@opencode_weave/weave 0.7.1 → 0.7.4-preview.1

package/dist/features/evals/schema.d.ts

@@ -1,9 +1,9 @@
 import { z } from "zod";
 export declare const EvalPhaseSchema: z.ZodEnum<{
-    phase1: "phase1";
-    phase2: "phase2";
-    phase3: "phase3";
-    phase4: "phase4";
+    prompt: "prompt";
+    experimental: "experimental";
+    routing: "routing";
+    trajectory: "trajectory";
 }>;
 export declare const BuiltinAgentPromptVariantSchema: z.ZodObject<{
     disabledAgents: z.ZodOptional<z.ZodArray<z.ZodString>>;
@@ -14,6 +14,7 @@ export declare const BuiltinAgentPromptTargetSchema: z.ZodObject<{
         pattern: "pattern";
         loom: "loom";
         tapestry: "tapestry";
+        shuttle: "shuttle";
         thread: "thread";
         spindle: "spindle";
         weft: "weft";
@@ -43,6 +44,7 @@ export declare const EvalTargetSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
         pattern: "pattern";
         loom: "loom";
         tapestry: "tapestry";
+        shuttle: "shuttle";
         thread: "thread";
         spindle: "spindle";
         weft: "weft";
@@ -144,6 +146,39 @@ export declare const TrajectoryAssertionEvaluatorSchema: z.ZodObject<{
     weight: z.ZodOptional<z.ZodNumber>;
     kind: z.ZodLiteral<"trajectory-assertion">;
     assertionRef: z.ZodOptional<z.ZodString>;
+    expectedSequence: z.ZodOptional<z.ZodArray<z.ZodString>>;
+    requiredAgents: z.ZodOptional<z.ZodArray<z.ZodString>>;
+    forbiddenAgents: z.ZodOptional<z.ZodArray<z.ZodString>>;
+    minTurns: z.ZodOptional<z.ZodNumber>;
+    maxTurns: z.ZodOptional<z.ZodNumber>;
+}, z.core.$strip>;
+export declare const TrajectoryTurnSchema: z.ZodObject<{
+    turn: z.ZodNumber;
+    role: z.ZodEnum<{
+        user: "user";
+        assistant: "assistant";
+    }>;
+    agent: z.ZodOptional<z.ZodString>;
+    content: z.ZodString;
+    mockResponse: z.ZodOptional<z.ZodString>;
+    expectedDelegation: z.ZodOptional<z.ZodString>;
+}, z.core.$strip>;
+export declare const TrajectoryScenarioSchema: z.ZodObject<{
+    id: z.ZodString;
+    title: z.ZodString;
+    description: z.ZodOptional<z.ZodString>;
+    agents: z.ZodArray<z.ZodString>;
+    turns: z.ZodArray<z.ZodObject<{
+        turn: z.ZodNumber;
+        role: z.ZodEnum<{
+            user: "user";
+            assistant: "assistant";
+        }>;
+        agent: z.ZodOptional<z.ZodString>;
+        content: z.ZodString;
+        mockResponse: z.ZodOptional<z.ZodString>;
+        expectedDelegation: z.ZodOptional<z.ZodString>;
+    }, z.core.$strip>>;
 }, z.core.$strip>;
 export declare const EvaluatorSpecSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
     weight: z.ZodOptional<z.ZodNumber>;
@@ -192,15 +227,21 @@ export declare const EvaluatorSpecSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
     weight: z.ZodOptional<z.ZodNumber>;
     kind: z.ZodLiteral<"trajectory-assertion">;
     assertionRef: z.ZodOptional<z.ZodString>;
+    expectedSequence: z.ZodOptional<z.ZodArray<z.ZodString>>;
+    requiredAgents: z.ZodOptional<z.ZodArray<z.ZodString>>;
+    forbiddenAgents: z.ZodOptional<z.ZodArray<z.ZodString>>;
+    minTurns: z.ZodOptional<z.ZodNumber>;
+    maxTurns: z.ZodOptional<z.ZodNumber>;
 }, z.core.$strip>], "kind">;
 export declare const EvalCaseSchema: z.ZodObject<{
     id: z.ZodString;
     title: z.ZodString;
+    description: z.ZodOptional<z.ZodString>;
     phase: z.ZodEnum<{
-        phase1: "phase1";
-        phase2: "phase2";
-        phase3: "phase3";
-        phase4: "phase4";
+        prompt: "prompt";
+        experimental: "experimental";
+        routing: "routing";
+        trajectory: "trajectory";
     }>;
     target: z.ZodDiscriminatedUnion<[z.ZodObject<{
         kind: z.ZodLiteral<"builtin-agent-prompt">;
@@ -208,6 +249,7 @@ export declare const EvalCaseSchema: z.ZodObject<{
             pattern: "pattern";
             loom: "loom";
             tapestry: "tapestry";
+            shuttle: "shuttle";
             thread: "thread";
             spindle: "spindle";
             weft: "weft";
@@ -286,6 +328,11 @@ export declare const EvalCaseSchema: z.ZodObject<{
         weight: z.ZodOptional<z.ZodNumber>;
         kind: z.ZodLiteral<"trajectory-assertion">;
         assertionRef: z.ZodOptional<z.ZodString>;
+        expectedSequence: z.ZodOptional<z.ZodArray<z.ZodString>>;
+        requiredAgents: z.ZodOptional<z.ZodArray<z.ZodString>>;
+        forbiddenAgents: z.ZodOptional<z.ZodArray<z.ZodString>>;
+        minTurns: z.ZodOptional<z.ZodNumber>;
+        maxTurns: z.ZodOptional<z.ZodNumber>;
     }, z.core.$strip>], "kind">>;
     tags: z.ZodOptional<z.ZodArray<z.ZodString>>;
     notes: z.ZodOptional<z.ZodString>;
@@ -294,10 +341,10 @@ export declare const EvalSuiteManifestSchema: z.ZodObject<{
     id: z.ZodString;
     title: z.ZodString;
     phase: z.ZodEnum<{
-        phase1: "phase1";
-        phase2: "phase2";
-        phase3: "phase3";
-        phase4: "phase4";
+        prompt: "prompt";
+        experimental: "experimental";
+        routing: "routing";
+        trajectory: "trajectory";
     }>;
     caseFiles: z.ZodArray<z.ZodString>;
     tags: z.ZodOptional<z.ZodArray<z.ZodString>>;
@@ -342,6 +389,7 @@ export declare const EvalArtifactsSchema: z.ZodObject<{
 }, z.core.$strip>;
 export declare const EvalCaseResultSchema: z.ZodObject<{
     caseId: z.ZodString;
+    description: z.ZodOptional<z.ZodString>;
     status: z.ZodEnum<{
         error: "error";
         failed: "failed";
@@ -406,10 +454,10 @@ export declare const EvalRunResultSchema: z.ZodObject<{
     finishedAt: z.ZodString;
     suiteId: z.ZodString;
     phase: z.ZodEnum<{
-        phase1: "phase1";
-        phase2: "phase2";
-        phase3: "phase3";
-        phase4: "phase4";
+        prompt: "prompt";
+        experimental: "experimental";
+        routing: "routing";
+        trajectory: "trajectory";
     }>;
     summary: z.ZodObject<{
         totalCases: z.ZodNumber;
@@ -422,6 +470,7 @@ export declare const EvalRunResultSchema: z.ZodObject<{
     }, z.core.$strip>;
     caseResults: z.ZodArray<z.ZodObject<{
         caseId: z.ZodString;
+        description: z.ZodOptional<z.ZodString>;
         status: z.ZodEnum<{
             error: "error";
             failed: "failed";

package/dist/features/evals/storage.d.ts

@@ -5,3 +5,5 @@ export declare const EVAL_LATEST_FILE = ".weave/evals/latest.json";
 export declare function ensureEvalStorageDir(directory: string): string;
 export declare function getDefaultEvalRunPath(directory: string, runId: string): string;
 export declare function writeEvalRunResult(directory: string, result: EvalRunResult, outputPath?: string): string;
+export declare function getDefaultJsonlPath(directory: string, suiteId: string): string;
+export declare function appendEvalRunJsonl(directory: string, result: EvalRunResult, jsonlPath?: string): string;

package/dist/features/evals/types.d.ts

@@ -1,5 +1,5 @@
 import type { WeaveAgentName } from "../../agents/types";
-export declare const EVAL_PHASES: readonly ["phase1", "phase2", "phase3", "phase4"];
+export declare const EVAL_PHASES: readonly ["prompt", "routing", "trajectory", "experimental"];
 export type EvalPhase = (typeof EVAL_PHASES)[number];
 export declare const EVAL_TARGET_KINDS: readonly ["builtin-agent-prompt", "custom-agent-prompt", "single-turn-agent", "trajectory-agent"];
 export type EvalTargetKind = (typeof EVAL_TARGET_KINDS)[number];
@@ -7,7 +7,7 @@ export declare const EXECUTOR_KINDS: readonly ["prompt-render", "model-response"
 export type ExecutorKind = (typeof EXECUTOR_KINDS)[number];
 export declare const EVALUATOR_KINDS: readonly ["contains-all", "contains-any", "excludes-all", "section-contains-all", "ordered-contains", "xml-sections-present", "tool-policy", "min-length", "llm-judge", "baseline-diff", "trajectory-assertion"];
 export type EvaluatorKind = (typeof EVALUATOR_KINDS)[number];
-export type BuiltinEvalAgentName = Exclude<WeaveAgentName, "shuttle">;
+export type BuiltinEvalAgentName = WeaveAgentName;
 export interface BuiltinAgentPromptVariant {
     disabledAgents?: string[];
 }
@@ -94,6 +94,11 @@ export interface BaselineDiffEvaluator extends WeightedEvaluatorSpec {
 export interface TrajectoryAssertionEvaluator extends WeightedEvaluatorSpec {
     kind: "trajectory-assertion";
     assertionRef?: string;
+    expectedSequence?: string[];
+    requiredAgents?: string[];
+    forbiddenAgents?: string[];
+    minTurns?: number;
+    maxTurns?: number;
 }
 export type EvaluatorSpec = ContainsAllEvaluator | ContainsAnyEvaluator | ExcludesAllEvaluator | SectionContainsAllEvaluator | OrderedContainsEvaluator | XmlSectionsPresentEvaluator | ToolPolicyEvaluator | MinLengthEvaluator | LlmJudgeEvaluator | BaselineDiffEvaluator | TrajectoryAssertionEvaluator;
 export interface EvalSuiteManifest {
@@ -106,6 +111,7 @@ export interface EvalSuiteManifest {
 export interface EvalCase {
     id: string;
     title: string;
+    description?: string;
     phase: EvalPhase;
     target: EvalTarget;
     executor: ExecutorSpec;
@@ -145,6 +151,7 @@ export interface AssertionResult {
 }
 export interface EvalCaseResult {
     caseId: string;
+    description?: string;
     status: "passed" | "failed" | "error";
     score: number;
     normalizedScore: number;
@@ -180,6 +187,7 @@ export interface ExecutionContext {
     mode: "local" | "ci" | "hosted";
     directory: string;
     outputPath?: string;
+    modelOverride?: string;
 }
 export interface RunnerFilters {
     caseIds?: string[];
@@ -192,6 +200,7 @@ export interface RunEvalSuiteOptions {
     filters?: RunnerFilters;
     outputPath?: string;
     mode?: ExecutionContext["mode"];
+    modelOverride?: string;
 }
 export interface EvalLoadErrorContext {
     filePath: string;
@@ -221,3 +230,35 @@ export interface BaselineComparison {
     regressions: string[];
     informational: string[];
 }
+export interface TrajectoryTurn {
+    turn: number;
+    role: "user" | "assistant";
+    agent?: string;
+    content: string;
+    mockResponse?: string;
+    expectedDelegation?: string;
+}
+export interface TrajectoryScenario {
+    id: string;
+    title: string;
+    description?: string;
+    agents: string[];
+    turns: TrajectoryTurn[];
+}
+export interface TrajectoryTurnResult {
+    turn: number;
+    agent: string;
+    role: "user" | "assistant";
+    response: string;
+    expectedDelegation?: string;
+    observedDelegation?: string | null;
+    durationMs: number;
+}
+export interface TrajectoryTrace {
+    scenarioId: string;
+    turns: TrajectoryTurnResult[];
+    delegationSequence: string[];
+    totalTurns: number;
+    completedTurns: number;
+}
+export declare function isTrajectoryTrace(trace: unknown): trace is TrajectoryTrace;

package/dist/features/skill-loader/loader.d.ts

@@ -3,5 +3,7 @@ export interface LoadSkillsOptions {
     serverUrl: string | URL;
     directory?: string;
     disabledSkills?: string[];
+    /** Additional directories to scan for skills (from config `skill_directories`) */
+    customDirs?: string[];
 }
 export declare function loadSkills(options: LoadSkillsOptions): Promise<SkillDiscoveryResult>;

package/dist/features/workflow/context.d.ts

@@ -11,6 +11,7 @@ export declare function resolveTemplate(template: string, instance: WorkflowInst
 export declare function buildContextHeader(instance: WorkflowInstance, definition: WorkflowDefinition): string;
 /**
  * Build the full context-threaded prompt for a step.
- * Combines: (1) workflow context header, (2) resolved step prompt.
+ * Combines: (1) workflow context header, (2) delegation instruction if step
+ * targets a non-loom agent, (3) resolved step prompt.
  */
 export declare function composeStepPrompt(stepDef: WorkflowStepDefinition, instance: WorkflowInstance, definition: WorkflowDefinition): string;

package/dist/features/workflow/discovery.d.ts

@@ -13,7 +13,10 @@ export interface DiscoveredWorkflow {
  */
 export declare function loadWorkflowDefinition(filePath: string): WorkflowDefinition | null;
 /**
- * Discover all valid workflow definitions from project and user directories.
- * Project workflows override user workflows with the same name.
+ * Discover all valid workflow definitions from project, custom, and user directories.
+ * Override precedence (highest wins): project > custom > user.
+ *
+ * @param directory - Project root directory (used to resolve relative custom paths)
+ * @param customDirs - Optional extra directories to scan (from config `workflows.directories`)
  */
-export declare function discoverWorkflows(directory: string): DiscoveredWorkflow[];
+export declare function discoverWorkflows(directory: string, customDirs?: string[]): DiscoveredWorkflow[];

package/dist/features/workflow/hook.d.ts

@@ -30,6 +30,7 @@ export declare function handleRunWorkflow(input: {
     promptText: string;
     sessionId: string;
     directory: string;
+    workflowDirs?: string[];
 }): WorkflowHookResult;
 /**
  * Check workflow continuation on session.idle.
@@ -41,6 +42,7 @@ export declare function checkWorkflowContinuation(input: {
     directory: string;
     lastAssistantMessage?: string;
     lastUserMessage?: string;
+    workflowDirs?: string[];
 }): {
     continuationPrompt: string | null;
     switchAgent: string | null;

package/dist/hooks/compaction-todo-preserver.d.ts

@@ -0,0 +1,20 @@
+/**
+ * compaction-todo-preserver hook
+ *
+ * Snapshots todos before compaction, restores them if wiped by compaction.
+ * Defense-in-depth against OpenCode's context compaction clearing the todo list.
+ */
+import type { PluginContext } from "../plugin/types";
+export type TodoSnapshot = {
+    content: string;
+    status: string;
+    priority: string;
+};
+export declare function createCompactionTodoPreserver(client: PluginContext["client"]): {
+    capture: (sessionID: string) => Promise<void>;
+    handleEvent: (event: {
+        type: string;
+        properties?: unknown;
+    }) => Promise<void>;
+    getSnapshot: (sessionID: string) => TodoSnapshot[] | undefined;
+};

package/dist/hooks/create-hooks.d.ts

@@ -5,6 +5,7 @@ import { shouldApplyVariant, markApplied, markSessionCreated, clearSession } fro
 import { processMessageForKeywords } from "./keyword-detector";
 import { checkPatternWrite } from "./pattern-md-only";
 import { buildVerificationReminder } from "./verification-reminder";
+import { applyTodoDescriptionOverride } from "./todo-description-override";
 export type CreatedHooks = ReturnType<typeof createHooks>;
 export declare function createHooks(args: {
     pluginConfig: WeaveConfig;
@@ -36,5 +37,8 @@ export declare function createHooks(args: {
     }) | null;
     workflowCommand: ((message: string) => import("../features/workflow").WorkflowCommandResult) | null;
     verificationReminder: typeof buildVerificationReminder | null;
+    todoDescriptionOverride: typeof applyTodoDescriptionOverride | null;
+    compactionTodoPreserverEnabled: boolean;
+    todoContinuationEnforcerEnabled: boolean;
     analyticsEnabled: boolean;
 };

package/dist/hooks/index.d.ts

@@ -12,3 +12,9 @@ export { buildVerificationReminder } from "./verification-reminder";
 export type { VerificationInput, VerificationResult } from "./verification-reminder";
 export { setContextLimit, updateUsage, getState, clearSession as clearTokenSession, clear as clearAllTokenState, } from "./session-token-state";
 export type { SessionTokenEntry } from "./session-token-state";
+export { applyTodoDescriptionOverride, TODOWRITE_DESCRIPTION } from "./todo-description-override";
+export { createCompactionTodoPreserver } from "./compaction-todo-preserver";
+export type { TodoSnapshot } from "./compaction-todo-preserver";
+export { createTodoContinuationEnforcer, FINALIZE_TODOS_MARKER } from "./todo-continuation-enforcer";
+export { resolveTodoWriter } from "./todo-writer";
+export type { TodoItem, TodoWriter } from "./todo-writer";

package/dist/hooks/todo-continuation-enforcer.d.ts

@@ -0,0 +1,25 @@
+/**
+ * todo-continuation-enforcer hook
+ *
+ * Ensures in_progress todos are finalized when a session goes idle.
+ * Extracted from the inline finalization logic in plugin-interface.ts.
+ *
+ * Primary path (zero-cost): If opencode/session/todo is available, directly
+ * mutates the todo list to mark in_progress items as completed.
+ *
+ * Fallback path (1 LLM turn): If unavailable, injects a prompt asking the LLM
+ * to finalize the todos.
+ */
+import type { PluginContext } from "../plugin/types";
+import { type TodoWriter } from "./todo-writer";
+export declare const FINALIZE_TODOS_MARKER = "<!-- weave:finalize-todos -->";
+export declare function createTodoContinuationEnforcer(client: PluginContext["client"], options?: {
+    /** Inject a mock todo writer for testing (bypasses dynamic import) */
+    todoWriterOverride?: TodoWriter | null;
+}): {
+    checkAndFinalize: (sessionID: string) => Promise<void>;
+    markFinalized: (sessionID: string) => void;
+    isFinalized: (sessionID: string) => boolean;
+    clearFinalized: (sessionID: string) => void;
+    clearSession: (sessionID: string) => void;
+};

package/dist/hooks/todo-description-override.d.ts

@@ -0,0 +1,18 @@
+/**
+ * todo-description-override hook
+ *
+ * Overrides the TodoWrite tool description with stronger language emphasizing
+ * that it is a destructive full-array replacement and items must NEVER be dropped.
+ */
+export declare const TODOWRITE_DESCRIPTION = "Manages the sidebar todo list. CRITICAL: This tool performs a FULL ARRAY REPLACEMENT \u2014 every call completely DELETES all existing todos and replaces them with whatever you send. NEVER drop existing items. ALWAYS include ALL current todos in EVERY call. If unsure what todos currently exist, call todoread BEFORE calling this tool. Rules: max 35 chars per item, encode WHERE + WHAT (e.g. \"src/foo.ts: add error handler\"). Status values: \"pending\", \"in_progress\", \"completed\", \"cancelled\". Priority values: \"high\", \"medium\", \"low\".";
+/**
+ * Applies the enhanced TodoWrite description override.
+ * Mutates `output.description` when `input.toolID === "todowrite"`.
+ * Pure function — no side effects, no async, no state.
+ */
+export declare function applyTodoDescriptionOverride(input: {
+    toolID: string;
+}, output: {
+    description: string;
+    parameters?: unknown;
+}): void;

package/dist/hooks/todo-writer.d.ts

@@ -0,0 +1,17 @@
+/**
+ * Shared todo-writer resolution utility.
+ *
+ * Dynamically imports opencode/session/todo to get Todo.update().
+ * Uses a variable for the module specifier to prevent bundler inlining.
+ * Returns null if the module is unavailable (non-fatal).
+ */
+export type TodoItem = {
+    content: string;
+    status: string;
+    priority?: string;
+};
+export type TodoWriter = (input: {
+    sessionID: string;
+    todos: TodoItem[];
+}) => void;
+export declare function resolveTodoWriter(): Promise<TodoWriter | null>;