experimental-ash 0.24.2 → 0.25.0

package/dist/src/evals/cli/eval.d.ts

@@ -1,4 +1,3 @@
-import type { Command } from "commander";
 interface EvalCliOptions {
     suite?: string[];
     url?: string;
@@ -6,17 +5,12 @@ interface EvalCliOptions {
     maxConcurrency?: string;
     json?: boolean;
     all?: boolean;
-    listSuites?: boolean;
     skipReport?: boolean;
 }
 type EvalCliLogger = {
     log(message: string): void;
     error(message: string): void;
 };
-/**
- * Registers the `ash eval` command on the given Commander program.
- */
-export declare function registerEvalCommand(program: Command, logger: EvalCliLogger): void;
 /**
  * Runs the `ash eval` command with already-parsed Commander options.
  */

package/dist/src/evals/cli/eval.js

@@ -1 +1 @@
-import{n as e}from"../../chunks/paths-CebY5GCi.js";import{loadDevelopmentEnvironmentFiles as t}from"../../cli/dev/environment.js";import{a as n,n as r,t as i}from"../../chunks/client-DLHAGI2g.js";import{n as a}from"../../chunks/host-zBy9FyyX.js";import{discoverAndImportSuites as o,discoverSuiteFiles as s,importSuiteFile as c}from"../runner/discover.js";import{executeSuite as l}from"../runner/execute-suite.js";import{ConsoleReporter as u}from"../runner/reporters/console.js";var d=n();function f(e,t){e.command(`eval`).description(`Run eval suites against an Ash agent.`).option(`--suite <id...>`,`Suite IDs to run (repeatable)`).option(`--all`,`Run all discovered suites`).option(`--url <url>`,`Remote agent URL (skip local host startup)`).option(`--timeout <ms>`,`Per-case timeout in milliseconds`).option(`--max-concurrency <n>`,`Max concurrent case executions per suite`).option(`--json`,`Output results as JSON`).option(`--list-suites`,`List discovered suites and exit`).option(`--skip-report`,`Skip suite-defined reporters (e.g. Braintrust)`).action(async e=>{await p(e,t)})}async function p(n,r){let i=e();if(t(i),n.listSuites){await y(i,r);return}let s=n.suite,c=await o(i,s);if(c.length===0){s&&s.length>0?r.error(`No suites found matching: ${s.join(`, `)}`):r.error(`No eval suites found. Create suite files under evals/ with the *.eval.ts extension.`),process.exitCode=1;return}let u,d;n.url?d={kind:`remote`,url:n.url}:(u=await a(i,{host:`127.0.0.1`,port:0}),d={kind:`local`,url:u.url});let f=m(d);try{let e=[];for(let t of c){let r=_(t,n),a=v(r,{json:n.json===!0,skipReport:n.skipReport===!0}),o=await l({suite:r,target:d,reporters:a,appRoot:i,client:f});e.push(o)}n.json&&r.log(JSON.stringify(e,null,2)),e.some(e=>e.errored>0)&&(process.exitCode=1)}finally{u&&await u.close()}process.exit(process.exitCode??0)}function m(e){if(e.kind===`local`)return new i({host:e.url});let t={},n=process.env.VERCEL_AUTOMATION_BYPASS_SECRET?.trim();return n&&(t[r]=n),new i({auth:h(),headers:Object.keys(t).length>0?t:void 0,host:e.url})}function h(){let e=process.env.ASH_EVAL_AUTH_TOKEN?.trim();return e?{bearer:e}:{bearer:g}}async function g(){try{let e=(await(0,d.getVercelOidcToken)()).trim();if(e.length>0)return e}catch{}return process.env.VERCEL_OIDC_TOKEN?.trim()??``}function _(e,t){let n=t.maxConcurrency?Number.parseInt(t.maxConcurrency,10):void 0,r=t.timeout?Number.parseInt(t.timeout,10):void 0;if(n===void 0&&r===void 0)return e;let i={...e};return n!==void 0&&(i.maxConcurrency=n),r!==void 0&&(i.timeoutMs=r),i}function v(e,t){let n=t.json?[]:[new u];return!t.skipReport&&e.reporters&&n.push(...e.reporters),n}async function y(e,t){let n=await s(e);if(n.length===0){t.log(`No eval suites found.`);return}t.log(`Found ${n.length} eval suite file(s):\n`);for(let r of n){let n=await c(e,r);t.log(`  ${n.id}${n.description?`  - ${n.description}`:``}`)}}export{f as registerEvalCommand,p as runEvalCommand};
+import{n as e}from"../../chunks/paths-DNjq5JOy.js";import{loadDevelopmentEnvironmentFiles as t}from"../../cli/dev/environment.js";import{n,s as r,t as i}from"../../chunks/client-BShLWzR6.js";import{n as a}from"../../chunks/host-JVy7fewA.js";import{discoverAndImportSuites as o}from"../runner/discover.js";import{executeSuite as s}from"../runner/execute-suite.js";import{ConsoleReporter as c}from"../runner/reporters/console.js";var l=r();async function u(n,r){let i=e();t(i);let c=n.suite,l=await o(i,c);if(l.length===0){c&&c.length>0?r.error(`No suites found matching: ${c.join(`, `)}`):r.error(`No eval suites found. Create suite files under evals/ with the *.eval.ts extension.`),process.exitCode=1;return}let u,f;n.url?f={kind:`remote`,url:n.url}:(u=await a(i,{host:`127.0.0.1`,port:0}),f={kind:`local`,url:u.url});let p=d(f);try{let e=[];for(let t of l){let r=m(t,n),a=h(r,{json:n.json===!0,skipReport:n.skipReport===!0}),o=await s({suite:r,target:f,reporters:a,appRoot:i,client:p});e.push(o)}n.json&&r.log(JSON.stringify(e,null,2)),e.some(e=>e.errored>0)&&(process.exitCode=1)}finally{u&&await u.close()}process.exit(process.exitCode??0)}function d(e){if(e.kind===`local`)return new i({host:e.url});let t={},r=process.env.VERCEL_AUTOMATION_BYPASS_SECRET?.trim();return r&&(t[n]=r),new i({auth:f(),headers:Object.keys(t).length>0?t:void 0,host:e.url})}function f(){let e=process.env.ASH_EVAL_AUTH_TOKEN?.trim();return e?{bearer:e}:{bearer:p}}async function p(){try{let e=(await(0,l.getVercelOidcToken)()).trim();if(e.length>0)return e}catch{}return process.env.VERCEL_OIDC_TOKEN?.trim()??``}function m(e,t){let n=t.maxConcurrency?Number.parseInt(t.maxConcurrency,10):void 0,r=t.timeout?Number.parseInt(t.timeout,10):void 0;if(n===void 0&&r===void 0)return e;let i={...e};return n!==void 0&&(i.maxConcurrency=n),r!==void 0&&(i.timeoutMs=r),i}function h(e,t){let n=t.json?[]:[new c];return!t.skipReport&&e.reporters&&n.push(...e.reporters),n}export{u as runEvalCommand};

package/dist/src/evals/loaders/index.d.ts

@@ -1,22 +1,2 @@
-/**
- * Loads and parses a JSON file.
- *
- * @example
- * ```ts
- * import { loadJson } from "experimental-ash/evals/loaders";
- *
- * const data = await loadJson("evals/data/cases.json");
- * ```
- */
-export declare function loadJson(filePath: string): Promise<unknown>;
-/**
- * Loads and parses a YAML file.
- *
- * @example
- * ```ts
- * import { loadYaml } from "experimental-ash/evals/loaders";
- *
- * const doc = await loadYaml("evals/data/cases.yaml");
- * ```
- */
-export declare function loadYaml(filePath: string): Promise<Record<string, unknown>>;
+export { loadJson } from "#evals/loaders/json.js";
+export { loadYaml } from "#evals/loaders/yaml.js";

package/dist/src/evals/loaders/index.js

@@ -1 +1 @@
-import{loadJson as e}from"./json.js";import{loadYaml as t}from"./yaml.js";async function n(t){return e(t)}async function r(e){return t(e)}export{n as loadJson,r as loadYaml};
+import{loadJson as e}from"./json.js";import{loadYaml as t}from"./yaml.js";export{e as loadJson,t as loadYaml};

package/dist/src/evals/reporters/index.d.ts

@@ -1,14 +1,2 @@
-import { type BraintrustReporterConfig } from "#evals/runner/reporters/braintrust.js";
-import type { EvalReporter } from "#evals/runner/reporters/types.js";
-/**
- * Creates a Braintrust reporter for use in a suite's `reporters` array.
- *
- * @example
- * ```ts
- * import { Braintrust } from "experimental-ash/evals/reporters";
- *
- * Braintrust({ projectName: "My Project" });
- * ```
- */
-export declare function Braintrust(config?: BraintrustReporterConfig): EvalReporter;
-export type { BraintrustReporterConfig, EvalReporter };
+export { Braintrust, type BraintrustReporterConfig } from "#evals/runner/reporters/braintrust.js";
+export type { EvalReporter } from "#evals/runner/reporters/types.js";

package/dist/src/evals/reporters/index.js

@@ -1 +1 @@
-import{Braintrust as e}from"../runner/reporters/braintrust.js";function t(t={}){return e(t)}export{t as Braintrust};
+import{Braintrust as e}from"../runner/reporters/braintrust.js";export{e as Braintrust};

package/dist/src/evals/runner/discover.d.ts

@@ -6,12 +6,6 @@ import type { AshEvalSuite } from "#evals/types.js";
  * Returns absolute paths sorted alphabetically by relative path.
  */
 export declare function discoverSuiteFiles(appRoot: string): Promise<string[]>;
-/**
- * Derives the canonical suite id from one absolute eval-suite file path.
- *
- * `<appRoot>/evals/sub/weather.eval.ts` → `"sub/weather"`.
- */
-export declare function deriveSuiteId(appRoot: string, filePath: string): string;
 /**
  * Imports a discovered suite file and stamps the path-derived id onto
  * the suite definition.

package/dist/src/evals/runner/discover.js

@@ -1 +1 @@
-import{t as e}from"../../chunks/authored-module-loader-DcCfCiBm.js";import{join as t,relative as n}from"node:path";import{readdir as r}from"node:fs/promises";const i=`.eval.ts`;async function a(e){let r=t(e,`evals`),i=[];try{await u(r,i)}catch(e){if(d(e))return[];throw e}return i.sort((e,t)=>n(r,e).localeCompare(n(r,t))),i}function o(e,r){let a=n(t(e,`evals`),r).split(/[\\/]/u).join(`/`);return a.endsWith(i)?a.slice(0,-8):a}async function s(t,n){let r=(await e(n)).default;if(!l(r))throw Error(`Suite file "${n}" does not export a valid AshEvalSuite as its default export. Use defineEvalSuite() to create the suite.`);return{...r,id:o(t,n)}}async function c(e,t){let n=await a(e);if(n.length===0)return[];let r=[];for(let i of n){let n=await s(e,i);(t===void 0||t.length===0||t.includes(n.id))&&r.push(n)}return r}function l(e){return typeof e==`object`&&!!e&&`_tag`in e&&e._tag===`AshEvalSuite`}async function u(e,n){let a=await r(e,{withFileTypes:!0});for(let r of a){let a=t(e,r.name);r.isDirectory()?await u(a,n):r.isFile()&&r.name.endsWith(i)&&n.push(a)}}function d(e){return typeof e==`object`&&!!e&&`code`in e&&e.code===`ENOENT`}export{o as deriveSuiteId,c as discoverAndImportSuites,a as discoverSuiteFiles,s as importSuiteFile};
+import{t as e}from"../../chunks/authored-module-loader-DcCfCiBm.js";import{join as t,relative as n}from"node:path";import{readdir as r}from"node:fs/promises";const i=`.eval.ts`;async function a(e){let r=t(e,`evals`),i=[];try{await u(r,i)}catch(e){if(d(e))return[];throw e}return i.sort((e,t)=>n(r,e).localeCompare(n(r,t))),i}function o(e,r){let a=n(t(e,`evals`),r).split(/[\\/]/u).join(`/`);return a.endsWith(i)?a.slice(0,-8):a}async function s(t,n){let r=(await e(n)).default;if(!l(r))throw Error(`Suite file "${n}" does not export a valid AshEvalSuite as its default export. Use defineEvalSuite() to create the suite.`);return{...r,id:o(t,n)}}async function c(e,t){let n=await a(e);if(n.length===0)return[];let r=[];for(let i of n){let n=await s(e,i);(t===void 0||t.length===0||t.includes(n.id))&&r.push(n)}return r}function l(e){return typeof e==`object`&&!!e&&`_tag`in e&&e._tag===`AshEvalSuite`}async function u(e,n){let a=await r(e,{withFileTypes:!0});for(let r of a){let a=t(e,r.name);r.isDirectory()?await u(a,n):r.isFile()&&r.name.endsWith(i)&&n.push(a)}}function d(e){return typeof e==`object`&&!!e&&`code`in e&&e.code===`ENOENT`}export{c as discoverAndImportSuites,a as discoverSuiteFiles,s as importSuiteFile};

package/dist/src/evals/runner/execute-case.d.ts

@@ -3,7 +3,7 @@ import type { AshEvalCase, AshEvalTask, AshEvalTaskResult } from "#evals/types.j
 /**
  * Options for executing one eval case.
  */
-export interface ExecuteCaseOptions {
+interface ExecuteCaseOptions {
     readonly client: Client;
     readonly testCase: AshEvalCase;
     readonly task?: AshEvalTask;
@@ -20,3 +20,4 @@ export interface ExecuteCaseOptions {
  * and status come from the final turn only.
  */
 export declare function executeCase(options: ExecuteCaseOptions): Promise<AshEvalTaskResult>;
+export {};

package/dist/src/evals/runner/execute-suite.d.ts

@@ -4,7 +4,7 @@ import type { EvalReporter } from "#evals/runner/reporters/types.js";
 /**
  * Options for executing one eval suite.
  */
-export interface ExecuteSuiteOptions {
+interface ExecuteSuiteOptions {
     readonly suite: AshEvalSuite;
     readonly target: AshEvalTarget;
     readonly reporters: readonly EvalReporter[];
@@ -21,3 +21,4 @@ export interface ExecuteSuiteOptions {
  * reports results.
  */
 export declare function executeSuite(options: ExecuteSuiteOptions): Promise<AshEvalSuiteResult>;
+export {};

package/dist/src/evals/runner/reporters/braintrust.js

@@ -1,2 +1,2 @@
-import{resolveLocalGitMetadata as e}from"../resolve-git-metadata.js";function t(e={}){return new n(e)}var n=class{#e;#t;#n;#r;#i;constructor(e){this.#e=e}async onSuiteStart(t,n){let o=await r();this.#t=o;let s=e(process.cwd());this.#i=t.metadata;let c=i(t,n),l=a(t,n);this.#n=await o.init({project:this.#e.projectName??t.id,projectId:this.#e.projectId,experiment:this.#e.experimentName,baseExperiment:this.#e.baseExperimentName,baseExperimentId:this.#e.baseExperimentId,update:this.#e.update,tags:c,metadata:l,noExitFlush:!0,setCurrent:!1,repoInfo:s.sha?{commit:s.sha,branch:s.branch}:null})}onCaseComplete(e){if(!this.#n)return;!this.#r&&e.result.runtimeIdentity&&(this.#r=e.result.runtimeIdentity);let t={};for(let n of e.scores)n.score!==null&&(t[n.name]=n.score);let n={...this.#i,...e.case.metadata,ashSessionId:e.result.sessionId,ashStatus:e.result.status,ashToolCalls:e.result.derived.toolCalls,ashSubagentCalls:e.result.derived.subagentCalls};e.result.derived.failureCode&&(n.ashFailureCode=e.result.derived.failureCode);let r={toolCallCount:e.result.derived.toolCallCount,subagentCallCount:e.result.derived.subagentCallCount,messageCount:e.result.derived.messageCount,reasoningBlockCount:e.result.derived.reasoningBlockCount};this.#n.log({id:e.case.id,input:e.case.input,output:e.result.output,expected:e.case.expected,error:e.error??void 0,scores:t,metadata:n,metrics:r,tags:e.case.tags?[...e.case.tags]:void 0})}async onSuiteComplete(e){if(this.#n)try{this.#t&&await this.#t.flush();let e=await this.#n.summarize();e.experimentUrl&&console.log(`Braintrust experiment: ${e.experimentUrl}\n\n`)}finally{await this.#n.close(),this.#n=void 0,this.#t=void 0}}};async function r(){try{return await import(`braintrust`)}catch{throw Error([`The 'braintrust' package is required for Braintrust reporting but was not found.`,``,`Install it with:`,`  npm install braintrust`].join(`
+import{resolveLocalGitMetadata as e}from"../resolve-git-metadata.js";function t(e={}){return new n(e)}var n=class{#e;#t;#n;#r;constructor(e){this.#e=e}async onSuiteStart(t,n){let o=await r();this.#t=o;let s=e(process.cwd());this.#r=t.metadata;let c=i(t,n),l=a(t,n);this.#n=await o.init({project:this.#e.projectName??t.id,projectId:this.#e.projectId,experiment:this.#e.experimentName,baseExperiment:this.#e.baseExperimentName,baseExperimentId:this.#e.baseExperimentId,update:this.#e.update,tags:c,metadata:l,noExitFlush:!0,setCurrent:!1,repoInfo:s.sha?{commit:s.sha,branch:s.branch}:null})}onCaseComplete(e){if(!this.#n)return;let t={};for(let n of e.scores)n.score!==null&&(t[n.name]=n.score);let n={...this.#r,...e.case.metadata,ashSessionId:e.result.sessionId,ashStatus:e.result.status,ashToolCalls:e.result.derived.toolCalls,ashSubagentCalls:e.result.derived.subagentCalls};e.result.derived.failureCode&&(n.ashFailureCode=e.result.derived.failureCode);let r={toolCallCount:e.result.derived.toolCallCount,subagentCallCount:e.result.derived.subagentCallCount,messageCount:e.result.derived.messageCount,reasoningBlockCount:e.result.derived.reasoningBlockCount};this.#n.log({id:e.case.id,input:e.case.input,output:e.result.output,expected:e.case.expected,error:e.error??void 0,scores:t,metadata:n,metrics:r,tags:e.case.tags?[...e.case.tags]:void 0})}async onSuiteComplete(e){if(this.#n)try{this.#t&&await this.#t.flush();let e=await this.#n.summarize();e.experimentUrl&&console.log(`Braintrust experiment: ${e.experimentUrl}\n\n`)}finally{await this.#n.close(),this.#n=void 0,this.#t=void 0}}};async function r(){try{return await import(`braintrust`)}catch{throw Error([`The 'braintrust' package is required for Braintrust reporting but was not found.`,``,`Install it with:`,`  npm install braintrust`].join(`
 `))}}function i(e,t){let n=[`ash`,`suite:${e.id}`,`target:${t.kind}`];return Array.isArray(e.tags)&&n.push(...e.tags),n}function a(e,t){return{ashSuiteId:e.id,ashTargetKind:t.kind,ashTargetUrl:t.url,ashTimestamp:new Date().toISOString()}}export{t as Braintrust};

package/dist/src/evals/runner/resolve-git-metadata.d.ts

@@ -5,9 +5,9 @@ interface GitMetadata {
 /**
  * Resolves local git metadata for the eval run context.
  *
- * Used to populate {@link AshEvalRunContext.localGitSha} and
- * {@link AshEvalRunContext.localGitBranch} for Braintrust experiment
- * metadata. This describes the eval code, not the remote target.
+ * Used to populate `repoInfo` on the Braintrust experiment so the dashboard
+ * shows which sha/branch produced the run. This describes the eval code,
+ * not the remote target.
  *
  * Returns an empty object when git is unavailable or the directory is
  * not a git repository.

package/dist/src/evals/scorers/autoevals-client.d.ts

@@ -4,9 +4,9 @@ import { Factuality } from "autoevals";
  * The OpenAI-shaped client surface autoevals expects. Extracted from the
  * library so we don't take a direct dependency on `openai` types.
  */
-export type AutoevalsClient = NonNullable<Parameters<typeof Factuality>[0]["client"]>;
+type AutoevalsClient = NonNullable<Parameters<typeof Factuality>[0]["client"]>;
 type ProviderOptions = Parameters<typeof generateText>[0]["providerOptions"];
-export interface AutoevalsClientConfig {
+interface AutoevalsClientConfig {
     readonly languageModel: LanguageModel;
     readonly providerOptions?: ProviderOptions;
 }

package/dist/src/execution/connection-auth-steps.d.ts

@@ -4,8 +4,6 @@
  */
 import type { HarnessEmissionState } from "#harness/emission.js";
 import type { HarnessSession } from "#harness/types.js";
-import { type ConnectionAuthorizationOutcome } from "#protocol/message.js";
-import type { ConnectionAuthorizationChallenge } from "#public/connections/errors.js";
 import type { JsonValue } from "#public/types/json.js";
 import type { AuthorizationCallbackRequest, ConnectionPrincipal, TokenResult } from "#runtime/connections/types.js";
 import { type PendingConnectionToolCall } from "#runtime/framework-tools/pending-connection-tool-calls.js";
@@ -13,7 +11,7 @@ import { type PendingConnectionToolCall } from "#runtime/framework-tools/pending
  * Result of one `startAuthorization` step.
  *
  * On success the runtime journals
- * `{ ok: true, principal, state, challenge, serializedContext }`. The
+ * `{ ok: true, principal, state, serializedContext }`. The
  * `principal` is the framework-resolved {@link ConnectionPrincipal}
  * captured at `startAuthorization` time; the orchestrator carries it
  * forward so `completeAuthorization` and the post-resume retry observe
@@ -32,7 +30,6 @@ import { type PendingConnectionToolCall } from "#runtime/framework-tools/pending
  * `outcome: "failed"` so channels clean up their UI.
  */
 export type StartAuthorizationStepResult = {
-    readonly challenge: ConnectionAuthorizationChallenge;
     readonly ok: true;
     readonly principal: ConnectionPrincipal;
     readonly serializedContext: Record<string, unknown>;
@@ -79,7 +76,6 @@ export type CompleteAuthorizationStepResult = {
     readonly token: TokenResult;
 } | {
     readonly ok: false;
-    readonly outcome: Exclude<ConnectionAuthorizationOutcome, "authorized">;
     readonly reason: string;
     /**
      * When `true`, downstream pending tool calls for this connection

package/dist/src/execution/connection-auth-steps.js

@@ -4,7 +4,7 @@
  */
 import { buildAdapterContext } from "#channel/adapter-context.js";
 import { callAdapterEventHandler } from "#channel/adapter.js";
-import { contextStorage, toContextAccessor } from "#context/container.js";
+import { contextStorage } from "#context/container.js";
 import { BundleKey, ChannelKey } from "#context/keys.js";
 import { getActiveRuntimeNode } from "#context/node.js";
 import { deserializeContext, serializeContext } from "#context/serialize.js";
@@ -82,7 +82,6 @@ export async function startAuthorizationForConnectionStep(input) {
     // the handler made (e.g. tracked Slack message ts) survive the
     // step boundary into the matching `completeAuthorization` step.
     return {
-        challenge,
         ok: true,
         principal,
         serializedContext: serializeContext(ctx),
@@ -130,7 +129,7 @@ export async function completeAuthorizationForConnectionStep(input) {
             stepIndex: input.emissionState.stepIndex,
             turnId: input.emissionState.turnId,
         }), input.parentWritable);
-        return { ok: false, outcome: "failed", reason, retryable: false };
+        return { ok: false, reason, retryable: false };
     }
     try {
         const token = await connection.authorization.completeAuthorization({
@@ -168,7 +167,7 @@ export async function completeAuthorizationForConnectionStep(input) {
             stepIndex: input.emissionState.stepIndex,
             turnId: input.emissionState.turnId,
         }), input.parentWritable);
-        return { ok: false, outcome: "failed", reason, retryable };
+        return { ok: false, reason, retryable };
     }
 }
 /**
@@ -308,7 +307,7 @@ function findConnection(ctx, connectionName) {
 }
 async function emitAuthorizationEvent(ctx, event, parentWritable) {
     const adapter = ctx.require(ChannelKey);
-    const adapterCtx = buildAdapterContext(adapter, toContextAccessor(ctx));
+    const adapterCtx = buildAdapterContext(adapter, ctx);
     const finalEvent = await callAdapterEventHandler(adapter, event, adapterCtx);
     // Pin any handler-driven channel-state mutations back onto ctx so
     // they survive the durable step boundary when the caller

package/dist/src/execution/node-step.d.ts

@@ -1,6 +1,6 @@
 import type { Runtime, SessionCapabilities } from "#channel/types.js";
 import type { HarnessEmitFn, StepFn } from "#harness/types.js";
-import type { RunMode } from "#run-mode.js";
+import type { RunMode } from "#shared/run-mode.js";
 import type { RuntimeCompiledArtifactsSource } from "#runtime/compiled-artifacts-source.js";
 import type { ResolvedRuntimeAgentNode } from "#runtime/graph.js";
 /**

package/dist/src/execution/skills/instructions.d.ts

@@ -1,4 +1,4 @@
-export interface AvailableSkillDescription {
+interface AvailableSkillDescription {
     readonly description: string;
     readonly name: string;
 }
@@ -15,3 +15,4 @@ export interface AvailableSkillDescription {
  * reuse the same formatter for durable history announcements.
  */
 export declare function formatAvailableSkillsSection(skills: readonly AvailableSkillDescription[]): string | null;
+export {};

package/dist/src/execution/subagent-hitl-proxy.d.ts

@@ -1,6 +1,6 @@
 import type { DeliverPayload, SubagentInputRequestHookPayload } from "#channel/types.js";
 import type { HarnessEmitFn, HarnessSession } from "#harness/types.js";
-import type { RunMode } from "#run-mode.js";
+import type { RunMode } from "#shared/run-mode.js";
 import type { InputResponse } from "#runtime/input/types.js";
 /**
  * Runs the parent-side work when a `subagent-input-request` arrives at

package/dist/src/execution/subagent-tool.d.ts

@@ -8,14 +8,6 @@ interface BatchEventMetadata {
     readonly sequence: number;
     readonly turnId: string;
 }
-/**
- * Returns the deterministic continuation token used for one delegated child
- * subagent run.
- */
-export declare function createSubagentContinuationToken(input: {
-    readonly callId: string;
-    readonly parentSessionId: string;
-}): string;
 /**
  * Result of {@link buildSubagentRunInput}.
  *

package/dist/src/execution/subagent-tool.js

@@ -1,22 +1,11 @@
 import { SUBAGENT_ADAPTER_KIND } from "#execution/subagent-adapter.js";
 import { mintSubagentContinuationToken } from "#execution/session.js";
-import { formatSubagentInvocation } from "#execution/subagent-invocation.js";
-/**
- * Returns the deterministic continuation token used for one delegated child
- * subagent run.
- */
-export function createSubagentContinuationToken(input) {
-    return mintSubagentContinuationToken(`${input.parentSessionId}:${input.callId}`);
-}
 /**
  * Builds the {@link RunInput} for one delegated subagent child run.
  */
 export function buildSubagentRunInput(input) {
     const { action, auth, batchEvent, capabilities, initiatorAuth, session } = input;
-    const childContinuationToken = createSubagentContinuationToken({
-        callId: action.callId,
-        parentSessionId: session.sessionId,
-    });
+    const childContinuationToken = mintSubagentContinuationToken(`${session.sessionId}:${action.callId}`);
     const runInput = {
         adapter: {
             kind: SUBAGENT_ADAPTER_KIND,
@@ -50,9 +39,13 @@ export function buildSubagentRunInput(input) {
  */
 function formatSubagentCallInputMessage(action) {
     const { message } = action.input;
-    return formatSubagentInvocation({
-        description: action.description,
+    return [
+        `You are the subagent "${action.subagentName}".`,
+        `Description: ${action.description}`,
+        "",
+        "The caller delegated the following task to you. Complete it and return the final result directly.",
+        "",
+        "Caller message:",
         message,
-        name: action.subagentName,
-    }).message;
+    ].join("\n");
 }

package/dist/src/execution/tool-compaction.js

@@ -38,8 +38,6 @@ export function createToolCompactionHandler(tools) {
                 });
                 continue;
             }
-            if (result === undefined)
-                continue;
             if (result.messages !== undefined && result.messages.length > 0) {
                 messages.push(...result.messages);
             }

package/dist/src/execution/turn-workflow.d.ts

@@ -1,6 +1,6 @@
 import type { HookPayload, SessionCapabilities } from "#channel/types.js";
 import type { HarnessSession } from "#harness/types.js";
-import type { RunMode } from "#run-mode.js";
+import type { RunMode } from "#shared/run-mode.js";
 export interface TurnResultPayload {
     readonly action: "done" | "park";
     readonly kind: "turn-result";

package/dist/src/execution/turn-workflow.js

@@ -1,9 +1,9 @@
 import { hasPendingInputBatch } from "#harness/input-requests.js";
 import { hasPendingRuntimeActionBatch } from "#harness/runtime-actions.js";
 import { awaitAuthorizationAndResolve } from "#execution/await-authorization-orchestrator.js";
-import { createTaskModeWaitError } from "#execution/task-mode.js";
 import { normalizeSerializableError } from "#execution/workflow-errors.js";
 import { turnStep } from "#execution/workflow-steps.js";
+const TASK_MODE_WAIT_ERROR_MESSAGE = "Task mode cannot wait for follow-up input (`next: null`).";
 /**
  * Short-lived workflow that owns one runtime turn for the durable
  * driver.
@@ -56,7 +56,7 @@ export async function turnWorkflow(input) {
                     });
                     return;
                 }
-                throw createTaskModeWaitError();
+                throw new Error(TASK_MODE_WAIT_ERROR_MESSAGE);
             }
             if (result.action === "await-authorization") {
                 const resolved = await awaitAuthorizationAndResolve({

package/dist/src/execution/web-fetch/tool.js

@@ -1,4 +1,4 @@
-import { ASH_PACKAGE_NAME } from "#package-name.js";
+import { ASH_PACKAGE_NAME } from "#internal/package-name.js";
 import { truncateHead } from "#execution/sandbox/truncate-output.js";
 import { convertHtmlToMarkdown, extractTextFromHtml } from "#execution/web-fetch/html.js";
 const MAX_RESPONSE_SIZE = 5 * 1024 * 1024; // 5 MB

package/dist/src/execution/workflow-steps.d.ts

@@ -2,8 +2,24 @@ import type { DeliverPayload, HookPayload, SessionAuthContext, SubagentInputRequ
 import { deserializeContext } from "#context/serialize.js";
 import type { HarnessSession } from "#harness/types.js";
 import type { RuntimeCompiledArtifactsSource } from "#runtime/compiled-artifacts-source.js";
+import { type PendingConnectionAuthorization } from "#runtime/framework-tools/connection-search.js";
+import { type PendingConnectionToolCall } from "#runtime/framework-tools/pending-connection-tool-calls.js";
 import { type TurnWorkflowInput } from "#execution/turn-workflow.js";
-import type { DurableStepResult } from "#execution/types.js";
+/**
+ * Serializable projection of a step result for workflow persistence.
+ */
+export type DurableStepResult = {
+    readonly action: "continue" | "park" | "done";
+    readonly output?: string;
+    readonly serializedContext: Record<string, unknown>;
+    readonly session: HarnessSession;
+} | {
+    readonly action: "await-authorization";
+    readonly pendingToolCalls: readonly PendingConnectionToolCall[];
+    readonly pendingAuths: readonly PendingConnectionAuthorization[];
+    readonly serializedContext: Record<string, unknown>;
+    readonly session: HarnessSession;
+};
 /**
  * Input for one atomic harness step inside a durable `"use step"`
  * boundary.

package/dist/src/execution/workflow-steps.js

@@ -1,6 +1,5 @@
 import { buildAdapterContext } from "#channel/adapter-context.js";
 import { callAdapterEventHandler, defaultDeliverResult } from "#channel/adapter.js";
-import { toContextAccessor } from "#context/container.js";
 import { dispatchStreamEventHooks, runHookLifecycleStep } from "#context/hook-lifecycle.js";
 import { AuthKey, BundleKey, CapabilitiesKey, ChannelKey, ContinuationTokenKey, InitiatorAuthKey, ModeKey, } from "#context/keys.js";
 import { runStep } from "#context/run-step.js";
@@ -12,7 +11,7 @@ import { getPendingRuntimeActionBatch, recordPendingSubagentChildToken, } from "
 import { createLogger, formatError } from "#internal/logging.js";
 import { createSessionFailedEvent, createSubagentCalledEvent, encodeMessageStreamEvent, timestampHandleMessageStreamEvent, } from "#protocol/message.js";
 import { drainPendingConnectionAuthorizations, PendingConnectionAuthorizationsKey, } from "#runtime/framework-tools/connection-search.js";
-import { PendingConnectionToolCallsKey } from "#runtime/framework-tools/pending-connection-tool-calls.js";
+import { PendingConnectionToolCallsKey, } from "#runtime/framework-tools/pending-connection-tool-calls.js";
 import { getCompiledRuntimeAgentBundle } from "#runtime/sessions/compiled-agent-cache.js";
 import { createExecutionNodeStep } from "#execution/node-step.js";
 import { emitProxiedInputRequest, routeDeliverPayload } from "#execution/subagent-hitl-proxy.js";
@@ -36,7 +35,7 @@ export async function turnStep(input) {
     }
     // Build the adapter context for deliver and event handlers.
     // Slack adapters override this to inject ctx.thread and ctx.slack.
-    const adapterCtx = buildAdapterContext(adapter, toContextAccessor(ctx));
+    const adapterCtx = buildAdapterContext(adapter, ctx);
     // Run the adapter's deliver hook for each queued payload — sets
     // custom context keys and optionally transforms the message.
     // Coalesces the resulting StepInput values.
@@ -220,7 +219,7 @@ export async function dispatchPendingRuntimeActionsStep(input) {
     const capabilities = ctx.get(CapabilitiesKey);
     const initiatorAuth = ctx.get(InitiatorAuthKey) ?? null;
     const writer = input.parentWritable.getWriter();
-    const adapterCtx = buildAdapterContext(adapter, toContextAccessor(ctx));
+    const adapterCtx = buildAdapterContext(adapter, ctx);
     let nextSession = input.session;
     try {
         for (const action of batch.actions) {
@@ -288,7 +287,7 @@ export async function emitTerminalSessionFailureStep(input) {
         const ctx = await deserializeContext(input.serializedContext);
         const adapter = ctx.get(ChannelKey);
         if (adapter !== undefined) {
-            const adapterCtx = buildAdapterContext(adapter, toContextAccessor(ctx));
+            const adapterCtx = buildAdapterContext(adapter, ctx);
             await callAdapterEventHandler(adapter, event, adapterCtx);
         }
     }
@@ -327,7 +326,7 @@ export async function runProxyInputRequestStep(input) {
     "use step";
     const ctx = await deserializeContext(input.serializedContext);
     const adapter = ctx.require(ChannelKey);
-    const adapterCtx = buildAdapterContext(adapter, toContextAccessor(ctx));
+    const adapterCtx = buildAdapterContext(adapter, ctx);
     const mode = ctx.require(ModeKey);
     const writer = input.parentWritable.getWriter();
     let proxyResult;

package/dist/src/harness/action-result-helpers.d.ts

@@ -1,23 +1,11 @@
 import type { ModelMessage, ToolSet, TypedToolResult } from "ai";
 import type { RuntimeToolResultActionResult } from "#runtime/actions/types.js";
-import type { JsonValue } from "#shared/json.js";
 type ToolResponsePart = Extract<ModelMessage, {
     role: "tool";
 }>["content"][number];
 type ToolResultPart = Extract<ToolResponsePart, {
     type: "tool-result";
 }>;
-/**
- * Coerces an arbitrary value to a JSON-safe {@link JsonValue} without
- * premature stringification.
- *
- * - Strings, numbers, booleans, and `null` pass through as primitives.
- * - `Error` instances surface only their message (no stack leak).
- * - Plain objects and arrays pass through structurally.
- * - Non-JSON-representable values (functions, symbols, BigInts) fall
- *   back to `String(value)`.
- */
-export declare function toJsonValue(value: unknown): JsonValue;
 /**
  * Builds a `RuntimeToolResultActionResult` from one AI SDK
  * {@link TypedToolResult}. Used for tool results captured on the AI SDK

package/dist/src/harness/action-result-helpers.js

@@ -8,7 +8,7 @@
  * - Non-JSON-representable values (functions, symbols, BigInts) fall
  *   back to `String(value)`.
  */
-export function toJsonValue(value) {
+function toJsonValue(value) {
     if (value === null ||
         typeof value === "string" ||
         typeof value === "number" ||

package/dist/src/harness/emission.d.ts

@@ -6,7 +6,7 @@ type InlineToolResultPart = Extract<ToolResponsePart, {
     type: "tool-result";
 }>;
 import type { AssistantStepFinishReason, RuntimeIdentity } from "#protocol/message.js";
-import type { RunMode } from "#run-mode.js";
+import type { RunMode } from "#shared/run-mode.js";
 import type { JsonObject } from "#shared/json.js";
 import type { HarnessEmitFn, HarnessSession, StepInput } from "#harness/types.js";
 /**
@@ -103,7 +103,7 @@ export declare function normalizeAssistantStepFinishReason(value: string | undef
  * harness splices them into persisted history to keep the prior turn's
  * `tool_use` block balanced with a matching `tool_result` on replay.
  */
-export interface EmittedStreamContent {
+interface EmittedStreamContent {
     readonly inlineActionResultCallIds: ReadonlySet<string>;
     readonly inlineToolResultParts: readonly InlineToolResultPart[];
 }

package/dist/src/harness/execute-tool.d.ts

@@ -6,7 +6,7 @@ import type { NeedsApprovalContext } from "#public/definitions/tool.js";
  * These tools are surfaced to the model without a local `execute` function.
  * The harness records the tool call and the runtime executes it later.
  */
-export type HarnessRuntimeActionDefinition = {
+type HarnessRuntimeActionDefinition = {
     readonly kind: "subagent-call";
     readonly nodeId: string;
     readonly subagentName: string;
@@ -24,3 +24,4 @@ export interface HarnessToolDefinition {
     readonly runtimeAction?: HarnessRuntimeActionDefinition;
     readonly toModelOutput?: (output: unknown) => unknown;
 }
+export {};

package/dist/src/harness/input-requests.d.ts

@@ -42,7 +42,7 @@ export declare function resolvePendingInput(input: {
     readonly session: HarnessSession;
     readonly stepInput?: StepInput;
 }): ResolvePendingInputResult;
-export type ResolvePendingInputResult = {
+type ResolvePendingInputResult = {
     readonly deferredMessage?: boolean;
     readonly outcome: "resolved" | "continue" | "unresolved";
     readonly messages: ModelMessage[];
@@ -72,3 +72,4 @@ export declare function getApprovedTools(session: HarnessSession): ReadonlySet<s
 export declare function createRuntimeToolCallActionFromToolCall(input: {
     readonly toolCall: TypedToolCall<ToolSet>;
 }): RuntimeToolCallActionRequest;
+export {};

package/dist/src/harness/prompt-cache.d.ts

@@ -1,4 +1,4 @@
-import type { LanguageModel, ModelMessage, PrepareStepFunction, ToolSet } from "ai";
+import type { LanguageModel, ModelMessage, ToolSet } from "ai";
 /**
  * The caching strategy to apply for one harness step.
  */
@@ -71,11 +71,3 @@ export declare function applyLastToolCacheBreakpoint(tools: ToolSet, marker: Ant
  * through any preceding tool results.
  */
 export declare function applyConversationCacheControl(messages: readonly ModelMessage[], marker: AnthropicCacheMarker): ModelMessage[];
-/**
- * Builds the AI SDK `prepareStep` callback for the `anthropic-direct` path.
- *
- * The returned function accepts the SDK's per-step input and returns a
- * `{ messages }` override with the last user and last assistant messages
- * marked with the Anthropic cache control breakpoint.
- */
-export declare function buildPromptCachePrepareStep(marker: AnthropicCacheMarker): PrepareStepFunction<ToolSet>;

package/dist/src/harness/prompt-cache.js

@@ -135,15 +135,3 @@ export function applyConversationCacheControl(messages, marker) {
     }
     return out;
 }
-/**
- * Builds the AI SDK `prepareStep` callback for the `anthropic-direct` path.
- *
- * The returned function accepts the SDK's per-step input and returns a
- * `{ messages }` override with the last user and last assistant messages
- * marked with the Anthropic cache control breakpoint.
- */
-export function buildPromptCachePrepareStep(marker) {
-    return ({ messages }) => ({
-        messages: applyConversationCacheControl(messages, marker),
-    });
-}

package/dist/src/harness/runtime-actions.d.ts

@@ -31,7 +31,7 @@ interface PendingRuntimeActionBatch {
 /**
  * Outcome of resolving a pending runtime-action batch.
  */
-export interface ResolvePendingRuntimeActionsResult {
+interface ResolvePendingRuntimeActionsResult {
     readonly messages: ModelMessage[];
     readonly outcome: "continue" | "resolved" | "unresolved";
     readonly session: HarnessSession;
@@ -72,7 +72,7 @@ export declare function recordPendingSubagentChildToken(input: {
  * accumulation loop process both kinds without coupling to the concrete
  * `HookPayload` shape.
  */
-export type RuntimeActionAccumulatorItem<TDeliver> = {
+type RuntimeActionAccumulatorItem<TDeliver> = {
     readonly kind: "deliver";
     readonly value: TDeliver;
 } | {
@@ -88,15 +88,6 @@ export declare function accumulateRuntimeActionResults<TDeliver>(input: {
     readonly getNext: () => Promise<RuntimeActionAccumulatorItem<TDeliver> | null>;
     readonly session: HarnessSession;
 }): Promise<RuntimeActionResult[] | null>;
-/**
- * Returns the stable ordered runtime-action results for the current pending
- * batch when every action has a matching result. Unknown and duplicate results
- * are ignored.
- */
-export declare function resolveReadyRuntimeActionResults(input: {
-    readonly results: readonly RuntimeActionResult[];
-    readonly session: HarnessSession;
-}): RuntimeActionResult[] | undefined;
 /**
  * Resolves one pending runtime-action batch back into model history.
  *