@cat-factory/orchestration 0.30.0 → 0.32.0

package/dist/modules/execution/ExecutionService.js

@@ -7,7 +7,7 @@ import { reviewableArtifactOutput } from './artifact-review.logic.js';
 import { resolveIndividualVendors, } from './individualVendors.logic.js';
 import { assertFound, ConflictError, getErrorMessage, isModelUsable, NotFoundError, parseLocalModelId, resolveModelRef, sameSubtasks, subscriptionOptionFor, ValidationError, } from '@cat-factory/kernel';
 import { DEFAULT_MERGE_PRESET } from '@cat-factory/kernel';
-import { CONFLICTS_AGENT_KIND, MERGER_AGENT_KIND, REQUIREMENTS_REVIEW_AGENT_KIND, CLARITY_REVIEW_AGENT_KIND, REQUIREMENTS_BRAINSTORM_AGENT_KIND, ARCHITECTURE_BRAINSTORM_AGENT_KIND, BUG_INVESTIGATOR_AGENT_KIND, TRACKER_AGENT_KIND, ANALYSIS_AGENT_KIND, TESTER_AGENT_KIND, HUMAN_TEST_AGENT_KIND, HUMAN_REVIEW_AGENT_KIND, BLUEPRINTS_AGENT_KIND, SPEC_WRITER_AGENT_KIND, } from './ci.logic.js';
+import { CONFLICTS_AGENT_KIND, MERGER_AGENT_KIND, REQUIREMENTS_REVIEW_AGENT_KIND, CLARITY_REVIEW_AGENT_KIND, REQUIREMENTS_BRAINSTORM_AGENT_KIND, ARCHITECTURE_BRAINSTORM_AGENT_KIND, BUG_INVESTIGATOR_AGENT_KIND, TRACKER_AGENT_KIND, ANALYSIS_AGENT_KIND, TESTER_AGENT_KIND, UI_TESTER_AGENT_KIND, isTesterKind, HUMAN_TEST_AGENT_KIND, VISUAL_CONFIRM_AGENT_KIND, HUMAN_REVIEW_AGENT_KIND, BLUEPRINTS_AGENT_KIND, SPEC_WRITER_AGENT_KIND, } from './ci.logic.js';
 import { DEFAULT_FOLLOW_UP_MAX_LOOPS, FOLLOW_UP_PRODUCER_KIND, followUpsToSendBack, hasPendingFollowUps, renderFollowUpRework, shouldLoopCoder, } from './followUp.logic.js';
 import { AgentContextBuilder } from './AgentContextBuilder.js';
 import { CompanionController } from './CompanionController.js';
@@ -16,6 +16,7 @@ import { MergeResolver } from './MergeResolver.js';
 import { ReviewGateController } from './ReviewGateController.js';
 import { TesterController } from './TesterController.js';
 import { HumanTestController } from './HumanTestController.js';
+import { VisualConfirmationController } from './VisualConfirmationController.js';
 import { getProvider, recordGateAttempt, registeredGateFactories, registeredStepResolverFactories, requireProvider, } from '@cat-factory/kernel';
 import { isAsyncAgentExecutor } from '@cat-factory/kernel';
 import { isDeployStep } from '@cat-factory/integrations';
@@ -46,7 +47,12 @@ const EXECUTION_FAILURE_HINTS = {
  * the per-poll env projection so the `getByBlock` read never hits the hot path for
  * the many container steps that have no env to show (see attachEnvironmentProjection).
  */
-const ENV_PROJECTION_KINDS = new Set(['deployer', 'tester', 'playwright']);
+const ENV_PROJECTION_KINDS = new Set([
+    'deployer',
+    TESTER_AGENT_KIND,
+    UI_TESTER_AGENT_KIND,
+    'playwright',
+]);
 /**
  * Parse `owner`/`repo` from a GitHub pull-request URL (`https://github.com/o/r/pull/42`).
  * Returns undefined for any URL that doesn't carry both segments. Host-agnostic on
@@ -98,6 +104,8 @@ export class ExecutionService {
     testerController;
     /** Drives the human-testing gate: provision env → park → confirm / fix / pull-main / recreate. */
     humanTestController;
+    /** Drives the visual-confirmation gate: gather screenshots → park → approve / fix / recapture. */
+    visualConfirmationController;
     /** Drives both iterative review gates (requirements + clarity); kind-parameterised. */
     reviewGate;
     /** The requirements subject for {@link reviewGate}. */
@@ -139,7 +147,7 @@ export class ExecutionService {
      * {@link stepResolverFor} and {@link StepCompletionResolver}.
      */
     stepResolverCache;
-    constructor({ workspaceRepository, blockRepository, pipelineRepository, executionRepository, accountRepository, idGenerator, clock, agentExecutor, workRunner, executionEventPublisher, boardService, spendService, documentRepository, taskRepository, requirementReviewRepository, requirementReviewService, kaizenScheduler, clarityReviewRepository, clarityReviewService, brainstormServices, brainstormSessionRepository, fragmentResolver, environmentProvisioning, environmentTeardown, branchUpdater, blueprintReconciler, notificationService, workspaceSettingsService, llmObservability, pullRequestMerger, mergePresetRepository, ticketTrackerProvider, issueWriteback, subscriptionActivationRepository, resolveWorkspaceModelDefault, resolveProviderCapabilities, localTestInfraSupported, resolveRunRepoContext, resolveTesterFallbackDefault, resolveRequireEnvironmentProvider, assertAgentBackendConfigured, runInitiatorScope, }) {
+    constructor({ workspaceRepository, blockRepository, pipelineRepository, executionRepository, accountRepository, idGenerator, clock, agentExecutor, workRunner, executionEventPublisher, boardService, spendService, documentRepository, taskRepository, requirementReviewRepository, requirementReviewService, kaizenScheduler, clarityReviewRepository, clarityReviewService, brainstormServices, brainstormSessionRepository, fragmentResolver, environmentProvisioning, environmentTeardown, branchUpdater, blueprintReconciler, notificationService, binaryArtifactStore, workspaceSettingsService, llmObservability, pullRequestMerger, mergePresetRepository, ticketTrackerProvider, issueWriteback, subscriptionActivationRepository, resolveWorkspaceModelDefault, resolveProviderCapabilities, localTestInfraSupported, resolveRunRepoContext, resolveTesterFallbackDefault, resolveRequireEnvironmentProvider, assertAgentBackendConfigured, runInitiatorScope, }) {
         this.runInitiatorScope = runInitiatorScope ?? ((_initiatedBy, fn) => fn());
         this.workspaceRepository = workspaceRepository;
         this.blockRepository = blockRepository;
@@ -249,6 +257,25 @@ export class ExecutionService {
             emitInstance: (ws, i) => this.emitInstance(ws, i),
             clockNow: () => this.clock.now(),
         });
+        this.visualConfirmationController = new VisualConfirmationController({
+            blockRepository,
+            executionRepository,
+            workRunner,
+            agentExecutor,
+            contextBuilder: this.contextBuilder,
+            notificationService,
+            ...(binaryArtifactStore ? { binaryArtifactStore } : {}),
+            resolveMergePreset: (ws, block) => this.resolveMergePreset(ws, block),
+            parkStepOnDecision: (ws, i, s, p) => this.parkStepOnDecision(ws, i, s, p),
+            finishStep: (s) => this.finishStep(s),
+            startStep: (s) => this.startStep(s),
+            updateBlockProgress: (ws, i, st) => this.updateBlockProgress(ws, i, st),
+            finalizeBlock: (ws, i, c) => this.finalizeBlock(ws, i, c),
+            stopRunContainer: (ws, i) => this.stopRunContainer(ws, i),
+            persistInstance: (ws, i) => this.executionRepository.upsert(ws, i),
+            emitInstance: (ws, i) => this.emitInstance(ws, i),
+            clockNow: () => this.clock.now(),
+        });
         this.reviewGate = new ReviewGateController({
             blockRepository,
             executionRepository,
@@ -472,7 +499,7 @@ export class ExecutionService {
         // configured (a docker-compose path, or an explicit "no infra dependencies"
         // flag). Block the start with a clear, actionable error otherwise — before any
         // side effects (activation mint / prior-run teardown).
-        if (pipeline.agentKinds.includes(TESTER_AGENT_KIND)) {
+        if (pipeline.agentKinds.some(isTesterKind)) {
             await this.assertTesterInfraConfigured(workspaceId, block);
         }
         // Block the start when the workspace delegates container agents to a runner pool that
@@ -756,7 +783,9 @@ export class ExecutionService {
             // the driver. Fall through so the gate re-evaluates and acts on it (dispatch a helper,
             // rebuild the env, or advance) instead of immediately re-parking.
             const reentrantHumanTest = step.agentKind === HUMAN_TEST_AGENT_KIND && !!step.humanTest?.pendingAction;
-            if (!reentrantRequirements && !reentrantHumanTest) {
+            // The visual-confirmation gate is likewise re-entrant on a human action.
+            const reentrantVisualConfirm = step.agentKind === VISUAL_CONFIRM_AGENT_KIND && !!step.visualConfirm?.pendingAction;
+            if (!reentrantRequirements && !reentrantHumanTest && !reentrantVisualConfirm) {
                 // Parked on either an agent-raised decision or a human approval gate; both
                 // are addressed by the same durable event id.
                 const pendingId = step.decision?.id ?? step.approval?.id;
@@ -829,6 +858,13 @@ export class ExecutionService {
         if (step.agentKind === HUMAN_TEST_AGENT_KIND) {
             return this.humanTestController.evaluate(workspaceId, instance, step, block, isFinalStep);
         }
+        // A `visual-confirmation` gate gathers the UI tester's screenshots + the uploaded
+        // reference designs and PARKS for a human to review actual-vs-reference, then on demand
+        // dispatches the Tester's `fixer`. Passes through (auto-advances) when no binary-artifact
+        // store is wired. See {@link VisualConfirmationController}.
+        if (step.agentKind === VISUAL_CONFIRM_AGENT_KIND) {
+            return this.visualConfirmationController.evaluate(workspaceId, instance, step, block, isFinalStep);
+        }
         // A polling gate step (`ci` / `conflicts`) runs a programmatic precheck and only
         // escalates to a helper container agent (`ci-fixer` / `conflict-resolver`) on a
         // negative verdict — no LLM of its own. Pass-through when the gate's provider is
@@ -1149,7 +1185,7 @@ export class ExecutionService {
         // step's own work: when it finishes (or fails) we drop the handle, return to
         // `testing`, and re-dispatch the Tester against the (now-fixed) branch — its
         // fresh report then drives greenlight-or-loop again. Mirrors the CI gate.
-        if (step.agentKind === TESTER_AGENT_KIND && step.test?.phase === 'fixing') {
+        if (isTesterKind(step.agentKind) && step.test?.phase === 'fixing') {
             step.jobId = undefined;
             step.subtasks = undefined;
             step.test.phase = 'testing';
@@ -1173,6 +1209,13 @@ export class ExecutionService {
                 state: update.state === 'failed' ? 'failed' : 'done',
             });
         }
+        // A `visual-confirmation` gate in its `fixing` phase has a `fixer` job in flight: when it
+        // settles, record the round, refresh the screenshot pairs, and re-park the human.
+        if (step.agentKind === VISUAL_CONFIRM_AGENT_KIND && step.visualConfirm?.phase === 'fixing') {
+            return this.visualConfirmationController.onHelperComplete(workspaceId, instance, step, {
+                state: update.state === 'failed' ? 'failed' : 'done',
+            });
+        }
         if (update.state === 'failed') {
             // A container eviction (the per-run container vanished, its in-memory job is
             // gone) is usually transient. Recover it by dropping the dead handle and
@@ -1488,7 +1531,7 @@ export class ExecutionService {
         // NOT finish the step: we loop the `fixer` (within the attempt budget) and
         // re-test, mirroring the CI gate. A greenlight (or no provider) falls through to
         // the normal finish/advance below. Records the report on the step either way.
-        if (step.agentKind === TESTER_AGENT_KIND && result.testReport !== undefined) {
+        if (isTesterKind(step.agentKind) && result.testReport !== undefined) {
             const looped = await this.testerController.resolveTesterResult(workspaceId, instance, step, result);
             if (looped)
                 return looped;
@@ -2655,6 +2698,9 @@ export class ExecutionService {
         if (step.agentKind === HUMAN_TEST_AGENT_KIND) {
             throw new ConflictError('Resolve the human-testing gate through its window (confirm / request a fix), not the approval gate');
         }
+        if (step.agentKind === VISUAL_CONFIRM_AGENT_KIND) {
+            throw new ConflictError('Resolve the visual-confirmation gate through its window (approve / request a fix), not the approval gate');
+        }
         if (step.companion?.exceeded) {
             throw new ConflictError('Resolve this companion review through its iteration-cap prompt, not the approval gate');
         }
@@ -3034,6 +3080,21 @@ export class ExecutionService {
     destroyHumanTestEnv(workspaceId, blockId) {
         return this.humanTestController.destroyEnvironment(workspaceId, blockId);
     }
+    // ---- visual-confirmation gate actions (driven from the dedicated window) --
+    // Each mutates the parked gate step and wakes the durable driver; see
+    // {@link VisualConfirmationController}.
+    /** Approve the reviewed screenshots: advance the run. */
+    approveVisualConfirm(workspaceId, blockId) {
+        return this.visualConfirmationController.approve(workspaceId, blockId);
+    }
+    /** Submit findings and request a fix: dispatch the Tester's `fixer`, then re-park. */
+    requestVisualConfirmFix(workspaceId, blockId, findings) {
+        return this.visualConfirmationController.requestFix(workspaceId, blockId, findings);
+    }
+    /** Refresh the screenshot pairs from the latest UI-tester report. */
+    recaptureVisualConfirm(workspaceId, blockId) {
+        return this.visualConfirmationController.recapture(workspaceId, blockId);
+    }
     /**
      * Dispatch the `fixer` against the human-review gate's PR branch from a human's freeform
      * instructions — bypassing the precheck + grace window. Parks a `pendingFix` on the gate step,
@@ -3148,6 +3209,7 @@ export class ExecutionService {
                 step.metrics = {
                     calls: s.calls,
                     promptTokens: s.promptTokens,
+                    cachedPromptTokens: s.cachedPromptTokens,
                     completionTokens: s.completionTokens,
                     peakCompletionTokens: s.peakCompletionTokens,
                     maxOutputTokens: s.maxOutputTokens,