@openllmsh/sdk 0.0.0-stage → 2.10.12-alpha.0

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 (245) hide show
  1. package/ARCHITECTURE.md +43 -0
  2. package/LICENSE +202 -0
  3. package/README.md +142 -2
  4. package/assets/openllm-light.svg +18 -0
  5. package/assets/openllm.svg +18 -0
  6. package/client/chat.ts +140 -0
  7. package/client/devices.ts +269 -0
  8. package/client/http.ts +72 -0
  9. package/client/index.ts +129 -0
  10. package/client/relay-mux.ts +547 -0
  11. package/client/types.ts +51 -0
  12. package/client/vault.ts +239 -0
  13. package/connect/chat.ts +58 -0
  14. package/connect/client.ts +240 -0
  15. package/connect/config.ts +47 -0
  16. package/connect/errors.ts +121 -0
  17. package/connect/index.ts +23 -0
  18. package/connect/oauth.ts +265 -0
  19. package/connect/routes.ts +234 -0
  20. package/connect/store.ts +107 -0
  21. package/mcp/handle.ts +24 -0
  22. package/mcp/index.ts +76 -0
  23. package/mcp/server.ts +319 -0
  24. package/mcp/tools/marketplace.ts +439 -0
  25. package/mcp/tools/memory-client.ts +297 -0
  26. package/mcp/tools/memory-config.ts +32 -0
  27. package/mcp/tools/memory.ts +369 -0
  28. package/mcp/tools/model-priority.ts +88 -0
  29. package/mcp/tools/openllm.ts +883 -0
  30. package/mcp/types.ts +28 -0
  31. package/openapi/call.ts +90 -0
  32. package/openapi/index.ts +8 -0
  33. package/openapi/openapi.json +80213 -0
  34. package/openapi/operations.ts +2075 -0
  35. package/package.json +39 -4
  36. package/protocol/CONTRIBUTING.md +21 -0
  37. package/protocol/README.md +134 -0
  38. package/protocol/access.ts +186 -0
  39. package/protocol/agent-prompt-text.ts +32 -0
  40. package/protocol/agent.ts +991 -0
  41. package/protocol/api-key-syntax.ts +105 -0
  42. package/protocol/assets/openllm-light.svg +18 -0
  43. package/protocol/assets/openllm.svg +18 -0
  44. package/protocol/audio.ts +29 -0
  45. package/protocol/auth.ts +142 -0
  46. package/protocol/changelog.ts +155 -0
  47. package/protocol/chat.ts +458 -0
  48. package/protocol/client-registry.ts +356 -0
  49. package/protocol/config.ts +345 -0
  50. package/protocol/control-channel.ts +115 -0
  51. package/protocol/cooldown-reason.ts +78 -0
  52. package/protocol/cost.ts +10 -0
  53. package/protocol/daemon-port.ts +31 -0
  54. package/protocol/daemon.ts +1524 -0
  55. package/protocol/default-chains.ts +66 -0
  56. package/protocol/device-state.ts +85 -0
  57. package/protocol/doctor-report-local.ts +443 -0
  58. package/protocol/doctor-report-policy.ts +119 -0
  59. package/protocol/doctor-report.ts +947 -0
  60. package/protocol/errors.ts +14 -0
  61. package/protocol/execution-selection.ts +141 -0
  62. package/protocol/hop-grant.ts +33 -0
  63. package/protocol/image-edit-parse.ts +263 -0
  64. package/protocol/images.ts +298 -0
  65. package/protocol/index.ts +74 -0
  66. package/protocol/inference-headers.ts +3 -0
  67. package/protocol/keepalive.ts +69 -0
  68. package/protocol/local-auth.ts +300 -0
  69. package/protocol/local-device-control.ts +133 -0
  70. package/protocol/local-provider.ts +288 -0
  71. package/protocol/local-session.ts +118 -0
  72. package/protocol/marketplace.ts +187 -0
  73. package/protocol/media-attempt.ts +81 -0
  74. package/protocol/media-bytes.ts +28 -0
  75. package/protocol/media-default.ts +357 -0
  76. package/protocol/media-dimensions.ts +132 -0
  77. package/protocol/media-input.ts +61 -0
  78. package/protocol/media-model.ts +142 -0
  79. package/protocol/media-option-issue.ts +90 -0
  80. package/protocol/media-types.ts +70 -0
  81. package/protocol/model-match.ts +1051 -0
  82. package/protocol/model-settings.ts +327 -0
  83. package/protocol/models.ts +767 -0
  84. package/protocol/mux.ts +722 -0
  85. package/protocol/plan-surface.ts +59 -0
  86. package/protocol/provider-models.ts +16 -0
  87. package/protocol/provider-status.ts +206 -0
  88. package/protocol/provider-tier.ts +67 -0
  89. package/protocol/provider-usage.ts +263 -0
  90. package/protocol/providers.ts +901 -0
  91. package/protocol/public-usage-index.ts +354 -0
  92. package/protocol/realtime-generated.ts +31 -0
  93. package/protocol/realtime.ts +324 -0
  94. package/protocol/registry-pointer.ts +118 -0
  95. package/protocol/relay.ts +702 -0
  96. package/protocol/replay-session.ts +11 -0
  97. package/protocol/responses.ts +280 -0
  98. package/protocol/semver.ts +182 -0
  99. package/protocol/setup-mode.ts +22 -0
  100. package/protocol/sha256-hex.ts +98 -0
  101. package/protocol/stats.ts +406 -0
  102. package/protocol/streaming.ts +19 -0
  103. package/protocol/subscription-daemon-floors.ts +33 -0
  104. package/protocol/subscription-provider-compat.ts +104 -0
  105. package/protocol/subscription-provider.ts +32 -0
  106. package/protocol/team-grant.ts +17 -0
  107. package/protocol/team-shareable.ts +61 -0
  108. package/protocol/transcription-encoding.ts +76 -0
  109. package/protocol/url-locality-dns.ts +37 -0
  110. package/protocol/url-locality.ts +288 -0
  111. package/protocol/usage-snapshots.ts +9 -0
  112. package/protocol/vendor-cli-install.ts +29 -0
  113. package/protocol/video-job-plan.ts +50 -0
  114. package/protocol/videos.ts +227 -0
  115. package/tsconfig.json +14 -0
  116. package/tunnel/ARCHITECTURE.md +111 -0
  117. package/tunnel/LICENSE +33 -0
  118. package/tunnel/README.md +45 -0
  119. package/tunnel/assets/openllm-light.svg +18 -0
  120. package/tunnel/assets/openllm.svg +18 -0
  121. package/tunnel/channel-envelope.ts +139 -0
  122. package/tunnel/codec.ts +191 -0
  123. package/tunnel/device-anchor.ts +229 -0
  124. package/tunnel/device-control.ts +118 -0
  125. package/tunnel/device-grant.ts +173 -0
  126. package/tunnel/fleet.ts +77 -0
  127. package/tunnel/index.ts +22 -0
  128. package/tunnel/mux.ts +580 -0
  129. package/tunnel/peer-link.ts +390 -0
  130. package/tunnel/rtc-auth.ts +415 -0
  131. package/tunnel/rtc-duplex.ts +187 -0
  132. package/tunnel/rtc-ice.ts +136 -0
  133. package/tunnel/secure.ts +1298 -0
  134. package/tunnel/stream-reset-error.ts +14 -0
  135. package/tunnel/streams.ts +1430 -0
  136. package/tunnel/team-grant.ts +302 -0
  137. package/tunnel/wire-bytes.ts +102 -0
  138. package/vault/ARCHITECTURE.md +470 -0
  139. package/vault/LICENSE +33 -0
  140. package/vault/api-key.ts +190 -0
  141. package/vault/argon2.ts +106 -0
  142. package/vault/cli-auth-seal.ts +157 -0
  143. package/vault/cred-seal.ts +146 -0
  144. package/vault/credential.ts +55 -0
  145. package/vault/crypto.ts +210 -0
  146. package/vault/daemon-attest.ts +229 -0
  147. package/vault/derive.ts +113 -0
  148. package/vault/device-access.ts +87 -0
  149. package/vault/device-anchor.ts +185 -0
  150. package/vault/device-control.ts +174 -0
  151. package/vault/handoff.ts +396 -0
  152. package/vault/index.ts +21 -0
  153. package/vault/passkey.ts +243 -0
  154. package/vault/phrase.ts +65 -0
  155. package/vault/pin.ts +153 -0
  156. package/vault/public-key-seal.ts +131 -0
  157. package/vault/seed.ts +272 -0
  158. package/vault/team.ts +355 -0
  159. package/vault/vault-item.ts +131 -0
  160. package/vault/vault-migrate-digest.ts +29 -0
  161. package/vault/wrapped.ts +77 -0
  162. package/wire/ARCHITECTURE.md +90 -0
  163. package/wire/CONTRIBUTING.md +19 -0
  164. package/wire/LICENSE +33 -0
  165. package/wire/README.md +51 -0
  166. package/wire/adapters/messages/anthropic-map.ts +37 -0
  167. package/wire/adapters/messages/index.ts +7 -0
  168. package/wire/adapters/messages/reasoning-from-items.ts +38 -0
  169. package/wire/adapters/messages/reasoning-signature.ts +243 -0
  170. package/wire/adapters/messages/request.ts +574 -0
  171. package/wire/adapters/messages/response.ts +215 -0
  172. package/wire/adapters/messages/streaming.ts +1027 -0
  173. package/wire/adapters/messages/usage.ts +43 -0
  174. package/wire/adapters/responses/index.ts +3 -0
  175. package/wire/adapters/responses/request.ts +395 -0
  176. package/wire/adapters/responses/response.ts +88 -0
  177. package/wire/adapters/responses/streaming.ts +344 -0
  178. package/wire/assets/openllm-light.svg +18 -0
  179. package/wire/assets/openllm.svg +18 -0
  180. package/wire/features/account-rank.ts +166 -0
  181. package/wire/features/compaction/compact-request.ts +862 -0
  182. package/wire/features/compaction/compaction-text.ts +43 -0
  183. package/wire/features/context-demote.ts +47 -0
  184. package/wire/features/context-overflow-strategy.ts +26 -0
  185. package/wire/features/context-skip.ts +170 -0
  186. package/wire/features/conversation-affinity.ts +144 -0
  187. package/wire/features/max-tokens-backfill.ts +141 -0
  188. package/wire/features/model-caps.ts +53 -0
  189. package/wire/features/model-settings.ts +24 -0
  190. package/wire/features/output-limit-clamp.ts +64 -0
  191. package/wire/features/quota-gate.ts +85 -0
  192. package/wire/index.ts +44 -0
  193. package/wire/lib/attempt-outcome.ts +43 -0
  194. package/wire/lib/canonical/content-part.ts +38 -0
  195. package/wire/lib/canonical/count-tokens-capability.ts +138 -0
  196. package/wire/lib/canonical/encoding-select.ts +82 -0
  197. package/wire/lib/canonical/image-signature.ts +202 -0
  198. package/wire/lib/canonical/image-sse.ts +216 -0
  199. package/wire/lib/canonical/message.ts +43 -0
  200. package/wire/lib/canonical/token-estimate.ts +266 -0
  201. package/wire/lib/encrypted-content.ts +66 -0
  202. package/wire/lib/error-class.ts +154 -0
  203. package/wire/lib/error-envelope.ts +56 -0
  204. package/wire/lib/forwarded-headers.ts +180 -0
  205. package/wire/lib/provider-error-rules.ts +48 -0
  206. package/wire/lib/refusal.ts +38 -0
  207. package/wire/lib/responses-apply-patch.ts +104 -0
  208. package/wire/lib/responses-namespace.ts +53 -0
  209. package/wire/lib/secret-scrub.ts +101 -0
  210. package/wire/lib/streaming/accumulate.ts +221 -0
  211. package/wire/lib/streaming/finish-reason.ts +33 -0
  212. package/wire/lib/streaming/peek.ts +154 -0
  213. package/wire/lib/streaming/provider-decode.ts +259 -0
  214. package/wire/lib/streaming/response-stream.ts +71 -0
  215. package/wire/lib/streaming/sse.ts +699 -0
  216. package/wire/lib/streaming/strip-tool-calls.ts +95 -0
  217. package/wire/lib/streaming/upstream-error.ts +58 -0
  218. package/wire/lib/strip-subagent-isolation.ts +284 -0
  219. package/wire/lib/thinking-signature.ts +156 -0
  220. package/wire/lib/token-ruler/bpe-core.ts +426 -0
  221. package/wire/lib/token-ruler/index.ts +2 -0
  222. package/wire/lib/token-ruler/ruler.ts +79 -0
  223. package/wire/lib/token-ruler/vocab/claude-ranks.ts +16 -0
  224. package/wire/lib/token-ruler/vocab/claude.ts +22 -0
  225. package/wire/lib/token-ruler/vocab/o200k-ranks.ts +18 -0
  226. package/wire/lib/token-ruler/vocab/o200k.ts +19 -0
  227. package/wire/lib/tool-name-map.ts +271 -0
  228. package/wire/lib/tool-schema.ts +353 -0
  229. package/wire/lib/usage-observer.ts +125 -0
  230. package/wire/providers/anthropic/adaptive-thinking.ts +291 -0
  231. package/wire/providers/anthropic/beta-headers.ts +165 -0
  232. package/wire/providers/anthropic/index.ts +10 -0
  233. package/wire/providers/anthropic/request.ts +607 -0
  234. package/wire/providers/anthropic/response.ts +155 -0
  235. package/wire/providers/anthropic/streaming.ts +315 -0
  236. package/wire/providers/anthropic/web-search-domains.ts +49 -0
  237. package/wire/providers/chatgpt/common.ts +136 -0
  238. package/wire/providers/chatgpt/index.ts +9 -0
  239. package/wire/providers/chatgpt/request.ts +959 -0
  240. package/wire/providers/chatgpt/streaming.ts +1007 -0
  241. package/wire/providers/chatgpt/web-search.ts +63 -0
  242. package/wire/providers/grok/web-search.ts +87 -0
  243. package/wire/providers/kimi/web-search.ts +145 -0
  244. package/wire/providers/upstream-deny.ts +355 -0
  245. package/wire/providers/upstream-request.ts +619 -0
@@ -0,0 +1,64 @@
1
+ /**
2
+ * Lower a wire request's output limit to a cap hold's budget — the one
3
+ * field-level operation the cloud tunnel path and the daemon walker share
4
+ * (review round 3, E1). Every output spelling present is clamped; when none
5
+ * is, the surface's own spelling is added. The core dispatch chain clamps
6
+ * its canonical request separately, because it is also bounded by the
7
+ * client's own fields after model settings and output backfill.
8
+ */
9
+ export type TOutputLimitSurface = "chat_completions" | "messages" | "responses";
10
+
11
+ const OUTPUT_FIELDS = [
12
+ "max_tokens",
13
+ "max_completion_tokens",
14
+ "max_output_tokens",
15
+ ] as const;
16
+
17
+ /** A copy of `body` with its output limit at most `limit`. */
18
+ export const clampOutputLimitFields = (
19
+ body: Readonly<Record<string, unknown>>,
20
+ surface: TOutputLimitSurface,
21
+ limit: number,
22
+ ): Record<string, unknown> => {
23
+ const result: Record<string, unknown> = { ...body };
24
+ let lowest = limit;
25
+ for (const key of OUTPUT_FIELDS) {
26
+ const value = result[key];
27
+ if (value === undefined || value === null) continue;
28
+ const clamped =
29
+ typeof value === "number" && Number.isFinite(value)
30
+ ? Math.min(value, limit)
31
+ : limit;
32
+ result[key] = clamped;
33
+ lowest = Math.min(lowest, clamped);
34
+ }
35
+ const target = surface === "responses" ? "max_output_tokens" : "max_tokens";
36
+ const hasTarget =
37
+ result[target] != null ||
38
+ (surface === "chat_completions" && result.max_completion_tokens != null);
39
+ if (!hasTarget) result[target] = lowest;
40
+ return result;
41
+ };
42
+
43
+ export const isOutputLimitSurface = (
44
+ surface: string,
45
+ ): surface is TOutputLimitSurface =>
46
+ surface === "chat_completions" ||
47
+ surface === "messages" ||
48
+ surface === "responses";
49
+
50
+ /** Parse and clamp a JSON wire body, rejecting malformed or non-object input. */
51
+ export const clampOutputLimitBytes = (
52
+ body: Uint8Array,
53
+ surface: TOutputLimitSurface,
54
+ limit: number,
55
+ ): Uint8Array => {
56
+ const parsed: unknown = JSON.parse(new TextDecoder().decode(body));
57
+ if (parsed === null || typeof parsed !== "object" || Array.isArray(parsed))
58
+ throw new Error("Invalid team request body");
59
+ return new TextEncoder().encode(
60
+ JSON.stringify(
61
+ clampOutputLimitFields(parsed as Record<string, unknown>, surface, limit),
62
+ ),
63
+ );
64
+ };
@@ -0,0 +1,85 @@
1
+ import type {
2
+ TProviderUsageSnapshot,
3
+ TSubscriptionMeterMatch,
4
+ } from "@openllmsh/sdk/protocol";
5
+ import { matchesSubscriptionMeter } from "@openllmsh/sdk/protocol";
6
+
7
+ /** Maximum age for a stale quota snapshot to remain eligible for routing. */
8
+ export const GATE_STALE_CAP_MS = 10 * 60_000;
9
+
10
+ export type TQuotaGateDecision =
11
+ | { readonly kind: "allow" }
12
+ | { readonly kind: "skip"; readonly reason: string };
13
+
14
+ type TQuotaPool = Extract<
15
+ TProviderUsageSnapshot,
16
+ { readonly kind: "quota" }
17
+ >["windows"][number];
18
+
19
+ /**
20
+ * May a STALE snapshot's exhausted `pool` still gate routing?
21
+ *
22
+ * A pool whose reset instant is KNOWN and still ahead is self-validating: the
23
+ * window cannot have refilled before it resets, so the read's age is
24
+ * irrelevant — it gates until that instant, then stops. Age-capping this case
25
+ * made the gate useless for precisely the provider that needs it: an exhausted
26
+ * account serves no successful request, so nothing re-samples its usage, so its
27
+ * snapshot sits permanently past the cap and the dead hop is dialled on every
28
+ * request (grok weekly at 100% with a 6-day-out reset, read 20 min earlier →
29
+ * dialled → HTTP 402 "usage balance exhausted").
30
+ *
31
+ * With NO reset instant there is nothing to validate against, so the freshness
32
+ * cap remains the only evidence and an aged read is not trusted.
33
+ */
34
+ const stalePoolIsGateable = (
35
+ snapshot: Extract<TProviderUsageSnapshot, { readonly kind: "quota" }>,
36
+ pool: TQuotaPool | undefined,
37
+ staleCapMs: number,
38
+ now: number,
39
+ ): boolean => {
40
+ if (!snapshot.stale) return true;
41
+ if (pool?.reset_at_ms !== null && pool?.reset_at_ms !== undefined) {
42
+ return now < pool.reset_at_ms;
43
+ }
44
+ return (
45
+ snapshot.as_of_ms !== undefined && now - snapshot.as_of_ms <= staleCapMs
46
+ );
47
+ };
48
+
49
+ export const quotaGateDecision = (params: {
50
+ readonly snapshot: TProviderUsageSnapshot | null;
51
+ readonly meter: TSubscriptionMeterMatch | undefined;
52
+ readonly finalHop: boolean;
53
+ readonly staleCapMs: number;
54
+ readonly now: number;
55
+ }): TQuotaGateDecision => {
56
+ if (params.finalHop || params.snapshot?.kind !== "quota") {
57
+ return { kind: "allow" };
58
+ }
59
+
60
+ const { snapshot } = params;
61
+ // A rejected quota snapshot derives from exhausted shared provider windows.
62
+ // For fresh snapshots that vendor verdict is current. For stale snapshots,
63
+ // assess EACH exhausted window independently: an old 5-hour window that has
64
+ // reset cannot erase the still-future 7-day exhaustion in the same snapshot.
65
+ const exhaustedProviderWindow = snapshot.windows.find(
66
+ (window) =>
67
+ window.percent_used >= 100 &&
68
+ stalePoolIsGateable(snapshot, window, params.staleCapMs, params.now),
69
+ );
70
+ if (snapshot.status === "rejected" && exhaustedProviderWindow !== undefined) {
71
+ return { kind: "skip", reason: "quota_skip: provider window exhausted" };
72
+ }
73
+
74
+ const exhaustedPool = snapshot.extra_pools?.find(
75
+ (pool) =>
76
+ matchesSubscriptionMeter(params.meter, pool.meter_id) &&
77
+ pool.percent_used >= 100 &&
78
+ stalePoolIsGateable(snapshot, pool, params.staleCapMs, params.now),
79
+ );
80
+ if (exhaustedPool !== undefined) {
81
+ return { kind: "skip", reason: "quota_skip: model meter exhausted" };
82
+ }
83
+
84
+ return { kind: "allow" };
85
+ };
package/wire/index.ts ADDED
@@ -0,0 +1,44 @@
1
+ /**
2
+ * `@openllmsh/sdk/wire` — pure wire-format transforms shared by the cloud
3
+ * pipeline (`@openllm/core`) and the local daemon. Depends only on
4
+ * `@openllmsh/sdk/protocol`: no Effect-DI, no provider registry, no `fetch`.
5
+ *
6
+ * These are the functions a coreless daemon needs to adapt a request
7
+ * to a hop's wire, encode the response back, and synthesize usage —
8
+ * carried verbatim from `core` so the matrix tests pin them unchanged.
9
+ */
10
+ export * from "./adapters/messages/reasoning-from-items";
11
+ export * from "./adapters/messages/reasoning-signature";
12
+ export * from "./adapters/messages/request";
13
+ export * from "./adapters/messages/response";
14
+ export * from "./adapters/messages/streaming";
15
+ export * from "./features/account-rank";
16
+ export * from "./features/compaction/compact-request";
17
+ export * from "./features/compaction/compaction-text";
18
+ export * from "./features/context-demote";
19
+ export * from "./features/context-overflow-strategy";
20
+ export * from "./features/context-skip";
21
+ export * from "./features/conversation-affinity";
22
+ export * from "./features/max-tokens-backfill";
23
+ export * from "./features/model-caps";
24
+ export * from "./features/output-limit-clamp";
25
+ export * from "./features/quota-gate";
26
+ export * from "./lib/attempt-outcome";
27
+ export * from "./lib/canonical/content-part";
28
+ export * from "./lib/canonical/count-tokens-capability";
29
+ export * from "./lib/canonical/message";
30
+ export * from "./lib/canonical/token-estimate";
31
+ export * from "./lib/encrypted-content";
32
+ export * from "./lib/error-class";
33
+ export * from "./lib/error-envelope";
34
+ export * from "./lib/provider-error-rules";
35
+ export * from "./lib/refusal";
36
+ export * from "./lib/secret-scrub";
37
+ export * from "./lib/streaming/accumulate";
38
+ export * from "./lib/streaming/peek";
39
+ export * from "./lib/streaming/sse";
40
+ export * from "./lib/streaming/strip-tool-calls";
41
+ export * from "./lib/streaming/upstream-error";
42
+ export * from "./lib/strip-subagent-isolation";
43
+ export * from "./lib/thinking-signature";
44
+ export * from "./lib/usage-observer";
@@ -0,0 +1,43 @@
1
+ export type TAttemptUsage = {
2
+ readonly in: number;
3
+ readonly out: number;
4
+ /** Cache detail for logging and cost only; `in` already includes it where
5
+ * the provider counts it separately. Never part of the cap charge. */
6
+ readonly cached?: number;
7
+ readonly cacheCreation?: number;
8
+ };
9
+
10
+ export type TAttemptOutcome = {
11
+ readonly sent: boolean;
12
+ readonly usage: TAttemptUsage | null;
13
+ readonly definiteProviderError: boolean;
14
+ };
15
+
16
+ /** A successfully read nonempty body is evidence, distinct from JSON validity. */
17
+ export const isReadableProviderBody = (text: string): boolean =>
18
+ text.trim().length > 0;
19
+
20
+ /** Classify an attempt only after validating any reported usage. */
21
+ export const classifyAttemptOutcome = (
22
+ sent: boolean,
23
+ status: number | null,
24
+ bodyRead: boolean,
25
+ usage: TAttemptUsage | null,
26
+ ): TAttemptOutcome => ({
27
+ sent,
28
+ usage,
29
+ definiteProviderError:
30
+ bodyRead &&
31
+ status !== null &&
32
+ [400, 401, 403, 404, 413, 415, 422].includes(status),
33
+ });
34
+
35
+ /** The charge for one attempt; an unsent attempt always refunds its hold. */
36
+ export const chargeForAttempt = (
37
+ outcome: TAttemptOutcome,
38
+ reservation: number,
39
+ ): number => {
40
+ if (!outcome.sent) return 0;
41
+ if (outcome.usage !== null) return outcome.usage.in + outcome.usage.out;
42
+ return outcome.definiteProviderError ? 0 : reservation;
43
+ };
@@ -0,0 +1,38 @@
1
+ import type { TChatCompletionRequest } from "@openllmsh/sdk/protocol";
2
+
3
+ /**
4
+ * The canonical content parts the inbound adapters PRODUCE — the
5
+ * OpenAI-native subset of `ContentPart` in `packages/sdk/protocol/chat.ts`
6
+ * (no `input_audio`: nothing inbound maps to it). Shared by the
7
+ * `adapters/messages` and `adapters/responses` request adapters so the
8
+ * two can't drift apart.
9
+ */
10
+ export type TCanonicalContentPart =
11
+ | { readonly type: "text"; readonly text: string }
12
+ | {
13
+ readonly type: "image_url";
14
+ readonly image_url: {
15
+ readonly url: string;
16
+ readonly detail?: "auto" | "low" | "high";
17
+ };
18
+ }
19
+ | {
20
+ readonly type: "file";
21
+ readonly file: {
22
+ readonly file_data?: string;
23
+ readonly file_id?: string;
24
+ readonly filename?: string;
25
+ };
26
+ };
27
+
28
+ /**
29
+ * True iff any `messages[].content[]` part is `type === "image_url"`.
30
+ * Walks every role (user / assistant / tool) — the quoted serde error
31
+ * is `messages[4]`. String content, null assistant content, and `file`
32
+ * parts are not images.
33
+ */
34
+ export const requestHasImageContent = (req: TChatCompletionRequest): boolean =>
35
+ req.messages.some((m) => {
36
+ if (typeof m.content === "string") return false;
37
+ return (m.content ?? []).some((part) => part.type === "image_url");
38
+ });
@@ -0,0 +1,138 @@
1
+ /**
2
+ * Bounded in-process memory of upstreams that do not expose
3
+ * `/v1/messages/count_tokens`. Client-facing preflight must still answer
4
+ * `{ input_tokens }` via the local BPE; it must not re-dial an endpoint
5
+ * already classified as unsupported.
6
+ *
7
+ * Skipping *auth/decrypt* is only possible when the count URL is known
8
+ * without a credential (cloud catalog Anthropic `defaultBaseUrl`). The
9
+ * daemon's captured inference URL is learned from the signed-plan
10
+ * delegate, so `acquireUpstream` still runs; the memory then skips the
11
+ * count HTTP. Identity is never taken from an unsigned caller URL.
12
+ *
13
+ * Scope is provider + endpoint identity (origin + pathname only).
14
+ * Secrets, userinfo, query strings, and API keys never enter the key.
15
+ * Custom endpoints stay isolated from each other and from catalog
16
+ * providers. Model id is NOT part of the key: a missing count *route*
17
+ * is an endpoint property; a missing *model* is not, and must not
18
+ * poison the endpoint.
19
+ */
20
+
21
+ export const COUNT_TOKENS_CAPABILITY_LIMIT = 256;
22
+
23
+ export type TCountTokensCapabilityScope = {
24
+ readonly provider: string;
25
+ readonly endpointIdentity: string;
26
+ };
27
+
28
+ /**
29
+ * Stable identity for a count URL: origin + pathname. Unparseable input
30
+ * (and anything that is not http/https) yields `""` — never the raw
31
+ * string, which may carry query tokens or userinfo.
32
+ */
33
+ export const countTokensEndpointIdentity = (countUrl: string): string => {
34
+ try {
35
+ const u = new URL(countUrl);
36
+ if (u.protocol !== "http:" && u.protocol !== "https:") return "";
37
+ return `${u.origin}${u.pathname}`;
38
+ } catch {
39
+ return "";
40
+ }
41
+ };
42
+
43
+ export type TCountTokensCapabilityMemory = {
44
+ readonly isUnsupported: (scope: TCountTokensCapabilityScope) => boolean;
45
+ readonly remember: (scope: TCountTokensCapabilityScope) => void;
46
+ };
47
+
48
+ /**
49
+ * Build a capability memory with its own LRU. Production uses the module
50
+ * default via {@link isCountTokensUnsupported} /
51
+ * {@link rememberCountTokensUnsupported}; tests build a fresh one per case.
52
+ */
53
+ export const createCountTokensCapabilityMemory = (
54
+ limit: number = COUNT_TOKENS_CAPABILITY_LIMIT,
55
+ ): TCountTokensCapabilityMemory => {
56
+ const unsupported = new Map<string, true>();
57
+ const scopeKey = (scope: TCountTokensCapabilityScope): string =>
58
+ `${scope.provider}\0${scope.endpointIdentity}`;
59
+
60
+ return {
61
+ isUnsupported(scope) {
62
+ if (scope.endpointIdentity.length === 0) return false;
63
+ const k = scopeKey(scope);
64
+ if (!unsupported.has(k)) return false;
65
+ unsupported.delete(k);
66
+ unsupported.set(k, true);
67
+ return true;
68
+ },
69
+ remember(scope) {
70
+ if (scope.endpointIdentity.length === 0) return;
71
+ const k = scopeKey(scope);
72
+ if (unsupported.has(k)) {
73
+ unsupported.delete(k);
74
+ unsupported.set(k, true);
75
+ return;
76
+ }
77
+ if (unsupported.size >= limit) {
78
+ const oldest = unsupported.keys().next().value;
79
+ if (oldest !== undefined) unsupported.delete(oldest);
80
+ }
81
+ unsupported.set(k, true);
82
+ },
83
+ };
84
+ };
85
+
86
+ const defaultMemory = createCountTokensCapabilityMemory();
87
+
88
+ export const isCountTokensUnsupported = (
89
+ scope: TCountTokensCapabilityScope,
90
+ ): boolean => defaultMemory.isUnsupported(scope);
91
+
92
+ export const rememberCountTokensUnsupported = (
93
+ scope: TCountTokensCapabilityScope,
94
+ ): void => {
95
+ defaultMemory.remember(scope);
96
+ };
97
+
98
+ const modelMissing = (t: string): boolean =>
99
+ /\bmodel\b.{0,120}\b(not found|unknown|does not exist|unsupported|invalid)\b/.test(
100
+ t,
101
+ ) ||
102
+ /\b(not found|unknown|does not exist|unsupported|invalid).{0,120}\bmodel\b/.test(
103
+ t,
104
+ );
105
+
106
+ const countRouteMissing = (t: string): boolean =>
107
+ /\b(unknown|no such|missing)\s+(endpoint|route|path|url)\b/.test(t) ||
108
+ /\b(endpoint|route|path|url)\s+(is\s+)?(not found|unknown|does not exist|missing)\b/.test(
109
+ t,
110
+ ) ||
111
+ /\bcount_tokens\s+(endpoint|route|path)\s+(not found|does not exist|unknown|not implemented|missing)\b/.test(
112
+ t,
113
+ ) ||
114
+ (t.includes("not implemented") &&
115
+ (t.includes("count_tokens") || t.includes("count tokens"))) ||
116
+ t.includes("method not allowed");
117
+
118
+ /**
119
+ * Classify an already-attempted count_tokens HTTP response as "this
120
+ * endpoint does not provide the route" vs everything else.
121
+ *
122
+ * 501 / 405 are unambiguous. 404 is only unsupported when the body
123
+ * names a missing *route/endpoint/method* — not a missing *model*,
124
+ * even if the same sentence mentions `count_tokens`.
125
+ * Auth, rate-limit, and server errors must never be remembered.
126
+ */
127
+ export const countTokensUnsupportedFromUpstream = (
128
+ status: number,
129
+ bodyText: string,
130
+ ): boolean => {
131
+ if (status === 401 || status === 403 || status === 429) return false;
132
+ if (status === 501 || status === 405) return true;
133
+ if (status >= 500) return false;
134
+ if (status !== 404) return false;
135
+ const t = bodyText.toLowerCase();
136
+ if (modelMissing(t)) return false;
137
+ return countRouteMissing(t);
138
+ };
@@ -0,0 +1,82 @@
1
+ /**
2
+ * The token-encoding ruler behind the estimators in `./token-estimate.ts`.
3
+ *
4
+ * A token count is only defined relative to ONE model's vocabulary, so the
5
+ * estimators need to know which family they're counting for. `wire` can't
6
+ * depend on the catalog (it's `schema`-only), so callers resolve their hop's
7
+ * provider to one of these coarse families and pass it in; the map from
8
+ * provider → family lives with the catalog in `packages/api`.
9
+ *
10
+ * The ruler itself is our own count-only BPE core over vendored vocab data —
11
+ * see `../token-ruler`. It replaced the `ai-tokenizer` dependency: same exact
12
+ * counts, but a ~15× smaller resident footprint (o200k: +47 MB heap / +42 MB
13
+ * external → +0.1 MB / +4.8 MB) and no native/WASM glue, so it compiles into
14
+ * the Bun daemon binary unchanged. This module is a thin family ↔ surface
15
+ * adapter over that core; the heavy vocab is imported lazily inside the ruler,
16
+ * so an estimate that never runs never pays the parse cost.
17
+ */
18
+ import type { TRulerCache, TRulerFamily } from "../token-ruler";
19
+ import { createRulerCache, getRuler, peekRuler } from "../token-ruler";
20
+
21
+ /**
22
+ * The coarse tokenizer families we select between. NOT provider ids — several
23
+ * providers share a ruler:
24
+ * - `claude` — Anthropic's reverse-engineered BPE (Claude 3/4.x).
25
+ * - `o200k` — OpenAI o200k_base; also the best local stand-in for the
26
+ * tiktoken-derived vocabs (Qwen / Kimi / DeepSeek / GLM), which
27
+ * land within a few %. The DEFAULT when a caller has no better
28
+ * information.
29
+ *
30
+ * This is exactly the ruler's {@link TRulerFamily}; the alias keeps the
31
+ * estimator-facing name stable.
32
+ */
33
+ export type TTokenEncoding = TRulerFamily;
34
+
35
+ export const DEFAULT_ENCODING: TTokenEncoding = "o200k";
36
+
37
+ /**
38
+ * Pick the ruler family for a request surface. `messages` is Anthropic-wire, so
39
+ * the Claude ruler is the accurate one; `chat_completions` and `responses` are
40
+ * OpenAI-family (o200k also being the best local stand-in for the tiktoken-
41
+ * derived Qwen/Kimi/DeepSeek vocabs). Shared by the cloud handler and the daemon
42
+ * so the same body is measured with the same ruler on both paths.
43
+ */
44
+ export const encodingForSurface = (
45
+ surface: "messages" | "chat_completions" | "responses",
46
+ ): TTokenEncoding => (surface === "messages" ? "claude" : "o200k");
47
+
48
+ /**
49
+ * A minimal counter — just the `count(text)` we need — so the rest of the
50
+ * module doesn't couple to the ruler's class shape.
51
+ */
52
+ export type TTokenCounter = {
53
+ readonly count: (text: string) => number;
54
+ };
55
+
56
+ export type { TRulerCache };
57
+ export { createRulerCache };
58
+
59
+ /**
60
+ * Resolve (and cache) the counter for a family. Async because the vocab data is
61
+ * imported lazily — the first call per family builds a ~65k–200k-entry lookup;
62
+ * every call after is a cache hit. Concurrent first-calls are deduped inside the
63
+ * ruler. Pass a {@link TRulerCache} to use a non-default cache (tests).
64
+ */
65
+ export const getTokenCounter = (
66
+ encoding: TTokenEncoding,
67
+ cache?: TRulerCache,
68
+ ): Promise<TTokenCounter> =>
69
+ (cache ?? { get: getRuler, peek: peekRuler }).get(encoding);
70
+
71
+ /**
72
+ * Synchronously return an already-loaded counter, or `null` if the family
73
+ * hasn't been warmed yet. Used by the cheap routing helpers
74
+ * (`estimateBodyTokensExact`) that must stay sync. Client-facing preflight
75
+ * (`estimateAnthropicInputTokens`) always awaits {@link getTokenCounter}
76
+ * instead of peeking.
77
+ */
78
+ export const peekTokenCounter = (
79
+ encoding: TTokenEncoding,
80
+ cache?: TRulerCache,
81
+ ): TTokenCounter | null =>
82
+ (cache ?? { get: getRuler, peek: peekRuler }).peek(encoding);