conversation's own identity. Missing or `#false` keeps per-conversation routing. Two consumers read it against the model each request actually sends (primary turns, handoff, side requests, and remote compaction under its `remoteCompaction.api`/`model` override):
249
+
250
+ - The coding agent gives a Task child the derived sibling key `<parent cache key>:<agent>` only on opted-in models; an explicit key (`--prompt-cache-key`, an SDK key, or a full fork's inherited key) always wins. The candidate is persisted on the session header at its first eligible request and survives full forks, `/tan`, and reopening; new and partial sessions discard it.
251
+ - OpenCode Go (`providerType ?? provider` is `opencode-go`) sends any explicit prompt-cache key as `x-opencode-session`, so siblings, forks, and their parent share one gateway cache. Without an explicit key the header keeps the conversation id. Credentials and response state stay per conversation.
252
+
253
+ Opted in: Muse Code `muse-spark-1.3`, `muse-spark-1.3-contributor`, and the four Wonder variants over `openai-responses` (Meta documents shared-prefix `prompt_cache_key` routing; sibling reuse measured on Spark 1.3 Contributor). OpenCode Go `muse-spark-1.3-contributor`, `muse-spark-1.2-contributor`, `gpt-5.6-luna`, `gpt-6-luna`, and `grok-4.6` over `openai-responses`, and `deepseek-v4-pro` and `omen-alpha` over `openai-completions`. A Go lane qualifies only when a cold / different-family / same-family probe warms on the repeated key and stays cold under the different key; [`providers/opencode-go.kdl`](providers/opencode-go.kdl) records the per-lane tallies and the lanes that warmed under any key (and so gain nothing). The standard Go `muse-spark-1.3`/`1.2` ids stay off: the gateway answers "Model is unavailable". Custom providers opt in only by declaring a matching `providerType`.
249
254
 
250
255
  `image-max-edge` is the longest image edge, in pixels, a host passes to the model; it downscales larger images to fit first (Cursor: 1024 for Claude and GPT). Snapcompact reads it through `resolveImageMaxEdge` and renders frames no wider, so their glyphs reach the model at the size they were drawn; unassigned means the host keeps the image's size.
251
256
 
@@ -322,6 +327,7 @@ behavior {
322
327
  exclude-substring "embedding"
323
328
  }
324
329
  model-operations provider="openai" { exact "o3"; prefix "gpt-"; operation "generate_image" }
330
+ append-only-context-providers "deepseek" "ollama" "google-antigravity" "google" "google-vertex"
325
331
  cursor-effort family-marker="gpt-" { tier "minimal" "low" "medium" "high" "xhigh" "max" }
326
332
  cursor-model-parameter model="composer-2.5" id="fast" value="false"
327
333
  quota-tiers provider="openai-codex" {
@@ -350,6 +356,8 @@ behavior {
350
356
 
351
357
  `exclude-discovery-modes` takes one or more exact, case-sensitive upstream mode strings plus `provider=`; discovery mappers preserve missing, malformed, and unknown modes unless the provider policy explicitly lists them. Matcher properties on `route` / `exclude-models` / `tier` nodes are `exact=` / `prefix=` / `substring=` / `glob=`, repeatable. `strip-prefix=#true` on a prefix route strips the matched prefix off the wire id. Values are copied verbatim from the TS constants they replaced; runtime accessors live in `src/compat/behavior.ts`.
352
358
 
359
+ `append-only-context-providers` names hosts whose prefix caches benefit from append-only message history. The coding agent uses this list for `provider.appendOnlyContext: auto`; user `on`/`off` always takes precedence, and URL/explicit stored-request signals remain separate.
360
+
353
361
  `role-defaults "<role>"` declares one agent role's built-in default model selector chain, one `model "<selector>"` child per entry, most preferred first. Selectors are copied verbatim (`provider/id`, bare ids, fuzzy fragments such as `-mini`, and `@role` references); duplicates within a role and repeated roles are compile errors. `roleDefaultModels()` returns every chain keyed by role in declaration order.
354
362
 
355
363
  ## Auth grammar
@@ -7,7 +7,8 @@ auth "google-antigravity" {
7
7
  authorize-url "https://accounts.google.com/o/oauth2/v2/auth"
8
8
  scopes "https://www.googleapis.com/auth/cloud-platform" "https://www.googleapis.com/auth/userinfo.email" \
9
9
  "https://www.googleapis.com/auth/userinfo.profile" "https://www.googleapis.com/auth/cclog" \
10
- "https://www.googleapis.com/auth/experimentsandconfigs"
10
+ "https://www.googleapis.com/auth/experimentsandconfigs" "https://www.googleapis.com/auth/aicode"
11
+ pkce #true
11
12
  authorize-params { access_type "offline"; prompt "consent" }
12
13
  instructions "Complete the sign-in in your browser."
13
14
  callback port=51121 path="/oauth-callback" hostname="127.0.0.1"
@@ -3,7 +3,7 @@ auth "llama.cpp" {
3
3
  env "LLAMA_CPP_API_KEY"
4
4
  login "api-key" {
5
5
  auth-url "https://github.com/ggml-org/llama.cpp#quick-start"
6
- instructions "Paste your llama.cpp API key if your server requires auth. Leave empty for local no-auth mode (default base URL: http://127.0.0.1:8080; set LLAMA_CPP_BASE_URL to customize)."
6
+ instructions "Paste your llama.cpp API key if your server requires auth. Leave empty for local no-auth mode (default base URLs: http://127.0.0.1:8080, then http://127.0.0.1:9931; set LLAMA_CPP_BASE_URL to customize)."
7
7
  prompt "Paste your llama.cpp API key (optional for local no-auth)" placeholder="llama-cpp-local"
8
8
  empty-fallback "llama-cpp-local"
9
9
  }
@@ -4,6 +4,6 @@ auth "xiaomi-token-plan-ams" {
4
4
  auth-url "https://platform.xiaomimimo.com/console/plan-manage"
5
5
  instructions "Copy your token-plan API key for the Europe region"
6
6
  prompt "Paste your Xiaomi Token Plan Europe API key (tp-...)" placeholder="tp-..."
7
- validate "chat-completions" label="xiaomi" base-url="https://token-plan-ams.xiaomimimo.com/v1" model="mimo-v2.5"
7
+ validate "chat-completions" label="xiaomi" base-url="https://token-plan-ams.xiaomimimo.com/v1" model="mimo-v2.6-flash"
8
8
  }
9
9
  }
@@ -4,6 +4,6 @@ auth "xiaomi-token-plan-cn" {
4
4
  auth-url "https://platform.xiaomimimo.com/console/plan-manage"
5
5
  instructions "Copy your token-plan API key for the China region"
6
6
  prompt "Paste your Xiaomi Token Plan China API key (tp-...)" placeholder="tp-..."
7
- validate "chat-completions" label="xiaomi" base-url="https://token-plan-cn.xiaomimimo.com/v1" model="mimo-v2.5"
7
+ validate "chat-completions" label="xiaomi" base-url="https://token-plan-cn.xiaomimimo.com/v1" model="mimo-v2.6-flash"
8
8
  }
9
9
  }
@@ -4,6 +4,6 @@ auth "xiaomi-token-plan-sgp" {
4
4
  auth-url "https://platform.xiaomimimo.com/console/plan-manage"
5
5
  instructions "Copy your token-plan API key for the Singapore region"
6
6
  prompt "Paste your Xiaomi Token Plan Singapore API key (tp-...)" placeholder="tp-..."
7
- validate "chat-completions" label="xiaomi" base-url="https://token-plan-sgp.xiaomimimo.com/v1" model="mimo-v2.5"
7
+ validate "chat-completions" label="xiaomi" base-url="https://token-plan-sgp.xiaomimimo.com/v1" model="mimo-v2.6-flash"
8
8
  }
9
9
  }
@@ -742,9 +742,14 @@ class "anthropic" {
742
742
  }
743
743
  on "google-antigravity" {
744
744
  drop-unsigned-thinking #true
745
- claude-thinking-beta-header #true
746
745
  antigravity-claude-tool-mode #true
747
746
  antigravity-usage-label "true"
747
+ // The interleaved-thinking beta gated Claude 4.x thinking here. The
748
+ // native 2.26.0 client sends no beta for Claude 5.5, whose thinking
749
+ // streams without it.
750
+ revision "<5" {
751
+ claude-thinking-beta-header #true
752
+ }
748
753
  }
749
754
  }
750
755
 
@@ -48,6 +48,16 @@ class "minimax" {
48
48
  }
49
49
  }
50
50
  }
51
+ // MiniMax's standard API doubles every rate above 512K input tokens;
52
+ // subscription hosts retain the same rates only as usage estimates.
53
+ on "minimax" "minimax-cn" "minimax-code" "minimax-code-cn" {
54
+ models "MiniMax-M3" {
55
+ long-context-cost {
56
+ input-threshold 512000
57
+ multiplier 2.0
58
+ }
59
+ }
60
+ }
51
61
  on "alibaba-coding-plan" "minimax-code" "minimax-code-cn" "wafer-serverless" {
52
62
  supports-developer-role #false
53
63
  }
@@ -116,13 +116,27 @@ provider "anthropic" {
116
116
  family "fable" {
117
117
  server-side-fallback-models "claude-opus-5"
118
118
  // Fallback-credit redemption targets: the refused model's permitted
119
- // fallback targets. Fable 5 and 5.1 → Opus 4.8 and Opus 5
119
+ // fallback targets
120
120
  // (https://platform.claude.com/docs/en/build-with-claude/fallback-credit).
121
+ // Fable 5 and 5.1 → Opus 4.8 and Opus 5.
121
122
  fallback-credit-targets "claude-opus-4-8" "claude-opus-5"
122
123
  }
123
124
  family "mythos" {
124
125
  server-side-fallback-models "claude-opus-5"
125
126
  }
127
+ family "opus" {
128
+ revision "=5.5" {
129
+ fallback-credit-targets "claude-opus-4-8" "claude-opus-5"
130
+ }
131
+ revision "=5" {
132
+ fallback-credit-targets "claude-opus-4-8"
133
+ }
134
+ }
135
+ family "sonnet" {
136
+ revision "=5.5" {
137
+ fallback-credit-targets "claude-sonnet-5"
138
+ }
139
+ }
126
140
  }
127
141
  // residue: taxonomy ranks and exact globs do not isolate these models.
128
142
  models "*20240229" "claude-3-7-sonnet-20250219" {
@@ -5,6 +5,18 @@ provider "deepseek" {
5
5
  env "DEEPSEEK_API_KEY"
6
6
  discovery label="DeepSeek"
7
7
 
8
+ // DeepSeek's vision guide documents 600 images per request and a 48 MiB
9
+ // request body. OMP reserves 8 MiB for text, tools, and the JSON envelope
10
+ // (48 MiB - 8 MiB = 40 MiB); Anthropic's 24,000,000-byte image budget under
11
+ // its 32 MB body uses an 8 MB reserve. DeepSeek documents MiB, so this uses
12
+ // an 8 MiB reserve. The split is OMP's conservative margin, not a
13
+ // DeepSeek-documented figure. A custom baseUrl keeps the conservative
14
+ // catalog floor because the gateway's own limits are unknown.
15
+ // Source: https://api-docs.deepseek.com/guides/vision#limits
16
+ catalog-endpoint-budgets #true
17
+ official-image-count-budget 600
18
+ inline-image-byte-budget 41943040
19
+
8
20
  extra-body {
9
21
  thinking {
10
22
  type "enabled"
@@ -52,7 +64,7 @@ provider "deepseek" {
52
64
  max-tokens 384000
53
65
  }
54
66
  }
55
- // residue: the dated price transition applies to the documented exact Pro SKU only.
67
+ // DeepSeek kept Pro billing unchanged after September 14 (2026-09-10 change log).
56
68
  models "deepseek-v4-pro" {
57
69
  cost-patch {
58
70
  input 1.32
@@ -74,15 +86,6 @@ provider "deepseek" {
74
86
  end-minute 600
75
87
  }
76
88
  }
77
- effective-rates {
78
- flash-pricing {
79
- effective-from "2026-09-14T04:00:00Z"
80
- input 0.30
81
- output 1.20
82
- cache-read 0.006
83
- cache-write 0
84
- }
85
- }
86
89
  }
87
90
  }
88
91
  // residue: taxonomy ranks and exact globs do not isolate these models.
@@ -405,6 +405,26 @@ provider "factory-droid" {
405
405
  cost input=0 output=0 cache-read=0 cache-write=0
406
406
  limits context=196600 max-tokens=64000
407
407
  }
408
+ // `generate_image` models, in the CLI's tool preference order. GPT Image 2
409
+ // is retired from the CLI's selectable chain and is not listed.
410
+ model "gpt-image-2.5-sunburst" name="GPT Image 2.5" {
411
+ reasoning #false
412
+ input "text" "image"
413
+ cost input=0 output=0 cache-read=0 cache-write=0
414
+ limits
415
+ }
416
+ model "gpt-image-2.5-flare" name="GPT Image 2.5 Flare" {
417
+ reasoning #false
418
+ input "text" "image"
419
+ cost input=0 output=0 cache-read=0 cache-write=0
420
+ limits
421
+ }
422
+ model "gemini-3-pro-image-preview" name="Gemini 3 Pro Image" {
423
+ reasoning #false
424
+ input "text" "image"
425
+ cost input=0 output=0 cache-read=0 cache-write=0
426
+ limits
427
+ }
408
428
  }
409
429
 
410
430
  // Upstream serving regions (the CLI's upstream region table). A model's own
@@ -836,6 +856,23 @@ provider "factory-droid" {
836
856
  thinking-requires-effort #true
837
857
  thinking-default-level "high"
838
858
  }
859
+ // Image models: the org's `imageGenerationEnabled` and
860
+ // `imageGenerationAllowedModelIds` managed settings gate them
861
+ // (discovery/factory-droid.ts), not feature flags. Credits follow the CLI's
862
+ // image registry: OpenAI output is image tokens at 6x input, Gemini's 60x.
863
+ models "gpt-image-2.5-sunburst" "gpt-image-2.5-flare" {
864
+ kind "image"
865
+ routing-family "openai"
866
+ upstream-rotation "openai"
867
+ credit-rates { input 2.0; output 6 }
868
+ }
869
+ models "gemini-3-pro-image-preview" {
870
+ kind "image"
871
+ routing-family "google"
872
+ upstream-rotation "google"
873
+ region-upstreams-us
874
+ credit-rates { input 0.8; output 60 }
875
+ }
839
876
  models "inkling" {
840
877
  routing-family "factory"
841
878
  upstream-rotation "fireworks"
@@ -1130,6 +1167,12 @@ provider "factory-droid" {
1130
1167
  on-upstream "snowflake" {
1131
1168
  disabled-thinking "disabled"
1132
1169
  }
1170
+ // Factory's proxy honors `ttl: "1h"` without the extended-cache-ttl beta
1171
+ // and the Anthropic and Bedrock upstreams reject the beta.
1172
+ extended-cache-ttl-beta #false
1173
+ on-upstream "anthropic" "vertex_anthropic" "bedrock_anthropic" "azure_anthropic" {
1174
+ supports-long-cache-retention #true
1175
+ }
1133
1176
  }
1134
1177
  // Adaptive models the registry marks `supportsDisabledThinking` send the
1135
1178
  // explicit disabled form at Off on every upstream.
@@ -1137,8 +1180,13 @@ provider "factory-droid" {
1137
1180
  "claude-opus-4-8-fast" "claude-opus-5" "claude-opus-5-fast" "claude-haiku-5-5" {
1138
1181
  disabled-thinking "disabled"
1139
1182
  }
1183
+ // Cache entry lifetimes per retention tier, in seconds, for the cache warmer.
1140
1184
  models "claude-*" "atlas-07-21" "aster-07-15" {
1141
1185
  anthropic-thinking "adaptive-summarized"
1186
+ prompt-cache {
1187
+ short 300
1188
+ long 3600
1189
+ }
1142
1190
  }
1143
1191
  models "claude-opus-4-6" "claude-sonnet-4-6" {
1144
1192
  anthropic-thinking "adaptive"
@@ -52,6 +52,13 @@ provider "google-antigravity" {
52
52
  prompt-cache {
53
53
  short 300
54
54
  }
55
+ // Cloud Code Assist rejects Claude 4.x requests over 64,000 output
56
+ // tokens with a 400 (`Request contains an invalid argument`) even when
57
+ // the roster advertises 65,536 (#3067). The ceiling also lowers the
58
+ // discovered per-SKU wire caps the transport sends verbatim.
59
+ revision "<5" {
60
+ max-tokens-ceiling 64000
61
+ }
55
62
  family "opus" {
56
63
  thinking-mode "budget"
57
64
  }