crosscheck-mcp 0.2.28 → 0.2.30

package/dist/embed.d.cts

@@ -1,3 +1,6 @@
+import { C as ChatMessage, I as ImagePart, S as SendArgs } from './types-HA3W47O4.cjs';
+export { a as IMAGE_MEDIA_TYPES } from './types-HA3W47O4.cjs';
+
 /**
  * A record of which engine processes are alive, and what they are RUNNING.
  *
@@ -96,6 +99,30 @@ declare const DEFAULT_FALLBACKS: Readonly<Record<string, readonly string[]>>;
  */
 declare const DEFAULT_EXCLUDED_PROVIDERS: readonly string[];
 
+/**
+ * The super lineup, as ORDERED CHAINS: index 0 is the model we want, the rest
+ * are fallbacks tried only on a model-access failure (never on auth, rate
+ * limit, or a 500 — a different model doesn't fix a bad key).
+ *
+ * Chains rather than single ids for one specific reason: it lets us name a
+ * model that hasn't shipped yet. `gpt-6-astra` is not in OpenAI's model list
+ * as of 2026-09-04 — verified against the live endpoint, no `gpt-6*` exists —
+ * so it 404s today and the chain silently serves gpt-5.6 instead. The day
+ * Astra becomes available it starts being used with no release on our side.
+ *
+ * The cost of that convenience, stated plainly: until Astra ships, an OpenAI
+ * seat in super mode pays one extra failed round trip per process per 5
+ * minutes (dead-models.ts caches the 404 for that long). A model-not-found is
+ * an immediate error rather than a timeout, so it's a few hundred ms, not a
+ * hang. If that ever becomes annoying, drop "gpt-6-astra" from the chain —
+ * nothing else depends on it.
+ *
+ * anthropic stays FIRST: resolveProviders preserves key order into `selected`,
+ * and debate's moderator / coordinate's proposer default to `selected[0]`.
+ * Reordering this object silently moves those seats to another provider.
+ */
+declare const SUPER_MODELS: Readonly<Record<string, readonly string[]>>;
+
 /**
  * Embedded host entrypoint — the API Crosscheck Desktop runs panels through.
  *
@@ -128,6 +155,7 @@ declare const DEFAULT_EXCLUDED_PROVIDERS: readonly string[];
  * inherits `process.env`. Keeping keys out of it is what stops a user's own
  * `env | grep KEY` from printing their Anthropic key inside our app.
  */
+
 /**
  * Vault provider id → engine env var.
  *
@@ -208,7 +236,62 @@ type EngineEvent =
     runId: string;
     at: number;
     message: string;
+}
+/** ACTIVITY — one provider request went out. Emitted for EVERY model call
+ *  a run makes, whatever the tool: panelists, the planner, workers,
+ *  reviewers, judges, synthesis, the co-reasoner. `purpose` is the
+ *  engine's role tag; `asked` is the start of the last user message. */
+ | {
+    type: "call.start";
+    runId: string;
+    at: number;
+    callId: number;
+    provider: string;
+    model: string;
+    purpose: string;
+    asked: string;
+}
+/** ACTIVITY — that request resolved. `text` is what the model said (capped
+ *  at ACTIVITY_TEXT_CAP; `truncated` says so). Finished text, not a token
+ *  stream: no adapter streams, and pretending otherwise would be a lie. */
+ | {
+    type: "call.done";
+    runId: string;
+    at: number;
+    callId: number;
+    provider: string;
+    model: string;
+    purpose: string;
+    ok: boolean;
+    ms: number;
+    text: string;
+    truncated: boolean;
+    error: string | null;
+    usage: EventUsage | null;
+    /** Set when the host stopped waiting (embed only): "user" from
+     *  skip(), "timeout" from the run's per-call time limit. The request
+     *  was aborted; the provider may still bill for work already done. */
+    stopped?: "user" | "timeout";
+}
+/** ACTIVITY — a step message from a multi-stage tool (orchestrate's
+ *  levels, recombine). */
+ | {
+    type: "progress";
+    runId: string;
+    at: number;
+    message: string;
 };
+/** How much of each model reply an activity event carries. Enough to read
+ *  the back-and-forth; the full text is in the envelope. */
+declare const ACTIVITY_TEXT_CAP = 8000;
+/** Providers whose models can see images (vault ids). Everyone else is told
+ *  an image exists that it can't see, rather than being sent one it would
+ *  reject — or silently answering as if there were none. */
+declare const VISION_PROVIDERS: readonly string[];
+declare const MAX_RUN_IMAGES = 8;
+/** Put the run's images on the newest user message — or, for a provider
+ *  that can't see images, a plain note saying they exist. */
+declare function withRunImages(args: SendArgs, provider: string, images: readonly ImagePart[] | undefined): SendArgs;
 type EngineEventSink = (event: EngineEvent) => void;
 interface EmbeddedEngineOptions {
     /** Vault provider id → raw API key. Only providers present here are built.
@@ -217,6 +300,9 @@ interface EmbeddedEngineOptions {
     /** Vault provider id → pinned model. Omit a provider to take the engine
      *  default. Mirrors what `/account/models` and `crosscheck models set` do. */
     models?: Readonly<Record<string, string>>;
+    /** Vault provider id → pinned SUPER-tier model (what `confer super` uses).
+     *  Omit a provider to take the engine's super lineup (SUPER_MODELS). */
+    superModels?: Readonly<Record<string, string>>;
     /** Default moderator/synthesis provider. */
     moderator?: string;
     /** Explicit pricing.json path. Defaults to the copy bundled in `dist/`. */
@@ -224,6 +310,37 @@ interface EmbeddedEngineOptions {
     /** Directory to archive run transcripts into. Omit to write none. */
     transcriptsDir?: string;
 }
+interface RunOptions {
+    /** Stop waiting on any single model call after this long (ms). The run
+     *  continues with the answers already in. Off when 0/undefined. */
+    callTimeoutMs?: number;
+    /** Images for this run (base64, png/jpeg/webp/gif; at most 8). Sent with
+     *  the question to every provider that can see images; the others are
+     *  told they exist. Embed-only — the MCP tools don't take images. */
+    images?: ImagePart[];
+}
+interface AskOptions {
+    /** Vault id ("grok", not "xai"). Must be a provider this engine built. */
+    provider: string;
+    /** Model override; defaults to the provider's effective model. */
+    model?: string;
+    messages: ChatMessage[];
+    schema?: Record<string, unknown>;
+    /** Role tag for usage + the activity feed, e.g. "unify" or "code". */
+    purpose: string;
+    maxTokens?: number;
+    runId?: string;
+    onEvent?: EngineEventSink;
+}
+interface AskResult {
+    ok: boolean;
+    /** The model's raw reply. */
+    text: string;
+    /** Schema-validated value when a schema was given and the reply passed. */
+    json: unknown | null;
+    error: string | null;
+    model: string;
+}
 interface EmbeddedEngine {
     readonly version: string;
     /** Providers actually built, by vault id — i.e. the ones with a usable key
@@ -233,12 +350,27 @@ interface EmbeddedEngine {
     readonly models: Readonly<Record<string, string>>;
     /** Every registered tool name. */
     listTools(): string[];
+    /** A tool's required input fields, from its own schema — or null if there
+     *  is no such tool. Lets a host test its argument mapping against the
+     *  engine instead of guessing field names (the desktop guessed five of ten
+     *  wrong, and every one of those tools failed on missing input). */
+    requiredArgs(name: string): string[] | null;
     /** Invoke any tool and wait for the final envelope. No events. */
     call(name: string, args: Record<string, unknown>): Promise<unknown>;
     /** Invoke a tool, emitting lifecycle events as the panel progresses.
      *  `confer` and `debate` stream; every other tool emits `run.start` and
      *  `run.done` around a plain `call`, so a host can use one code path. */
-    run(name: string, args: Record<string, unknown>, onEvent: EngineEventSink): Promise<unknown>;
+    run(name: string, args: Record<string, unknown>, onEvent: EngineEventSink, opts?: RunOptions): Promise<unknown>;
+    /** Stop waiting on one provider for the rest of a run: its in-flight
+     *  requests are aborted and later ones (e.g. debate rounds) aren't sent.
+     *  The run carries on with everyone else. The run's moderator is never
+     *  stopped mid-synthesis. Returns how many in-flight calls were stopped;
+     *  0 (and a no-op) once the run has finished. */
+    skip(runId: string, provider: string): number;
+    /** One call to one chosen provider/model — the host's own moderator or
+     *  coding step. With `runId` + `onEvent` it reports into that run's
+     *  activity feed. With `schema`, returns validated JSON when it can. */
+    ask(opts: AskOptions): Promise<AskResult>;
     /** Flush queued telemetry. Call before the host process exits. */
     close(): Promise<void>;
 }
@@ -259,4 +391,4 @@ declare const __test_internals: {
     vaultId: typeof vaultId;
 };
 
-export { DEFAULT_EXCLUDED_PROVIDERS, DEFAULT_FALLBACKS, DEFAULT_MODELS, type EmbeddedEngine, type EmbeddedEngineOptions, type EngineEvent, type EngineEventSink, type EventUsage, type ProcRecord, type ProcStatus, KEY_ENV as __KEY_ENV, MODEL_ENV as __MODEL_ENV, __test_internals, createEmbeddedEngine, isAlive, listProcesses, processRegistryPath };
+export { ACTIVITY_TEXT_CAP, type AskOptions, type AskResult, DEFAULT_EXCLUDED_PROVIDERS, DEFAULT_FALLBACKS, DEFAULT_MODELS, type EmbeddedEngine, type EmbeddedEngineOptions, type EngineEvent, type EngineEventSink, type EventUsage, ImagePart, MAX_RUN_IMAGES, type ProcRecord, type ProcStatus, type RunOptions, SUPER_MODELS, VISION_PROVIDERS, KEY_ENV as __KEY_ENV, MODEL_ENV as __MODEL_ENV, __test_internals, createEmbeddedEngine, isAlive, listProcesses, processRegistryPath, withRunImages };

package/dist/embed.d.ts

@@ -1,3 +1,6 @@
+import { C as ChatMessage, I as ImagePart, S as SendArgs } from './types-HA3W47O4.js';
+export { a as IMAGE_MEDIA_TYPES } from './types-HA3W47O4.js';
+
 /**
  * A record of which engine processes are alive, and what they are RUNNING.
  *
@@ -96,6 +99,30 @@ declare const DEFAULT_FALLBACKS: Readonly<Record<string, readonly string[]>>;
  */
 declare const DEFAULT_EXCLUDED_PROVIDERS: readonly string[];
 
+/**
+ * The super lineup, as ORDERED CHAINS: index 0 is the model we want, the rest
+ * are fallbacks tried only on a model-access failure (never on auth, rate
+ * limit, or a 500 — a different model doesn't fix a bad key).
+ *
+ * Chains rather than single ids for one specific reason: it lets us name a
+ * model that hasn't shipped yet. `gpt-6-astra` is not in OpenAI's model list
+ * as of 2026-09-04 — verified against the live endpoint, no `gpt-6*` exists —
+ * so it 404s today and the chain silently serves gpt-5.6 instead. The day
+ * Astra becomes available it starts being used with no release on our side.
+ *
+ * The cost of that convenience, stated plainly: until Astra ships, an OpenAI
+ * seat in super mode pays one extra failed round trip per process per 5
+ * minutes (dead-models.ts caches the 404 for that long). A model-not-found is
+ * an immediate error rather than a timeout, so it's a few hundred ms, not a
+ * hang. If that ever becomes annoying, drop "gpt-6-astra" from the chain —
+ * nothing else depends on it.
+ *
+ * anthropic stays FIRST: resolveProviders preserves key order into `selected`,
+ * and debate's moderator / coordinate's proposer default to `selected[0]`.
+ * Reordering this object silently moves those seats to another provider.
+ */
+declare const SUPER_MODELS: Readonly<Record<string, readonly string[]>>;
+
 /**
  * Embedded host entrypoint — the API Crosscheck Desktop runs panels through.
  *
@@ -128,6 +155,7 @@ declare const DEFAULT_EXCLUDED_PROVIDERS: readonly string[];
  * inherits `process.env`. Keeping keys out of it is what stops a user's own
  * `env | grep KEY` from printing their Anthropic key inside our app.
  */
+
 /**
  * Vault provider id → engine env var.
  *
@@ -208,7 +236,62 @@ type EngineEvent =
     runId: string;
     at: number;
     message: string;
+}
+/** ACTIVITY — one provider request went out. Emitted for EVERY model call
+ *  a run makes, whatever the tool: panelists, the planner, workers,
+ *  reviewers, judges, synthesis, the co-reasoner. `purpose` is the
+ *  engine's role tag; `asked` is the start of the last user message. */
+ | {
+    type: "call.start";
+    runId: string;
+    at: number;
+    callId: number;
+    provider: string;
+    model: string;
+    purpose: string;
+    asked: string;
+}
+/** ACTIVITY — that request resolved. `text` is what the model said (capped
+ *  at ACTIVITY_TEXT_CAP; `truncated` says so). Finished text, not a token
+ *  stream: no adapter streams, and pretending otherwise would be a lie. */
+ | {
+    type: "call.done";
+    runId: string;
+    at: number;
+    callId: number;
+    provider: string;
+    model: string;
+    purpose: string;
+    ok: boolean;
+    ms: number;
+    text: string;
+    truncated: boolean;
+    error: string | null;
+    usage: EventUsage | null;
+    /** Set when the host stopped waiting (embed only): "user" from
+     *  skip(), "timeout" from the run's per-call time limit. The request
+     *  was aborted; the provider may still bill for work already done. */
+    stopped?: "user" | "timeout";
+}
+/** ACTIVITY — a step message from a multi-stage tool (orchestrate's
+ *  levels, recombine). */
+ | {
+    type: "progress";
+    runId: string;
+    at: number;
+    message: string;
 };
+/** How much of each model reply an activity event carries. Enough to read
+ *  the back-and-forth; the full text is in the envelope. */
+declare const ACTIVITY_TEXT_CAP = 8000;
+/** Providers whose models can see images (vault ids). Everyone else is told
+ *  an image exists that it can't see, rather than being sent one it would
+ *  reject — or silently answering as if there were none. */
+declare const VISION_PROVIDERS: readonly string[];
+declare const MAX_RUN_IMAGES = 8;
+/** Put the run's images on the newest user message — or, for a provider
+ *  that can't see images, a plain note saying they exist. */
+declare function withRunImages(args: SendArgs, provider: string, images: readonly ImagePart[] | undefined): SendArgs;
 type EngineEventSink = (event: EngineEvent) => void;
 interface EmbeddedEngineOptions {
     /** Vault provider id → raw API key. Only providers present here are built.
@@ -217,6 +300,9 @@ interface EmbeddedEngineOptions {
     /** Vault provider id → pinned model. Omit a provider to take the engine
      *  default. Mirrors what `/account/models` and `crosscheck models set` do. */
     models?: Readonly<Record<string, string>>;
+    /** Vault provider id → pinned SUPER-tier model (what `confer super` uses).
+     *  Omit a provider to take the engine's super lineup (SUPER_MODELS). */
+    superModels?: Readonly<Record<string, string>>;
     /** Default moderator/synthesis provider. */
     moderator?: string;
     /** Explicit pricing.json path. Defaults to the copy bundled in `dist/`. */
@@ -224,6 +310,37 @@ interface EmbeddedEngineOptions {
     /** Directory to archive run transcripts into. Omit to write none. */
     transcriptsDir?: string;
 }
+interface RunOptions {
+    /** Stop waiting on any single model call after this long (ms). The run
+     *  continues with the answers already in. Off when 0/undefined. */
+    callTimeoutMs?: number;
+    /** Images for this run (base64, png/jpeg/webp/gif; at most 8). Sent with
+     *  the question to every provider that can see images; the others are
+     *  told they exist. Embed-only — the MCP tools don't take images. */
+    images?: ImagePart[];
+}
+interface AskOptions {
+    /** Vault id ("grok", not "xai"). Must be a provider this engine built. */
+    provider: string;
+    /** Model override; defaults to the provider's effective model. */
+    model?: string;
+    messages: ChatMessage[];
+    schema?: Record<string, unknown>;
+    /** Role tag for usage + the activity feed, e.g. "unify" or "code". */
+    purpose: string;
+    maxTokens?: number;
+    runId?: string;
+    onEvent?: EngineEventSink;
+}
+interface AskResult {
+    ok: boolean;
+    /** The model's raw reply. */
+    text: string;
+    /** Schema-validated value when a schema was given and the reply passed. */
+    json: unknown | null;
+    error: string | null;
+    model: string;
+}
 interface EmbeddedEngine {
     readonly version: string;
     /** Providers actually built, by vault id — i.e. the ones with a usable key
@@ -233,12 +350,27 @@ interface EmbeddedEngine {
     readonly models: Readonly<Record<string, string>>;
     /** Every registered tool name. */
     listTools(): string[];
+    /** A tool's required input fields, from its own schema — or null if there
+     *  is no such tool. Lets a host test its argument mapping against the
+     *  engine instead of guessing field names (the desktop guessed five of ten
+     *  wrong, and every one of those tools failed on missing input). */
+    requiredArgs(name: string): string[] | null;
     /** Invoke any tool and wait for the final envelope. No events. */
     call(name: string, args: Record<string, unknown>): Promise<unknown>;
     /** Invoke a tool, emitting lifecycle events as the panel progresses.
      *  `confer` and `debate` stream; every other tool emits `run.start` and
      *  `run.done` around a plain `call`, so a host can use one code path. */
-    run(name: string, args: Record<string, unknown>, onEvent: EngineEventSink): Promise<unknown>;
+    run(name: string, args: Record<string, unknown>, onEvent: EngineEventSink, opts?: RunOptions): Promise<unknown>;
+    /** Stop waiting on one provider for the rest of a run: its in-flight
+     *  requests are aborted and later ones (e.g. debate rounds) aren't sent.
+     *  The run carries on with everyone else. The run's moderator is never
+     *  stopped mid-synthesis. Returns how many in-flight calls were stopped;
+     *  0 (and a no-op) once the run has finished. */
+    skip(runId: string, provider: string): number;
+    /** One call to one chosen provider/model — the host's own moderator or
+     *  coding step. With `runId` + `onEvent` it reports into that run's
+     *  activity feed. With `schema`, returns validated JSON when it can. */
+    ask(opts: AskOptions): Promise<AskResult>;
     /** Flush queued telemetry. Call before the host process exits. */
     close(): Promise<void>;
 }
@@ -259,4 +391,4 @@ declare const __test_internals: {
     vaultId: typeof vaultId;
 };
 
-export { DEFAULT_EXCLUDED_PROVIDERS, DEFAULT_FALLBACKS, DEFAULT_MODELS, type EmbeddedEngine, type EmbeddedEngineOptions, type EngineEvent, type EngineEventSink, type EventUsage, type ProcRecord, type ProcStatus, KEY_ENV as __KEY_ENV, MODEL_ENV as __MODEL_ENV, __test_internals, createEmbeddedEngine, isAlive, listProcesses, processRegistryPath };
+export { ACTIVITY_TEXT_CAP, type AskOptions, type AskResult, DEFAULT_EXCLUDED_PROVIDERS, DEFAULT_FALLBACKS, DEFAULT_MODELS, type EmbeddedEngine, type EmbeddedEngineOptions, type EngineEvent, type EngineEventSink, type EventUsage, ImagePart, MAX_RUN_IMAGES, type ProcRecord, type ProcStatus, type RunOptions, SUPER_MODELS, VISION_PROVIDERS, KEY_ENV as __KEY_ENV, MODEL_ENV as __MODEL_ENV, __test_internals, createEmbeddedEngine, isAlive, listProcesses, processRegistryPath, withRunImages };