@vitest-agent/sdk 1.0.1 → 1.2.0

package/testing.js

@@ -1,7 +1,17 @@
+import { AgentNotFoundError, RegistrationConflictError } from "./errors/AgentErrors.js";
 import { DataStoreError } from "./errors/DataStoreError.js";
 import { BehaviorNotFoundError, GoalNotFoundError, IllegalStatusTransitionError, TddTaskAlreadyEndedError, TddTaskNotFoundError } from "./errors/TddErrors.js";
 import { DataReader } from "./services/DataReader.js";
+import { AgentId, ChatId } from "./schemas/Identity.js";
+import { Agent, IdempotencyHit } from "./schemas/Agent.js";
 import { DataStore } from "./services/DataStore.js";
+import { CoverageReport, FileCoverageReport } from "./schemas/Coverage.js";
+import { AgentReport } from "./schemas/AgentReport.js";
+import { CoverageBaselines } from "./schemas/Baselines.js";
+import { CacheManifest } from "./schemas/CacheManifest.js";
+import { HistoryRecord } from "./schemas/History.js";
+import { BehaviorDetail, BehaviorRow, BehaviorStatus, GoalDetail, GoalRow, GoalStatus } from "./schemas/Tdd.js";
+import { TrendEntry, TrendRecord } from "./schemas/Trends.js";
 import { DataStoreTestLayer, makeTestLayer } from "./testing/layers.js";
 import { Effect, Layer } from "effect";
 
@@ -240,4 +250,4 @@ function withTddTask(filename) {
 }
 
 //#endregion
-export { BehaviorNotFoundError, DataReader, DataStore, DataStoreError, DataStoreTestLayer, GoalNotFoundError, IllegalStatusTransitionError, TddTaskAlreadyEndedError, TddTaskNotFoundError, empty, flaky, makeTestLayer, singlePassingRun, withFailures, withTddTask };
+export { Agent, AgentId, AgentNotFoundError, AgentReport, BehaviorDetail, BehaviorNotFoundError, BehaviorRow, BehaviorStatus, CacheManifest, ChatId, CoverageBaselines, CoverageReport, DataReader, DataStore, DataStoreError, DataStoreTestLayer, FileCoverageReport, GoalDetail, GoalNotFoundError, GoalRow, GoalStatus, HistoryRecord, IdempotencyHit, IllegalStatusTransitionError, RegistrationConflictError, TddTaskAlreadyEndedError, TddTaskNotFoundError, TrendEntry, TrendRecord, empty, flaky, makeTestLayer, singlePassingRun, withFailures, withTddTask };

package/utils/console-leaks.js

@@ -0,0 +1,77 @@
+//#region src/utils/console-leaks.ts
+const SAMPLE_MAX_CHARS = 160;
+const MAX_FILES = 25;
+const MAX_TESTS_PER_FILE = 10;
+function truncateSample(content) {
+	const trimmed = (content.split("\n", 1)[0] ?? "").trim();
+	return trimmed.length > SAMPLE_MAX_CHARS ? `${trimmed.slice(0, SAMPLE_MAX_CHARS)}…` : trimmed;
+}
+/**
+* Aggregate raw console-leak entries into a ConsoleLeaks signal:
+* bucket by file, split stdout/stderr, collect attributable test names,
+* capture one truncated sample per file, sort by total writes descending,
+* and cap the file list. Returns `undefined` when there are no entries so a
+* clean run attaches nothing.
+* @public
+*/
+function buildConsoleLeaks(entries) {
+	if (entries.length === 0) return void 0;
+	const byFile = /* @__PURE__ */ new Map();
+	for (const e of entries) {
+		let acc = byFile.get(e.file);
+		if (acc === void 0) {
+			acc = {
+				stdout: 0,
+				stderr: 0,
+				tests: /* @__PURE__ */ new Set()
+			};
+			byFile.set(e.file, acc);
+		}
+		if (e.type === "stdout") acc.stdout++;
+		else acc.stderr++;
+		if (e.test !== void 0 && e.test !== "") acc.tests.add(e.test);
+		if (acc.sample === void 0) {
+			const candidate = truncateSample(e.content);
+			if (candidate !== "") acc.sample = candidate;
+		}
+	}
+	const files = [];
+	for (const [file, acc] of byFile) files.push({
+		file,
+		stdout: acc.stdout,
+		stderr: acc.stderr,
+		...acc.tests.size > 0 ? { tests: [...acc.tests].slice(0, MAX_TESTS_PER_FILE) } : {},
+		...acc.sample !== void 0 ? { sample: acc.sample } : {}
+	});
+	files.sort((a, b) => b.stdout + b.stderr - (a.stdout + a.stderr));
+	const truncated = files.length > MAX_FILES;
+	return {
+		total: entries.length,
+		byFile: truncated ? files.slice(0, MAX_FILES) : files,
+		...truncated ? { truncated: true } : {}
+	};
+}
+/**
+* Walk a Vitest `File[]` task tree (from `vitest.state.getFiles()`) into flat
+* {@link ConsoleLeakEntry} values. Each task `log` becomes one entry attributed
+* to its enclosing file and, when the log sits on a test task, that test's name.
+* @public
+*/
+function collectConsoleLeakEntries(files) {
+	const entries = [];
+	const visit = (task, file, test) => {
+		const currentTest = task.type === "test" ? task.fullTestName ?? task.name ?? test : test;
+		for (const log of task.logs ?? []) entries.push({
+			file,
+			...currentTest !== void 0 ? { test: currentTest } : {},
+			type: log.type,
+			content: log.content
+		});
+		for (const child of task.tasks ?? []) visit(child, file, currentTest);
+	};
+	for (const file of files) visit(file, file.name ?? "(unknown)", void 0);
+	return entries;
+}
+
+//#endregion
+export { buildConsoleLeaks, collectConsoleLeakEntries };

package/utils/validate-phase-transition.js

@@ -14,6 +14,10 @@ const requiredArtifactForTransition = (from, to) => {
 		kind: "test_failed_run",
 		humanHint: "Run the failing test via run_tests, then record the test_failed_run artifact before requesting red→green."
 	};
+	if (from === "red.triangulate" && to === "green") return {
+		kind: "test_failed_run",
+		humanHint: "A triangulation batch must still have produced at least one real failing run; record the batch's test_failed_run before requesting red.triangulate→green."
+	};
 	if (from === "green" && to === "refactor") return {
 		kind: "test_passed_run",
 		humanHint: "Run the test via run_tests and confirm it passes; record test_passed_run before requesting green→refactor."
@@ -24,6 +28,21 @@ const requiredArtifactForTransition = (from, to) => {
 	};
 	return null;
 };
+/**
+* Whether D2 binding rule 2 (behavior-match) applies to a given transition.
+*
+* Only the transitions whose cited evidence must belong to the behavior being
+* transitioned enforce behavior-match: `red→green` (the failing test for this
+* behavior) and `green→refactor` (its passing test). It deliberately excludes
+* `red.triangulate→green` (the cited failing run belongs to an earlier batch
+* behavior) and `refactor→red` (the required `test_passed_run` is the
+* just-finished behavior's, never the new target). Exported so the MCP
+* `tdd_phase_transition_request` tool scopes artifact auto-resolution by
+* behavior on exactly the same transitions the validator will accept —
+* keeping the two in lockstep (issue #115).
+* @public
+*/
+const transitionEnforcesBehaviorMatch = (from, to) => from === "red" && to === "green" || from === "green" && to === "refactor";
 /** @public */
 const validatePhaseTransition = (ctx) => {
 	if (ctx.requested_phase === "green" && ctx.current_phase !== "red" && ctx.current_phase !== "red.triangulate" && ctx.current_phase !== "green.fake-it") return {
@@ -36,6 +55,7 @@ const validatePhaseTransition = (ctx) => {
 			humanHint: `Cannot transition from '${ctx.current_phase}' directly to 'green'. The red phase must be entered explicitly first (${ctx.current_phase}→red), then a failing test written and run, then red→green requested with a test_failed_run artifact.`
 		}
 	};
+	const isTriangulateGreen = ctx.current_phase === "red.triangulate" && ctx.requested_phase === "green";
 	const expected = requiredArtifactForTransition(ctx.current_phase, ctx.requested_phase);
 	if (expected === null) return {
 		accepted: true,
@@ -62,7 +82,7 @@ const validatePhaseTransition = (ctx) => {
 		}
 	};
 	if (expected.kind === "test_failed_run") {
-		if (ctx.cited_artifact.test_case_created_turn_at !== null && ctx.cited_artifact.test_case_created_turn_at < ctx.phase_started_at) return {
+		if (!isTriangulateGreen && ctx.cited_artifact.test_case_created_turn_at !== null && ctx.cited_artifact.test_case_created_turn_at < ctx.phase_started_at) return {
 			accepted: false,
 			phase: ctx.current_phase,
 			denialReason: "evidence_not_in_phase_window",
@@ -83,7 +103,7 @@ const validatePhaseTransition = (ctx) => {
 			}
 		};
 	}
-	if (ctx.requested_behavior_id !== null && ctx.cited_artifact.behavior_id !== ctx.requested_behavior_id) return {
+	if (transitionEnforcesBehaviorMatch(ctx.current_phase, ctx.requested_phase) && ctx.requested_behavior_id !== null && ctx.cited_artifact.behavior_id !== ctx.requested_behavior_id) return {
 		accepted: false,
 		phase: ctx.current_phase,
 		denialReason: "evidence_not_for_behavior",
@@ -110,4 +130,4 @@ const validatePhaseTransition = (ctx) => {
 };
 
 //#endregion
-export { requiredArtifactForTransition, validatePhaseTransition };
+export { requiredArtifactForTransition, transitionEnforcesBehaviorMatch, validatePhaseTransition };