@straiffi/archon 1.10.9 → 1.10.12

package/dist/server/workers/review.js

@@ -12,7 +12,7 @@ import { resolveReviewerConfig } from '../lib/reviewerConfig.js';
 import { enqueue, isRunning } from '../lib/queue.js';
 import { appendLifecycleEvent, getExecutionRun, transitionExecutionRun, withExecutionRunTransaction } from '../lib/executionRuns.js';
 import { buildReviewContext } from '../lib/reviewContext.js';
-import { computeDiffSignature, constrainReviewOutputToDiff, createReviewRun, finalizeReviewRun, formatExistingReviewFindingsForPrompt, formatReviewFindingsMarkdown, getLatestCompletedReviewRunForContext, hasPriorReviewRunForBundle, insertReviewFindings, listReviewFindings, mergeReviewFindings, parseStructuredReviewOutput, reconcileReviewFindingsToDiff, updateReviewRunPreparation, } from '../lib/reviewFindings.js';
+import { computeDiffSignature, computeReviewEvaluatorFingerprint, constrainReviewOutputToDiff, createReviewRun, finalizeReviewRun, formatExistingReviewFindingsForPrompt, formatReviewFindingsMarkdown, getLatestCompletedReviewRunForContext, hasPriorReviewRunForBundle, insertReviewFindings, listReviewFindings, mergeReviewFindings, parseStructuredReviewOutput, reconcileReviewFindingsToDiff, updateReviewRunPreparation, } from '../lib/reviewFindings.js';
 import { buildSpecCoverageContext } from '../lib/specCoverageContext.js';
 import { listTickets } from '../lib/tickets.js';
 import { registerBundleActivity, unregisterBundleActivity } from '../lib/bundleActivity.js';
@@ -23,12 +23,12 @@ import { captureStageBaseline, captureStageSuccess } from '../lib/ticketUndo.js'
 import { beginTicketThinkingRun, bindTicketThinkingMessage } from '../lib/thinkingState.js';
 import { autoParkTicketIfStale, getTicket, resolveTicketBranch, resolveTicketTool, resolveTicketWorkerKey } from '../lib/tickets.js';
 import { createAgentRun, emitAgentRunsUpdated, emitExecutionAgentRunsUpdated, getAgentRun, listAgentRunsForExecutionRun, terminalizeAgentRunsForExecutionRun, } from '../lib/agentRuns.js';
-import { buildAutomationFixInstruction, buildAutomationFixStatusMessage, countBlockingFindings, decideReviewAutomation, depthAtLeast, getBundleAutomationState, markBundleAutomationDeliveryReady, resetBundleAutomationLoop, resolveAutomationConfig, } from '../lib/ticketAutomation.js';
+import { buildAutomationFixInstruction, buildAutomationFixStatusMessage, depthAtLeast, resetBundleAutomationLoop, resolveAutomationConfig, } from '../lib/ticketAutomation.js';
 import { runAutomationCommitChain } from '../lib/ticketAutomationCommit.js';
 import { submitTicketFollowUp } from '../lib/ticketFollowUpOperations.js';
-import { completeAutomationProgressionInTransaction, completeAutomationProgression, completeAutomationReviewFixActionInTransaction, createAutomationReviewFixInTransaction, ensureAutomationProgression, finishAutomationReviewFix, pauseAutomationProgression, resolveProgressionForChild, waitAutomationProgression, } from '../lib/automationProgression.js';
+import { completeAutomationProgressionInTransaction, completeAutomationProgression, completeAutomationReviewFixActionInTransaction, ensureAutomationProgression, finishAutomationReviewFix, pauseAutomationProgression, resolveProgressionForChild, waitAutomationProgression, } from '../lib/automationProgression.js';
 import { assertAutomationAuthorizationCurrent, assertAutomationBoundaryNotRequired, AutomationCheckGateError, } from '../lib/automationCheckGate.js';
-import { attachReviewRepairActionEnvelopeInTransaction, attachReviewRepairActionInTransaction, appendReviewRepairIterationDecisionInTransaction, attachReviewRepairResultInTransaction, claimReviewRepairIterationInTransaction, getReviewRepairIteration, getReviewRepairIterationByVerification, pauseReviewRepairIteration, pauseReviewRepairIterationInTransaction, terminalizeReviewRepairIterationInTransaction, } from '../lib/reviewRepairIterations.js';
+import { attachReviewRepairActionInTransaction, attachReviewRepairResultInTransaction, getReviewRepairIteration, getReviewRepairIterationByVerification, pauseReviewRepairIteration, terminalizeReviewRepairIterationInTransaction, } from '../lib/reviewRepairIterations.js';
 import { formatBuildOpenQuestionAnswersPrompt, listAnsweredBuildOpenQuestionsForBundle } from '../lib/buildOpenQuestions.js';
 import { validateWorkflowContext } from '../lib/types.js';
 import { isModelResolutionError } from '../lib/modelResolutionErrors.js';
@@ -97,6 +97,11 @@ const stripSystemReminderBlocks = (output) => {
 const normalizeReviewOutput = (output) => {
     return extractProjectMemorySuggestionsFromOutput(stripSystemReminderBlocks(output));
 };
+const getEffectiveReviewOutputProvenance = (outputs) => {
+    const provenances = outputs.map(output => output.outputProvenance ?? 'parse_failed');
+    return ['parse_failed', 'markdown_fallback', 'subagent_aggregation', 'structured_transcript', 'structured_primary']
+        .find(provenance => provenances.includes(provenance)) ?? 'parse_failed';
+};
 const resolveStructuredReviewOutput = (parentResponse, transcriptRuns, mainRole = 'code-review') => {
     const normalizedParent = typeof parentResponse === 'string'
         ? normalizeReviewOutput(parentResponse)
@@ -114,7 +119,7 @@ const resolveStructuredReviewOutput = (parentResponse, transcriptRuns, mainRole
         if (!normalized.cleanedOutput) {
             return [];
         }
-        const parsed = parseStructuredReviewOutput(normalized.cleanedOutput);
+        const parsed = parseStructuredReviewOutput(normalized.cleanedOutput, 'structured_transcript');
         return parsed.status === 'completed' ? [parsed] : [];
     })
         .at(0);
@@ -129,7 +134,7 @@ const resolveStructuredReviewOutput = (parentResponse, transcriptRuns, mainRole
         if (!normalized.cleanedOutput) {
             return [];
         }
-        const parsed = parseStructuredReviewOutput(normalized.cleanedOutput);
+        const parsed = parseStructuredReviewOutput(normalized.cleanedOutput, 'subagent_aggregation');
         return parsed.status === 'completed' ? [parsed] : [];
     });
     if (parsedSubagentOutputs.length === 0) {
@@ -148,12 +153,18 @@ const resolveStructuredReviewOutput = (parentResponse, transcriptRuns, mainRole
             findings,
             resolvedExistingFindingIds: [],
             status: 'completed',
+            outputProvenance: parsedSubagentOutputs.some(output => output.outputProvenance === 'markdown_fallback')
+                ? 'markdown_fallback'
+                : 'subagent_aggregation',
         },
     };
 };
 const MAX_INLINE_DIFF_LINES = 400;
 const MAX_INLINE_DIFF_CHARS = 20_000;
 const SUBAGENT_REVIEW_GUIDANCE = 'If you delegate any part of the review to subagents, you remain responsible for evaluating their results and consolidating every accepted finding into your final response using the exact format below.';
+const REVIEW_PROMPT_VERSION = 'ticket-review-v2';
+const REVIEW_ROLE_POLICY_VERSION = 'ticket-review-role-policy-v1';
+const REVIEW_OUTPUT_CONTRACT_VERSION = 'archon-review-findings-v1';
 const countLines = (value) => {
     if (!value) {
         return 0;
@@ -372,89 +383,23 @@ Every finding must set \`file_path\` to a path in \`<review_scope>\` and set at
 
 Do not include planning text, implementation steps, or commentary about changing modes.`, ticket.skills);
 };
-const runReviewCompletionAutomation = async (ticket, project, reviewRunId, parsedStatus, findings, io, executionContext, initiation, persistedAutomation) => {
-    if (!ticket.worktree_bundle_id || !ticket.project_id) {
+const runReviewCompletionAutomation = async (ticket, project, reviewRunId, io, executionContext, initiation, persistedAutomation) => {
+    if (!ticket.worktree_bundle_id || !ticket.project_id || !persistedAutomation) {
         return false;
     }
     const projectId = ticket.project_id;
     const bundleId = ticket.worktree_bundle_id;
-    const liveTicket = getTicket(ticket.id) ?? ticket;
-    const liveProject = getEffectiveProject(liveTicket) ?? project;
-    const config = persistedAutomation?.config ?? resolveAutomationConfig(liveTicket.bundle ?? null, liveProject);
-    const bundleState = getBundleAutomationState(projectId, bundleId);
-    const decision = persistedAutomation?.decision
-        ?? decideReviewAutomation(config, bundleState, reviewRunId, parsedStatus, findings);
-    const progressionRunId = persistedAutomation?.progressionRunId
-        ?? (initiation.initiator === 'automation'
-            ? resolveProgressionForChild({ projectId, bundleId, childRunId: executionContext.runId }).id
-            : null);
+    const { config, decision, progressionRunId, disposition } = persistedAutomation;
     const progression = progressionRunId ? { id: progressionRunId } : null;
-    if (progression) {
-        const iteration = getReviewRepairIterationByVerification(projectId, executionContext.runId);
-        if (iteration?.phase === 'verification_active'
-            && iteration.resulting_review_execution_run_id === executionContext.runId) {
-            const resultingBlockingCount = parsedStatus === 'completed'
-                ? countBlockingFindings(findings, config)
-                : null;
-            withExecutionRunTransaction(transaction => {
-                if (parsedStatus === 'completed' && reviewRunId) {
-                    attachReviewRepairResultInTransaction(transaction, {
-                        projectId,
-                        iterationId: iteration.id,
-                        expectedPhase: 'verification_active',
-                        expectedVerificationRunId: executionContext.runId,
-                        resultingReviewRunId: reviewRunId,
-                    });
-                }
-                appendReviewRepairIterationDecisionInTransaction(transaction, {
-                    projectId,
-                    iterationId: iteration.id,
-                    decisionPoint: 'verification_completion',
-                    effectiveAutomationConfig: config,
-                    decisionCode: parsedStatus === 'completed' ? 'verification_completed' : 'verification_failed',
-                    operationRunId: executionContext.runId,
-                });
-                const terminal = decision.kind === 'fix_iteration'
-                    ? { outcome: 'continuing', stopReason: null }
-                    : decision.kind === 'clean'
-                        ? { outcome: 'repaired', stopReason: 'no_blocking_findings' }
-                        : decision.kind === 'exhausted'
-                            ? { outcome: 'exhausted', stopReason: 'max_iterations' }
-                            : decision.kind === 'pause_blocking'
-                                ? { outcome: 'interrupted', stopReason: 'automation_disabled' }
-                                : parsedStatus !== 'completed'
-                                    ? { outcome: 'failed', stopReason: 'verification_failed' }
-                                    : { outcome: 'interrupted', stopReason: 'configuration_changed' };
-                if (parsedStatus !== 'completed') {
-                    pauseReviewRepairIterationInTransaction(transaction, {
-                        projectId,
-                        iterationId: iteration.id,
-                        expectedPhase: 'verification_active',
-                        expectedCurrentRunId: executionContext.runId,
-                        stopReason: 'verification_failed',
-                    });
-                }
-                else {
-                    terminalizeReviewRepairIterationInTransaction(transaction, {
-                        projectId,
-                        iterationId: iteration.id,
-                        expectedPhase: 'verification_active',
-                        expectedCurrentRunId: executionContext.runId,
-                        outcome: terminal.outcome,
-                        stopReason: terminal.stopReason,
-                        resultingBlockingCount,
-                    });
-                    if (decision.kind === 'exhausted' && progression) {
-                        completeAutomationProgressionInTransaction(transaction, {
-                            projectId,
-                            bundleId,
-                            progressionRunId: progression.id,
-                            deliveryReadyStageRunId: executionContext.runId,
-                        });
-                    }
-                }
-            }, asGenericEmitter(io), 'immediate');
+    if (disposition === 'handled' || decision.kind === 'no_progress') {
+        return true;
+    }
+    if (disposition === 'commit') {
+        if (!progressionRunId) {
+            throw new Error('Persisted Review commit has no owning progression');
         }
+        await runAutomationCommitChain(ticket, project, ticket.bundle ?? null, config, io, progressionRunId, undefined, executionContext.runId);
+        return true;
     }
     switch (decision.kind) {
         case 'manual':
@@ -511,61 +456,14 @@ const runReviewCompletionAutomation = async (ticket, project, reviewRunId, parse
                 review_finding_run_id: reviewRunId,
                 review_severity: config.reviewSeverity,
             });
-            const repairProgressionRunId = progression?.id ?? ensureAutomationProgression({
-                projectId,
-                bundleId,
-                ticketId: ticket.id,
-                ticketTitle: ticket.title,
-                parentRunId: executionContext.runId,
-                workflowContext: initiation.workflowContext,
-                broadcaster: asGenericEmitter(io),
-            }).runId;
-            let reviewFix = persistedAutomation?.repair?.reviewFix ?? null;
-            let iterationId = persistedAutomation?.repair?.iterationId ?? null;
-            if (persistedAutomation && !persistedAutomation.repair) {
+            if (disposition !== 'dispatch_repair' || !persistedAutomation.repair) {
                 return true;
             }
-            if (!reviewFix) {
-                const claimed = withExecutionRunTransaction(transaction => {
-                    const iteration = claimReviewRepairIterationInTransaction(transaction, {
-                        projectId,
-                        worktreeBundleId: bundleId,
-                        ownerTicketId: ticket.id,
-                        automationProgressionRunId: repairProgressionRunId,
-                        triggeringReviewExecutionRunId: executionContext.runId,
-                        triggeringReviewRunId: reviewRunId,
-                        triggeringBlockingCount: countBlockingFindings(findings, config),
-                        effectiveAutomationConfig: config,
-                        decisionCode: 'repair_admitted',
-                    });
-                    if (iteration.phase !== 'observed') {
-                        return { iteration, reviewFix: null };
-                    }
-                    const createdReviewFix = createAutomationReviewFixInTransaction(transaction, {
-                        projectId,
-                        bundleId,
-                        ticketId: ticket.id,
-                        ticketTitle: ticket.title,
-                        progressionRunId: repairProgressionRunId,
-                        workflowContext: reviewFixWorkflowContext,
-                        triggeringReviewRunId: executionContext.runId,
-                        reviewFindingRunId: reviewRunId,
-                        reviewSeverity: config.reviewSeverity,
-                    });
-                    attachReviewRepairActionEnvelopeInTransaction(transaction, {
-                        projectId,
-                        iterationId: iteration.id,
-                        expectedPhase: 'observed',
-                        runId: createdReviewFix.runId,
-                    });
-                    return { iteration, reviewFix: createdReviewFix };
-                }, asGenericEmitter(io), 'immediate');
-                iterationId = claimed.iteration.id;
-                reviewFix = claimed.reviewFix;
-            }
-            if (!reviewFix) {
-                return true;
+            const { reviewFix, iterationId } = persistedAutomation.repair;
+            if (!progressionRunId) {
+                throw new Error('Persisted Review repair has no owning progression');
             }
+            const repairProgressionRunId = progressionRunId;
             const started = submitTicketFollowUp({
                 ticketId: ticket.id,
                 requestedProjectId: projectId,
@@ -580,7 +478,7 @@ const runReviewCompletionAutomation = async (ticket, project, reviewRunId, parse
                 onExecutionCreated: (createdContext, transaction) => {
                     attachReviewRepairActionInTransaction(transaction, {
                         projectId,
-                        iterationId: iterationId,
+                        iterationId,
                         expectedPhase: 'action_claimed',
                         runId: createdContext.runId,
                     });
@@ -674,8 +572,28 @@ const runReviewCompletionAutomation = async (ticket, project, reviewRunId, parse
             return true;
         }
         case 'exhausted': {
-            if (!progression) {
-                markBundleAutomationDeliveryReady(projectId, bundleId, executionContext.runId);
+            if (progression) {
+                waitAutomationProgression({
+                    projectId,
+                    bundleId,
+                    progressionRunId: progression.id,
+                    pauseCode: 'review',
+                    pauseReason: 'Blocking Review findings remain after repair verification.',
+                    reasonCode: 'precondition_failed',
+                    pausedStageRunId: executionContext.runId,
+                    broadcaster: asGenericEmitter(io),
+                });
+            }
+            else {
+                pauseAutomationProgression({
+                    projectId,
+                    bundleId,
+                    pauseCode: 'review',
+                    pauseReason: 'Blocking Review findings remain after repair verification.',
+                    reasonCode: 'precondition_failed',
+                    pausedStageRunId: null,
+                    broadcaster: asGenericEmitter(io),
+                });
             }
             const pausedTicket = getTicket(ticket.id);
             if (pausedTicket) {
@@ -938,6 +856,16 @@ export function startReview(ticket, io, options) {
                     model: acceptedReviewExecution.model,
                     variant: acceptedReviewExecution.variant,
                 };
+                const evaluatorInput = {
+                    tool,
+                    ...reviewerConfig,
+                    skills,
+                    promptVersion: REVIEW_PROMPT_VERSION,
+                    rolePolicyVersion: REVIEW_ROLE_POLICY_VERSION,
+                    outputContractVersion: REVIEW_OUTPUT_CONTRACT_VERSION,
+                    rolePolicy: 'preparing',
+                    changeScope: null,
+                };
                 reviewExecutionProvenance = { tool, model: reviewerConfig.model, variant: reviewerConfig.variant };
                 const reviewProcessState = registerReviewProcess(acceptedTicket.id);
                 registeredReviewProcess = reviewProcessState;
@@ -1112,6 +1040,7 @@ export function startReview(ticket, io, options) {
                         messageId: null,
                         model: reviewerConfig.model,
                         variant: reviewerConfig.variant,
+                        evaluatorFingerprint: computeReviewEvaluatorFingerprint(evaluatorInput),
                         status: 'running',
                     });
                     transaction.addExecutionRunLink({
@@ -1181,13 +1110,15 @@ export function startReview(ticket, io, options) {
                 const supplementalRequirements = ticket.project_id && ticket.worktree_bundle_id
                     ? formatBuildOpenQuestionAnswersPrompt(listAnsweredBuildOpenQuestionsForBundle(ticket.project_id, ticket.worktree_bundle_id))
                     : '';
-                const previousCompletedReviewRun = getLatestCompletedReviewRunForContext({
-                    ticketId: ticket.id,
-                    projectId: ticket.project_id,
-                    worktreeBundleId: ticket.worktree_bundle_id,
-                });
-                const existingReviewFindings = previousCompletedReviewRun
-                    ? reconcileReviewFindingsToDiff(listReviewFindings(previousCompletedReviewRun.id), diffText)
+                const previousReviewRunId = verificationIteration
+                    ? verificationIteration.triggering_review_run_id
+                    : getLatestCompletedReviewRunForContext({
+                        ticketId: ticket.id,
+                        projectId: ticket.project_id,
+                        worktreeBundleId: ticket.worktree_bundle_id,
+                    })?.id;
+                const existingReviewFindings = previousReviewRunId
+                    ? reconcileReviewFindingsToDiff(listReviewFindings(previousReviewRunId), diffText)
                     : [];
                 const specCoverageEligible = Boolean(project.spec_aware_code_review_enabled)
                     && Boolean(ticket.origin_spec_snapshot_id)
@@ -1204,6 +1135,14 @@ export function startReview(ticket, io, options) {
                 }
                 const runSpecCoverage = Boolean(specContext) && !hasPriorReview;
                 const specGuardContext = hasPriorReview ? specContext : null;
+                const preparedEvaluatorInput = {
+                    ...evaluatorInput,
+                    rolePolicy: runSpecCoverage ? 'code-review+spec-coverage' : specGuardContext ? 'code-review+spec-guard' : 'code-review',
+                    changeScope,
+                };
+                const evaluatorFingerprint = computeReviewEvaluatorFingerprint(preparedEvaluatorInput);
+                db.prepare('UPDATE review_runs SET evaluator_fingerprint = ? WHERE id = ? AND status = ?')
+                    .run(evaluatorFingerprint, activeReviewRunId, 'running');
                 let specCoverageRunStreamTarget = null;
                 if (runSpecCoverage) {
                     const runId = createAgentRun({
@@ -1338,11 +1277,16 @@ export function startReview(ticket, io, options) {
                 let persistedMessageId = null;
                 let reviewContent = null;
                 let findingsToPersist = [];
+                let outputProvenance = constrainedBlindParsed?.outputProvenance ?? 'parse_failed';
                 if (constrainedBlindParsed?.status === 'completed') {
                     const specOutput = resolveStructuredReviewOutput(specCoverageResult?.response ?? null, specCoverageResult?.transcriptRuns ?? [], 'spec-coverage').parsed;
                     const specParsed = specOutput
                         ? constrainReviewOutputToDiff(specOutput, diffText)
                         : null;
+                    outputProvenance = getEffectiveReviewOutputProvenance([
+                        constrainedBlindParsed,
+                        ...(specParsed?.status === 'completed' ? [specParsed] : []),
+                    ]);
                     const newlyReportedFindings = specParsed && specParsed.status === 'completed'
                         ? [...constrainedBlindParsed.findings, ...specParsed.findings.map(finding => ({
                                 ...finding,
@@ -1353,11 +1297,15 @@ export function startReview(ticket, io, options) {
                         existingFindings: existingReviewFindings,
                         newFindings: newlyReportedFindings,
                         resolvedExistingFindingIds: constrainedBlindParsed.resolvedExistingFindingIds,
+                        outputProvenance,
+                        evaluatorFingerprint,
                     });
                     const blindFindings = mergeReviewFindings({
                         existingFindings: existingReviewFindings,
                         newFindings: constrainedBlindParsed.findings,
                         resolvedExistingFindingIds: constrainedBlindParsed.resolvedExistingFindingIds,
+                        outputProvenance: constrainedBlindParsed.outputProvenance ?? 'parse_failed',
+                        evaluatorFingerprint,
                     });
                     parsedFindings = mergedFindings;
                     parsedStatus = constrainedBlindParsed.status;
@@ -1393,6 +1341,7 @@ export function startReview(ticket, io, options) {
                             status: parsedStatus ?? 'parse_failed',
                             messageId,
                             failureMessage: parseFailed ? 'Review output could not be parsed' : null,
+                            outputProvenance,
                         });
                         if (findingsToPersist.length > 0) {
                             insertReviewFindings(activeReviewRunId, findingsToPersist);
@@ -1460,7 +1409,7 @@ export function startReview(ticket, io, options) {
                     let handled = false;
                     try {
                         await releaseWorkloadEnvironment();
-                        handled = await runReviewCompletionAutomation(updated, project, reviewRunId, parsedStatus, parsedFindings, io, executionContext, initiation, persisted.persistedAutomation);
+                        handled = await runReviewCompletionAutomation(updated, project, reviewRunId, io, executionContext, initiation, persisted.persistedAutomation);
                     }
                     catch (automationError) {
                         if (ticket.project_id && ticket.worktree_bundle_id) {