stable-harness 0.0.18 → 0.0.20

package/dist/index.js

@@ -1 +1 @@
-import{createDeepAgentsAdapter as e}from"@stable-harness/adapter-deepagents";import{createLangGraphRuntimeAdapter as r,createLangGraphWorkflowAdapter as a,createRegistrySkillResolverProvider as t}from"@stable-harness/adapter-langgraph";import{createStableHarnessRuntime as o}from"@stable-harness/core";import{createModuleToolGateway as n}from"@stable-harness/tool-gateway";import{loadWorkspaceFromYaml as s}from"@stable-harness/workspace-yaml";export{createDeepAgentsAdapter,createDeepAgentsMemoryMaintenanceTarget}from"@stable-harness/adapter-deepagents";export{createDeepAgentsMiddlewareSkillProvider,createLangGraphRuntimeAdapter,createLangGraphWorkflowAdapter,createRegistrySkillResolverProvider}from"@stable-harness/adapter-langgraph";export{createLangMemServiceProvider}from"@stable-harness/memory";export{applySpecDrivenPhaseTransition,containsRecoverableResultOutput,createSpecDrivenArtifact,createSpecDrivenArtifactEvent,createSpecDrivenPhaseEvent,createSpecDrivenWorkflowPolicy,createSpecDrivenWorkflowState,projectRuntimeTrace}from"@stable-harness/core";export{loadWorkspaceFromYaml}from"@stable-harness/workspace-yaml";export{createInMemoryToolGateway,createModuleToolGateway}from"@stable-harness/tool-gateway";export function createStableHarnessRuntime(e){return"string"==typeof e?createStableRuntime({workspaceRoot:e}):"workspaceRoot"in e?createStableRuntime(e):o(e)}export async function createStableRuntime(e){const r=await s(e.workspaceRoot),a=e.toolGateway??await n({tools:r.tools.values()});return o({workspace:r,toolGateway:a,adapters:e.adapters??createRuntimeAdapters(r,e),workflowAdapters:e.workflowAdapters??createWorkflowAdapters(r,e)})}export async function requestStableRuntime(e,r){return e.request(r)}function createRuntimeAdapters(a,t){const o={deepagents:({policy:r})=>e(r.config?{config:r.config}:{}),langgraph:({policy:e})=>r({...readLangGraphOptions(e.config),name:e.name}),...t.adapterFactories},n=function runtimeAdapterPolicies(e){const r=e.runtime.adapters?.filter(e=>!1!==e.enabled);return r&&r.length>0?r:[...new Set([...e.agents.values()].map(e=>e.backend))].map(e=>({name:e}))}(a);return n.map(e=>{const r=o[e.name];if(r)return r({policy:e,workspace:a});throw new Error(`Unsupported runtime adapter: ${e.name}`)})}function createWorkflowAdapters(e,r){const t={langgraph:({name:e,options:r})=>a({...readLangGraphOptions(r),name:e}),...r.workflowAdapterFactories};return[...new Set([...e.workflows.values()].map(e=>e.adapter??"").filter(Boolean))].map(a=>{const o=t[a];return o?.({name:a,workspace:e,options:readWorkflowAdapterOptions(r,a)})}).filter(e=>Boolean(e))}function readWorkflowAdapterOptions(e,r){return e.workflowAdapterOptions?.[r]??{}}function readLangGraphOptions(e){return isRecord(e)?{...e,...void 0!==readLangGraphSkillProvider(e)?{skillProvider:readLangGraphSkillProvider(e)}:{}}:{}}function readLangGraphSkillProvider(e){if(!1===e.skillProvider)return!1;const r=function readSkillProviderConfig(e){return isRecord(e.skills)?e.skills:isRecord(e.skillProvider)?e.skillProvider:void 0}(e);if(!r)return;const a=readString(r.provider)??readString(r.name)??"registry-resolver";if(["none","disabled","false"].includes(a))return!1;if("registry-resolver"!==a)throw new Error(`Unsupported LangGraph skill provider: ${a}`);return t({..."boolean"==typeof r.includeContent?{includeContent:r.includeContent}:{},..."number"==typeof r.maxBytes&&Number.isFinite(r.maxBytes)?{maxBytes:r.maxBytes}:{}})}function readString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}
+import{createBackendModel as e,createDeepAgentsAdapter as r}from"@stable-harness/adapter-deepagents";import{createLangGraphRuntimeAdapter as t,createLangGraphWorkflowAdapter as a,createRegistrySkillResolverProvider as o}from"@stable-harness/adapter-langgraph";import{createStableHarnessRuntime as n}from"@stable-harness/core";import{createModuleToolGateway as i}from"@stable-harness/tool-gateway";import{loadWorkspaceFromYaml as s}from"@stable-harness/workspace-yaml";export{createDeepAgentsAdapter,createDeepAgentsMemoryMaintenanceTarget}from"@stable-harness/adapter-deepagents";export{createDeepAgentsMiddlewareSkillProvider,createLangGraphRuntimeAdapter,createLangGraphWorkflowAdapter,createRegistrySkillResolverProvider}from"@stable-harness/adapter-langgraph";export{createLangMemServiceProvider}from"@stable-harness/memory";export{applySpecDrivenPhaseTransition,containsRecoverableResultOutput,createSpecDrivenArtifact,createSpecDrivenArtifactEvent,createSpecDrivenPhaseEvent,createSpecDrivenWorkflowPolicy,createSpecDrivenWorkflowState,projectRuntimeTrace}from"@stable-harness/core";export{loadWorkspaceFromYaml}from"@stable-harness/workspace-yaml";export{createInMemoryToolGateway,createModuleToolGateway}from"@stable-harness/tool-gateway";export function createStableHarnessRuntime(e){return"string"==typeof e?createStableRuntime({workspaceRoot:e}):"workspaceRoot"in e?createStableRuntime(e):n(e)}export async function createStableRuntime(e){const r=await s(e.workspaceRoot),t=e.toolGateway??await i({tools:r.tools.values()});return n({workspace:r,toolGateway:t,qualityReviewModel:createQualityReviewModel(r),adapters:e.adapters??createRuntimeAdapters(r,e),workflowAdapters:e.workflowAdapters??createWorkflowAdapters(r,e)})}function createQualityReviewModel(r){const t=function readQualityModelRef(e){const r=isRecord(e)?e:{};return readString((isRecord(r.reviewer)?r.reviewer:r).modelRef)}(r.runtime.quality),a=t?r.models.get(t):void 0,o=a?e(a):void 0;return function isQualityReviewModel(e){return isRecord(e)&&"function"==typeof e.invoke}(o)?o:void 0}export async function requestStableRuntime(e,r){return e.request(r)}function createRuntimeAdapters(e,a){const o={deepagents:({policy:e})=>r(e.config?{config:e.config}:{}),langgraph:({policy:e})=>t({...readLangGraphOptions(e.config),name:e.name}),...a.adapterFactories},n=function runtimeAdapterPolicies(e){const r=e.runtime.adapters?.filter(e=>!1!==e.enabled);return r&&r.length>0?r:[...new Set([...e.agents.values()].map(e=>e.backend))].map(e=>({name:e}))}(e);return n.map(r=>{const t=o[r.name];if(t)return t({policy:r,workspace:e});throw new Error(`Unsupported runtime adapter: ${r.name}`)})}function createWorkflowAdapters(e,r){const t={langgraph:({name:e,options:r})=>a({...readLangGraphOptions(r),name:e}),...r.workflowAdapterFactories};return[...new Set([...e.workflows.values()].map(e=>e.adapter??"").filter(Boolean))].map(a=>{const o=t[a];return o?.({name:a,workspace:e,options:readWorkflowAdapterOptions(r,a)})}).filter(e=>Boolean(e))}function readWorkflowAdapterOptions(e,r){return e.workflowAdapterOptions?.[r]??{}}function readLangGraphOptions(e){return isRecord(e)?{...e,...void 0!==readLangGraphSkillProvider(e)?{skillProvider:readLangGraphSkillProvider(e)}:{}}:{}}function readLangGraphSkillProvider(e){if(!1===e.skillProvider)return!1;const r=function readSkillProviderConfig(e){return isRecord(e.skills)?e.skills:isRecord(e.skillProvider)?e.skillProvider:void 0}(e);if(!r)return;const t=readString(r.provider)??readString(r.name)??"registry-resolver";if(["none","disabled","false"].includes(t))return!1;if("registry-resolver"!==t)throw new Error(`Unsupported LangGraph skill provider: ${t}`);return o({..."boolean"==typeof r.includeContent?{includeContent:r.includeContent}:{},..."number"==typeof r.maxBytes&&Number.isFinite(r.maxBytes)?{maxBytes:r.maxBytes}:{}})}function readString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

package/dist/tools.d.ts

@@ -2,11 +2,11 @@ export declare function tool(input: {
     name?: string;
     description?: string;
     schema?: unknown;
-    func?: (args: unknown) => unknown | Promise<unknown>;
-    invoke?: (args: unknown) => unknown | Promise<unknown>;
+    func?: (args: unknown, context?: unknown) => unknown | Promise<unknown>;
+    invoke?: (args: unknown, context?: unknown) => unknown | Promise<unknown>;
 }): {
     name: string | undefined;
     description: string | undefined;
     schema: unknown;
-    invoke(args: unknown): Promise<unknown>;
+    invoke(args: unknown, context?: unknown): Promise<unknown>;
 };

package/dist/tools.js

@@ -1 +1 @@
-export function tool(n){return{name:n.name,description:n.description,schema:n.schema,async invoke(o){if(n.invoke)return n.invoke(o);if(n.func)return n.func(o);throw new Error(`Tool ${n.name??"unknown"} has no function`)}}}
+export function tool(n){return{name:n.name,description:n.description,schema:n.schema,async invoke(o,e){if(n.invoke)return n.invoke(o,e);if(n.func)return n.func(o,e);throw new Error(`Tool ${n.name??"unknown"} has no function`)}}}

package/docs/guides/index.md

@@ -12,6 +12,8 @@ embed it, operate it, or explain why it exists.
   workflows, memory, and protocol exposure in YAML.
 - [Integration guide](integration-guide.md): embed the runtime inside an app,
   run it through CLI, or expose it through HTTP/OpenAI-compatible clients.
+- [Quality gates](quality-gates.md): enable plan review, execution evidence
+  review, and configured recovery loops without replacing upstream planning.
 - [Operator runbook](operator-runbook.md): validate a workspace, inspect
   events, run smoke tests, and keep the runtime operable.
 

package/docs/guides/operator-runbook.md

@@ -20,6 +20,12 @@ npm run release:pack
 npm run release:smoke
 ```
 
+Stable Harness publishes compiled npm artifacts only. The package must not
+include TypeScript sources, source maps, source-map content markers, tests,
+scripts, examples, benchmarks, or development-only source directories. The
+release package may include compiled JavaScript, declaration files, package
+metadata, docs, README, NOTICE, and LICENSE.
+
 `release:smoke` installs the packed package into an isolated temporary project,
 initializes a workspace, and runs a CLI tool call. It catches missing published
 dependencies that monorepo tests can hide.

package/docs/guides/quality-gates.md

@@ -0,0 +1,80 @@
+# Quality Gates
+
+Stable Harness can supervise agent work without replacing the backend agent
+framework. Quality gates are runtime control-plane checks over plan evidence,
+execution evidence, blockers, and recovery loops.
+
+The normal user-facing configuration is intentionally small:
+
+```yaml
+runtime:
+  quality: balanced
+```
+
+Profiles:
+
+- `fast`: keep latency low. Run lightweight final-output checks and blocker
+  handling only when explicitly configured.
+- `balanced`: require upstream planning evidence for agents with declared
+  tools, skills, or subagents; review final execution evidence; allow limited
+  recovery.
+- `strict`: use the balanced checks and also require successful tool or
+  delegated-task evidence before final synthesis.
+
+Advanced configuration can override individual pieces:
+
+```yaml
+runtime:
+  quality:
+    profile: balanced
+    reviewer:
+      modelRef: local-dev
+      prompts:
+        planning: |
+          Review whether the plan is executable, complete, and grounded in the
+          declared tools, skills, and subagents. Return only the JSON verdict.
+        execution: |
+          Review whether the final answer is supported by executed evidence and
+          whether the agent should continue with ReAct. Return only the JSON
+          verdict.
+    executionReview:
+      requireToolEvidence: true
+    recovery:
+      maxLoops: 2
+```
+
+## Runtime Boundary
+
+Quality gates do not create a second planner. Planning still belongs to the
+upstream backend, such as DeepAgents. Stable Harness only observes upstream
+planning events, reviews generic evidence quality, and feeds structured gaps
+back into the same upstream agent loop.
+
+When `reviewer.modelRef` is configured, Stable Harness uses that LLM as the
+primary reviewer. The model must return JSON:
+
+```json
+{
+  "verdict": "pass | revise_plan | continue_react | blocked",
+  "issues": [
+    { "code": "missing_evidence", "message": "Explain the gap.", "recoverable": true }
+  ]
+}
+```
+
+The built-in deterministic reviewers remain as fallback and cover:
+
+- missing upstream planning evidence
+- empty final answers
+- missing successful tool evidence when required
+- control blockers such as approval blocks, invalid input, failed argument
+  repair, or repeated-tool limits
+
+If plan review fails, the runtime asks the agent to revise the plan using the
+declared inventory. If execution review fails, the runtime asks the agent to
+continue from the current state with ReAct: inspect the evidence gap, choose the
+next declared tool or subagent action, execute it, and synthesize from evidence.
+
+BetterCall remains focused on individual tool-call reliability. Quality gates
+consume tool and repair diagnostics as evidence, but they live in Stable Harness
+because they govern the run lifecycle rather than a single tool call.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "stable-harness",
-  "version": "0.0.18",
+  "version": "0.0.20",
   "type": "module",
   "description": "Stable application runtime and operator control plane for agent workspaces.",
   "license": "Apache-2.0",
@@ -62,7 +62,7 @@
     "benchmark:tool-guard:matrix": "node scripts/benchmarks/tool-argument-guard-matrix.mjs",
     "test:langmem:sqlite:e2e": "node scripts/run-langmem-sqlite-e2e.mjs",
     "validate:workspace": "node scripts/validate-workspace.mjs",
-    "test": "rm -rf dist/test && tsc -b test/tsconfig.json && node --test dist/test/*.test.js dist/test/adapter/*.test.js dist/test/compat/*.test.js dist/test/memory/*.test.js dist/test/protocol/*.test.js dist/test/retry/*.test.js dist/test/runtime/*.test.js dist/test/sdk/*.test.js dist/test/workspace/*.test.js",
+    "test": "rm -rf dist/test && tsc -b test/tsconfig.json && node --test dist/test/*.test.js dist/test/adapter/*.test.js dist/test/adapter/*/*.test.js dist/test/compat/*.test.js dist/test/memory/*.test.js dist/test/protocol/*.test.js dist/test/retry/*.test.js dist/test/runtime/*.test.js dist/test/runtime/*/*.test.js dist/test/sdk/*.test.js dist/test/workspace/*.test.js",
     "test:langmem:maintenance:e2e": "node scripts/run-langmem-maintenance-e2e.mjs",
     "test:skill-mining:e2e": "node scripts/run-skill-candidate-mining-e2e.mjs",
     "prepublishOnly": "npm run build && npm run release:check-package",
@@ -78,7 +78,7 @@
     "packages/*"
   ],
   "dependencies": {
-    "@botbotgo/better-call": "^0.1.54",
+    "@botbotgo/better-call": "^0.1.58",
     "@langchain/core": "^1.1.43",
     "@langchain/langgraph": "^1.3.0",
     "@langchain/langgraph-api": "^1.2.1",

package/packages/adapter-deepagents/dist/src/index.d.ts

@@ -1,3 +1,4 @@
 export { createDeepAgentsAdapter } from "./adapter.js";
 export { createDeepAgentsMemoryMaintenanceTarget, resolveDeepAgentsNativeMemories } from "./memory.js";
+export { createBackendModel } from "./model-providers.js";
 export type { DeepAgentFactory, DeepAgentsAdapterRunner } from "./types.js";

package/packages/adapter-deepagents/dist/src/index.js

@@ -1 +1 @@
-export{createDeepAgentsAdapter}from"./adapter.js";export{createDeepAgentsMemoryMaintenanceTarget,resolveDeepAgentsNativeMemories}from"./memory.js";
+export{createDeepAgentsAdapter}from"./adapter.js";export{createDeepAgentsMemoryMaintenanceTarget,resolveDeepAgentsNativeMemories}from"./memory.js";export{createBackendModel}from"./model-providers.js";

package/packages/adapter-deepagents/dist/src/internal/builtin-args.d.ts

@@ -1,4 +1,5 @@
 export declare function normalizeArgsRecord(args: unknown): Record<string, unknown>;
 export declare function normalizeWriteTodosArgs(args: unknown): Record<string, unknown>;
 export declare function normalizeFilesystemArgs(toolId: string, args: unknown, workspaceRoot: string): Record<string, unknown>;
+export declare function normalizeExecuteArgs(args: unknown, workspaceRoot: string): Record<string, unknown>;
 export declare function shallowEqualRecord(left: Record<string, unknown>, right: Record<string, unknown>): boolean;

package/packages/adapter-deepagents/dist/src/internal/builtin-args.js

@@ -1 +1 @@
-import r from"node:path";export function normalizeArgsRecord(r){if(isRecord(r))return r;if("string"!=typeof r||!r.trim())return{};try{const t=JSON.parse(r);return isRecord(t)?t:{}}catch{return{}}}export function normalizeWriteTodosArgs(r){if(Array.isArray(r))return{todos:r.map(normalizeTodoItem)};if("string"==typeof r&&r.trim())try{const t=JSON.parse(r);if(Array.isArray(t))return{todos:t.map(normalizeTodoItem)}}catch{return{todos:[{content:r.trim()}]}}const t=normalizeArgsRecord(r),e=readTodoList(t);return e?{...t,todos:e.map(normalizeTodoItem)}:t}export function normalizeFilesystemArgs(r,t,e){const n={...normalizeArgsRecord(t)};return normalizePathField(n,e,"path"),normalizePathField(n,e,"file_path"),normalizePathField(n,e,"pattern","glob"===r),n}export function shallowEqualRecord(r,t){const e=Object.keys(r),n=Object.keys(t);return e.length===n.length&&e.every(e=>r[e]===t[e])}function readTodoList(r){const t=r.todos??r.items??r.tasks??r.plan;if(Array.isArray(t))return t;if("string"==typeof t&&t.trim())try{const r=JSON.parse(t);return Array.isArray(r)?r:isRecord(r)?readTodoList(r):void 0}catch{return}}function normalizeTodoItem(r){if("string"==typeof r&&r.trim())return{content:r.trim()};if(!isRecord(r))return r;const t=readString(r.content)??readString(r.description)??readString(r.task)??readString(r.title)??readString(r.name),e=function normalizeTodoStatus(r){if(!r)return;const t=r.toLowerCase().replaceAll("-","_").replaceAll(" ","_");return"todo"===t||"not_started"===t?"pending":"doing"===t||"in_progress"===t?"in_progress":"done"===t||"complete"===t?"completed":t}(readString(r.status)??readString(r.state));return{...t?{content:t}:{},...e?{status:e}:{}}}function readString(r){return"string"==typeof r&&r.trim()?r.trim():void 0}function normalizePathField(t,e,n,o=!0){const i=o?readString(t[n]):void 0;i&&(t[n]=function workspaceBackendPath(t,e){if(!e.startsWith("/"))return e;if(function isPathInside(t,e){const n=r.relative(t,e);return""===n||!!n&&!n.startsWith("..")&&!r.isAbsolute(n)}(t,e)){const n=r.relative(t,e).split(r.sep).filter(Boolean).join("/");return n?`/${n}`:"/"}return e}(e,i))}function isRecord(r){return"object"==typeof r&&null!==r&&!Array.isArray(r)}
+import t from"node:path";export function normalizeArgsRecord(t){if(isRecord(t))return t;if("string"!=typeof t||!t.trim())return{};try{const r=JSON.parse(t);return isRecord(r)?r:{}}catch{return{}}}export function normalizeWriteTodosArgs(t){if(Array.isArray(t))return{todos:t.map(normalizeTodoItem)};if("string"==typeof t&&t.trim())try{const r=JSON.parse(t);if(Array.isArray(r))return{todos:r.map(normalizeTodoItem)}}catch{return{todos:[{content:t.trim()}]}}const r=normalizeArgsRecord(t),e=readTodoList(r);return e?{...r,todos:e.map(normalizeTodoItem)}:r}export function normalizeFilesystemArgs(t,r,e){const i={...normalizeArgsRecord(r)};return copyFirstStringAlias(i,"path",function filesystemPathAliases(t){return"grep"===t?["file","filename","filepath","filePath","directory","dir"]:["file_path","file","filename","filepath","filePath","target","directory","dir"]}(t)),copyFirstStringAlias(i,"pattern",function filesystemPatternAliases(t){return"glob"!==t?[]:["query","regex","search"]}(t)),copyFirstStringAlias(i,"content","write_file"===t?["text","body"]:[]),copyFirstStringAlias(i,"old_string","edit_file"===t?["oldText","old","replace","from"]:[]),copyFirstStringAlias(i,"new_string","edit_file"===t?["newText","new","with","to"]:[]),normalizePathField(i,e,"path"),normalizePathField(i,e,"file_path"),normalizePathField(i,e,"pattern","glob"===t),i}export function normalizeExecuteArgs(t,r){const e={...normalizeArgsRecord(t)};return copyFirstStringAlias(e,"command",["cmd","shell","input"]),normalizePathField(e,r,"cwd"),e}export function shallowEqualRecord(t,r){const e=Object.keys(t),i=Object.keys(r);return e.length===i.length&&e.every(e=>t[e]===r[e])}function readTodoList(t){const r=t.todos??t.items??t.tasks??t.plan;if(Array.isArray(r))return r;if(isRecord(r))return readTodoList(r)??[r];if("string"==typeof r&&r.trim())try{const t=JSON.parse(r);return Array.isArray(t)?t:isRecord(t)?readTodoList(t):void 0}catch{return}}function normalizeTodoItem(t){if("string"==typeof t&&t.trim())return{content:t.trim()};if(!isRecord(t))return t;const r=readString(t.content)??readString(t.description)??readString(t.task)??readString(t.step)??readString(t.gap)??readString(t.evidenceGap)??readString(t.evidence_gap)??readString(t.action)??readString(t.title)??readString(t.name),e=function normalizeTodoStatus(t){if(!t)return;const r=t.toLowerCase().replaceAll("-","_").replaceAll(" ","_");return"todo"===r||"not_started"===r||"planned"===r||"blocked"===r?"pending":"doing"===r||"in_progress"===r||"partial"===r?"in_progress":"done"===r||"complete"===r||"filled"===r?"completed":r}(readString(t.status)??readString(t.state));return{...r?{content:r}:{},...e?{status:e}:{}}}function readString(t){return"string"==typeof t&&t.trim()?t.trim():void 0}function copyFirstStringAlias(t,r,e){if(!readString(t[r]))for(const i of e){const e=readString(t[i]);if(e)return void(t[r]=e)}}function normalizePathField(r,e,i,n=!0){const o=n?readString(r[i]):void 0;o&&(r[i]=function workspaceBackendPath(r,e){if(!e.startsWith("/"))return e;if(function isPathInside(r,e){const i=t.relative(r,e);return""===i||!!i&&!i.startsWith("..")&&!t.isAbsolute(i)}(r,e)){const i=t.relative(r,e).split(t.sep).filter(Boolean).join("/");return i?`/${i}`:"/"}return e}(e,o))}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}

package/packages/adapter-deepagents/dist/src/internal/builtin-call-repair.d.ts

@@ -0,0 +1,9 @@
+export declare function repairBuiltinToolRequest<T extends {
+    toolCall?: {
+        args?: unknown;
+    };
+}>(input: {
+    toolId: string;
+    request: T;
+    workspaceRoot: string;
+}): Promise<T>;

package/packages/adapter-deepagents/dist/src/internal/builtin-call-repair.js

@@ -0,0 +1 @@
+import{repairCallSelection as t}from"@botbotgo/better-call";import{normalizeArgsRecord as e,normalizeExecuteArgs as r,normalizeFilesystemArgs as o,normalizeWriteTodosArgs as i,shallowEqualRecord as s}from"./builtin-args.js";const a=new Set(["ls","read_file","write_file","edit_file","glob","grep"]);export async function repairBuiltinToolRequest(e){const r=normalizeBuiltinArgs(e.toolId,e.request.toolCall?.args,e.workspaceRoot),o=function builtinCandidate(t){const e=n[t];return e?{id:t,description:`DeepAgents builtin tool ${t}`,schema:e}:void 0}(e.toolId);if(!o)return updateRequestArgs(e.request,e.request.toolCall?.args,r);const i=await t({call:{id:e.toolId,args:r},candidates:[o],mode:"repair"}),s=i.ok?normalizeBuiltinArgs(e.toolId,i.args,e.workspaceRoot):r;return updateRequestArgs(e.request,e.request.toolCall?.args,s)}function normalizeBuiltinArgs(t,s,n){return"write_todos"===t?i(s):"read_todos"===t||"task"===t?e(s):"execute"===t?r(s,n):a.has(t)?o(t,s,n):e(s)}function updateRequestArgs(t,e,r){return function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}(e)&&s(e,r)?t:{...t,toolCall:{...t.toolCall,args:r}}}const n={write_todos:objectSchema({todos:{type:"array",items:{type:"object",additionalProperties:!0}}}),read_todos:objectSchema({}),task:objectSchema({subagent_type:{type:"string"},subagentType:{type:"string"},description:{type:"string"},task:{type:"string"}}),execute:objectSchema({command:{type:"string"},cwd:{type:"string"},timeoutMs:{type:"number"}}),ls:objectSchema({path:{type:"string"}}),read_file:objectSchema({path:{type:"string"},file_path:{type:"string"}}),write_file:objectSchema({path:{type:"string"},file_path:{type:"string"},content:{type:"string"}}),edit_file:objectSchema({path:{type:"string"},file_path:{type:"string"},old_string:{type:"string"},new_string:{type:"string"}}),glob:objectSchema({path:{type:"string"},pattern:{type:"string"}}),grep:objectSchema({path:{type:"string"},pattern:{type:"string"}})};function objectSchema(t){return{type:"object",properties:t,additionalProperties:!0}}

package/packages/adapter-deepagents/dist/src/internal/builtin-tool-policy.js

@@ -1 +1 @@
-import{realpathSync as e}from"node:fs";import{repairCallSelection as t}from"@botbotgo/better-call";import{ToolMessage as o}from"@langchain/core/messages";import{normalizeArgsRecord as n,normalizeFilesystemArgs as r,normalizeWriteTodosArgs as i,shallowEqualRecord as a}from"./builtin-args.js";import{afterToolInvoke as s,beforeToolInvoke as l,createToolRepeatState as u,stringifyDeepAgentResult as c,toolControlProjection as d}from"./gateway-tools.js";import{validateSkillFileBuiltinCall as g}from"./skill-file-policy.js";import{filterRepeatLimitedTools as p}from"./tool-repeat-visibility.js";import{traceProjectionForBuiltinTool as f}from"./trace-projection.js";const m=new Set(["ls","read_file","write_file","edit_file","glob","grep"]),h=new Set(["write_todos","task",...m]);export function createBuiltinToolPolicyMiddleware(e,t={}){return{name:"StableHarnessBuiltinToolPolicy",async wrapModelCall(o,n){const r=Array.isArray(o.tools)?p(o.tools.filter(t=>!function hasHiddenBuiltins(e){return isFilesystemDisabled(e)||!isTaskVisible(e)}(e)||isModelVisibleBuiltin(e,t.name)),t.repeatState):o.tools,i=function normalizeToolChoice(e,t,o){return"required"===t?o&&o.length>0?t:"auto":function isForcedHiddenTool(e,t){return"string"==typeof t?.function?.name&&!isModelVisibleBuiltin(e,t.function.name)}(e,t)?"auto":t}(e,o.toolChoice,r);return n({...o,tools:r,toolChoice:i})}}}export function validateFilesystemBuiltinCall(e,t,n){if(isFilesystemDisabled(e)&&m.has(t))return new o({tool_call_id:n.toolCall?.id??`stable-harness-${t}-policy`,name:t,content:`Filesystem builtin tool '${t}' is disabled for this agent. Do not retry filesystem tools; use the agent's registered workspace tools and already collected evidence instead.`})}export function createObserverMiddleware(e,d={}){const p=d.repeatState??u(e.workspace.runtime.toolGateway);return{name:"StableHarnessObserver",async wrapToolCall(u,f){const y=u.toolCall?.name;if(!y||!h.has(y))return f(u);const b=function normalizeBuiltinToolRequest(e,t,o){return"write_todos"===t?{...o,toolCall:{...o.toolCall,args:i(o.toolCall?.args)}}:"task"===t?{...o,toolCall:{...o.toolCall,args:n(o.toolCall?.args)}}:m.has(t)?function normalizeFilesystemToolRequest(e,t,o){const i=n(o.toolCall?.args),s=r(t,i,e.workspace.root);return a(i,s)?o:{...o,toolCall:{...o.toolCall,args:s}}}(e,t,o):o}(e,y,u);emitToolEvent(e,y,"agent.tool.start",b.toolCall?.args);const v=p?l(y,b.toolCall?.args,p):void 0;if(v)return emitToolEvent(e,y,"agent.tool.result",b.toolCall?.args,{output:v.eventOutput}),builtinToolMessage(u,y,v.modelOutput);const C="task"===y?await async function repairTaskCall(e,r){const i=function allowedTaskTypes(e){const t=readConfigRecord(e.agent.config,"deepagents");if(!0===t?.generalPurposeAgent)return;const o=readConfigRecord(e.agent.config,"builtinTools")?.modelExposed;return!1===o?[]:Array.isArray(o)?o.includes("task")?e.agent.subagents:[]:0===e.agent.subagents.length?[]:void 0}(e);if(void 0===i)return{request:r};const a=n(r.toolCall?.args),s=function readTaskSubagentType(e){const t=isRecord(e)?e:{};return readString(t.subagent_type)??readString(t.subagentType)}(a);if(s&&i.includes(s))return{request:r};const l=await t({call:{id:s,args:a},candidates:taskCallCandidates(e,i),mode:"repair"});if(l.ok){const e={...a,...l.args,subagent_type:l.candidateId};return{request:{...r,toolCall:{...r.toolCall,args:e}}}}const u=s?`: ${s}`:"",c=i.length>0?i.join(", "):"none";return{request:r,blocked:new o({tool_call_id:r.toolCall?.id??"stable-harness-task-policy",name:"task",status:"error",content:[`Task delegation target is not in the workspace inventory${u}. Allowed task targets: ${c}.`,"Retry with an allowed target only when that target semantically owns the original user request.","Do not substitute another specialist just to continue the same evidence need; synthesize from already collected evidence when no allowed target is a semantic match."].join(" ")})}}(e,b):{request:b},k=C.request,w=C.blocked;if(w)return emitToolEvent(e,y,"agent.tool.result",k.toolCall?.args,{output:w.content}),w;const T=validateFilesystemBuiltinCall(e,y,k);if(T)return emitToolEvent(e,y,"agent.tool.result",k.toolCall?.args,{output:T.content}),T;const R=g(e,y,k);if(R)return emitToolEvent(e,y,"agent.tool.result",k.toolCall?.args,{output:R.content}),R;try{const t=await f(k),o=function observedToolOutput(e,t,o){return"write_todos"===e?JSON.stringify({status:"recorded",args:t.toolCall?.args}):c(o)}(y,k,t),n=p?s(y,k.toolCall?.args,o,t,p):{};return emitToolEvent(e,y,"agent.tool.result",k.toolCall?.args,{output:n.eventOutput??o}),d.observedToolIds?.add(y),void 0===n.modelOutput?t:builtinToolMessage(u,y,n.modelOutput)}catch(t){const n=function recoverableBuiltinToolError(e,t,n){const r=formatError(n);return"task"===e&&/repeat limit reached for tool/iu.test(r)?new o({tool_call_id:t.toolCall?.id??"stable-harness-task-repeat-limit",name:"task",content:JSON.stringify({status:"delegated_task_repeat_limit",finalizationRequired:!0,instruction:"The delegated agent reached a configured tool repeat limit. Stop delegating this evidence need and do not send a synthesis task to another subagent for the same need. Finalize only from evidence that is already visible in this run. If the visible evidence is insufficient for a requested claim, report an explicit blocker or evidence gap instead of estimating, inventing, or using generic knowledge.",error:previewError(r)})}):/Received tool input did not match expected schema|Invalid input:/iu.test(r)?new o({tool_call_id:t.toolCall?.id??`stable-harness-${e}-argument-error`,name:e,status:"error",content:JSON.stringify({status:"tool_argument_error",toolId:e,instruction:"The upstream builtin tool rejected these arguments. Fix the tool arguments according to the tool schema, or choose a more appropriate available tool.",error:previewError(r)})}):void 0}(y,u,t);if(n)return emitToolEvent(e,y,"agent.tool.result",b.toolCall?.args,{output:n.content}),n;throw emitToolEvent(e,y,"agent.tool.result",b.toolCall?.args,{error:formatError(t)}),t}}}}function builtinToolMessage(e,t,n){return new o({tool_call_id:e.toolCall?.id??`stable-harness-${t}-repeat-guard`,name:t,content:n})}export function resolveFilesystemPermissions(e,t){const o=readConfigRecord(t?.config,"builtinTools"),n=[];if(n.push(...function skillReadPermissions(e,t){const o=[...new Set((t?.skills??[]).flatMap(t=>function skillReadPaths(e,t){return t?[...new Set([t,canonicalPath(t),backendSkillPath(e,t)])].flatMap(e=>function skillReadPathCandidates(e){const t=e.replace(/\/+$/u,""),o=t.endsWith("/SKILL.md")?t.slice(0,-9):t,n=function parentPath(e){const t=e.lastIndexOf("/");return t>0?e.slice(0,t):void 0}(o);return[t,o,`${o}/**`,...n?[n,`${n}/**`]:[]]}(e)):[]}(e.workspace.root,e.workspace.skills.get(t)?.path)).filter(e=>e.startsWith("/")))];return o.length>0?[{operations:["read"],paths:o,mode:"allow"}]:[]}(e,t)),!1!==o?.filesystem){if(deepagentsMemoryWritable(e))return n.length>0?n:void 0}else n.push({operations:["read"],paths:["/memories/**"],mode:"allow"}),n.push({operations:["read","write"],paths:["/**"],mode:"deny"});return deepagentsMemoryWritable(e)||n.push({operations:["write"],paths:["/memories/**"],mode:"deny"}),n.length>0?n:void 0}function emitToolEvent(e,t,o,n,r={}){e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,event:{adapter:"deepagents",phase:o,toolId:t,..."agent.tool.start"===o?{args:n}:{},...r,..."string"==typeof r.output?d(r.output):{},...f(t,o,n)}})}function backendSkillPath(e,t){const o=function pathRelative(e,t){const o=e.split("/").filter(Boolean),n=t.split("/").filter(Boolean);if(!(n.length<o.length)){for(let e=0;e<o.length;e+=1)if(o[e]!==n[e])return;return n.slice(o.length).join("/")}}(e,t);return void 0===o?t:`/${o.split("/").filter(Boolean).join("/")}`}function canonicalPath(t){try{return e.native(t)}catch{return t}}function deepagentsMemoryWritable(e){const t=readConfigRecord(e.workspace.runtime.memory,"deepagentsMem");return!1!==t?.write}function taskCallCandidates(e,t){return t.map(t=>({id:t,description:e.workspace.agents.get(t)?.description,schema:{type:"object",properties:{subagent_type:{type:"string"},subagentType:{type:"string"},description:{type:"string"}},required:["description"],additionalProperties:!0}}))}function isFilesystemDisabled(e){const t=readConfigRecord(e.agent.config,"builtinTools");return!1===t?.filesystem}function isTaskVisible(e){const t=readConfigRecord(e.agent.config,"builtinTools")?.modelExposed;return!1!==t&&(!Array.isArray(t)||t.includes("task"))}function isModelVisibleBuiltin(e,t){return(!isFilesystemDisabled(e)||!function isFilesystemTool(e){return"string"==typeof e&&m.has(e)}(t))&&("task"!==t||isTaskVisible(e))}function readConfigRecord(e,t){const o=isRecord(e)?e:{};return isRecord(o[t])?o[t]:void 0}function readString(e){return"string"==typeof e&&e.trim()?e:void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function formatError(e){return e instanceof Error?e.message:String(e)}function previewError(e){const t=e.replace(/\s+/gu," ").trim();return t.length>800?`${t.slice(0,797)}...`:t}
+import{realpathSync as e}from"node:fs";import{repairCallSelection as t}from"@botbotgo/better-call";import{ToolMessage as o}from"@langchain/core/messages";import{normalizeArgsRecord as n}from"./builtin-args.js";import{repairBuiltinToolRequest as r}from"./builtin-call-repair.js";import{afterToolInvoke as i,beforeToolInvoke as a,createToolRepeatState as s,stringifyDeepAgentResult as l,toolControlProjection as u}from"./gateway-tools.js";import{observedToolEvidence as c,recordObservedToolEvidence as d}from"./gateway/tool-evidence.js";import{validateSkillFileBuiltinCall as g}from"./skill-file-policy.js";import{filterRepeatLimitedTools as p}from"./tool-repeat-visibility.js";import{traceProjectionForBuiltinTool as f}from"./trace-projection.js";const m=new Set(["ls","read_file","write_file","edit_file","glob","grep"]),h=new Set(["write_todos","read_todos","task","execute",...m]);export function createBuiltinToolPolicyMiddleware(e,t={}){return{name:"StableHarnessBuiltinToolPolicy",async wrapModelCall(o,n){const r=Array.isArray(o.tools)?p(o.tools.filter(t=>!function hasHiddenBuiltins(e){return isFilesystemDisabled(e)||!isTaskVisible(e)}(e)||isModelVisibleBuiltin(e,t.name)),t.repeatState):o.tools,i=function normalizeToolChoice(e,t,o){return"required"===t?o&&o.length>0?t:"auto":function isForcedHiddenTool(e,t){return"string"==typeof t?.function?.name&&!isModelVisibleBuiltin(e,t.function.name)}(e,t)?"auto":t}(e,o.toolChoice,r);return n({...o,tools:r,toolChoice:i})}}}export function validateFilesystemBuiltinCall(e,t,n){if(isFilesystemDisabled(e)&&m.has(t))return new o({tool_call_id:n.toolCall?.id??`stable-harness-${t}-policy`,name:t,content:`Filesystem builtin tool '${t}' is disabled for this agent. Do not retry filesystem tools; use the agent's registered workspace tools and already collected evidence instead.`})}export function createObserverMiddleware(e,u={}){const d=u.repeatState??s(e.workspace.runtime.toolGateway);return{name:"StableHarnessObserver",async wrapToolCall(s,p){const f=s.toolCall?.name;if(!f||!h.has(f))return p(s);const m=await r({toolId:f,request:s,workspaceRoot:e.workspace.root});emitToolEvent(e,f,"agent.tool.start",m.toolCall?.args);const y="task"===f?await async function repairTaskCall(e,r){const i=function allowedTaskTypes(e){const t=readConfigRecord(e.agent.config,"deepagents");if(!0===t?.generalPurposeAgent)return;const o=readConfigRecord(e.agent.config,"builtinTools")?.modelExposed;return!1===o?[]:Array.isArray(o)?o.includes("task")?e.agent.subagents:[]:0===e.agent.subagents.length?[]:void 0}(e);if(void 0===i)return{request:r};const a=n(r.toolCall?.args),s=function readTaskSubagentType(e){const t=isRecord(e)?e:{};return readString(t.subagent_type)??readString(t.subagentType)}(a);if(s&&i.includes(s))return{request:r};const l=await t({call:{id:s,args:a},candidates:taskCallCandidates(e,i),mode:"repair"});if(l.ok){const e={...a,...l.args,subagent_type:l.candidateId};return{request:{...r,toolCall:{...r.toolCall,args:e}}}}const u=s?`: ${s}`:"",c=i.length>0?i.join(", "):"none";return{request:r,blocked:new o({tool_call_id:r.toolCall?.id??"stable-harness-task-policy",name:"task",status:"error",content:[`Task delegation target is not in the workspace inventory${u}. Allowed task targets: ${c}.`,"Retry with an allowed target only when that target semantically owns the original user request.","Do not substitute another specialist just to continue the same evidence need; synthesize from already collected evidence when no allowed target is a semantic match."].join(" ")})}}(e,m):{request:m},b=y.request,v=y.blocked;if(v)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:v.content}),v;const w=validateFilesystemBuiltinCall(e,f,b);if(w)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:w.content}),w;const k=g(e,f,b);if(k)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:k.content}),k;try{const t=d?a(f,b.toolCall?.args,d):void 0;if(t)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:t.eventOutput}),builtinToolMessage(s,f,t.modelOutput);const o=await p(b),n=function observedToolOutput(e,t,o){return"write_todos"===e?JSON.stringify({status:"recorded",args:t.toolCall?.args}):l(o)}(f,b,o),r=d?i(f,b.toolCall?.args,n,o,d):{};return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:r.eventOutput??n}),u.observedToolIds?.add(f),void 0===r.modelOutput?o:builtinToolMessage(s,f,r.modelOutput)}catch(t){const n=function recoverableBuiltinToolError(e,t,n,r){const i=formatError(r);if("task"===t&&/repeat limit reached for tool/iu.test(i)){const t=function formatObservedEvidence(e){const t=c(e);if(0===t.length)return"";const o=t.map(e=>[`Agent: ${e.agentId}`,`Tool: ${e.toolId}`,e.output].join("\n")).join("\n\n---\n\n");return o.length>12e3?`${o.slice(0,12e3)}\n[truncated]`:o}(e);return new o({tool_call_id:n.toolCall?.id??"stable-harness-task-repeat-limit",name:"task",content:JSON.stringify({status:"delegated_task_repeat_limit",finalizationRequired:!0,instruction:"The delegated agent reached a configured tool repeat limit. Stop delegating this evidence need and do not send a synthesis task to another subagent for the same need. Finalize only from observedEvidence and other evidence that is already visible in this run. If the visible evidence is insufficient for a requested claim, report an explicit blocker or evidence gap instead of estimating, inventing, or using generic knowledge.",...t?{observedEvidence:t}:{},error:previewError(i)})})}if(/Received tool input did not match expected schema|Invalid input:/iu.test(i))return new o({tool_call_id:n.toolCall?.id??`stable-harness-${t}-argument-error`,name:t,status:"error",content:JSON.stringify({status:"tool_argument_error",toolId:t,instruction:"The upstream builtin tool rejected these arguments. Fix the tool arguments according to the tool schema, or choose a more appropriate available tool.",error:previewError(i)})})}(e,f,s,t);if(n)return emitToolEvent(e,f,"agent.tool.result",m.toolCall?.args,{output:n.content}),n;throw emitToolEvent(e,f,"agent.tool.result",m.toolCall?.args,{error:formatError(t)}),t}}}}function builtinToolMessage(e,t,n){return new o({tool_call_id:e.toolCall?.id??`stable-harness-${t}-repeat-guard`,name:t,content:n})}export function resolveFilesystemPermissions(e,t){const o=readConfigRecord(t?.config,"builtinTools"),n=[];if(n.push(...function skillReadPermissions(e,t){const o=[...new Set((t?.skills??[]).flatMap(t=>function skillReadPaths(e,t){return t?[...new Set([t,canonicalPath(t),backendSkillPath(e,t)])].flatMap(e=>function skillReadPathCandidates(e){const t=e.replace(/\/+$/u,""),o=t.endsWith("/SKILL.md")?t.slice(0,-9):t,n=function parentPath(e){const t=e.lastIndexOf("/");return t>0?e.slice(0,t):void 0}(o);return[t,o,`${o}/**`,...n?[n,`${n}/**`]:[]]}(e)):[]}(e.workspace.root,e.workspace.skills.get(t)?.path)).filter(e=>e.startsWith("/")))];return o.length>0?[{operations:["read"],paths:o,mode:"allow"}]:[]}(e,t)),!1!==o?.filesystem){if(deepagentsMemoryWritable(e))return n.length>0?n:void 0}else n.push({operations:["read"],paths:["/memories/**"],mode:"allow"}),n.push({operations:["read","write"],paths:["/**"],mode:"deny"});return deepagentsMemoryWritable(e)||n.push({operations:["write"],paths:["/memories/**"],mode:"deny"}),n.length>0?n:void 0}function emitToolEvent(e,t,o,n,r={}){"string"==typeof r.output&&d(e,e.agent.id,t,r.output),e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,event:{adapter:"deepagents",phase:o,toolId:t,..."agent.tool.start"===o?{args:n}:{},...r,..."string"==typeof r.output?u(r.output):{},...f(t,o,n)}})}function backendSkillPath(e,t){const o=function pathRelative(e,t){const o=e.split("/").filter(Boolean),n=t.split("/").filter(Boolean);if(!(n.length<o.length)){for(let e=0;e<o.length;e+=1)if(o[e]!==n[e])return;return n.slice(o.length).join("/")}}(e,t);return void 0===o?t:`/${o.split("/").filter(Boolean).join("/")}`}function canonicalPath(t){try{return e.native(t)}catch{return t}}function deepagentsMemoryWritable(e){const t=readConfigRecord(e.workspace.runtime.memory,"deepagentsMem");return!1!==t?.write}function taskCallCandidates(e,t){return t.map(t=>({id:t,description:e.workspace.agents.get(t)?.description,schema:{type:"object",properties:{subagent_type:{type:"string"},subagentType:{type:"string"},description:{type:"string"}},required:["description"],additionalProperties:!0}}))}function isFilesystemDisabled(e){const t=readConfigRecord(e.agent.config,"builtinTools");return!1===t?.filesystem}function isTaskVisible(e){const t=readConfigRecord(e.agent.config,"builtinTools")?.modelExposed;return!1!==t&&(!Array.isArray(t)||t.includes("task"))}function isModelVisibleBuiltin(e,t){return(!isFilesystemDisabled(e)||!function isFilesystemTool(e){return"string"==typeof e&&m.has(e)}(t))&&("task"!==t||isTaskVisible(e))}function readConfigRecord(e,t){const o=isRecord(e)?e:{};return isRecord(o[t])?o[t]:void 0}function readString(e){return"string"==typeof e&&e.trim()?e:void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function formatError(e){return e instanceof Error?e.message:String(e)}function previewError(e){const t=e.replace(/\s+/gu," ").trim();return t.length>800?`${t.slice(0,797)}...`:t}

package/packages/adapter-deepagents/dist/src/internal/gateway/tool-evidence.d.ts

@@ -0,0 +1,10 @@
+import type { RuntimeAdapter } from "@stable-harness/core";
+export type ObservedToolEvidence = {
+    agentId: string;
+    toolId: string;
+    output: string;
+};
+export declare function observedToolEvidence(input: Parameters<RuntimeAdapter["run"]>[0]): ObservedToolEvidence[];
+export declare function recordObservedToolEvidence(input: Parameters<RuntimeAdapter["run"]>[0], agentId: string, toolId: string, output: string): void;
+export declare function isControlOnlyOutput(output: string): boolean;
+export declare function isSuccessfulEvidenceOutput(output: string): boolean;

package/packages/adapter-deepagents/dist/src/internal/gateway/tool-evidence.js

@@ -0,0 +1 @@
+const e="deepagents.observedToolEvidence";export function observedToolEvidence(t){const r=t.requestState?.get(e);return Array.isArray(r)?r:[]}export function recordObservedToolEvidence(t,r,o,n){if(!t.requestState||!isSuccessfulEvidenceOutput(n)||isControlOnlyOutput(n))return;const s=observedToolEvidence(t);t.requestState.set(e,[...s,{agentId:r,toolId:o,output:n}].slice(-12))}export function isControlOnlyOutput(e){return/"status"\s*:\s*"(?:duplicate_tool_call|repeated_tool_call_limit|tool_argument_error|delegated_task_repeat_limit|invalid_input|blocked|failed|error|plan_required|dependency_required)"|Status:\s*(?:invalid_input|blocked|failed|error|plan_required|dependency_required)\b|Filesystem builtin tool .* is disabled/iu.test(e)}export function isSuccessfulEvidenceOutput(e){const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}(t)?t:void 0}catch{return}}(e),r="string"==typeof t?.status?t.status:function readTextStatus(e){return String(e).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(e);return!r||/^(?:completed|success|ok|recorded)$/iu.test(r)}

package/packages/adapter-deepagents/dist/src/internal/gateway-tools.js

@@ -1 +1 @@
-import{ToolMessage as t}from"@langchain/core/messages";import{tool as e}from"@langchain/core/tools";export function buildGatewayTools(o,r,n,s,a=createToolRepeatState(o.workspace.runtime.toolGateway)){return o.toolGateway?n.flatMap(n=>{const l=o.toolGateway?.get(n);if(!l)return[];const i=o.workspace.tools.get(n),u=i?.schema??l.schema;return[e(async e=>async function invokeGuardedGatewayTool(e){emitToolResult(e.input,e.agentId,e.toolId,void 0);const o=e.repeatState?beforeToolInvoke(e.toolId,e.args,e.repeatState):void 0;if(o)return emitToolResult(e.input,e.agentId,e.toolId,o.eventOutput),o.modelOutput;const r=await async function invokeGatewayTool(e,o,r,n,s){try{return await e.toolGateway.invoke({toolId:r,args:n,repairModel:s,context:{workspaceRoot:e.workspace.root,requestId:e.requestId,sessionId:e.sessionId,agentId:o}})}catch(o){if(function isToolArgumentValidationError(t){return t instanceof Error&&"ToolArgumentValidationError"===t.name&&"string"==typeof t.toolId&&Array.isArray(t.issues)}(o))return new t({tool_call_id:`stable-harness-${r}-argument-guard`,name:r,status:"error",content:formatToolArgumentError(o)});if(e.workspace.runtime.retry?.tools?.enabled)throw o;return new t({tool_call_id:`stable-harness-${r}-execution-error`,name:r,status:"error",content:JSON.stringify({error:"tool_execution_failed",toolId:r,message:formatError(o),retry:"Use the error as evidence, adjust the tool arguments if possible, or return a final answer with the blocker."})})}}(e.input,e.agentId,e.toolId,e.args,e.repairModel),n=r instanceof t?String(r.content):stringifyDeepAgentResult(r.output),s=e.repeatState?afterToolInvoke(e.toolId,e.args,n,r,e.repeatState):{};return emitToolResult(e.input,e.agentId,e.toolId,s.eventOutput??n),void 0!==s.modelOutput?s.modelOutput:r instanceof t?r:n}({input:o,agentId:r,toolId:n,args:e,repairModel:s,repeatState:a}),{name:n,description:buildToolDescription(i?.description??l.description??n,u,o.workspace.runtime.toolGateway,n),schema:{type:"object",additionalProperties:!0}})]}):[]}export function createToolRepeatState(t){if(function repeatGuardEnabled(t){return!0===repeatGuardConfig(t).enabled}(t))return{successfulCalls:new Map,duplicateCallCounts:new Map,latestSuccessfulOutputByTool:new Map,successfulToolCounts:new Map,toolCallCounts:new Map,maxDuplicateCallsPerTool:readPositiveInteger(repeatGuardConfig(t).maxDuplicateCallsPerTool)??3,maxCallsPerTool:readPositiveInteger(repeatGuardConfig(t).maxCallsPerTool),maxSuccessfulCallsPerTool:readPositiveInteger(repeatGuardConfig(t).maxSuccessfulCallsPerTool),maxCallsByTool:readPositiveIntegerMap(repeatGuardConfig(t).maxCallsByTool),maxSuccessfulCallsByTool:readPositiveIntegerMap(repeatGuardConfig(t).maxSuccessfulCallsByTool)}}function repeatGuardConfig(t){return isRecord(t)&&isRecord(t.repeatGuard)?t.repeatGuard:{}}function readPositiveInteger(t){return"number"==typeof t&&Number.isInteger(t)&&t>0?t:void 0}function readPositiveIntegerMap(t){return isRecord(t)?new Map(Object.entries(t).map(([t,e])=>[t,readPositiveInteger(e)]).filter(t=>void 0!==t[1])):new Map}export function beforeToolInvoke(t,e,o){const r=o.toolCallCounts.get(t)??0;o.toolCallCounts.set(t,r+1);const n=o.maxCallsByTool.get(t)??o.maxCallsPerTool;if(void 0!==n&&r>=n)throw new Error(`Stable Harness repeat limit reached for tool '${t}' after ${n} call(s).`);const s=o.maxSuccessfulCallsByTool.get(t)??o.maxSuccessfulCallsPerTool;if(void 0!==s&&(o.successfulToolCounts.get(t)??0)>=s){const e=repeatedToolCallLimitContent(t,o.latestSuccessfulOutputByTool.get(t));return{eventOutput:e,modelOutput:e}}const a=stableToolCallKey(t,e),l=o.successfulCalls.get(a);if(void 0!==l){const e=o.duplicateCallCounts.get(a)??0;if(o.duplicateCallCounts.set(a,e+1),void 0!==o.maxDuplicateCallsPerTool&&e>=o.maxDuplicateCallsPerTool){const e=repeatedToolCallLimitContent(t);return{eventOutput:e,modelOutput:e}}const r=function duplicateToolCallContent(t,e){return JSON.stringify({status:"duplicate_tool_call",toolId:t,instruction:"This agent already completed an equivalent tool call. Use the prior evidence instead of calling the tool again.",previousOutput:e})}(t,l);return{eventOutput:r,modelOutput:l}}}export function isToolRepeatLimitReached(t,e){if(!e)return!1;const o=e.maxCallsByTool.get(t)??e.maxCallsPerTool;if(void 0!==o&&(e.toolCallCounts.get(t)??0)>=o)return!0;const r=e.maxSuccessfulCallsByTool.get(t)??e.maxSuccessfulCallsPerTool;return void 0!==r&&(e.successfulToolCounts.get(t)??0)>=r}export function afterToolInvoke(e,o,r,n,s){return n instanceof t&&"error"===n.status||(s.successfulCalls.set(stableToolCallKey(e,o),r),s.latestSuccessfulOutputByTool.set(e,r),s.successfulToolCounts.set(e,(s.successfulToolCounts.get(e)??0)+1)),{}}function emitToolResult(t,e,o,r){t.emit({type:"runtime.adapter.event",requestId:t.requestId,sessionId:t.sessionId,agentId:e,event:void 0===r?{adapter:"deepagents",phase:"agent.tool.start",toolId:o}:{adapter:"deepagents",phase:"agent.tool.result",toolId:o,output:previewToolOutput(r),...toolControlProjection(r)}})}function repeatedToolCallLimitContent(t,e){return JSON.stringify({status:"repeated_tool_call_limit",toolId:t,instruction:"This tool reached the configured repeat limit for this request. Do not call this tool or a substitute tool for the same evidence need again. Use previousOutput and the collected evidence to produce the final answer now, or report the remaining gap explicitly.",...void 0!==e?{previousOutput:e}:{}})}export function stringifyDeepAgentResult(e){if(e instanceof t)return function stringifyToolMessageContent(t){return"string"==typeof t?t:JSON.stringify(t)}(e.content);if("string"==typeof e)return e;if(isRecord(e)){const t=e.structuredResponse??e.structured_response;if(void 0!==t)return"string"==typeof t?t:JSON.stringify(t);const o=(Array.isArray(e.messages)?e.messages:[]).at(-1);if(isRecord(o)&&"string"==typeof o.content)return o.content;const r=(isRecord(e.update)&&Array.isArray(e.update.messages)?e.update.messages:[]).at(-1);if(isRecord(r)&&isRecord(r.kwargs)&&"string"==typeof r.kwargs.content)return r.kwargs.content;if(isRecord(r)&&"string"==typeof r.content)return r.content}return JSON.stringify(e)}function buildToolDescription(t,e,o,r){const n=function toolRepeatPolicyDescription(t,e){const o=repeatGuardConfig(t),r=readPositiveIntegerMap(o.maxSuccessfulCallsByTool).get(e)??readPositiveInteger(o.maxSuccessfulCallsPerTool);return void 0===r?"":`Stable runtime repeat policy: call this tool at most ${r} successful time(s) for this request. If more detail is needed, include the dimensions in the first call and synthesize after the result returns.`}(o,r),s=n?`${t}\n\n${n}`:t;return e?`${s}\n\nStable tool input schema:\n${previewToolOutput(JSON.stringify(e))}`:s}function previewToolOutput(t){const e=t.replace(/\s+/gu," ").trim();return e.length>500?`${e.slice(0,497)}...`:e}export function toolControlProjection(t){const e=function parseJsonRecord(t){try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t);return"string"==typeof e?.status?{controlStatus:e.status}:"string"==typeof e?.error?{controlStatus:e.error}:t.startsWith("Task delegation target is not in the workspace inventory")?{controlStatus:"task_inventory_blocked"}:{}}function stableToolCallKey(t,e){return`${t}:${stableJson(e)}`}function stableJson(t){return Array.isArray(t)?`[${t.map(stableJson).join(",")}]`:isRecord(t)?`{${Object.keys(t).sort().map(e=>`${JSON.stringify(e)}:${stableJson(t[e])}`).join(",")}}`:JSON.stringify(t)}function formatToolArgumentError(t){return JSON.stringify({error:"tool_argument_validation_failed",toolId:t.toolId,issues:t.issues,retry:"Call the same tool again with arguments that satisfy the reported schema and semantic issues."})}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}function formatError(t){return t instanceof Error?t.message:String(t)}
+import{ToolMessage as e}from"@langchain/core/messages";import{tool as t}from"@langchain/core/tools";import{isSuccessfulEvidenceOutput as o,observedToolEvidence as r,recordObservedToolEvidence as n}from"./gateway/tool-evidence.js";export function buildGatewayTools(o,n,s,a,i=createToolRepeatState(o.workspace.runtime.toolGateway)){return o.toolGateway?s.flatMap(s=>{const l=o.toolGateway?.get(s);if(!l)return[];const u=o.workspace.tools.get(s),c=u?.schema??l.schema;return[t(async t=>async function invokeGuardedGatewayTool(t){emitToolResult(t.input,t.agentId,t.toolId,void 0);const o=function missingRequiredPlanContent(e,t){const o=readRecord(e.agent.config.executionContract);if(!0!==o?.requiresPlan)return"";const n=readStringArray(o.planEvidenceTools);if(0===n.length||n.includes(t))return"";const s=new Set(r(e).map(e=>e.toolId));return n.some(e=>s.has(e))?"":["Status: plan_required",`Evidence tool: ${t}`,`Blocker: execution contract requires a planning checkpoint from one of: ${n.join(", ")} before evidence tools run.`,"Instruction: call the planning tool first, then retry this atomic evidence tool with repaired arguments."].join("\n")}(t.input,t.toolId);if(o)return emitToolResult(t.input,t.agentId,t.toolId,o),o;const n=function missingToolDependencyContent(e,t){const o=readRecord(e.agent.config.executionContract),n=readStringArray(readRecord(o.toolDependencies)[t]);if(0===n.length)return"";const s=new Set(r(e).map(e=>e.toolId)),a=n.filter(e=>!s.has(e));return 0===a.length?"":["Status: dependency_required",`Evidence tool: ${t}`,`Blocker: this atomic evidence tool requires completed dependency evidence from: ${a.join(", ")}.`,"Instruction: complete the dependency tool first, evaluate it, then retry this atomic evidence tool."].join("\n")}(t.input,t.toolId);if(n)return emitToolResult(t.input,t.agentId,t.toolId,n),n;const s=t.repeatState?beforeToolInvoke(t.toolId,t.args,t.repeatState):void 0;if(s)return emitToolResult(t.input,t.agentId,t.toolId,s.eventOutput),s.modelOutput;const a=await async function invokeGatewayTool(t,o,r,n,s){try{return await t.toolGateway.invoke({toolId:r,args:n,repairModel:s,context:{workspaceRoot:t.workspace.root,requestId:t.requestId,sessionId:t.sessionId,agentId:o,requestInput:t.request.input,observedEvidence:formatObservedEvidenceForToolContext(t)}})}catch(o){if(function isToolArgumentValidationError(e){return e instanceof Error&&"ToolArgumentValidationError"===e.name&&"string"==typeof e.toolId&&Array.isArray(e.issues)}(o))return new e({tool_call_id:`stable-harness-${r}-argument-guard`,name:r,status:"error",content:formatToolArgumentError(o)});if(t.workspace.runtime.retry?.tools?.enabled)throw o;return new e({tool_call_id:`stable-harness-${r}-execution-error`,name:r,status:"error",content:JSON.stringify({error:"tool_execution_failed",toolId:r,message:formatError(o),retry:"Use the error as evidence, adjust the tool arguments if possible, or return a final answer with the blocker."})})}}(t.input,t.agentId,t.toolId,t.args,t.repairModel),i=a instanceof e?String(a.content):stringifyDeepAgentResult(a.output),l=t.repeatState?afterToolInvoke(t.toolId,t.args,i,a,t.repeatState):{};return emitToolResult(t.input,t.agentId,t.toolId,l.eventOutput??i),void 0!==l.modelOutput?l.modelOutput:a instanceof e?a:i}({input:o,agentId:n,toolId:s,args:t,repairModel:a,repeatState:i}),{name:s,description:buildToolDescription(u?.description??l.description??s,c,o.workspace.runtime.toolGateway,s),schema:{type:"object",additionalProperties:!0}})]}):[]}export function createToolRepeatState(e){if(function repeatGuardEnabled(e){return!0===repeatGuardConfig(e).enabled}(e))return{successfulCalls:new Map,duplicateCallCounts:new Map,latestSuccessfulOutputByTool:new Map,successfulToolCounts:new Map,toolCallCounts:new Map,maxDuplicateCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxDuplicateCallsPerTool)??3,maxCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxCallsPerTool),maxSuccessfulCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxSuccessfulCallsPerTool),maxCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxCallsByTool),maxSuccessfulCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxSuccessfulCallsByTool)}}function repeatGuardConfig(e){return isRecord(e)&&isRecord(e.repeatGuard)?e.repeatGuard:{}}function readPositiveInteger(e){return"number"==typeof e&&Number.isInteger(e)&&e>0?e:void 0}function readPositiveIntegerMap(e){return isRecord(e)?new Map(Object.entries(e).map(([e,t])=>[e,readPositiveInteger(t)]).filter(e=>void 0!==e[1])):new Map}export function beforeToolInvoke(e,t,o){const r=o.toolCallCounts.get(e)??0;o.toolCallCounts.set(e,r+1);const n=o.maxCallsByTool.get(e)??o.maxCallsPerTool;if(void 0!==n&&r>=n){const t=repeatedToolCallLimitContent(e,o.latestSuccessfulOutputByTool.get(e));return{eventOutput:t,modelOutput:t}}const s=o.maxSuccessfulCallsByTool.get(e)??o.maxSuccessfulCallsPerTool;if(void 0!==s&&(o.successfulToolCounts.get(e)??0)>=s){const t=repeatedToolCallLimitContent(e,o.latestSuccessfulOutputByTool.get(e));return{eventOutput:t,modelOutput:t}}const a=stableToolCallKey(e,t),i=o.successfulCalls.get(a);if(void 0!==i){const t=o.duplicateCallCounts.get(a)??0;if(o.duplicateCallCounts.set(a,t+1),void 0!==o.maxDuplicateCallsPerTool&&t>=o.maxDuplicateCallsPerTool){const t=repeatedToolCallLimitContent(e);return{eventOutput:t,modelOutput:t}}const r=function duplicateToolCallContent(e,t){return JSON.stringify({status:"duplicate_tool_call",toolId:e,instruction:"This agent already completed an equivalent tool call. Use the prior evidence instead of calling the tool again.",previousOutput:t})}(e,i);return{eventOutput:r,modelOutput:i}}}export function isToolRepeatLimitReached(e,t){if(!t)return!1;const o=t.maxCallsByTool.get(e)??t.maxCallsPerTool;if(void 0!==o&&(t.toolCallCounts.get(e)??0)>=o)return!0;const r=t.maxSuccessfulCallsByTool.get(e)??t.maxSuccessfulCallsPerTool;return void 0!==r&&(t.successfulToolCounts.get(e)??0)>=r}export function afterToolInvoke(t,r,n,s,a){return s instanceof e&&"error"===s.status?{}:o(n)?(a.successfulCalls.set(stableToolCallKey(t,r),n),a.latestSuccessfulOutputByTool.set(t,n),a.successfulToolCounts.set(t,(a.successfulToolCounts.get(t)??0)+1),{}):{}}function emitToolResult(e,t,o,r){void 0!==r&&n(e,t,o,r),e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:t,event:void 0===r?{adapter:"deepagents",phase:"agent.tool.start",toolId:o}:{adapter:"deepagents",phase:"agent.tool.result",toolId:o,output:previewToolOutput(r),...toolControlProjection(r)}})}function repeatedToolCallLimitContent(e,t){return JSON.stringify({status:"repeated_tool_call_limit",toolId:e,instruction:"This tool reached the configured repeat limit for this request. Do not call this tool or a substitute tool for the same evidence need again. Use previousOutput and the collected evidence to produce the final answer now, or report the remaining gap explicitly.",...void 0!==t?{previousOutput:t}:{}})}export function stringifyDeepAgentResult(t){if(t instanceof e)return function stringifyToolMessageContent(e){return"string"==typeof e?e:JSON.stringify(e)}(t.content);if("string"==typeof t)return t;if(isRecord(t)){const e=t.structuredResponse??t.structured_response;if(void 0!==e)return"string"==typeof e?e:JSON.stringify(e);const o=(Array.isArray(t.messages)?t.messages:[]).at(-1);if(isRecord(o)&&"string"==typeof o.content)return o.content;const r=(isRecord(t.update)&&Array.isArray(t.update.messages)?t.update.messages:[]).at(-1);if(isRecord(r)&&isRecord(r.kwargs)&&"string"==typeof r.kwargs.content)return r.kwargs.content;if(isRecord(r)&&"string"==typeof r.content)return r.content}return JSON.stringify(t)}function buildToolDescription(e,t,o,r){const n=function toolRepeatPolicyDescription(e,t){const o=repeatGuardConfig(e),r=readPositiveIntegerMap(o.maxSuccessfulCallsByTool).get(t)??readPositiveInteger(o.maxSuccessfulCallsPerTool);return void 0===r?"":`Stable runtime repeat policy: call this tool at most ${r} successful time(s) for this request. If more detail is needed, include the dimensions in the first call and synthesize after the result returns.`}(o,r),s=n?`${e}\n\n${n}`:e;return t?`${s}\n\nStable tool input schema:\n${previewToolOutput(JSON.stringify(t))}`:s}function previewToolOutput(e){const t=e.replace(/\s+/gu," ").trim();return t.length>500?`${t.slice(0,497)}...`:t}export function toolControlProjection(e){const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);if("string"==typeof t?.status)return{controlStatus:t.status};const o=function readTextStatus(e){return String(e).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(e);return o?{controlStatus:o}:"string"==typeof t?.error?{controlStatus:t.error}:e.startsWith("Task delegation target is not in the workspace inventory")?{controlStatus:"task_inventory_blocked"}:{}}function stableToolCallKey(e,t){return`${e}:${stableJson(t)}`}function stableJson(e){return Array.isArray(e)?`[${e.map(stableJson).join(",")}]`:isRecord(e)?`{${Object.keys(e).sort().map(t=>`${JSON.stringify(t)}:${stableJson(e[t])}`).join(",")}}`:JSON.stringify(e)}function formatObservedEvidenceForToolContext(e){const t=r(e).map(e=>`Tool: ${e.toolId}\n${e.output}`).join("\n\n---\n\n");return t.length>12e3?`${t.slice(0,12e3)}\n[truncated]`:t}function formatToolArgumentError(e){return JSON.stringify({error:"tool_argument_validation_failed",toolId:e.toolId,issues:e.issues,retry:"Call the same tool again with arguments that satisfy the reported schema and semantic issues."})}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function readRecord(e){return isRecord(e)?e:{}}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}function formatError(e){return e instanceof Error?e.message:String(e)}

package/packages/adapter-deepagents/package.json

@@ -10,7 +10,7 @@
   "main": "dist/src/index.js",
   "types": "dist/src/index.d.ts",
   "dependencies": {
-    "@botbotgo/better-call": "^0.1.54",
+    "@botbotgo/better-call": "^0.1.58",
     "@langchain/core": "^1.1.43",
     "@langchain/ollama": "^1.2.7",
     "@langchain/openai": "^1.4.5",

package/packages/adapter-langgraph/dist/src/skill-providers.js

@@ -1 +1 @@
-import{readFile as e}from"node:fs/promises";import r from"node:path";export function createRegistrySkillResolverProvider(e={}){return{name:"registry-resolver",resolve:r=>async function resolveRegistrySkill(e,r){const t=e.workspace.skills.get(e.id);if(!t)throw new Error(`LangGraph skill resolver cannot find skill ${e.id}`);return{id:t.id,path:t.path,...t.description?{description:t.description}:{},allowedTools:t.allowedTools,...!1===r.includeContent?{}:{content:await readSkillContent(t,r.maxBytes)}}}(r,e)}}export function createDeepAgentsMiddlewareSkillProvider(e={}){return{name:"deepagents-middleware",async createMiddleware(r){const t=await async function loadDeepAgentsSkillsModule(e){return e.importDeepAgents?e.importDeepAgents():import("deepagents")}(e),o=t.createSkillsMiddleware;if("function"!=typeof o)throw new Error("deepagents does not export createSkillsMiddleware");return o({backend:e.backend??createDeepAgentsFilesystemBackend(t,r.workspace.root),sources:e.sources??deriveDeepAgentsSkillSources(r)})}}}export function resolveSkillProvider(e){if(!1!==e.skillProvider)return e.skillProvider??createRegistrySkillResolverProvider()}async function readSkillContent(r,t=1048576){const o=await e(r.path,"utf8");if(Buffer.byteLength(o,"utf8")>t)throw new Error(`Skill ${r.id} exceeds registry resolver size limit of ${t} bytes`);return o}function createDeepAgentsFilesystemBackend(e,r){if("function"!=typeof e.FilesystemBackend)throw new Error("deepagents does not export FilesystemBackend");return new e.FilesystemBackend({rootDir:r})}function deriveDeepAgentsSkillSources(e){const t=new Set([...(o=e.agent,o?.skills??[]),...skillIdsFromWorkflow(e)]);var o;const n=new Set;for(const o of t){const t=e.workspace.skills.get(o);t&&n.add(toPosixPath(r.relative(e.workspace.root,r.dirname(r.dirname(t.path)))))}return[...n].filter(e=>e&&!e.startsWith(".."))}function skillIdsFromWorkflow(e){return e.workflow.nodes.map(e=>e.use.match(/^skills\.([^./][^.]*)$/u)?.[1]).filter(e=>Boolean(e))}function toPosixPath(e){return e.split(r.sep).join("/")}
+import{readFile as e}from"node:fs/promises";import r from"node:path";import{repairRuntimeSelection as t}from"@stable-harness/core";export function createRegistrySkillResolverProvider(e={}){return{name:"registry-resolver",resolve:r=>async function resolveRegistrySkill(e,r){const o=await async function resolveWorkspaceSkill(e){const r=e.workspace.skills.get(e.id);if(r)return r;const o=await t({id:e.id,candidates:[...e.workspace.skills.values()].map(e=>({id:e.id,description:e.description}))});return o.ok?e.workspace.skills.get(o.id):void 0}(e);if(!o)throw new Error(`LangGraph skill resolver cannot find skill ${e.id}`);return{id:o.id,path:o.path,...o.description?{description:o.description}:{},allowedTools:o.allowedTools,...!1===r.includeContent?{}:{content:await readSkillContent(o,r.maxBytes)}}}(r,e)}}export function createDeepAgentsMiddlewareSkillProvider(e={}){return{name:"deepagents-middleware",async createMiddleware(r){const t=await async function loadDeepAgentsSkillsModule(e){return e.importDeepAgents?e.importDeepAgents():import("deepagents")}(e),o=t.createSkillsMiddleware;if("function"!=typeof o)throw new Error("deepagents does not export createSkillsMiddleware");return o({backend:e.backend??createDeepAgentsFilesystemBackend(t,r.workspace.root),sources:e.sources??deriveDeepAgentsSkillSources(r)})}}}export function resolveSkillProvider(e){if(!1!==e.skillProvider)return e.skillProvider??createRegistrySkillResolverProvider()}async function readSkillContent(r,t=1048576){const o=await e(r.path,"utf8");if(Buffer.byteLength(o,"utf8")>t)throw new Error(`Skill ${r.id} exceeds registry resolver size limit of ${t} bytes`);return o}function createDeepAgentsFilesystemBackend(e,r){if("function"!=typeof e.FilesystemBackend)throw new Error("deepagents does not export FilesystemBackend");return new e.FilesystemBackend({rootDir:r})}function deriveDeepAgentsSkillSources(e){const t=new Set([...(o=e.agent,o?.skills??[]),...skillIdsFromWorkflow(e)]);var o;const i=new Set;for(const o of t){const t=e.workspace.skills.get(o);t&&i.add(toPosixPath(r.relative(e.workspace.root,r.dirname(r.dirname(t.path)))))}return[...i].filter(e=>e&&!e.startsWith(".."))}function skillIdsFromWorkflow(e){return e.workflow.nodes.map(e=>e.use.match(/^skills\.([^./][^.]*)$/u)?.[1]).filter(e=>Boolean(e))}function toPosixPath(e){return e.split(r.sep).join("/")}

package/packages/cli/dist/src/cli.js

@@ -1,2 +1,2 @@
 #!/usr/bin/env node
-import{realpathSync as e}from"node:fs";import{fileURLToPath as t}from"node:url";import{createDeepAgentsAdapter as r}from"@stable-harness/adapter-deepagents";import{createLangGraphWorkflowAdapter as o}from"@stable-harness/adapter-langgraph";import{createStableHarnessRuntime as s}from"@stable-harness/core";import{projectEvent as a,projectRuntimeTrace as i}from"@stable-harness/core";import{createModuleToolGateway as n}from"@stable-harness/tool-gateway";import{loadWorkspaceFromYaml as u}from"@stable-harness/workspace-yaml";import{helpText as l,parseArgs as d}from"./args.js";import{formatCliRuntimeEvent as p,readCliEventViewConfig as c,shouldEnableCliProgressNarration as m}from"./event-view.js";import{initWorkspace as w}from"./init.js";import{ensureCliMemoryServices as f}from"./memory/lifecycle.js";import{createCliMemoryProviders as g}from"./memory/providers.js";import{formatDetail as v,inspectWorkflow as I,renderAgent as k,renderWorkflow as y,workspaceStatus as h}from"./output.js";import{serveProtocol as b,stopProtocol as q}from"./server.js";export async function runCli(e=process.argv.slice(2)){const t=d(e);if(t.help)return void process.stdout.write(l());const o=setTimeout(()=>{process.stderr.write(`stable-harness request timed out after ${t.timeoutMs}ms\n`),process.exit(124)},t.timeoutMs),R=t.workspaceRoot;try{if("init"===t.command)return void process.stdout.write(await w(t.prompt||R));const e=await u(R);if(t.workflowRenderId)return void process.stdout.write(y(e,t.workflowRenderId));if(t.workflowInspectId)return void process.stdout.write(I(e,t.workflowInspectId));if(t.agentRenderId)return void process.stdout.write(k(e,t.agentRenderId));if("stop"===t.command)return clearTimeout(o),void await q(e,t);const l=await n({tools:e.tools.values(),options:{betterCall:{mode:"repair"}}});await f(e);const d=g(e),C=c(e.runtime);let $;if($=s({workspace:e,toolGateway:l,memoryProviders:d,adapters:[r()],workflowAdapters:[createCliWorkflowAdapter(l,()=>$)],progressNarration:m(C,e.runtime)?{enabled:!0,style:"cli"}:void 0}),t.serveOpenAi)return clearTimeout(o),void await b($,t);if(!t.shouldRunRequest)return void process.stdout.write(h(e,R));t.trace&&$.subscribe(e=>{const t=a(e);t&&process.stdout.write(`trace:${t.agentId}:${t.label}${v(t.detail)}\n`)}),$.subscribe(e=>{const t=p(e,C);t&&process.stdout.write(`${t}\n`)});const j=await $.request({input:t.prompt,agentId:t.agentId,sessionId:t.sessionId,toolCall:t.toolId?{toolId:t.toolId,args:t.toolArgs}:void 0,workflow:t.workflowRunId?{workflowId:t.workflowRunId,input:t.prompt}:void 0});if(t.trace||t.traceJson){const e=$.getRun(j.requestId),r=e?i(e):[];t.traceJson&&process.stdout.write(`${JSON.stringify({trace:r})}\n`)}process.stdout.write(`${j.output}\n`)}finally{clearTimeout(o)}}function createCliWorkflowAdapter(e,t){return o({nodeResolvers:{tools:async({id:t,node:r,request:o,requestId:s,sessionId:a,state:i,workspace:n})=>{return(await e.invoke({toolId:t,args:(u=r.config,l=o.input,d=i.outputs,!0===u?.inputFromState?{...u,requestInput:l,outputs:d}:u&&"requiredInput"in u?u.requiredInput:u&&("args"in u||"cwd"in u||"timeoutMs"in u)?u:"object"==typeof l&&null!==l?l:{}),context:{workspaceRoot:n.root,requestId:s,sessionId:a,agentId:`workflow:${r.id}`}})).output;var u,l,d},agents:async({id:e,node:r,request:o,sessionId:s,state:a})=>{var i,n,u,l;return(await t().request({input:(i=e,n=o.input,u=a.outputs,l=r.config,[`Workflow node agents.${i}: synthesize the workflow evidence into the requested final output.`,`Original request: ${"string"==typeof n?n:JSON.stringify(n)}`,"Requirements:","- Produce the final answer now; do not ask follow-up questions.","- Match the original request language unless workflow config explicitly says otherwise.","- Use only the workflow outputs as evidence; call out uncertainty directly.",...l?[`Workflow node config: ${JSON.stringify(l)}`]:[],"Prior workflow outputs:",JSON.stringify(u)].join("\n")),agentId:e,sessionId:s,metadata:o.metadata})).output}}})}(function isCliEntrypoint(){const r=process.argv[1];if(!r)return!1;try{return e(t(import.meta.url))===e(r)}catch{return!1}})()&&runCli().catch(e=>{process.stderr.write(`${e instanceof Error?e.message:String(e)}\n`),process.exitCode=1});
+import{realpathSync as e}from"node:fs";import{fileURLToPath as t}from"node:url";import{createBackendModel as o,createDeepAgentsAdapter as r}from"@stable-harness/adapter-deepagents";import{createLangGraphWorkflowAdapter as s}from"@stable-harness/adapter-langgraph";import{createStableHarnessRuntime as i}from"@stable-harness/core";import{projectEvent as a,projectRuntimeTrace as n}from"@stable-harness/core";import{createModuleToolGateway as u}from"@stable-harness/tool-gateway";import{loadWorkspaceFromYaml as l}from"@stable-harness/workspace-yaml";import{helpText as d,parseArgs as c}from"./args.js";import{formatCliRuntimeEvent as p,readCliEventViewConfig as f,shouldEnableCliProgressNarration as m}from"./event-view.js";import{initWorkspace as w}from"./init.js";import{ensureCliMemoryServices as g}from"./memory/lifecycle.js";import{createCliMemoryProviders as v}from"./memory/providers.js";import{formatDetail as y,inspectWorkflow as k,renderAgent as I,renderWorkflow as R,workspaceStatus as h}from"./output.js";import{serveProtocol as q,stopProtocol as b}from"./server.js";export async function runCli(e=process.argv.slice(2)){const t=c(e);if(t.help)return void process.stdout.write(d());const o=setTimeout(()=>{process.stderr.write(`stable-harness request timed out after ${t.timeoutMs}ms\n`),process.exit(124)},t.timeoutMs),s=t.workspaceRoot;try{if("init"===t.command)return void process.stdout.write(await w(t.prompt||s));const e=await l(s);if(t.workflowRenderId)return void process.stdout.write(R(e,t.workflowRenderId));if(t.workflowInspectId)return void process.stdout.write(k(e,t.workflowInspectId));if(t.agentRenderId)return void process.stdout.write(I(e,t.agentRenderId));if("stop"===t.command)return clearTimeout(o),void await b(e,t);const d=await u({tools:e.tools.values(),options:{betterCall:{mode:"repair"}}});await g(e);const c=v(e),C=f(e.runtime);let M;if(M=i({workspace:e,toolGateway:d,memoryProviders:c,adapters:[r()],workflowAdapters:[createCliWorkflowAdapter(d,()=>M)],progressNarration:m(C,e.runtime)?{enabled:!0,style:"cli"}:void 0,qualityReviewModel:createQualityReviewModel(e)}),t.serveOpenAi)return clearTimeout(o),void await q(M,t);if(!t.shouldRunRequest)return void process.stdout.write(h(e,s));t.trace&&M.subscribe(e=>{const t=a(e);t&&process.stdout.write(`trace:${t.agentId}:${t.label}${y(t.detail)}\n`)}),M.subscribe(e=>{const t=p(e,C);t&&process.stdout.write(`${t}\n`)});const j=await M.request({input:t.prompt,agentId:t.agentId,sessionId:t.sessionId,toolCall:t.toolId?{toolId:t.toolId,args:t.toolArgs}:void 0,workflow:t.workflowRunId?{workflowId:t.workflowRunId,input:t.prompt}:void 0});if(t.trace||t.traceJson){const e=M.getRun(j.requestId),o=e?n(e):[];t.traceJson&&process.stdout.write(`${JSON.stringify({trace:o})}\n`)}process.stdout.write(`${j.output}\n`)}finally{clearTimeout(o)}}function createQualityReviewModel(e){const t=function readQualityModelRef(e){const t=isRecord(e)?e:{},o=isRecord(t.reviewer)?t.reviewer:t;return"string"==typeof o.modelRef&&o.modelRef.trim()?o.modelRef.trim():void 0}(e.runtime.quality),r=t?e.models.get(t):void 0,s=r?o(r):void 0;return function isQualityReviewModel(e){return isRecord(e)&&"function"==typeof e.invoke}(s)?s:void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function createCliWorkflowAdapter(e,t){return s({nodeResolvers:{tools:async({id:t,node:o,request:r,requestId:s,sessionId:i,state:a,workspace:n})=>{return(await e.invoke({toolId:t,args:(u=o.config,l=r.input,d=a.outputs,!0===u?.inputFromState?{...u,requestInput:l,outputs:d}:u&&"requiredInput"in u?u.requiredInput:u&&("args"in u||"cwd"in u||"timeoutMs"in u)?u:"object"==typeof l&&null!==l?l:{}),context:{workspaceRoot:n.root,requestId:s,sessionId:i,agentId:`workflow:${o.id}`}})).output;var u,l,d},agents:async({id:e,node:o,request:r,sessionId:s,state:i})=>{var a,n,u,l;return(await t().request({input:(a=e,n=r.input,u=i.outputs,l=o.config,[`Workflow node agents.${a}: synthesize the workflow evidence into the requested final output.`,`Original request: ${"string"==typeof n?n:JSON.stringify(n)}`,"Requirements:","- Produce the final answer now; do not ask follow-up questions.","- Match the original request language unless workflow config explicitly says otherwise.","- Use only the workflow outputs as evidence; call out uncertainty directly.",...l?[`Workflow node config: ${JSON.stringify(l)}`]:[],"Prior workflow outputs:",JSON.stringify(u)].join("\n")),agentId:e,sessionId:s,metadata:r.metadata})).output}}})}(function isCliEntrypoint(){const o=process.argv[1];if(!o)return!1;try{return e(t(import.meta.url))===e(o)}catch{return!1}})()&&runCli().catch(e=>{process.stderr.write(`${e instanceof Error?e.message:String(e)}\n`),process.exitCode=1});

package/packages/core/dist/execution-contract.js

@@ -1 +1 @@
-export function assertExecutionContract(e){(function contractDisabled(e){const t=isRecord(e?.executionContract)?e.executionContract:void 0;return!1===t?.enabled})(e.metadata)||(function assertRequiredPlan(e){const t=readExecutionContract(e.agent);if(!t.requiresPlan)return;const n=readStringArray(t.planEvidenceTools);if(0===n.length)throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"invalid_plan_contract"}),new Error("Execution contract requires plan evidence tools when requiresPlan is enabled");const r=e.store.getRun(e.requestId),o=new Set((r?.events??[]).flatMap(readEvidenceToolId));if(!n.some(e=>o.has(e)))throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"missing_plan",missingEvidenceTools:n}),new Error(`Execution contract requires a planning checkpoint from one of: ${n.join(", ")}`)}(e),function assertRequiredEvidenceTools(e){const t=function readRequiredEvidenceTools(e){return readStringArray(readExecutionContract(e).requiredEvidenceTools)}(e.agent);if(0===t.length)return;const n=e.store.getRun(e.requestId),r=new Set((n?.events??[]).flatMap(readEvidenceToolId)),o=t.filter(e=>!r.has(e));if(0!==o.length)throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"missing_required_evidence",missingEvidenceTools:o}),new Error(`Execution contract missing required evidence tools: ${o.join(", ")}`)}(e))}function readExecutionContract(e){return isRecord(e.config.executionContract)?e.config.executionContract:{}}function readEvidenceToolId(e){return"runtime.tool.direct.completed"===e.type?[e.toolId]:"runtime.adapter.event"===e.type&&isRecord(e.event)&&"agent.tool.result"===e.event.phase&&"string"==typeof e.event.toolId?[e.event.toolId]:[]}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}
+export function assertExecutionContract(e){(function contractDisabled(e){const t=isRecord(e?.executionContract)?e.executionContract:void 0;return!1===t?.enabled})(e.metadata)||(function assertRequiredPlan(e){const t=readExecutionContract(e.agent);if(!t.requiresPlan)return;const n=readStringArray(t.planEvidenceTools);if(0===n.length)throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"invalid_plan_contract"}),new Error("Execution contract requires plan evidence tools when requiresPlan is enabled");const r=e.store.getRun(e.requestId),o=new Set((r?.events??[]).flatMap(readEvidenceToolId));if(!n.some(e=>o.has(e)))throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"missing_plan",missingEvidenceTools:n}),new Error(`Execution contract requires a planning checkpoint from one of: ${n.join(", ")}`)}(e),function assertRequiredEvidenceTools(e){const t=function readRequiredEvidenceTools(e){return readStringArray(readExecutionContract(e).requiredEvidenceTools)}(e.agent);if(0===t.length)return;const n=e.store.getRun(e.requestId),r=new Set((n?.events??[]).flatMap(readEvidenceToolId)),o=t.filter(e=>!r.has(e));if(0!==o.length)throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"missing_required_evidence",missingEvidenceTools:o}),new Error(`Execution contract missing required evidence tools: ${o.join(", ")}`)}(e),function assertToolDependencies(e){const t=readExecutionContract(e.agent),n=isRecord(t.toolDependencies)?t.toolDependencies:{},r=Object.entries(n).map(([e,t])=>[e,readStringArray(t)]).filter(e=>e[1].length>0);if(0===r.length)return;const o=e.store.getRun(e.requestId),i=new Set((o?.events??[]).flatMap(readEvidenceToolId)),s=r.filter(([e])=>i.has(e)).flatMap(([e,t])=>t.filter(e=>!i.has(e)).map(t=>`${e} requires ${t}`));if(0!==s.length)throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"missing_tool_dependency",missingEvidenceTools:s}),new Error(`Execution contract missing required tool dependencies: ${s.join(", ")}`)}(e))}function readExecutionContract(e){return isRecord(e.config.executionContract)?e.config.executionContract:{}}function readEvidenceToolId(e){return"runtime.tool.direct.completed"===e.type?[e.toolId]:"runtime.adapter.event"===e.type&&isRecord(e.event)?"agent.tool.result"!==e.event.phase||"string"!=typeof e.event.toolId?[]:function isSuccessfulEvidenceEvent(e){const t=function readString(e){return"string"==typeof e&&e.length>0?e:void 0}(e.controlStatus)??function readOutputStatus(e){if("string"!=typeof e)return;const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);return"string"==typeof t?.status?t.status:e.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(e.output);return!t||/^(?:completed|success|ok|recorded)$/iu.test(t)}(e.event)?[e.event.toolId]:[]:[]}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}

package/packages/core/dist/index.d.ts

@@ -6,9 +6,11 @@ export * from "./runtime/persistence/inspection.js";
 export * from "./memory-plugins.js";
 export * from "./runtime/persistence/queue.js";
 export * from "./runtime.js";
+export * from "./runtime/selection-repair.js";
 export * from "./runtime/persistence/stores.js";
 export * from "./trace.js";
 export * from "./types.js";
 export * from "./evaluations/index.js";
+export * from "./quality/index.js";
 export * from "./spec-driven/index.js";
 export * from "./workflows/index.js";

package/packages/core/dist/index.js

@@ -1 +1 @@
-export*from"./runtime/persistence/artifacts.js";export*from"./boundary-scan.js";export*from"./execution-contract.js";export*from"./recovery/tool-call.js";export*from"./runtime/persistence/inspection.js";export*from"./memory-plugins.js";export*from"./runtime/persistence/queue.js";export*from"./runtime.js";export*from"./runtime/persistence/stores.js";export*from"./trace.js";export*from"./types.js";export*from"./evaluations/index.js";export*from"./spec-driven/index.js";export*from"./workflows/index.js";
+export*from"./runtime/persistence/artifacts.js";export*from"./boundary-scan.js";export*from"./execution-contract.js";export*from"./recovery/tool-call.js";export*from"./runtime/persistence/inspection.js";export*from"./memory-plugins.js";export*from"./runtime/persistence/queue.js";export*from"./runtime.js";export*from"./runtime/selection-repair.js";export*from"./runtime/persistence/stores.js";export*from"./trace.js";export*from"./types.js";export*from"./evaluations/index.js";export*from"./quality/index.js";export*from"./spec-driven/index.js";export*from"./workflows/index.js";

package/packages/core/dist/quality/event-evidence.d.ts

@@ -0,0 +1,5 @@
+import type { RuntimeEvent } from "../types.js";
+export declare function hasPlanningEvidence(events: RuntimeEvent[]): boolean;
+export declare function successfulEvidenceToolIds(events: RuntimeEvent[]): string[];
+export declare function successfulEvidenceOutputs(events: RuntimeEvent[]): string[];
+export declare function controlBlockers(events: RuntimeEvent[]): string[];

package/packages/core/dist/quality/event-evidence.js

@@ -0,0 +1 @@
+export function hasPlanningEvidence(t){return t.some(t=>{const e=readAdapterEvent(t);return!!e&&("plan"===e.traceType||String(e.traceLabel??"").startsWith("plan.")||"write_todos"===e.toolId&&String(e.phase??"").startsWith("agent.tool."))})}export function successfulEvidenceToolIds(t){const e=t.flatMap(t=>{if("runtime.tool.direct.completed"===t.type)return[t.toolId];const e=readAdapterEvent(t);return e&&"agent.tool.result"===e.phase&&"string"==typeof e.toolId&&isSuccessfulEvidenceEvent(e)?[e.toolId]:[]});return[...new Set(e)]}export function successfulEvidenceOutputs(t){return t.flatMap(t=>{if("runtime.tool.direct.completed"===t.type)return stringifyEvidence(t.output);const e=readAdapterEvent(t);return e&&"agent.tool.result"===e.phase&&"string"==typeof e.toolId?!isSuccessfulEvidenceEvent(e)||function isPlanningTool(t){return/^(?:write_todos|read_todos)$/u.test(t)}(e.toolId)?[]:stringifyEvidence(e.output):[]})}export function controlBlockers(t){return t.flatMap(t=>{const e=readAdapterEvent(t),r=readString(e?.controlStatus)??readOutputStatus(e?.output);return r&&function isBlockerStatus(t){return/^(?:blocked|approval_required|schema_repair_failed|tool_argument_error|invalid_input)$/iu.test(t)}(r)?[`${readString(e?.toolId)??"tool"}:${r}`]:[]})}function stringifyEvidence(t){return"string"==typeof t?t.trim()?[t]:[]:null==t?[]:[JSON.stringify(t)]}function readAdapterEvent(t){return"runtime.adapter.event"===t.type&&isRecord(t.event)?t.event:void 0}function isSuccessfulEvidenceEvent(t){const e=readString(t.controlStatus)??readOutputStatus(t.output);return!e||/^(?:completed|success|ok|recorded)$/iu.test(e)}function readOutputStatus(t){if("string"!=typeof t)return;const e=function parseJsonRecord(t){try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t);return"string"==typeof e?.status?e.status:t.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}function readString(t){return"string"==typeof t&&t.trim()?t:void 0}

package/packages/core/dist/quality/execution-review.d.ts

@@ -0,0 +1,2 @@
+import type { QualityPolicy, QualityReviewInput, QualityReviewResult } from "./types.js";
+export declare function reviewExecutionEvidence(input: QualityReviewInput, policy: QualityPolicy): QualityReviewResult;

package/packages/core/dist/quality/execution-review.js

@@ -0,0 +1 @@
+import{controlBlockers as e,successfulEvidenceOutputs as n,successfulEvidenceToolIds as t}from"./event-evidence.js";const r=/(?<![\w.])(?:\d{1,3}(?:,\d{3})+|\d+)(?:\.\d+)?[%kKmMbBtTxX]?(?!\w)/gu;export function reviewExecutionEvidence(e,n){if(!n.enabled||!n.executionReview.enabled)return{verdict:"pass",issues:[]};const t=[...blockerIssues(e,n),...emptyFinalIssues(e,n),...toolEvidenceIssues(e,n),...ungroundedNumberIssues(e,n)];return 0===t.length?{verdict:"pass",issues:[]}:{verdict:t.some(e=>!e.recoverable)?"blocked":"continue_react",issues:t}}function ungroundedNumberIssues(e,t){if(!t.executionReview.rejectUngroundedNumbers||!e.output?.text.trim())return[];const r=numberSet(n(e.events).join("\n"));if(0===r.size)return[];const s=[...numberSet(e.output.text)].filter(e=>!function isSupportedNumber(e,n){if(n.has(e))return!0;const t=Number.parseFloat(e);if(!Number.isFinite(t))return!1;for(const e of n){const n=Number.parseFloat(e);if(Number.isFinite(n)&&Math.abs(n-t)<=roundingTolerance(t))return!0}return!1}(e,r));return 0===s.length?[]:[{code:"ungrounded_numeric_claim",message:`Final answer contains numeric claims not found in successful tool evidence: ${s.slice(0,12).join(", ")}`,recoverable:!1}]}function numberSet(e){const n=new Set;for(const t of e.matchAll(r)){const e=normalizeNumber(t[0]);e&&n.add(e)}return n}function normalizeNumber(e){const n=e.replace(/,/gu,"").replace(/^\+/u,"").replace(/[%kKmMbBtTxX]$/u,"").trim();if(n){if(/^\d+$/u.test(n)){const e=Number.parseInt(n,10);if(e>=1&&e<=20)return;return String(e)}return/^\d+\.\d+$/u.test(n)?n.replace(/0+$/u,"").replace(/\.$/u,""):void 0}}function roundingTolerance(e){return Math.abs(e)>=1e3?1:Math.abs(e)>=100?.1:Math.abs(e)>=10?.05:.005}function blockerIssues(n,t){return t.executionReview.stopOnBlocker?e(n.events).map(e=>({code:"control_blocker",message:`Execution produced a control blocker: ${e}`,recoverable:!1})):[]}function emptyFinalIssues(e,n){return!n.executionReview.rejectEmptyFinal||e.output?.text.trim()?[]:[{code:"empty_final_answer",message:"The final answer is empty.",recoverable:!0}]}function toolEvidenceIssues(e,n){return!n.executionReview.requireToolEvidence||t(e.events).length>0?[]:[{code:"missing_tool_evidence",message:"No successful tool or delegated-task evidence was observed.",recoverable:!0}]}

package/packages/core/dist/quality/index.d.ts

@@ -0,0 +1,7 @@
+export * from "./types.js";
+export * from "./profile.js";
+export * from "./planning-review.js";
+export * from "./execution-review.js";
+export * from "./llm-review.js";
+export * from "./recovery-policy.js";
+export * from "./runtime.js";

package/packages/core/dist/quality/index.js

@@ -0,0 +1 @@
+export*from"./types.js";export*from"./profile.js";export*from"./planning-review.js";export*from"./execution-review.js";export*from"./llm-review.js";export*from"./recovery-policy.js";export*from"./runtime.js";

package/packages/core/dist/quality/llm-review.d.ts

@@ -0,0 +1,7 @@
+import type { QualityPolicy, QualityReviewInput, QualityReviewModel, QualityReviewResult } from "./types.js";
+export declare function reviewWithLlm(input: {
+    phase: "planning" | "execution";
+    review: QualityReviewInput;
+    policy: QualityPolicy;
+    model?: QualityReviewModel;
+}): Promise<QualityReviewResult | undefined>;

package/packages/core/dist/quality/llm-review.js

@@ -0,0 +1 @@
+export async function reviewWithLlm(e){if(!e.model||"llm"!==e.policy.reviewer?.mode)return;const t="planning"===e.phase?e.policy.reviewer.planningPrompt??function defaultPlanningPrompt(){return["You are Stable Harness planning quality review.","Review only generic agent-work quality: goal coverage, executable steps, and declared inventory fit.",'Do not invent a domain workflow. Return only JSON: {"verdict":"pass|revise_plan|blocked","issues":[{"code":"...","message":"...","recoverable":true}]}'].join("\n")}():e.policy.reviewer.executionPrompt??function defaultExecutionPrompt(){return["You are Stable Harness execution evidence review.","Review only whether execution evidence supports the final answer and whether unresolved blockers or gaps remain.",'Do not answer the user request. Return only JSON: {"verdict":"pass|continue_react|blocked","issues":[{"code":"...","message":"...","recoverable":true}]}'].join("\n")}();return function parseReviewResult(e){const t=function readResponseText(e){if("string"==typeof e)return e;if(!isRecord(e))return;if("string"==typeof e.content)return e.content;const t=isRecord(e.message)?e.message:void 0;return"string"==typeof t?.content?t.content:void 0}(e),n="string"==typeof t?function parseJsonRecord(e){const t=e.trim().replace(/^```(?:json)?\s*/iu,"").replace(/\s*```$/u,"");try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t):isRecord(e)?e:void 0,r=function readVerdict(e){return"pass"===e||"revise_plan"===e||"continue_react"===e||"blocked"===e?e:void 0}(n?.verdict);if(n&&r)return{verdict:r,issues:readIssues(n.issues)}}(await e.model.invoke([{role:"system",content:t},{role:"user",content:JSON.stringify(buildReviewContext(e.review),null,2)}]))}function buildReviewContext(e){return{userRequest:e.request.input,agent:{id:e.agent.id,tools:e.agent.tools,skills:e.agent.skills??[],subagents:e.agent.subagents},finalOutput:e.output?.text,events:e.events.slice(-30).map(projectEvent)}}function projectEvent(e){return"runtime.adapter.event"===e.type?{type:e.type,event:e.event}:{...e}}function readIssues(e){return Array.isArray(e)?e.flatMap(e=>isRecord(e)&&"string"==typeof e.code&&"string"==typeof e.message?[{code:e.code,message:e.message,recoverable:!1!==e.recoverable}]:[]):[]}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

package/packages/core/dist/quality/planning-review.d.ts

@@ -0,0 +1,2 @@
+import type { QualityPolicy, QualityReviewInput, QualityReviewResult } from "./types.js";
+export declare function reviewPlanningEvidence(input: QualityReviewInput, policy: QualityPolicy): QualityReviewResult;

package/packages/core/dist/quality/planning-review.js

@@ -0,0 +1 @@
+import{hasPlanningEvidence as e}from"./event-evidence.js";export function reviewPlanningEvidence(n,s){return s.enabled&&s.planningReview.enabled&&s.planningReview.requirePlan?e(n.events)?{verdict:"pass",issues:[]}:{verdict:"revise_plan",issues:[{code:"missing_plan_evidence",message:"No upstream planning evidence was observed before the agent produced a result.",recoverable:!0}]}:{verdict:"pass",issues:[]}}

package/packages/core/dist/quality/profile.d.ts

@@ -0,0 +1,3 @@
+import type { WorkspaceAgent, WorkspaceRuntimePolicy } from "../types.js";
+import type { QualityPolicy } from "./types.js";
+export declare function resolveQualityPolicy(runtime: WorkspaceRuntimePolicy, agent: WorkspaceAgent): QualityPolicy;

package/packages/core/dist/quality/profile.js

@@ -0,0 +1 @@
+export function resolveQualityPolicy(e,n){const r=n.config.quality??e.quality,o=function readProfile(e){const n="string"==typeof e?e:isRecord(e)?e.profile:void 0;return"fast"===n||"balanced"===n||"strict"===n?n:"off"}(r),i=function profileDefaults(e,n){const r=n.tools.length>0||n.subagents.length>0||(n.skills?.length??0)>0;return"off"===e?function disabledPolicy(e){return{enabled:!1,profile:e,reviewer:{mode:"deterministic"},planningReview:{enabled:!1,requirePlan:!1},executionReview:{enabled:!1,requireToolEvidence:!1,rejectEmptyFinal:!1,stopOnBlocker:!1,rejectUngroundedNumbers:!1},recovery:{enabled:!1,maxLoops:0}}}(e):{enabled:!0,profile:e,planningReview:{enabled:"fast"!==e,requirePlan:"fast"!==e&&r},executionReview:{enabled:!0,requireToolEvidence:"strict"===e&&r,rejectEmptyFinal:!0,stopOnBlocker:!0,rejectUngroundedNumbers:"strict"===e},recovery:{enabled:"fast"!==e,maxLoops:"strict"===e?3:2}}}(o,n),t=isRecord(r)?r:{},a=function readReviewer(e){const n=isRecord(e.reviewer)?e.reviewer:e,r="llm"===n.mode||"string"==typeof n.modelRef?"llm":"deterministic",o=isRecord(n.prompts)?n.prompts:{},i=readString(n.modelRef);return{mode:r,...i?{modelRef:i}:{},...readString(n.planningPrompt)??readString(o.planning)?{planningPrompt:readString(n.planningPrompt)??readString(o.planning)}:{},...readString(n.executionPrompt)??readString(o.execution)?{executionPrompt:readString(n.executionPrompt)??readString(o.execution)}:{}}}(t);return{enabled:i.enabled,profile:o,...a?{reviewer:a}:{},planningReview:{enabled:readBoolean(t.planningReview,"enabled")??i.planningReview.enabled,requirePlan:readBoolean(t.planningReview,"requirePlan")??i.planningReview.requirePlan},executionReview:{enabled:readBoolean(t.executionReview,"enabled")??i.executionReview.enabled,requireToolEvidence:readBoolean(t.executionReview,"requireToolEvidence")??i.executionReview.requireToolEvidence,rejectEmptyFinal:readBoolean(t.executionReview,"rejectEmptyFinal")??i.executionReview.rejectEmptyFinal,stopOnBlocker:readBoolean(t.executionReview,"stopOnBlocker")??i.executionReview.stopOnBlocker,rejectUngroundedNumbers:readBoolean(t.executionReview,"rejectUngroundedNumbers")??i.executionReview.rejectUngroundedNumbers},recovery:{enabled:readBoolean(t.recovery,"enabled")??i.recovery.enabled,maxLoops:readPositiveInteger(t.recovery,"maxLoops")??i.recovery.maxLoops}}}function readString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function readBoolean(e,n){const r=isRecord(e)?e:{};return"boolean"==typeof r[n]?r[n]:void 0}function readPositiveInteger(e,n){const r=(isRecord(e)?e:{})[n];return"number"==typeof r&&Number.isInteger(r)&&r>0?r:void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

package/packages/core/dist/quality/recovery-policy.d.ts

@@ -0,0 +1,11 @@
+import type { RuntimeRequest } from "../types.js";
+import type { QualityPolicy, QualityReviewResult } from "./types.js";
+export declare function buildQualityRecoveryRequest(input: {
+    request: RuntimeRequest;
+    result: QualityReviewResult;
+    phase: "planning" | "execution";
+    policy: QualityPolicy;
+    availableToolIds: string[];
+    availableSubagentIds: string[];
+    observedEvidence?: string[];
+}): RuntimeRequest | undefined;

package/packages/core/dist/quality/recovery-policy.js

@@ -0,0 +1 @@
+export function buildQualityRecoveryRequest(e){if(!e.policy.recovery.enabled||"pass"===e.result.verdict||"blocked"===e.result.verdict)return;const n="planning"===e.phase?function planningLines(e){return["Stable runtime quality review: the previous plan was not ready to execute.",...issueLines(e.result),inventoryLine("Available configured tools",e.availableToolIds),inventoryLine("Available configured subagents",e.availableSubagentIds),"Revise the plan using only the declared inventory, then continue through the backend's normal planning and tool-calling flow.","Do not invent tools, subagents, or a separate planning language."].filter(e=>e.length>0)}(e):function executionLines(e){return["Stable runtime quality review: the execution evidence is not sufficient for a final answer.",...issueLines(e.result),inventoryLine("Available configured tools",e.availableToolIds),inventoryLine("Available configured subagents",e.availableSubagentIds),...(n=e.observedEvidence??[],0===n.length?[]:["Completed observed evidence available for synthesis:",...n.slice(-8).map((e,n)=>`Evidence ${n+1}: ${function compact(e){const n=e.replace(/\s+/gu," ").trim();return n.length>900?`${n.slice(0,897)}...`:n}(e)}`)]),"Continue from the current state using ReAct: inspect the evidence gap, then either synthesize directly from completed observed evidence or choose one remaining declared tool/subagent action.","Do not call tools that already produced completed evidence or a repeat-limit control result for the same evidence need.","Preserve blockers explicitly instead of converting control states into unsupported factual answers."].filter(e=>e.length>0);var n}(e);return{...e.request,input:[e.request.input,"",...n].join("\n"),metadata:{...e.request.metadata,stableHarnessRecovery:`quality_${e.phase}`}}}function issueLines(e){return e.issues.map(e=>`- ${e.code}: ${e.message}`)}function inventoryLine(e,n){return n.length>0?`${e}: ${n.join(", ")}`:""}

package/packages/core/dist/quality/runtime.d.ts

@@ -0,0 +1,13 @@
+import type { RuntimeMemoryContext, RuntimeOutput, RuntimeRequest } from "../types.js";
+import type { QualityPolicy, QualityReviewInput, QualityReviewModel } from "./types.js";
+export type QualityRuntimeInput = QualityReviewInput & {
+    requestId: string;
+    sessionId: string;
+    emit: (event: import("../types.js").RuntimeEvent) => void;
+    getEvents: () => import("../types.js").RuntimeEvent[];
+    runAdapter: (request: RuntimeRequest) => Promise<RuntimeOutput>;
+    reviewModel?: QualityReviewModel;
+    memory?: RuntimeMemoryContext;
+    pluginMemories: RuntimeMemoryContext[];
+};
+export declare function recoverQualityReview(input: QualityRuntimeInput, request: RuntimeRequest, result: RuntimeOutput, policy: QualityPolicy): Promise<RuntimeOutput>;

package/packages/core/dist/quality/runtime.js

@@ -0,0 +1 @@
+import{successfulEvidenceOutputs as e}from"./event-evidence.js";import{buildQualityRecoveryRequest as i}from"./recovery-policy.js";import{reviewExecutionEvidence as t}from"./execution-review.js";import{reviewWithLlm as n}from"./llm-review.js";import{reviewPlanningEvidence as r}from"./planning-review.js";export async function recoverQualityReview(e,i,t,n){if(!n.enabled)return t;let r=i,s=t;for(let i=0;i<n.recovery.maxLoops+1;i+=1){const t=await emitPlanningReview(e,r,s,n);if("blocked"===t.verdict)return qualityFailureOutput("planning",t);const u=buildQualityRecovery(e,r,t,"planning",n,i);if(u){r=u,s=await e.runAdapter(r);continue}const a=await emitExecutionReview(e,r,s,n),o=buildQualityRecovery(e,r,a,"execution",n,i);if(!o)return"pass"===a.verdict?s:qualityFailureOutput("execution",a);r=o,s=await e.runAdapter(r)}return qualityFailureOutput("execution",{verdict:"blocked",issues:[{code:"quality_recovery_exhausted",message:`Quality recovery exceeded maxLoops=${n.recovery.maxLoops}.`,recoverable:!1}]})}function emitPlanningReview(e,i,t,n){return emitReview(e,"planning",r,i,t,n)}function emitExecutionReview(e,i,n,r){return emitReview(e,"execution",t,i,n,r)}async function emitReview(e,i,t,r,s,u){const a={...reviewInputFor(e,r),output:s},o="planning"===i?u.planningReview.enabled:u.executionReview.enabled;if(!o)return t(a,u);const c=t(a,u),l=await n({phase:i,review:a,policy:u,model:e.reviewModel}),d="pass"===c.verdict?l??c:c;return o&&function emitReviewEvent(e,i,t){"planning"!==i?e.emit({type:"runtime.quality.execution.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:t.verdict,issues:t.issues}):e.emit({type:"runtime.quality.planning.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:t.verdict,issues:t.issues})}(e,i,d),d}function buildQualityRecovery(t,n,r,s,u,a){if(a>=u.recovery.maxLoops)return;const o=i({request:n,result:r,phase:s,policy:u,availableToolIds:t.agent.tools,availableSubagentIds:t.agent.subagents,observedEvidence:"execution"===s?e(t.getEvents()):[]});return o&&t.emit({type:"runtime.quality.recovery.started",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,phase:s,attempt:a+1,verdict:r.verdict}),o}function reviewInputFor(e,i){return{workspace:e.workspace,agent:e.agent,request:i,events:e.getEvents()}}function qualityFailureOutput(e,i){return{text:[`Stable runtime quality review blocked final delivery during ${e}.`,"",...i.issues.length>0?i.issues.map(e=>`- ${e.code}: ${e.message}`):["- quality_review_failed: Quality review did not pass."]].join("\n")}}

package/packages/core/dist/quality/types.d.ts

@@ -0,0 +1,47 @@
+import type { CompiledWorkspace, RuntimeEvent, RuntimeOutput, RuntimeRequest, WorkspaceAgent } from "../types.js";
+export type QualityProfile = "off" | "fast" | "balanced" | "strict";
+export type QualityReviewVerdict = "pass" | "revise_plan" | "continue_react" | "blocked";
+export type QualityReviewIssue = {
+    code: string;
+    message: string;
+    recoverable: boolean;
+};
+export type QualityReviewResult = {
+    verdict: QualityReviewVerdict;
+    issues: QualityReviewIssue[];
+};
+export type QualityPolicy = {
+    enabled: boolean;
+    profile: QualityProfile;
+    reviewer?: {
+        mode: "deterministic" | "llm";
+        modelRef?: string;
+        planningPrompt?: string;
+        executionPrompt?: string;
+    };
+    planningReview: {
+        enabled: boolean;
+        requirePlan: boolean;
+    };
+    executionReview: {
+        enabled: boolean;
+        requireToolEvidence: boolean;
+        rejectEmptyFinal: boolean;
+        stopOnBlocker: boolean;
+        rejectUngroundedNumbers: boolean;
+    };
+    recovery: {
+        enabled: boolean;
+        maxLoops: number;
+    };
+};
+export type QualityReviewInput = {
+    workspace: CompiledWorkspace;
+    agent: WorkspaceAgent;
+    request: RuntimeRequest;
+    events: RuntimeEvent[];
+    output?: RuntimeOutput;
+};
+export type QualityReviewModel = {
+    invoke(input: unknown): Promise<unknown> | unknown;
+};

package/packages/core/dist/quality/types.js

@@ -0,0 +1 @@
+export{};

package/packages/core/dist/runtime/direct-tool-call.js

@@ -1 +1 @@
-export async function runDirectToolCall(o){const t=o.request.toolCall;if(!t)throw new Error("Direct tool call request is missing");if(!o.gateway)throw new Error("Runtime tool gateway is not configured");const e=await async function resolveDirectToolCall(o){if(o.agent.tools.includes(o.toolId)&&o.gateway.get(o.toolId))return{toolId:o.toolId,args:o.args};const t=await(o.gateway.repairToolCall?.({toolId:o.toolId,args:o.args,allowedToolIds:o.agent.tools,context:{workspaceRoot:o.workspace.root,requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id}}));if(t&&o.agent.tools.includes(t.toolId)&&o.gateway.get(t.toolId))return t;if(!o.agent.tools.includes(o.toolId))throw new Error(`Tool ${o.toolId} is not assigned to agent ${o.agent.id}`);throw new Error(`Tool is not registered: ${o.toolId}`)}({gateway:o.gateway,workspace:o.workspace,requestId:o.requestId,sessionId:o.sessionId,agent:o.agent,toolId:t.toolId,args:t.args});o.emit({type:"runtime.tool.direct.started",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:e.toolId});const s=await o.gateway.invoke({toolId:e.toolId,args:e.args,context:{workspaceRoot:o.workspace.root,requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id}});return o.emit({type:"runtime.tool.direct.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:s.toolId,output:s.output}),{text:(r=s.output,"string"==typeof r?r:JSON.stringify(r)),metadata:{toolCall:{toolId:s.toolId}}};var r}
+export async function runDirectToolCall(t){const o=t.request.toolCall;if(!o)throw new Error("Direct tool call request is missing");if(!t.gateway)throw new Error("Runtime tool gateway is not configured");const e=await async function resolveDirectToolCall(t){if(t.agent.tools.includes(t.toolId)&&t.gateway.get(t.toolId))return{toolId:t.toolId,args:t.args};const o=await(t.gateway.repairToolCall?.({toolId:t.toolId,args:t.args,allowedToolIds:t.agent.tools,context:{workspaceRoot:t.workspace.root,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,requestInput:t.request.input}}));if(o&&t.agent.tools.includes(o.toolId)&&t.gateway.get(o.toolId))return o;if(!t.agent.tools.includes(t.toolId))throw new Error(`Tool ${t.toolId} is not assigned to agent ${t.agent.id}`);throw new Error(`Tool is not registered: ${t.toolId}`)}({gateway:t.gateway,workspace:t.workspace,requestId:t.requestId,sessionId:t.sessionId,agent:t.agent,request:t.request,toolId:o.toolId,args:o.args});t.emit({type:"runtime.tool.direct.started",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,toolId:e.toolId});const s=await t.gateway.invoke({toolId:e.toolId,args:e.args,context:{workspaceRoot:t.workspace.root,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,requestInput:t.request.input}});return t.emit({type:"runtime.tool.direct.completed",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,toolId:s.toolId,output:s.output}),{text:(r=s.output,"string"==typeof r?r:JSON.stringify(r)),metadata:{toolCall:{toolId:s.toolId}}};var r}

package/packages/core/dist/runtime/events.d.ts

@@ -192,6 +192,36 @@ export type RuntimeEvent = {
     sourceEventIds?: string[];
     model?: string;
     style?: string;
+} | {
+    type: "runtime.quality.planning.reviewed";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    verdict: string;
+    issues: Array<{
+        code: string;
+        message: string;
+        recoverable: boolean;
+    }>;
+} | {
+    type: "runtime.quality.execution.reviewed";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    verdict: string;
+    issues: Array<{
+        code: string;
+        message: string;
+        recoverable: boolean;
+    }>;
+} | {
+    type: "runtime.quality.recovery.started";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    phase: "planning" | "execution";
+    attempt: number;
+    verdict: string;
 } | {
     type: "runtime.adapter.event";
     requestId: string;

package/packages/core/dist/runtime/selection-repair.d.ts

@@ -0,0 +1,14 @@
+import type { CallCandidate, CallSelectionDiagnostics } from "@botbotgo/better-call";
+export type RuntimeSelectionRepairResult = {
+    ok: true;
+    id: string;
+    diagnostics?: CallSelectionDiagnostics;
+} | {
+    ok: false;
+    reason: string;
+    diagnostics?: CallSelectionDiagnostics;
+};
+export declare function repairRuntimeSelection(input: {
+    id: string;
+    candidates: CallCandidate[];
+}): Promise<RuntimeSelectionRepairResult>;

package/packages/core/dist/runtime/selection-repair.js

@@ -0,0 +1 @@
+import{repairCallSelection as a}from"@botbotgo/better-call";export async function repairRuntimeSelection(i){const t=await a({rawIntent:i.id,candidates:i.candidates,mode:"repair"});return t.ok?{ok:!0,id:t.candidateId,diagnostics:t.diagnostics}:{ok:!1,reason:t.reason,diagnostics:t.diagnostics}}

package/packages/core/dist/runtime/tool-gateway.d.ts

@@ -7,6 +7,8 @@ export type RuntimeToolGatewayContext = {
     requestId: string;
     sessionId: string;
     agentId: string;
+    requestInput?: string;
+    observedEvidence?: string;
 };
 export type RuntimeToolGatewayTool = {
     id: string;

package/packages/core/dist/runtime.d.ts

@@ -1,5 +1,6 @@
 import type { ApprovalQueue } from "@stable-harness/governance";
 import type { MemoryProvider, RuntimeMemoryStore } from "@stable-harness/memory";
+import type { QualityReviewModel } from "./quality/index.js";
 import type { CompiledWorkspace, RuntimeCapabilityModule, RuntimeToolGateway, RuntimeAdapter, RuntimeStore, RuntimeProgressNarrationOptions, RuntimeWorkflowAdapter, StableHarnessRuntime } from "./types.js";
 type RuntimeFactoryInput = {
     workspace: CompiledWorkspace;
@@ -11,6 +12,7 @@ type RuntimeFactoryInput = {
     toolGateway?: RuntimeToolGateway;
     store?: RuntimeStore;
     progressNarration?: RuntimeProgressNarrationOptions | false;
+    qualityReviewModel?: QualityReviewModel;
     capabilities?: RuntimeCapabilityModule[];
 };
 export declare function createStableHarnessRuntime(input: RuntimeFactoryInput): StableHarnessRuntime;

package/packages/core/dist/runtime.js

@@ -1 +1 @@
-import{randomUUID as e}from"node:crypto";import{assertExecutionContract as t}from"./execution-contract.js";import{projectRequestInspection as r,projectRequestSummary as s,projectSessionSummaries as o}from"./runtime/persistence/inspection.js";import{assertNoRawToolCallOutput as a,assertNoToolExecutionErrorOutput as n,buildAdapterErrorRecoveryPrompt as i,buildEvidenceSynthesisOutput as u,buildExecutionContractRecoveryRequest as c,buildResultRecoveryRequest as l,containsRawToolCallOutput as p,isRecoverableAdapterError as d,rawToolCallFailureMessage as m,toolCallRecoveryEnabled as w}from"./recovery/tool-call.js";import{completeRun as f,failRun as y}from"./runtime/completion.js";import{runDirectToolCall as R}from"./runtime/direct-tool-call.js";import{createRuntimeCapabilityRegistry as g,normalizeAdapterResult as q}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as I}from"./runtime/memory.js";import{createInMemoryRuntimeStore as k}from"./runtime/persistence/stores.js";import{createProgressNarrationCapability as v}from"./runtime/progress-narration.js";import{runWorkflowRequest as b}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const a=new Set,n=t.store??k(),u=g([I(t),v({options:t.progressNarration,policy:t.workspace.runtime}),...t.capabilities??[]]),emitBase=e=>{n.appendEvent(e);for(const t of a)t(e)},emit=e=>{emitBase(e),u.emitSideEffects(e,emitBase)};return{request:async r=>async function runRuntimeRequest(t){const{agent:r,adapter:s}=function resolveExecution(e,t){const r=t.agentId??e.workspace.runtime.defaultAgentId,s=e.workspace.agents.get(r);if(!s)throw new Error(`Agent ${r} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:s,adapter:void 0};const o=e.adapters.find(e=>e.canRun(s));if(!o)throw new Error(`No runtime adapter can run backend ${s.backend} for agent ${s.id}`);return{agent:s,adapter:o}}(t.input,t.request),o=t.request.requestId??e(),a=t.request.sessionId??e();t.store.createRun(function createRunRecord(e,t,r,s){return{requestId:t,sessionId:r,agentId:s.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,o,a,r)),t.emit({type:"runtime.request.started",requestId:o,sessionId:a,agentId:r.id,input:t.request.input});try{if(t.request.workflow){const e=await b({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:o,sessionId:a,agentId:r.id,emit:t.emit});return f({store:t.store,emit:t.emit,requestId:o,sessionId:a,agent:r,result:e})}if(t.request.toolCall){const e=await R({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:o,sessionId:a,agent:r});return f({store:t.store,emit:t.emit,requestId:o,sessionId:a,agent:r,result:e})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,r=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),s=r.memory,o=r.pluginMemories??[],a=e.input.workspace.runtime,n=new Map;let u;try{u=await runAdapterOnce(e,t,e.request,s,o,n)}catch(r){if(!d(r,a))throw r;u=await runAdapterOnce(e,t,i(e.request,r,a),s,o,n)}u=await recoverAdapterResultOutput(e,t,e.request,u,s,o,a,n),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:u});try{assertRequestExecutionContract(e)}catch(r){const i=c({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:a});if(!i)throw r;u=await runAdapterOnce(e,t,i,s,o,n),u=await recoverAdapterResultOutput(e,t,i,u,s,o,a,n),assertRequestExecutionContract(e)}const l=f({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:u});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:u,response:l}),l}({...t,adapter:s,requestId:o,sessionId:a,agent:r})}catch(e){return y({store:t.store,emit:t.emit,requestId:o,sessionId:a,agent:r,error:e})}}({input:t,capabilities:u,store:n,emit:emit,request:r}),subscribe:e=>(a.add(e),()=>a.delete(e)),inspect(){return{workspaceRoot:t.workspace.root,agents:[...t.workspace.agents.keys()].sort(),workflows:[...t.workspace.workflows.keys()].sort(),...t.workspace.runtime.specDrivenWorkflow?{specDrivenWorkflow:(e=t.workspace.runtime.specDrivenWorkflow,{enabled:e.enabled,artifactsDir:e.artifactsDir,...e.constitution?{constitution:e.constitution}:{},phases:e.phases.map(e=>e.id)})}:{},evaluations:[...t.workspace.evaluations?.keys()??[]].sort(),...t.workspace.runtime.workflowRouting?.defaultWorkflowId?{defaultWorkflowId:t.workspace.runtime.workflowRouting.defaultWorkflowId}:{},workflowRoutes:(t.workspace.runtime.workflowRouting?.routes??[]).map(e=>e.id).sort(),models:[...t.workspace.models.keys()].sort(),tools:[...t.workspace.tools.keys()].sort(),runs:n.listRuns()};var e},getRuntimePolicy:()=>t.workspace.runtime,getWorkflow:e=>t.workspace.workflows.get(e),getRun:e=>n.getRun(e),listRequests:e=>n.listRuns(e).map(s),listSessions:()=>o(n.listRuns()),inspectRequest(e){const s=n.getRun(e);return s?r(t.workspace,s):void 0},cancel(e,t){const r=n.getRun(e);r&&"running"===r.state&&(n.updateRun(e,{state:"cancelled",completedAt:(new Date).toISOString()}),emit({type:"runtime.request.cancelled",requestId:e,sessionId:r.sessionId,agentId:r.agentId,reason:t}))},async stop(){await u.stop(),a.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}async function recoverAdapterResultOutput(e,t,r,s,o,i,c,d){let f=r;const y=function resultRecoveryAttempts(e){const t="object"!=typeof e||null===e||Array.isArray(e)?void 0:e.recovery,r="object"!=typeof t||null===t||Array.isArray(t)?void 0:t.toolCall,s="object"!=typeof r||null===r||Array.isArray(r)?void 0:r.maxResultRecoveryAttempts;return"number"==typeof s&&Number.isInteger(s)&&s>0?s:3}(c);let R=0;for(let r=0;r<y;r+=1){const r=e.store.getRun(e.requestId)?.events??[],a=l({request:f,output:s.text,events:r.slice(R),availableToolIds:e.agent.tools,policy:c});if(!a)break;f=a,R=e.store.getRun(e.requestId)?.events.length??0,s=await runAdapterOnce(e,t,a,o,i,d)}if(w(c)){let t=!1;p(s.text,c)&&function rawToolCallFailureReturnsMessage(e){return"message"===("object"!=typeof e?.toolCallRecovery||null===e.toolCallRecovery||Array.isArray(e.toolCallRecovery)?{}:e.toolCallRecovery).onFailure}(r.metadata)&&(s={...s,text:m(),metadata:{...s.metadata,toolCallRecovery:{failed:!0,reason:"raw_tool_call_output"}}});const o=u({request:r,output:s.text,events:e.store.getRun(e.requestId)?.events??[],policy:c});o&&(t=!0,s={...s,text:o,metadata:{...s.metadata,toolCallRecovery:{synthesized:!0,reason:"raw_tool_call_output_with_evidence"}}}),t||(a(s.text,c),n(s.text,c))}return s}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,s,o,a){return q(await t.run({workspace:e.input.workspace,agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:s,pluginMemories:o,toolGateway:e.input.toolGateway,requestState:a,emit:e.emit}))}
+import{randomUUID as e}from"node:crypto";import{assertExecutionContract as t}from"./execution-contract.js";import{projectRequestInspection as r,projectRequestSummary as s,projectSessionSummaries as o}from"./runtime/persistence/inspection.js";import{assertNoRawToolCallOutput as a,assertNoToolExecutionErrorOutput as n,buildAdapterErrorRecoveryPrompt as i,buildEvidenceSynthesisOutput as u,buildExecutionContractRecoveryRequest as c,buildResultRecoveryRequest as l,containsRawToolCallOutput as p,isRecoverableAdapterError as d,rawToolCallFailureMessage as m,toolCallRecoveryEnabled as w}from"./recovery/tool-call.js";import{recoverQualityReview as f,resolveQualityPolicy as y}from"./quality/index.js";import{completeRun as R,failRun as g}from"./runtime/completion.js";import{runDirectToolCall as I}from"./runtime/direct-tool-call.js";import{createRuntimeCapabilityRegistry as q,normalizeAdapterResult as k}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as v}from"./runtime/memory.js";import{createInMemoryRuntimeStore as b}from"./runtime/persistence/stores.js";import{createProgressNarrationCapability as A}from"./runtime/progress-narration.js";import{repairRuntimeSelection as C}from"./runtime/selection-repair.js";import{runWorkflowRequest as x}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const a=new Set,n=t.store??b(),u=q([v(t),A({options:t.progressNarration,policy:t.workspace.runtime}),...t.capabilities??[]]),emitBase=e=>{n.appendEvent(e);for(const t of a)t(e)},emit=e=>{emitBase(e),u.emitSideEffects(e,emitBase)};return{request:async r=>async function runRuntimeRequest(t){const{agent:r,adapter:s}=await async function resolveExecution(e,t){const r=t.agentId?await async function resolveRequestedAgentId(e,t){if(e.agents.has(t))return t;const r=await C({id:t,candidates:[...e.agents.values()].map(e=>({id:e.id,description:e.description}))});return r.ok?r.id:t}(e.workspace,t.agentId):function resolveRoutedAgentId(e,t){for(const r of e.runtime.routes??[])if(routeMatches(r,t))return r.agentId;return e.runtime.defaultAgentId}(e.workspace,t.input),s=e.workspace.agents.get(r);if(!s)throw new Error(`Agent ${r} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:s,adapter:void 0};const o=e.adapters.find(e=>e.canRun(s));if(!o)throw new Error(`No runtime adapter can run backend ${s.backend} for agent ${s.id}`);return{agent:s,adapter:o}}(t.input,t.request),o=t.request.requestId??e(),a=t.request.sessionId??e();t.store.createRun(function createRunRecord(e,t,r,s){return{requestId:t,sessionId:r,agentId:s.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,o,a,r)),t.emit({type:"runtime.request.started",requestId:o,sessionId:a,agentId:r.id,input:t.request.input});try{if(t.request.workflow){const e=await x({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:o,sessionId:a,agentId:r.id,emit:t.emit});return R({store:t.store,emit:t.emit,requestId:o,sessionId:a,agent:r,result:e})}if(t.request.toolCall){const e=await I({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:o,sessionId:a,agent:r});return R({store:t.store,emit:t.emit,requestId:o,sessionId:a,agent:r,result:e})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,r=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),s=r.memory,o=r.pluginMemories??[],a=e.input.workspace.runtime,n=y(e.input.workspace.runtime,e.agent),u=new Map;let l;try{l=await runAdapterOnce(e,t,e.request,s,o,u)}catch(r){if(!d(r,a))throw r;l=await runAdapterOnce(e,t,i(e.request,r,a),s,o,u)}l=await recoverAdapterResultOutput(e,t,e.request,l,s,o,a,u),l=await f(createQualityRuntimeInput(e,s,o,u),e.request,l,n),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:l});try{assertRequestExecutionContract(e)}catch(r){const i=c({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:a});if(!i)throw r;l=await runAdapterOnce(e,t,i,s,o,u),l=await recoverAdapterResultOutput(e,t,i,l,s,o,a,u),l=await f(createQualityRuntimeInput(e,s,o,u),i,l,n),assertRequestExecutionContract(e)}const p=R({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:l});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:l,response:p}),p}({...t,adapter:s,requestId:o,sessionId:a,agent:r})}catch(e){return g({store:t.store,emit:t.emit,requestId:o,sessionId:a,agent:r,error:e})}}({input:t,capabilities:u,store:n,emit:emit,request:r}),subscribe:e=>(a.add(e),()=>a.delete(e)),inspect(){return{workspaceRoot:t.workspace.root,agents:[...t.workspace.agents.keys()].sort(),workflows:[...t.workspace.workflows.keys()].sort(),...t.workspace.runtime.specDrivenWorkflow?{specDrivenWorkflow:(e=t.workspace.runtime.specDrivenWorkflow,{enabled:e.enabled,artifactsDir:e.artifactsDir,...e.constitution?{constitution:e.constitution}:{},phases:e.phases.map(e=>e.id)})}:{},evaluations:[...t.workspace.evaluations?.keys()??[]].sort(),...t.workspace.runtime.workflowRouting?.defaultWorkflowId?{defaultWorkflowId:t.workspace.runtime.workflowRouting.defaultWorkflowId}:{},workflowRoutes:(t.workspace.runtime.workflowRouting?.routes??[]).map(e=>e.id).sort(),models:[...t.workspace.models.keys()].sort(),tools:[...t.workspace.tools.keys()].sort(),runs:n.listRuns()};var e},getRuntimePolicy:()=>t.workspace.runtime,getWorkflow:e=>t.workspace.workflows.get(e),getRun:e=>n.getRun(e),listRequests:e=>n.listRuns(e).map(s),listSessions:()=>o(n.listRuns()),inspectRequest(e){const s=n.getRun(e);return s?r(t.workspace,s):void 0},cancel(e,t){const r=n.getRun(e);r&&"running"===r.state&&(n.updateRun(e,{state:"cancelled",completedAt:(new Date).toISOString()}),emit({type:"runtime.request.cancelled",requestId:e,sessionId:r.sessionId,agentId:r.agentId,reason:t}))},async stop(){await u.stop(),a.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}function createQualityRuntimeInput(e,t,r,s){return{workspace:e.input.workspace,agent:e.agent,request:e.request,requestId:e.requestId,sessionId:e.sessionId,events:e.store.getRun(e.requestId)?.events??[],emit:e.emit,getEvents:()=>e.store.getRun(e.requestId)?.events??[],runAdapter:o=>runAdapterOnce(e,e.adapter,o,t,r,s),reviewModel:e.input.qualityReviewModel,memory:t,pluginMemories:r}}async function recoverAdapterResultOutput(e,t,r,s,o,i,c,d){let f=r;const y=function resultRecoveryAttempts(e){const t="object"!=typeof e||null===e||Array.isArray(e)?void 0:e.recovery,r="object"!=typeof t||null===t||Array.isArray(t)?void 0:t.toolCall,s="object"!=typeof r||null===r||Array.isArray(r)?void 0:r.maxResultRecoveryAttempts;return"number"==typeof s&&Number.isInteger(s)&&s>0?s:3}(c);let R=0;for(let r=0;r<y;r+=1){const r=e.store.getRun(e.requestId)?.events??[],a=l({request:f,output:s.text,events:r.slice(R),availableToolIds:e.agent.tools,policy:c});if(!a)break;f=a,R=e.store.getRun(e.requestId)?.events.length??0,s=await runAdapterOnce(e,t,a,o,i,d)}if(w(c)){let t=!1;p(s.text,c)&&function rawToolCallFailureReturnsMessage(e){return"message"===("object"!=typeof e?.toolCallRecovery||null===e.toolCallRecovery||Array.isArray(e.toolCallRecovery)?{}:e.toolCallRecovery).onFailure}(r.metadata)&&(s={...s,text:m(),metadata:{...s.metadata,toolCallRecovery:{failed:!0,reason:"raw_tool_call_output"}}});const o=u({request:r,output:s.text,events:e.store.getRun(e.requestId)?.events??[],policy:c});o&&(t=!0,s={...s,text:o,metadata:{...s.metadata,toolCallRecovery:{synthesized:!0,reason:"raw_tool_call_output_with_evidence"}}}),t||(a(s.text,c),n(s.text,c))}return s}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,s,o,a){return k(await t.run({workspace:e.input.workspace,agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:s,pluginMemories:o,toolGateway:e.input.toolGateway,requestState:a,emit:e.emit}))}function routeMatches(e,t){if(e.pattern)try{if(new RegExp(e.pattern,"iu").test(t))return!0}catch{return!1}const r=t.toLowerCase();return(e.keywords??[]).some(e=>r.includes(e.toLowerCase()))}

package/packages/core/dist/workflows/runtime.js

@@ -1 +1 @@
-export async function runWorkflowRequest(e){const o=function resolveWorkflow(e,o){const r=o.routeId?e.runtime.workflowRouting?.routes?.find(e=>e.id===o.routeId):void 0;if(o.routeId&&!r)throw new Error(`Workflow route is not defined: ${o.routeId}`);const t=o.workflowId??r?.workflowId??e.runtime.workflowRouting?.defaultWorkflowId;if(!t)throw new Error("Workflow request requires workflowId, routeId, or runtime.workflowRouting.defaultWorkflowId");const n=e.workflows.get(t);if(!n)throw new Error(`Workflow is not defined: ${t}`);return n}(e.workspace,e.request),r=function resolveWorkflowAdapter(e,o){const r=o?e.find(e=>e.name===o):e[0];if(!r)throw new Error("No workflow adapter is configured"+(o?` for ${o}`:""));return r}(e.adapters,o.adapter);e.emit({type:"runtime.workflow.started",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agentId,adapter:r.name,workflowId:o.id});const t=await r.run({workspace:e.workspace,workflow:o,request:e.request,requestId:e.requestId,sessionId:e.sessionId,toolGateway:e.toolGateway,emit:o=>e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agentId,event:o})});return e.emit({type:"runtime.workflow.completed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agentId,adapter:r.name,workflowId:o.id}),"string"==typeof t?{text:t}:t}
+import{repairRuntimeSelection as e}from"../runtime/selection-repair.js";export async function runWorkflowRequest(o){const r=await async function resolveWorkflow(o,r){const t=r.routeId?await async function resolveWorkflowRouteId(o,r){const t=o.runtime.workflowRouting?.routes??[];if(t.some(e=>e.id===r))return r;const n=await e({id:r,candidates:t.map(e=>({id:e.id,description:e.description,metadata:e.metadata}))});return n.ok?n.id:r}(o,r.routeId):void 0,n=t?o.runtime.workflowRouting?.routes?.find(e=>e.id===t):void 0;if(r.routeId&&!n)throw new Error(`Workflow route is not defined: ${r.routeId}`);const i=(r.workflowId?await async function resolveWorkflowId(o,r){if(o.workflows.has(r))return r;const t=await e({id:r,candidates:[...o.workflows.values()].map(e=>({id:e.id,description:e.description}))});return t.ok?t.id:r}(o,r.workflowId):void 0)??n?.workflowId??o.runtime.workflowRouting?.defaultWorkflowId;if(!i)throw new Error("Workflow request requires workflowId, routeId, or runtime.workflowRouting.defaultWorkflowId");const s=o.workflows.get(i);if(!s)throw new Error(`Workflow is not defined: ${i}`);return s}(o.workspace,o.request),t=function resolveWorkflowAdapter(e,o){const r=o?e.find(e=>e.name===o):e[0];if(!r)throw new Error("No workflow adapter is configured"+(o?` for ${o}`:""));return r}(o.adapters,r.adapter);o.emit({type:"runtime.workflow.started",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agentId,adapter:t.name,workflowId:r.id});const n=await t.run({workspace:o.workspace,workflow:r,request:o.request,requestId:o.requestId,sessionId:o.sessionId,toolGateway:o.toolGateway,emit:e=>o.emit({type:"runtime.adapter.event",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agentId,event:e})});return o.emit({type:"runtime.workflow.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agentId,adapter:t.name,workflowId:r.id}),"string"==typeof n?{text:n}:n}

package/packages/core/dist/workspace/types.d.ts

@@ -49,6 +49,7 @@ export type WorkspaceAgent = {
 };
 export type WorkspaceRuntimePolicy = {
     defaultAgentId: string;
+    routes?: WorkspaceAgentRoute[];
     workspaceId?: string;
     profile?: string;
     workspaceValidation?: WorkspaceValidationPolicy;
@@ -63,9 +64,17 @@ export type WorkspaceRuntimePolicy = {
     protocols?: Record<string, unknown>;
     progress?: Record<string, unknown>;
     cli?: Record<string, unknown>;
+    quality?: string | Record<string, unknown>;
     responseLanguage?: Record<string, unknown>;
     responsePresentation?: Record<string, unknown>;
 };
+export type WorkspaceAgentRoute = {
+    id: string;
+    agentId: string;
+    keywords?: string[];
+    pattern?: string;
+    description?: string;
+};
 export type WorkspaceValidationPolicy = {
     boundaryScan?: WorkspaceBoundaryScanPolicy;
 };

package/packages/core/package.json

@@ -11,6 +11,7 @@
     ".": "./dist/index.js"
   },
   "dependencies": {
+    "@botbotgo/better-call": "^0.1.58",
     "@stable-harness/governance": "0.0.1",
     "@stable-harness/memory": "0.0.1"
   }

package/packages/tool-gateway/dist/src/module-loader.js

@@ -1 +1 @@
-import o from"node:path";import{pathToFileURL as e}from"node:url";import{createInMemoryToolGateway as t}from"./in-memory.js";export async function createModuleToolGateway(o){const e=[];for(const t of o.tools){const o=await loadModuleTool(t);o&&e.push(o)}return t(e,o.options)}async function loadModuleTool(t){if(!t.sourcePath)return;const n=await import(e(o.resolve(t.sourcePath)).href),r=n[t.name??t.id]??n.default;return function hasInvoke(o){return"object"==typeof o&&null!==o&&"invoke"in o&&"function"==typeof o.invoke}(r)?{id:t.id,description:t.description??r.description,schema:t.schema??r.schema,validateArgs:r.validateArgs,invoke:(o,e)=>async function invokeWithWorkspaceCwd(o,e,t){const n=process.cwd();try{return process.chdir(t.workspaceRoot),await o(e)}finally{process.chdir(n)}}(r.invoke,o,e)}:void 0}
+import o from"node:path";import{pathToFileURL as e}from"node:url";import{createInMemoryToolGateway as t}from"./in-memory.js";export async function createModuleToolGateway(o){const e=[];for(const t of o.tools){const o=await loadModuleTool(t);o&&e.push(o)}return t(e,o.options)}async function loadModuleTool(t){if(!t.sourcePath)return;const n=await import(e(o.resolve(t.sourcePath)).href),r=n[t.name??t.id]??n.default;return function hasInvoke(o){return"object"==typeof o&&null!==o&&"invoke"in o&&"function"==typeof o.invoke}(r)?{id:t.id,description:t.description??r.description,schema:t.schema??r.schema,validateArgs:r.validateArgs,invoke:(o,e)=>async function invokeWithWorkspaceCwd(o,e,t){const n=process.cwd();try{return process.chdir(t.workspaceRoot),await o(e,t)}finally{process.chdir(n)}}(r.invoke,o,e)}:void 0}

package/packages/tool-gateway/dist/src/types.d.ts

@@ -3,6 +3,8 @@ export type ToolGatewayContext = {
     requestId: string;
     sessionId: string;
     agentId: string;
+    requestInput?: string;
+    observedEvidence?: string;
 };
 export type ToolGatewayTool = {
     id: string;

package/packages/tool-gateway/package.json

@@ -10,6 +10,6 @@
   "main": "dist/src/index.js",
   "types": "dist/src/index.d.ts",
   "dependencies": {
-    "@botbotgo/better-call": "^0.1.18"
+    "@botbotgo/better-call": "^0.1.58"
   }
 }

package/packages/workspace-yaml/dist/documents.js

@@ -1 +1 @@
-import{assertSpecDrivenWorkflowPolicy as e,createSpecDrivenWorkflowPolicy as r}from"@stable-harness/core";function assertRecord(e,r){if("object"!=typeof e||null===e||Array.isArray(e))throw new Error(`${r} must be an object`);return e}function readName(e,r){const t=e.metadata?.name;if("string"==typeof t&&t.trim())return t.trim();if(r)return r;throw new Error(`Document kind ${String(e.kind)} requires metadata.name`)}function readDescription(e){const r=e.metadata?.description;return"string"==typeof r&&r.trim()?r.trim():void 0}function readOptionalString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function toStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.trim().length>0):[]}function resolveValue(e){if("string"!=typeof e)return e;const r=e.match(/^\$\{env:([A-Za-z_][A-Za-z0-9_]*)(?::-(.*))?\}$/u);return r?process.env[r[1]]??r[2]??"":e}export function compileRuntime(e){const r=assertRecord(e.spec,"Runtime.spec"),t=assertRecord(r.routing??{},"Runtime.spec.routing");return{defaultAgentId:"string"==typeof t.defaultAgentId&&t.defaultAgentId.trim()?t.defaultAgentId.trim():"orchestra",...readOptionalString(r.workspaceId)?{workspaceId:readOptionalString(r.workspaceId)}:{},...readOptionalString(r.profile)?{profile:readOptionalString(r.profile)}:{},...void 0!==r.adapters?{adapters:readAdapters(r.adapters)}:{},..."object"==typeof r.workflowRouting&&r.workflowRouting?{workflowRouting:readWorkflowRouting(r.workflowRouting)}:{},..."object"==typeof r.specDrivenWorkflow&&r.specDrivenWorkflow?{specDrivenWorkflow:readSpecDrivenWorkflow(r.specDrivenWorkflow)}:{},..."object"==typeof r.approvals&&r.approvals?{approvals:r.approvals}:{},..."object"==typeof r.recovery&&r.recovery?{recovery:r.recovery}:{},..."object"==typeof r.retry&&r.retry?{retry:r.retry}:{},..."object"==typeof r.toolGateway&&r.toolGateway?{toolGateway:r.toolGateway}:{},..."object"==typeof r.memory&&r.memory?{memory:r.memory}:{},..."object"==typeof r.protocols&&r.protocols?{protocols:r.protocols}:{},..."object"==typeof r.progress&&r.progress?{progress:r.progress}:{},..."object"==typeof r.cli&&r.cli?{cli:r.cli}:{},..."object"==typeof r.workspaceValidation&&r.workspaceValidation?{workspaceValidation:r.workspaceValidation}:{},..."object"==typeof r.responseLanguage&&r.responseLanguage?{responseLanguage:r.responseLanguage}:{},..."object"==typeof r.responsePresentation&&r.responsePresentation?{responsePresentation:r.responsePresentation}:{}}}function readSpecDrivenWorkflow(t){const o=assertRecord(t,"Runtime.spec.specDrivenWorkflow"),n=r({enabled:!0===o.enabled,constitution:readOptionalString(o.constitution),artifactsDir:readOptionalString(o.artifactsDir),phases:void 0===o.phases?void 0:readSpecDrivenPhases(o.phases),..."object"==typeof o.gates&&o.gates?{gates:o.gates}:{},..."object"==typeof o.config&&o.config?{config:o.config}:{}});return e(n),n}function readSpecDrivenPhases(e){if(!Array.isArray(e))throw new Error("Runtime.spec.specDrivenWorkflow.phases must be an array");return e.map(e=>{if("string"==typeof e&&e.trim())return{id:e.trim()};const r=assertRecord(e,"Runtime.spec.specDrivenWorkflow.phases[]"),t=readOptionalString(r.id);if(!t)throw new Error("Runtime.spec.specDrivenWorkflow.phases[] requires id");return{id:t,...readOptionalString(r.artifactKind)?{artifactKind:readOptionalString(r.artifactKind)}:{},..."boolean"==typeof r.required?{required:r.required}:{},...readOptionalString(r.gate)?{gate:readOptionalString(r.gate)}:{},..."object"==typeof r.config&&r.config?{config:r.config}:{}}})}export function compileAgent(e,r){const t=assertRecord(e.spec,"Agent.spec"),o=readName(e),n=readOptionalString(t.backend);if(!n)throw new Error(`Agent ${o} requires spec.backend`);const i="object"==typeof t.config&&t.config?t.config:{},a="string"==typeof t.systemPrompt?t.systemPrompt:"string"==typeof i.systemPrompt?i.systemPrompt:void 0;return{id:o,...readDescription(e)?{description:readDescription(e)}:{},sourcePath:r,backend:n,..."string"==typeof t.modelRef&&t.modelRef.trim()?{modelRef:(s=t.modelRef,s.replace(/^[^/]+\//u,""))}:{},...void 0!==a?{systemPrompt:a}:{},tools:toStringArray(t.tools),skills:toStringArray(t.skills),memory:Array.isArray(t.memory)?t.memory:[],subagents:toStringArray(t.subagents),...void 0!==t.edges?{edges:readAgentEdges(t.edges,o)}:{},config:i};var s}export function compileModel(e){return compileModelSpec(assertRecord(e.spec,"Model.spec"),readName(e))}export function compileModelSpec(e,r){const t="string"==typeof e.name&&e.name.trim()?e.name.trim():r??"default",o=resolveValue(e.provider),n=resolveValue(e.model),i="string"==typeof o&&o.trim()?o.trim():"unknown",a="string"==typeof n&&n.trim()?n.trim():t,s={...e};return delete s.name,delete s.provider,delete s.model,{id:t,provider:i,model:a,config:Object.fromEntries(Object.entries(s).map(([e,r])=>[e,resolveValue(r)]))}}export function compileTool(e,r){const t=assertRecord(e.spec,"Tool.spec");return{id:readName(e),...r?{sourcePath:r}:{},..."string"==typeof t.description?{description:t.description}:{},...void 0!==t.schema?{schema:t.schema}:{},..."string"==typeof t.implementation?{implementation:t.implementation}:{}}}export function compileMemory(e){const r=assertRecord(e.spec,"Memory.spec"),t=readName(e),o={...r};return delete o.provider,delete o.profile,delete o.mode,delete o.enabled,delete o.prompts,{id:t,provider:readOptionalString(r.provider)??"langmem",...readOptionalString(r.profile)?{profile:readOptionalString(r.profile)}:{},...readOptionalString(r.mode)?{mode:readOptionalString(r.mode)}:{},enabled:!1!==r.enabled,..."object"==typeof r.prompts&&r.prompts?{prompts:readMemoryPrompts(r.prompts)}:{},...Object.keys(o).length>0?{config:o}:{}}}function readWorkflowRouting(e){const r=assertRecord(e,"Runtime.spec.workflowRouting"),t=void 0===r.routes?void 0:function readWorkflowRoutes(e){if(!Array.isArray(e))throw new Error("Runtime.spec.workflowRouting.routes must be an array");return e.map(e=>{const r=assertRecord(e,"Runtime.spec.workflowRouting.routes[]"),t=readOptionalString(r.id),o=readOptionalString(r.workflowId);if(!t||!o)throw new Error("Runtime.spec.workflowRouting.routes[] requires id and workflowId");return{id:t,workflowId:o,...readOptionalString(r.description)?{description:readOptionalString(r.description)}:{},..."object"==typeof r.metadata&&r.metadata?{metadata:r.metadata}:{}}})}(r.routes);return{...readOptionalString(r.defaultWorkflowId)?{defaultWorkflowId:readOptionalString(r.defaultWorkflowId)}:{},...t?{routes:t}:{}}}function readAdapters(e){if(!Array.isArray(e))throw new Error("Runtime.spec.adapters must be an array");return e.map(readAdapter)}function readAgentEdges(e,r){if(!Array.isArray(e))throw new Error(`Agent ${r} spec.edges must be an array`);return e.map(e=>{const t=assertRecord(e,`Agent ${r} spec.edges[]`),o=readOptionalString(t.from),n=readOptionalString(t.to);if(!o||!n)throw new Error(`Agent ${r} spec.edges[] requires from and to`);return{from:o,to:n,...readOptionalString(t.condition)?{condition:readOptionalString(t.condition)}:{}}})}function readAdapter(e){if("string"==typeof e&&e.trim())return{name:e.trim()};const r=assertRecord(e,"Runtime.spec.adapters[]"),t=readOptionalString(r.name)??readOptionalString(r.id)??readOptionalString(r.backend);if(!t)throw new Error("Runtime.spec.adapters[] requires name");return{name:t,..."boolean"==typeof r.enabled?{enabled:r.enabled}:{},..."object"==typeof r.config&&r.config?{config:r.config}:{}}}function readMemoryPrompts(e){const r=assertRecord(e,"Memory.spec.prompts");return{...readOptionalString(r.semantic)?{semantic:readOptionalString(r.semantic)}:{},...readOptionalString(r.episodic)?{episodic:readOptionalString(r.episodic)}:{},...readOptionalString(r.procedural)?{procedural:readOptionalString(r.procedural)}:{}}}
+import{assertSpecDrivenWorkflowPolicy as e,createSpecDrivenWorkflowPolicy as r}from"@stable-harness/core";function assertRecord(e,r){if("object"!=typeof e||null===e||Array.isArray(e))throw new Error(`${r} must be an object`);return e}function readName(e,r){const t=e.metadata?.name;if("string"==typeof t&&t.trim())return t.trim();if(r)return r;throw new Error(`Document kind ${String(e.kind)} requires metadata.name`)}function readDescription(e){const r=e.metadata?.description;return"string"==typeof r&&r.trim()?r.trim():void 0}function readOptionalString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function toStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.trim().length>0):[]}function resolveValue(e){if("string"!=typeof e)return e;const r=e.match(/^\$\{env:([A-Za-z_][A-Za-z0-9_]*)(?::-(.*))?\}$/u);return r?process.env[r[1]]??r[2]??"":e}export function compileRuntime(e){const r=assertRecord(e.spec,"Runtime.spec"),t=assertRecord(r.routing??{},"Runtime.spec.routing");return{defaultAgentId:"string"==typeof t.defaultAgentId&&t.defaultAgentId.trim()?t.defaultAgentId.trim():"orchestra",...void 0!==t.routes?{routes:readAgentRoutes(t.routes)}:{},...readOptionalString(r.workspaceId)?{workspaceId:readOptionalString(r.workspaceId)}:{},...readOptionalString(r.profile)?{profile:readOptionalString(r.profile)}:{},...void 0!==r.adapters?{adapters:readAdapters(r.adapters)}:{},..."object"==typeof r.workflowRouting&&r.workflowRouting?{workflowRouting:readWorkflowRouting(r.workflowRouting)}:{},..."object"==typeof r.specDrivenWorkflow&&r.specDrivenWorkflow?{specDrivenWorkflow:readSpecDrivenWorkflow(r.specDrivenWorkflow)}:{},..."object"==typeof r.approvals&&r.approvals?{approvals:r.approvals}:{},..."object"==typeof r.recovery&&r.recovery?{recovery:r.recovery}:{},..."object"==typeof r.retry&&r.retry?{retry:r.retry}:{},..."object"==typeof r.toolGateway&&r.toolGateway?{toolGateway:r.toolGateway}:{},..."object"==typeof r.memory&&r.memory?{memory:r.memory}:{},..."object"==typeof r.protocols&&r.protocols?{protocols:r.protocols}:{},..."object"==typeof r.progress&&r.progress?{progress:r.progress}:{},..."object"==typeof r.cli&&r.cli?{cli:r.cli}:{},..."string"==typeof r.quality||"object"==typeof r.quality&&r.quality?{quality:r.quality}:{},..."object"==typeof r.workspaceValidation&&r.workspaceValidation?{workspaceValidation:r.workspaceValidation}:{},..."object"==typeof r.responseLanguage&&r.responseLanguage?{responseLanguage:r.responseLanguage}:{},..."object"==typeof r.responsePresentation&&r.responsePresentation?{responsePresentation:r.responsePresentation}:{}}}function readAgentRoutes(e){if(!Array.isArray(e))throw new Error("Runtime.spec.routing.routes must be an array");return e.map(e=>{const r=assertRecord(e,"Runtime.spec.routing.routes[]"),t=readOptionalString(r.id),o=readOptionalString(r.agentId);if(!t||!o)throw new Error("Runtime.spec.routing.routes[] requires id and agentId");const n=void 0===r.keywords?void 0:function assertStringArray(e,r){if(!Array.isArray(e))throw new Error(`${r} must be an array`);return e.map(e=>{if("string"!=typeof e||!e.trim())throw new Error(`${r} must contain non-empty strings`);return e.trim()})}(r.keywords,"Runtime.spec.routing.routes[].keywords"),i=readOptionalString(r.pattern);if(!(n&&0!==n.length||i))throw new Error("Runtime.spec.routing.routes[] requires keywords or pattern");return{id:t,agentId:o,...n&&n.length>0?{keywords:n}:{},...i?{pattern:i}:{},...readOptionalString(r.description)?{description:readOptionalString(r.description)}:{}}})}function readSpecDrivenWorkflow(t){const o=assertRecord(t,"Runtime.spec.specDrivenWorkflow"),n=r({enabled:!0===o.enabled,constitution:readOptionalString(o.constitution),artifactsDir:readOptionalString(o.artifactsDir),phases:void 0===o.phases?void 0:readSpecDrivenPhases(o.phases),..."object"==typeof o.gates&&o.gates?{gates:o.gates}:{},..."object"==typeof o.config&&o.config?{config:o.config}:{}});return e(n),n}function readSpecDrivenPhases(e){if(!Array.isArray(e))throw new Error("Runtime.spec.specDrivenWorkflow.phases must be an array");return e.map(e=>{if("string"==typeof e&&e.trim())return{id:e.trim()};const r=assertRecord(e,"Runtime.spec.specDrivenWorkflow.phases[]"),t=readOptionalString(r.id);if(!t)throw new Error("Runtime.spec.specDrivenWorkflow.phases[] requires id");return{id:t,...readOptionalString(r.artifactKind)?{artifactKind:readOptionalString(r.artifactKind)}:{},..."boolean"==typeof r.required?{required:r.required}:{},...readOptionalString(r.gate)?{gate:readOptionalString(r.gate)}:{},..."object"==typeof r.config&&r.config?{config:r.config}:{}}})}export function compileAgent(e,r){const t=assertRecord(e.spec,"Agent.spec"),o=readName(e),n=readOptionalString(t.backend);if(!n)throw new Error(`Agent ${o} requires spec.backend`);const i="object"==typeof t.config&&t.config?t.config:{},a="string"==typeof t.systemPrompt?t.systemPrompt:"string"==typeof i.systemPrompt?i.systemPrompt:void 0;return{id:o,...readDescription(e)?{description:readDescription(e)}:{},sourcePath:r,backend:n,..."string"==typeof t.modelRef&&t.modelRef.trim()?{modelRef:(s=t.modelRef,s.replace(/^[^/]+\//u,""))}:{},...void 0!==a?{systemPrompt:a}:{},tools:toStringArray(t.tools),skills:toStringArray(t.skills),memory:Array.isArray(t.memory)?t.memory:[],subagents:toStringArray(t.subagents),...void 0!==t.edges?{edges:readAgentEdges(t.edges,o)}:{},config:i};var s}export function compileModel(e){return compileModelSpec(assertRecord(e.spec,"Model.spec"),readName(e))}export function compileModelSpec(e,r){const t="string"==typeof e.name&&e.name.trim()?e.name.trim():r??"default",o=resolveValue(e.provider),n=resolveValue(e.model),i="string"==typeof o&&o.trim()?o.trim():"unknown",a="string"==typeof n&&n.trim()?n.trim():t,s={...e};return delete s.name,delete s.provider,delete s.model,{id:t,provider:i,model:a,config:Object.fromEntries(Object.entries(s).map(([e,r])=>[e,resolveValue(r)]))}}export function compileTool(e,r){const t=assertRecord(e.spec,"Tool.spec");return{id:readName(e),...r?{sourcePath:r}:{},..."string"==typeof t.description?{description:t.description}:{},...void 0!==t.schema?{schema:t.schema}:{},..."string"==typeof t.implementation?{implementation:t.implementation}:{}}}export function compileMemory(e){const r=assertRecord(e.spec,"Memory.spec"),t=readName(e),o={...r};return delete o.provider,delete o.profile,delete o.mode,delete o.enabled,delete o.prompts,{id:t,provider:readOptionalString(r.provider)??"langmem",...readOptionalString(r.profile)?{profile:readOptionalString(r.profile)}:{},...readOptionalString(r.mode)?{mode:readOptionalString(r.mode)}:{},enabled:!1!==r.enabled,..."object"==typeof r.prompts&&r.prompts?{prompts:readMemoryPrompts(r.prompts)}:{},...Object.keys(o).length>0?{config:o}:{}}}function readWorkflowRouting(e){const r=assertRecord(e,"Runtime.spec.workflowRouting"),t=void 0===r.routes?void 0:function readWorkflowRoutes(e){if(!Array.isArray(e))throw new Error("Runtime.spec.workflowRouting.routes must be an array");return e.map(e=>{const r=assertRecord(e,"Runtime.spec.workflowRouting.routes[]"),t=readOptionalString(r.id),o=readOptionalString(r.workflowId);if(!t||!o)throw new Error("Runtime.spec.workflowRouting.routes[] requires id and workflowId");return{id:t,workflowId:o,...readOptionalString(r.description)?{description:readOptionalString(r.description)}:{},..."object"==typeof r.metadata&&r.metadata?{metadata:r.metadata}:{}}})}(r.routes);return{...readOptionalString(r.defaultWorkflowId)?{defaultWorkflowId:readOptionalString(r.defaultWorkflowId)}:{},...t?{routes:t}:{}}}function readAdapters(e){if(!Array.isArray(e))throw new Error("Runtime.spec.adapters must be an array");return e.map(readAdapter)}function readAgentEdges(e,r){if(!Array.isArray(e))throw new Error(`Agent ${r} spec.edges must be an array`);return e.map(e=>{const t=assertRecord(e,`Agent ${r} spec.edges[]`),o=readOptionalString(t.from),n=readOptionalString(t.to);if(!o||!n)throw new Error(`Agent ${r} spec.edges[] requires from and to`);return{from:o,to:n,...readOptionalString(t.condition)?{condition:readOptionalString(t.condition)}:{}}})}function readAdapter(e){if("string"==typeof e&&e.trim())return{name:e.trim()};const r=assertRecord(e,"Runtime.spec.adapters[]"),t=readOptionalString(r.name)??readOptionalString(r.id)??readOptionalString(r.backend);if(!t)throw new Error("Runtime.spec.adapters[] requires name");return{name:t,..."boolean"==typeof r.enabled?{enabled:r.enabled}:{},..."object"==typeof r.config&&r.config?{config:r.config}:{}}}function readMemoryPrompts(e){const r=assertRecord(e,"Memory.spec.prompts");return{...readOptionalString(r.semantic)?{semantic:readOptionalString(r.semantic)}:{},...readOptionalString(r.episodic)?{episodic:readOptionalString(r.episodic)}:{},...readOptionalString(r.procedural)?{procedural:readOptionalString(r.procedural)}:{}}}

package/packages/workspace-yaml/dist/loader.js

@@ -1 +1 @@
-import{readFile as o}from"node:fs/promises";import s from"node:path";import{parseAllDocuments as e}from"yaml";import{discoverModuleTools as a,discoverSkills as t,listYamlFiles as r}from"./discovery.js";import{compileAgent as i,compileMemory as n,compileModel as l,compileModelSpec as f,compileRuntime as c,compileTool as w}from"./documents.js";import{compileWorkflow as d,validateWorkflows as m}from"./workflows.js";import{compileEvaluation as k,validateEvaluations as p}from"./evaluations.js";import{assertWorkspaceBoundaryDiagnostics as u,scanWorkspaceBoundaries as g}from"./boundary-scan.js";export async function loadWorkspaceFromYaml(M){const y=s.join(M,"config"),v=await r(y),W=[],b=new Map,h=new Map,A=new Map,R=new Map,j=new Map,I=new Map,E=new Map;for(const s of v){const a=await o(s,"utf8"),t=e(a).map(o=>o.toJSON()).filter(o=>null!==o);for(const o of t)if("string"==typeof o.kind)switch(o.kind){case"Runtime":W.push(c(o));break;case"Agent":{const e=i(o,s);b.set(e.id,e);break}case"Model":{const s=l(o);h.set(s.id,s);break}case"Models":if(Array.isArray(o.spec))for(const s of o.spec)if("object"==typeof s&&null!==s&&!Array.isArray(s)){const o=f(s);h.set(o.id,o)}break;case"Tool":{const e=w(o,s);A.set(e.id,e);break}case"Memory":{const s=n(o);j.set(s.id,s);break}case"Workflow":{const e=d(o,s);I.set(e.id,e);break}case"Evaluation":{const e=k(o,s);E.set(e.id,e);break}}}for(const o of await a(M))A.has(o.id)||A.set(o.id,o);for(const o of await t(M))R.set(o.id,o);const F=W.at(-1)??{defaultAgentId:"orchestra"};m({workflows:I,agents:b,tools:A,skills:R}),p({evaluations:E,agents:b,tools:A,workflows:I}),function validateWorkflowRouting(o,s){const e=o.workflowRouting;if(e){if(e.defaultWorkflowId&&!s.has(e.defaultWorkflowId))throw new Error(`Runtime workflowRouting.defaultWorkflowId references unknown workflow ${e.defaultWorkflowId}`);for(const o of e.routes??[])if(!s.has(o.workflowId))throw new Error(`Runtime workflowRouting route ${o.id} references unknown workflow ${o.workflowId}`)}}(F,I);const S={root:M,runtime:F,agents:b,models:h,tools:A,skills:R,memories:j,workflows:I,evaluations:E},T=g(S);return u(T),{...S,...T.length>0?{diagnostics:T}:{}}}
+import{readFile as o}from"node:fs/promises";import e from"node:path";import{parseAllDocuments as t}from"yaml";import{discoverModuleTools as s,discoverSkills as r,listYamlFiles as n}from"./discovery.js";import{compileAgent as a,compileMemory as l,compileModel as i,compileModelSpec as c,compileRuntime as f,compileTool as u}from"./documents.js";import{compileWorkflow as w,validateWorkflows as m}from"./workflows.js";import{compileEvaluation as d,validateEvaluations as p}from"./evaluations.js";import{assertWorkspaceBoundaryDiagnostics as k,scanWorkspaceBoundaries as g}from"./boundary-scan.js";export async function loadWorkspaceFromYaml(a){const l=e.join(a,"config"),i=await n(l),c=[],f=new Map,u=new Map,w=new Map,d=new Map,v=new Map,M=new Map,y=new Map;for(const e of i){const s=await o(e,"utf8"),r=t(s).map(o=>o.toJSON()).filter(o=>null!==o);for(const o of r)collectWorkspaceDocument(o,e,{runtimeDocs:c,agents:f,models:u,tools:w,memories:v,workflows:M,evaluations:y})}for(const o of await s(a))w.has(o.id)||w.set(o.id,o);for(const o of await r(a))d.set(o.id,o);const W=c.at(-1)??{defaultAgentId:"orchestra"};m({workflows:M,agents:f,tools:w,skills:d}),p({evaluations:y,agents:f,tools:w,workflows:M}),function validateAgentRouting(o,e){for(const t of o.routes??[])if(!e.has(t.agentId))throw new Error(`Runtime routing route ${t.id} references unknown agent ${t.agentId}`)}(W,f),function validateWorkflowRouting(o,e){const t=o.workflowRouting;if(t){if(t.defaultWorkflowId&&!e.has(t.defaultWorkflowId))throw new Error(`Runtime workflowRouting.defaultWorkflowId references unknown workflow ${t.defaultWorkflowId}`);for(const o of t.routes??[])if(!e.has(o.workflowId))throw new Error(`Runtime workflowRouting route ${o.id} references unknown workflow ${o.workflowId}`)}}(W,M);const h={root:a,runtime:W,agents:f,models:u,tools:w,skills:d,memories:v,workflows:M,evaluations:y},R=g(h);return k(R),{...h,...R.length>0?{diagnostics:R}:{}}}function collectWorkspaceDocument(o,e,t){if("string"==typeof o.kind)switch(o.kind){case"Runtime":return void t.runtimeDocs.push(f(o));case"Agent":return collectOne(t.agents,a(o,e));case"Model":return collectOne(t.models,i(o));case"Models":return function collectModelSpecs(o,e){if(Array.isArray(o.spec))for(const t of o.spec)if("object"==typeof t&&null!==t&&!Array.isArray(t)){const o=c(t);e.set(o.id,o)}}(o,t.models);case"Tool":return collectOne(t.tools,u(o,e));case"Memory":return collectOne(t.memories,l(o));case"Workflow":return collectOne(t.workflows,w(o,e));case"Evaluation":return collectOne(t.evaluations,d(o,e));default:return}}function collectOne(o,e){o.set(e.id,e)}