pi-harness-runtime 1.1.451 → 1.1.452

package/packages/two-phase-review/src/state.ts

@@ -29,7 +29,8 @@ const PHASE_ORDER: Record<TwoPhasePhase, number> = {
 	review_pending: 1,
 	review: 2,
 	review_fixes: 3,
-	complete: 4,
+	test_pending: 4,
+	complete: 5,
 };
 
 /** Makes the name of a temporary state file unique. */
@@ -66,7 +67,14 @@ function listDir(dir: string): string {
 function isPhase(value: unknown): value is TwoPhasePhase {
 	return (
 		typeof value === "string" &&
-		["implementation", "review_pending", "review", "review_fixes", "complete"].includes(value)
+		[
+			"implementation",
+			"review_pending",
+			"review",
+			"review_fixes",
+			"test_pending",
+			"complete",
+		].includes(value)
 	);
 }
 
@@ -257,6 +265,18 @@ export function transitionToReviewFixes(state: TwoPhaseState): TwoPhaseState {
 	return moveToPhase(state, "review_fixes");
 }
 
+/**
+ * The review is done: the validation of the task is now the work.
+ *
+ * The phase only moves when the review really ran over every chunk: a test
+ * phase before the review would hide the work that is still open.
+ */
+export function transitionToTestPending(state: TwoPhaseState): TwoPhaseState {
+	if (state.phase !== "review" && state.phase !== "review_fixes") return state;
+	if (!allChunksDone(state)) return state;
+	return moveToPhase(state, "test_pending");
+}
+
 /** Every chunk is finished. */
 export function allChunksDone(state: TwoPhaseState): boolean {
 	return state.chunks.length > 0 && state.chunks.every((chunk) => chunk.status === "done");
@@ -274,6 +294,22 @@ export function markComplete(state: TwoPhaseState): TwoPhaseState {
 	return moveToPhase(state, "complete");
 }
 
+/**
+ * The task changed no file, so there is nothing to review.
+ *
+ * This is the one case in which the implementation phase completes the task
+ * on its own: a review without a diff has nothing to cover. A scope with
+ * files never completes here - that scope needs the review.
+ */
+export function completeWithoutReview(
+	state: TwoPhaseState,
+	scopeFiles: string[] = [],
+): TwoPhaseState {
+	if (state.phase !== "implementation") return state;
+	if (scopeFiles.length > 0) return state;
+	return moveToPhase(state, "complete");
+}
+
 /** Is a review queued? */
 export function hasQueuedReview(state: TwoPhaseState): boolean {
 	return state.signals.length > 0;

package/packages/two-phase-review/src/types.ts

@@ -12,10 +12,11 @@ export type TwoPhasePhase =
 	| "review_pending"
 	| "review"
 	| "review_fixes"
+	| "test_pending"
 	| "complete";
 
 /** What a todo item represents. */
-export type TodoKind = "implement" | "review" | "review_fix" | "other";
+export type TodoKind = "implement" | "review" | "review_fix" | "test" | "other";
 
 /** A "review now" signal that arrived while the implementation was running. */
 export interface ReviewSignal {

package/packages/two-phase-review/src/validation.ts

@@ -0,0 +1,86 @@
+/**
+ * The validation gate of the workflow.
+ *
+ * The workflow does not know how the tests are run: the task author declares
+ * the tiers in the metadata of the todo item and the runs are recorded in a
+ * ledger. This module only reads the *result* of that ledger, so the phase
+ * gate and the ledger can never disagree about whether a tier passed.
+ *
+ * The shape of the state is mirrored here on purpose: the workflow must not
+ * depend on the judge package, so the caller adapts its ledger to this
+ * shape.
+ */
+
+/** The part of a todo item the gate needs. */
+export interface GateTodo {
+	id?: number;
+	subject?: string;
+	status?: string;
+	metadata?: Record<string, unknown>;
+}
+
+/**
+ * The validation state of the test items of a session.
+ */
+export interface ValidationGateState {
+	/** Could the validation state be read at all? */
+	available: boolean;
+	/** Why the state could not be read. */
+	reason?: string;
+	/** Test items whose required final tier has not passed yet. */
+	pendingFinalTier: number;
+	/** Test items whose final tier failed, timed out or errored. */
+	failedFinalTier: number;
+	/** Test items whose plan requires a final tier but declares none. */
+	incompletePlan: number;
+}
+
+/** Reads the validation state of the test items of a session. */
+export type ValidationGateReader = (
+	todos: GateTodo[],
+) => Promise<ValidationGateState>;
+
+/** A gate state that says nothing. */
+export function emptyValidationGateState(): ValidationGateState {
+	return {
+		available: false,
+		pendingFinalTier: 0,
+		failedFinalTier: 0,
+		incompletePlan: 0,
+	};
+}
+
+/**
+ * Is the validation of the test items done?
+ *
+ * A gate that could not be read is not satisfied: the workflow must never
+ * report a test as passed because it does not know how it went.
+ */
+export function validationGateSatisfied(state: ValidationGateState): boolean {
+	if (!state.available) return false;
+	return (
+		state.pendingFinalTier === 0 &&
+		state.failedFinalTier === 0 &&
+		state.incompletePlan === 0
+	);
+}
+
+/**
+ * A short description of the gate, for the status line and the instruction.
+ */
+export function describeValidationGate(state: ValidationGateState): string {
+	if (!state.available) {
+		return `the validation state is not readable (${state.reason ?? "unknown reason"})`;
+	}
+	const parts: string[] = [];
+	if (state.pendingFinalTier > 0) {
+		parts.push(`${state.pendingFinalTier} final tier(s) pending`);
+	}
+	if (state.failedFinalTier > 0) {
+		parts.push(`${state.failedFinalTier} final tier(s) failed`);
+	}
+	if (state.incompletePlan > 0) {
+		parts.push(`${state.incompletePlan} plan(s) without a final tier`);
+	}
+	return parts.length > 0 ? parts.join(", ") : "every required tier passed";
+}

package/packages/types/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-harness/types",
-  "version": "1.1.451",
+  "version": "1.1.452",
   "description": "Shared TypeScript types for pi-harness-runtime packages",
   "type": "module",
   "main": "./dist/index.js",

package/packages/webmcp-tester/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-harness/webmcp-tester",
-  "version": "1.1.451",
+  "version": "1.1.452",
   "description": "WebMCP-powered E2E testing for pi-harness - natural language test scenarios",
   "type": "module",
   "main": "./dist/index.js",

package/packages/wiki/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-harness/wiki",
-  "version": "1.1.451",
+  "version": "1.1.452",
   "description": "LLM Wiki integration for pi-harness-runtime",
   "type": "module",
   "main": "./dist/index.js",

package/packages/write-review/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@moocoding/write-review",
-  "version": "1.1.451",
+  "version": "1.1.452",
   "description": "Two-agent write-review loop with blackboard coordination",
   "type": "module",
   "main": "src/index.ts",