cartethyia 2.7.0 → 2.7.1

package/src/protocol/gateway-error.ts

@@ -1,232 +1,232 @@
-// Typed gateway failures: the stable public error codes, the error class, and
-// the sanitizers that keep upstream payloads out of public envelopes.
-
-/** Stable typed error codes returned at gateway boundaries. */
-export type GatewayErrorCode =
-  | "capability_unsupported"
-  | "ambiguous_model"
-  | "model_not_found"
-  | "upstream_not_found"
-  | "context_length_exceeded"
-  | "request_too_large"
-  | "accounts_unavailable"
-  | "accounts_rate_limited"
-  | "capacity_exhausted"
-  | "provider_capacity_exhausted"
-  | "upstream_queued"
-  | "proxy_pool_capacity_exceeded"
-  | "proxy_pool_provider_capacity_exceeded"
-  | "proxy_pool_blocked"
-  | "proxy_pool_cooldown"
-  | "proxy_pool_unavailable"
-  | "proxy_pool_unhealthy"
-  | "invalid_pool_limits"
-  | "admission_unavailable"
-  | "slug_reserved"
-  | "quota_exceeded"
-  | "authentication_failed"
-  | "policy_rejected"
-  | "invalid_request"
-  | "not_found"
-  | "link_not_found"
-  | "link_expired"
-  | "link_revoked"
-  | "link_disabled"
-  | "invalid_sequence"
-  | "invalid_lifecycle"
-  | "unsupported_field"
-  | "unsupported_media_type"
-  | "upstream_conflict"
-  | "upstream_unprocessable"
-  | "internal_error"
-  | "transport_unavailable"
-  | "tunnel_setup_failed"
-  | "platform_unavailable"
-  | "tenant_capacity_exhausted"
-  | "tls_rejected"
-  | "proxy_auth_required"
-  | "deadline_exceeded"
-  | "transport_closed"
-  | "max_connections_exceeded"
-  | "overloaded"
-  | "proxy_unreachable"
-  | "tool_call_loop_detected"
-  | "client_router_denied"
-  | "shutting_down"
-  | "restart_for_update";
-
-/** Identifies which boundary produced a safe public error. */
-export type GatewayErrorOrigin = "cartethyia" | "upstream" | "network";
-
-/**
- * Error with a client-safe stable code, origin, and sanitized metadata.
- */
-export class GatewayError extends Error {
-  readonly code: GatewayErrorCode;
-  readonly status: number;
-  readonly origin: GatewayErrorOrigin;
-  readonly details: Readonly<Record<string, unknown>>;
-
-  constructor(
-    code: GatewayErrorCode,
-    status: number,
-    message: string,
-    details: Readonly<Record<string, unknown>> = {},
-    origin: GatewayErrorOrigin = "cartethyia",
-  ) {
-    super(message);
-    this.name = "GatewayError";
-    this.code = code;
-    this.status = status;
-    this.origin = origin;
-    this.details = details;
-  }
-}
-
-/**
- * Legacy origin brand prefixes. Older builds stamped these into the public
- * `message`; strip them so a value that crossed two shapers is not doubled
- * and so clients never see product branding in the error text.
- *
- * Blame lives in the structured `origin` field (`cartethyia` | `upstream` |
- * `network`). The public message is always `code: explanatory` for every
- * origin — gateway and upstream look the same on the wire.
- */
-const LEGACY_ORIGIN_LABELS: readonly string[] = [
-  "Cartethyia Error:",
-  "Upstream Error:",
-  "Network Error:",
-];
-
-function stripLegacyOriginLabel(message: string): string {
-  const trimmed = message.trim();
-  for (const prefix of LEGACY_ORIGIN_LABELS) {
-    if (trimmed.startsWith(prefix)) return trimmed.slice(prefix.length).trimStart();
-  }
-  return trimmed;
-}
-
-/**
- * Public client-facing error text: a plain sentence, with no machine code
- * stamped into it.
- *
- * The code is not lost — it travels in the envelope's `error.code` field, and
- * the HTTP status carries the class. Stamping it into the text
- * (`accounts_unavailable: Model 'x' has no available account (52 candidate(s)
- * unusable: credit_floor_reached, model_cooldown)`) would put internal enum
- * names in front of the operator's sentence and make the message unreadable
- * while adding nothing the field does not already carry.
- *
- * Stripping is idempotent, and that is the point: a message that already
- * carries a `code:` prefix — stamped by an older builder, or by a code path
- * that formats before it reaches here — loses it once instead of being
- * doubled. Origin is never written into the text either; clients that need the
- * layer read `error.origin`.
- */
-export function formatPublicErrorMessage(code: string, message: string): string {
-  const strippedOrigin = stripLegacyOriginLabel(message);
-  // A leading `<code>:` is the old stamping; drop it so the sentence stands on
-  // its own. Matched against the code this call is formatting, so a message
-  // that legitimately begins with some other word is left alone.
-  const prefix = `${code}:`;
-  const explanatory = strippedOrigin.startsWith(prefix)
-    ? strippedOrigin.slice(prefix.length).trimStart()
-    : strippedOrigin;
-  if (explanatory.length === 0) return code;
-  return explanatory;
-}
-
-/** Formats a gateway error for public/internal clients without mutating it. */
-export function explainGatewayError(error: GatewayError): string {
-  return formatPublicErrorMessage(error.code, error.message);
-}
-
-const PUBLIC_ERROR_DETAIL_KEYS = new Set([
-  "accountId",
-  "accountScope",
-  "available",
-  "capacity",
-  "configuredPools",
-  "cooldownReason",
-  "gateway_hint",
-  "currentInflight",
-  "credentialEvidence",
-  "model",
-  "poolId",
-  "pools",
-  "providerCode",
-  "providerId",
-  "providerStatus",
-  "providerScope",
-  "raw",
-  "rateLimitScope",
-  "reason",
-  "reasons",
-  "candidate_count",
-  "requestId",
-  "retryAfterMs",
-  "retryAt",
-  "safeMessage",
-  "scope",
-  "status",
-  "upstreamRequestId",
-  "upstreamStatus",
-  "maxInflight",
-  "weight",
-]);
-function sanitizePublicDetail(value: unknown, depth = 0): unknown {
-  if (typeof value === "string") return value.slice(0, 500);
-  if (typeof value !== "object" || value === null) return value;
-  if (depth >= 3) return "[redacted]";
-  if (Array.isArray(value))
-    return value.slice(0, 32).map((entry) => sanitizePublicDetail(entry, depth + 1));
-  return Object.fromEntries(
-    Object.entries(value).map(([key, entry]) => [
-      key,
-      sanitizePublicDetail(entry, depth + 1),
-    ]),
-  );
-}
-
-/** Returns bounded allowlisted details suitable for public error envelopes. */
-export function publicGatewayErrorDetails(
-  error: GatewayError,
-): Readonly<Record<string, unknown>> {
-  return Object.fromEntries(
-    Object.entries(error.details)
-      .filter(([key]) => PUBLIC_ERROR_DETAIL_KEYS.has(key))
-      .map(([key, value]) => [key, sanitizePublicDetail(value)]),
-  );
-}
-
-/** Serializes a GatewayError with the same public shape as transport errors. */
-export function publicGatewayErrorBody(error: GatewayError): {
-  readonly error: {
-    readonly origin: GatewayErrorOrigin;
-    readonly code: GatewayErrorCode;
-    readonly message: string;
-    readonly details: Readonly<Record<string, unknown>>;
-  };
-} {
-  return {
-    error: {
-      origin: error.origin,
-      code: error.code,
-      message: explainGatewayError(error),
-      details: publicGatewayErrorDetails(error),
-    },
-  };
-}
-
-/** Creates the typed rejection required when a semantic feature is unavailable. */
-export function capabilityUnsupported(
-  capability: string,
-  details: Readonly<Record<string, unknown>> = {},
-): GatewayError {
-  return new GatewayError(
-    "capability_unsupported",
-    400,
-    `The selected route does not support capability: ${capability}`,
-    { capability, ...details },
-  );
-}
+// Typed gateway failures: the stable public error codes, the error class, and
+// the sanitizers that keep upstream payloads out of public envelopes.
+
+/** Stable typed error codes returned at gateway boundaries. */
+export type GatewayErrorCode =
+  | "capability_unsupported"
+  | "ambiguous_model"
+  | "model_not_found"
+  | "upstream_not_found"
+  | "context_length_exceeded"
+  | "request_too_large"
+  | "accounts_unavailable"
+  | "accounts_rate_limited"
+  | "capacity_exhausted"
+  | "provider_capacity_exhausted"
+  | "upstream_queued"
+  | "proxy_pool_capacity_exceeded"
+  | "proxy_pool_provider_capacity_exceeded"
+  | "proxy_pool_blocked"
+  | "proxy_pool_cooldown"
+  | "proxy_pool_unavailable"
+  | "proxy_pool_unhealthy"
+  | "invalid_pool_limits"
+  | "admission_unavailable"
+  | "slug_reserved"
+  | "quota_exceeded"
+  | "authentication_failed"
+  | "policy_rejected"
+  | "invalid_request"
+  | "not_found"
+  | "link_not_found"
+  | "link_expired"
+  | "link_revoked"
+  | "link_disabled"
+  | "invalid_sequence"
+  | "invalid_lifecycle"
+  | "unsupported_field"
+  | "unsupported_media_type"
+  | "upstream_conflict"
+  | "upstream_unprocessable"
+  | "internal_error"
+  | "transport_unavailable"
+  | "tunnel_setup_failed"
+  | "platform_unavailable"
+  | "tenant_capacity_exhausted"
+  | "tls_rejected"
+  | "proxy_auth_required"
+  | "deadline_exceeded"
+  | "transport_closed"
+  | "max_connections_exceeded"
+  | "overloaded"
+  | "proxy_unreachable"
+  | "tool_call_loop_detected"
+  | "client_router_denied"
+  | "shutting_down"
+  | "restart_for_update";
+
+/** Identifies which boundary produced a safe public error. */
+export type GatewayErrorOrigin = "cartethyia" | "upstream" | "network";
+
+/**
+ * Error with a client-safe stable code, origin, and sanitized metadata.
+ */
+export class GatewayError extends Error {
+  readonly code: GatewayErrorCode;
+  readonly status: number;
+  readonly origin: GatewayErrorOrigin;
+  readonly details: Readonly<Record<string, unknown>>;
+
+  constructor(
+    code: GatewayErrorCode,
+    status: number,
+    message: string,
+    details: Readonly<Record<string, unknown>> = {},
+    origin: GatewayErrorOrigin = "cartethyia",
+  ) {
+    super(message);
+    this.name = "GatewayError";
+    this.code = code;
+    this.status = status;
+    this.origin = origin;
+    this.details = details;
+  }
+}
+
+/**
+ * Legacy origin brand prefixes. Older builds stamped these into the public
+ * `message`; strip them so a value that crossed two shapers is not doubled
+ * and so clients never see product branding in the error text.
+ *
+ * Blame lives in the structured `origin` field (`cartethyia` | `upstream` |
+ * `network`). The public message is always `code: explanatory` for every
+ * origin — gateway and upstream look the same on the wire.
+ */
+const LEGACY_ORIGIN_LABELS: readonly string[] = [
+  "Cartethyia Error:",
+  "Upstream Error:",
+  "Network Error:",
+];
+
+function stripLegacyOriginLabel(message: string): string {
+  const trimmed = message.trim();
+  for (const prefix of LEGACY_ORIGIN_LABELS) {
+    if (trimmed.startsWith(prefix)) return trimmed.slice(prefix.length).trimStart();
+  }
+  return trimmed;
+}
+
+/**
+ * Public client-facing error text: a plain sentence, with no machine code
+ * stamped into it.
+ *
+ * The code is not lost — it travels in the envelope's `error.code` field, and
+ * the HTTP status carries the class. Stamping it into the text
+ * (`accounts_unavailable: Model 'x' has no available account (52 candidate(s)
+ * unusable: credit_floor_reached, model_cooldown)`) would put internal enum
+ * names in front of the operator's sentence and make the message unreadable
+ * while adding nothing the field does not already carry.
+ *
+ * Stripping is idempotent, and that is the point: a message that already
+ * carries a `code:` prefix — stamped by an older builder, or by a code path
+ * that formats before it reaches here — loses it once instead of being
+ * doubled. Origin is never written into the text either; clients that need the
+ * layer read `error.origin`.
+ */
+export function formatPublicErrorMessage(code: string, message: string): string {
+  const strippedOrigin = stripLegacyOriginLabel(message);
+  // A leading `<code>:` is the old stamping; drop it so the sentence stands on
+  // its own. Matched against the code this call is formatting, so a message
+  // that legitimately begins with some other word is left alone.
+  const prefix = `${code}:`;
+  const explanatory = strippedOrigin.startsWith(prefix)
+    ? strippedOrigin.slice(prefix.length).trimStart()
+    : strippedOrigin;
+  if (explanatory.length === 0) return code;
+  return explanatory;
+}
+
+/** Formats a gateway error for public/internal clients without mutating it. */
+export function explainGatewayError(error: GatewayError): string {
+  return formatPublicErrorMessage(error.code, error.message);
+}
+
+const PUBLIC_ERROR_DETAIL_KEYS = new Set([
+  "accountId",
+  "accountScope",
+  "available",
+  "capacity",
+  "configuredPools",
+  "cooldownReason",
+  "gateway_hint",
+  "currentInflight",
+  "credentialEvidence",
+  "model",
+  "poolId",
+  "pools",
+  "providerCode",
+  "providerId",
+  "providerStatus",
+  "providerScope",
+  "raw",
+  "rateLimitScope",
+  "reason",
+  "reasons",
+  "candidate_count",
+  "requestId",
+  "retryAfterMs",
+  "retryAt",
+  "safeMessage",
+  "scope",
+  "status",
+  "upstreamRequestId",
+  "upstreamStatus",
+  "maxInflight",
+  "weight",
+]);
+function sanitizePublicDetail(value: unknown, depth = 0): unknown {
+  if (typeof value === "string") return value.slice(0, 500);
+  if (typeof value !== "object" || value === null) return value;
+  if (depth >= 3) return "[redacted]";
+  if (Array.isArray(value))
+    return value.slice(0, 32).map((entry) => sanitizePublicDetail(entry, depth + 1));
+  return Object.fromEntries(
+    Object.entries(value).map(([key, entry]) => [
+      key,
+      sanitizePublicDetail(entry, depth + 1),
+    ]),
+  );
+}
+
+/** Returns bounded allowlisted details suitable for public error envelopes. */
+export function publicGatewayErrorDetails(
+  error: GatewayError,
+): Readonly<Record<string, unknown>> {
+  return Object.fromEntries(
+    Object.entries(error.details)
+      .filter(([key]) => PUBLIC_ERROR_DETAIL_KEYS.has(key))
+      .map(([key, value]) => [key, sanitizePublicDetail(value)]),
+  );
+}
+
+/** Serializes a GatewayError with the same public shape as transport errors. */
+export function publicGatewayErrorBody(error: GatewayError): {
+  readonly error: {
+    readonly origin: GatewayErrorOrigin;
+    readonly code: GatewayErrorCode;
+    readonly message: string;
+    readonly details: Readonly<Record<string, unknown>>;
+  };
+} {
+  return {
+    error: {
+      origin: error.origin,
+      code: error.code,
+      message: explainGatewayError(error),
+      details: publicGatewayErrorDetails(error),
+    },
+  };
+}
+
+/** Creates the typed rejection required when a semantic feature is unavailable. */
+export function capabilityUnsupported(
+  capability: string,
+  details: Readonly<Record<string, unknown>> = {},
+): GatewayError {
+  return new GatewayError(
+    "capability_unsupported",
+    400,
+    `The selected route does not support capability: ${capability}`,
+    { capability, ...details },
+  );
+}

package/src/{providers/provider-reasoning.ts → protocol/reasoning.ts}

@@ -1,6 +1,6 @@
-import type { CanonicalRequest, ReasoningEffort, ReasoningIntent } from "../protocol/canonical-model";
-import { budgetToLevel } from "../transport/translation/thinking";
-import { isRecord } from "../protocol/codecs";
+import type { CanonicalRequest, ReasoningEffort, ReasoningIntent } from "./canonical-model";
+import { budgetToLevel } from "./translation/thinking";
+import { isRecord } from "./codecs";
 
 /**
  * The effective reasoning effort for a canonical intent, deriving one from an
@@ -132,12 +132,13 @@ const CLAMPED_EFFORT_MIN_OUTPUT_TOKENS = 16_384;
 const BUDDY_LEVEL_RANK: Readonly<Record<string, number>> = {
   off: 0,
   none: 0,
-  minimal: 1,
-  low: 2,
-  medium: 3,
-  high: 4,
-  xhigh: 5,
-  max: 6,
+  instant: 1,
+  minimal: 2,
+  low: 3,
+  medium: 4,
+  high: 5,
+  xhigh: 6,
+  max: 7,
 };
 
 /** Builds the Anthropic thinking block from canonical reasoning intent. */

package/src/protocol/request/chat.ts

@@ -3,18 +3,18 @@
  */
 import type { CanonicalRequest, ResponseFormat } from "../canonical-model";
 import { isRecord } from "../codecs";
-import { markLatestBreakpoint } from "../../transport/translation/cache-controls";
-import { isWebSearchTool } from "../../transport/translation/search";
-import { pickWireSupportedControls } from "../../transport/translation/capabilities";
-import { normalizeWireMaxTokens } from "../../transport/translation/normalize";
-import { resolveWireMaxTokens } from "../../transport/translation/quirks";
+import { markLatestBreakpoint } from "../../protocol/translation/cache-controls";
+import { isWebSearchTool } from "../../protocol/translation/search";
+import { pickWireSupportedControls } from "../../protocol/translation/capabilities";
+import { normalizeWireMaxTokens } from "../../protocol/translation/normalize";
+import { resolveWireMaxTokens } from "../../protocol/translation/quirks";
 import {
   clampReasoningEffort,
   resolveSupportedReasoningEfforts,
-} from "../../transport/translation/thinking";
+} from "../../protocol/translation/thinking";
 import { isClaudeBillingHeaderText, resolveImageSource } from "../codecs";
 import { resolvePromptCacheKey } from "../../providers/operations/session-resolution";
-import { reasoningEffortFromIntent } from "../../providers/provider-reasoning";
+import { reasoningEffortFromIntent } from "../../protocol/reasoning";
 import { parseChatResponseToEvents, decodeChatSseStream } from "../response/chat";
 
 /**

package/src/protocol/request/codex.ts

@@ -9,8 +9,8 @@
  */
 import type { CanonicalMessage, CanonicalRequest, ContentPart } from "../canonical-model";
 import { canContainToolResult, toolResultParts } from "../canonical-model";
-import { isWebSearchTool } from "../../transport/translation/search";
-import { reasoningEffortFromIntent } from "../../providers/provider-reasoning";
+import { isWebSearchTool } from "../../protocol/translation/search";
+import { reasoningEffortFromIntent } from "../../protocol/reasoning";
 import {
   decodeCodexToolCallId,
   encodeCodexToolCallId,

package/src/protocol/request/gemini.ts

@@ -14,7 +14,7 @@ import {
   splitDataUrl,
   type RewriteWalk,
 } from "../codecs";
-import { geminiThinkingOutputFloor } from "../../providers/provider-reasoning";
+import { geminiThinkingOutputFloor } from "../../protocol/reasoning";
 import { base64DecodedByteLength, clampImage } from "../image-clamp";
 
 export function geminiModelUrl(base: string, model: string, action: string): string {

package/src/protocol/request/messages.ts

@@ -2,15 +2,15 @@
  * Canonical → Claude Messages request builder.
  */
 import { allParts, type CanonicalRequest, type ContentPart } from "../canonical-model";
-import { buildAnthropicThinkingPayload } from "../../providers/provider-reasoning";
-import { anthropicCacheControl } from "../../transport/translation/cache-controls";
+import { buildAnthropicThinkingPayload } from "../../protocol/reasoning";
+import { anthropicCacheControl } from "../../protocol/translation/cache-controls";
 import {
   THINKING_STRIPPED_SAMPLING_CONTROLS,
   ensureMaxTokensForThinking,
   isThinkingEnabled,
-} from "../../transport/translation/capabilities";
-import { normalizeWireMaxTokens } from "../../transport/translation/normalize";
-import { resolveWireMaxTokens, WIRE_TOKEN_BOUNDS } from "../../transport/translation/quirks";
+} from "../../protocol/translation/capabilities";
+import { normalizeWireMaxTokens } from "../../protocol/translation/normalize";
+import { resolveWireMaxTokens, WIRE_TOKEN_BOUNDS } from "../../protocol/translation/quirks";
 import {
   isClaudeBillingHeaderText,
   normalizeAnthropicToolCallId,
@@ -28,7 +28,7 @@ import {
 } from "../codecs";
 import { log } from "../../observability/logger";
 import { base64DecodedByteLength, clampImage } from "../image-clamp";
-import { clampReasoningEffort, claudeUsesAdaptiveThinking, LEVEL_TO_BUDGET, resolveSupportedReasoningEfforts } from "../../transport/translation/thinking";
+import { clampReasoningEffort, claudeUsesAdaptiveThinking, LEVEL_TO_BUDGET, resolveSupportedReasoningEfforts } from "../../protocol/translation/thinking";
 import { claudeResponseToEvents, parseClaudeSseStream } from "../response/messages";
 
 /**

package/src/protocol/request/responses.ts

@@ -3,18 +3,18 @@
  */
 import type { CanonicalRequest } from "../canonical-model";
 import { clampToolCallId, isRecord } from "../codecs";
-import { markLatestBreakpoint } from "../../transport/translation/cache-controls";
-import { pickWireSupportedControls } from "../../transport/translation/capabilities";
-import { normalizeGenerationControls } from "../../transport/translation/normalize";
-import { resolveWireMaxTokens } from "../../transport/translation/quirks";
+import { markLatestBreakpoint } from "../../protocol/translation/cache-controls";
+import { pickWireSupportedControls } from "../../protocol/translation/capabilities";
+import { normalizeGenerationControls } from "../../protocol/translation/normalize";
+import { resolveWireMaxTokens } from "../../protocol/translation/quirks";
 import {
   clampReasoningEffort,
   resolveSupportedReasoningEfforts,
-} from "../../transport/translation/thinking";
+} from "../../protocol/translation/thinking";
 import { isClaudeBillingHeaderText, resolveImageSource, resolveOutputVerbosity } from "../codecs";
 import { resolvePromptCacheKey } from "../../providers/operations/session-resolution";
-import { reasoningEffortFromIntent } from "../../providers/provider-reasoning";
-import { isWebSearchTool } from "../../transport/translation/search";
+import { reasoningEffortFromIntent } from "../../protocol/reasoning";
+import { isWebSearchTool } from "../../protocol/translation/search";
 import { parseResponsesResponseToEvents, decodeResponsesSseStream } from "../response/responses";
 import { repairToolArguments } from "../../transport/surface/chat/encode";