@straiffi/archon 1.8.12 → 1.8.14

package/dist/server/workers/review.js

@@ -2,15 +2,17 @@ import db from '../db.js';
 import { assertNoActiveAutomationChild } from '../lib/activeAutomationChild.js';
 import { formatAgentErrorForTranscript, getAgentErrorTranscriptRuns, registerReviewProcess, runAgent, storeTicketUsageSnapshot, storeTicketUsageSnapshotDirect, untrackReviewProcess, } from '../lib/agent.js';
 import config from '../lib/config.js';
+import { getBundlePullRequestRecord } from '../lib/bundlePullRequests.js';
 import { persistProjectMemorySuggestions } from '../lib/projectMemory.js';
 import { getReviewDiffDataAsync, resolveWorktreePath } from '../lib/git.js';
 import { getProjectMemoryPromptBlock } from '../lib/projectMemoryPrompt.js';
 import { extractProjectMemorySuggestionsFromOutput, getProjectMemorySuggestionPromptRules } from '../lib/projectMemorySuggestions.js';
 import { getEffectiveProject } from '../lib/projects.js';
+import { resolveReviewerConfig } from '../lib/reviewerConfig.js';
 import { enqueue, isRunning } from '../lib/queue.js';
 import { getExecutionRun, transitionExecutionRun, withExecutionRunTransaction } from '../lib/executionRuns.js';
 import { buildReviewContext } from '../lib/reviewContext.js';
-import { computeDiffSignature, createReviewRun, hasPriorReviewRunForBundle, insertReviewFindings, parseStructuredReviewOutput, } from '../lib/reviewFindings.js';
+import { computeDiffSignature, constrainReviewOutputToDiff, createReviewRun, finalizeReviewRun, hasPriorReviewRunForBundle, insertReviewFindings, parseStructuredReviewOutput, updateReviewRunPreparation, } from '../lib/reviewFindings.js';
 import { buildSpecCoverageContext } from '../lib/specCoverageContext.js';
 import { listTickets } from '../lib/tickets.js';
 import { registerBundleActivity, unregisterBundleActivity } from '../lib/bundleActivity.js';
@@ -26,9 +28,17 @@ import { runAutomationCommitChain } from '../lib/ticketAutomationCommit.js';
 import { submitTicketFollowUp } from '../lib/ticketFollowUpOperations.js';
 import { completeAutomationProgression, createAutomationReviewFix, ensureAutomationProgression, finishAutomationReviewFix, pauseAutomationProgression, resolveProgressionForChild, waitAutomationProgression, } from '../lib/automationProgression.js';
 import { formatBuildOpenQuestionAnswersPrompt, listAnsweredBuildOpenQuestionsForBundle } from '../lib/buildOpenQuestions.js';
+import { validateWorkflowContext } from '../lib/types.js';
+import { isOpencodeModelResolutionError } from '../lib/opencodeErrors.js';
 const getErrorMessage = (error) => {
     return error instanceof Error ? error.message : String(error);
 };
+const getReviewerExecutionError = (error, model, variant) => {
+    const reviewer = variant ? `${model} (${variant})` : model;
+    const wrapped = new Error(`Configured reviewer ${reviewer} could not run: ${getErrorMessage(error)}`, { cause: error });
+    wrapped.transcriptRuns = getAgentErrorTranscriptRuns(error);
+    return wrapped;
+};
 export const resolveReviewContinuation = (executionRunId) => {
     const runs = listAgentRunsForExecutionRun(executionRunId);
     return {
@@ -73,7 +83,7 @@ const stripSystemReminderBlocks = (output) => {
 const normalizeReviewOutput = (output) => {
     return extractProjectMemorySuggestionsFromOutput(stripSystemReminderBlocks(output));
 };
-const resolveStructuredReviewOutput = (parentResponse, transcriptRuns) => {
+const resolveStructuredReviewOutput = (parentResponse, transcriptRuns, mainRole = 'code-review') => {
     const normalizedParent = typeof parentResponse === 'string'
         ? normalizeReviewOutput(parentResponse)
         : { cleanedOutput: '', suggestions: [] };
@@ -84,7 +94,7 @@ const resolveStructuredReviewOutput = (parentResponse, transcriptRuns) => {
         return { normalizedParent, parsed: parsedParent };
     }
     const parsedMainTranscript = transcriptRuns
-        .filter(run => run.role === 'code-review' && run.agent_status === 'done' && run.response_text?.trim())
+        .filter(run => run.role === mainRole && run.agent_status === 'done' && run.response_text?.trim())
         .flatMap(run => {
         const normalized = normalizeReviewOutput(run.response_text ?? '');
         if (!normalized.cleanedOutput) {
@@ -126,6 +136,15 @@ const resolveStructuredReviewOutput = (parentResponse, transcriptRuns) => {
         },
     };
 };
+const formatParsedReviewMarkdown = (parsed) => {
+    if (parsed.cleanedMarkdown) {
+        return parsed.cleanedMarkdown;
+    }
+    if (parsed.findings.length === 0) {
+        return '## Findings\nNo genuine issues found.';
+    }
+    return `## Findings\n${parsed.findings.map((finding, index) => (`${index + 1}. ${finding.title}\n   - Severity: ${finding.severity}\n   - What: ${finding.what_text}${finding.why_text ? `\n   - Why: ${finding.why_text}` : ''}${finding.fix_text ? `\n   - Fix: ${finding.fix_text}` : ''}`)).join('\n\n')}`;
+};
 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.';
@@ -205,7 +224,9 @@ const reviewPrompt = ({ ticket, diff, memoryEnabled, reviewContext, agentRunId,
     const continuationBlock = buildReviewContinuationBlock(continuation, 'blind');
     return prependSkillPromptBlock(`Review the following code changes.
 
-This is a code review task, not an implementation task. Focus on genuine defects: bugs, regressions, broken assumptions, missing or incorrect tests, type or runtime issues, and risky behavior changes. Infer the intended behavior from the patch itself instead of relying on any external ticket context. Review only the current project worktree changes and treat any linked projects or other repositories as read-only context.
+This is a code review task, not an implementation task. Focus on genuine defects: bugs, regressions, broken assumptions, missing or incorrect tests, type or runtime issues, and risky behavior changes. Infer the intended behavior from the patch itself instead of relying on any external ticket context. The supplied diff is the authoritative review scope. Treat linked projects and other repositories as read-only context.
+
+You may inspect unchanged files, callers, helpers, tests, and project memory only to understand the behavior and impact of changed code. Unchanged code is supporting evidence, not an independent review target. Report a finding only when it is introduced, exposed, or materially worsened by this diff; it can be anchored to a changed path and changed line responsible for the issue; and fixing it requires changing this patch rather than unrelated pre-existing code. Do not report pre-existing defects, general cleanup, architectural improvements, or issues from another branch or ticket. If unchanged code demonstrates a regression, cite it as evidence but anchor the finding to the changed line that caused the regression. When causality is uncertain, omit the finding.
 
 Before writing findings, perform this investigation workflow:
 1. ${diffPromptBlock.investigationInstruction}
@@ -213,6 +234,7 @@ Before writing findings, perform this investigation workflow:
 3. Inspect directly related code paths such as nearby helpers, callers, consumers, and tests that could be affected.
 4. Check whether the change is consistent with the project's accepted decisions and conventions.
 5. Only then report genuine issues that are supported by the code you inspected.
+6. Before returning the review, discard every candidate finding that cannot be traced directly to a changed line in the supplied diff.
 
 ${SUBAGENT_REVIEW_GUIDANCE}
 
@@ -250,7 +272,7 @@ After the markdown findings, append a machine-readable block using these exact m
 {"findings":[{"ordinal":1,"title":"...","severity":"high","file_path":"path/to/file.ts","old_line_number":12,"new_line_number":15,"line_content":"...","hunk_header":"@@ -12,3 +15,4 @@","what_text":"...","why_text":"...","fix_text":"..."}]}
 </archon-review-findings>
 
-Use an empty findings array when there are no genuine issues. The JSON must be valid and must not be wrapped in code fences. Escape all JSON string contents correctly. If a snippet would be awkward to escape, set \`line_content\` to null. For pure additions, set only \`new_line_number\` and use \`old_line_number\`: null. For pure deletions, set only \`old_line_number\` and use \`new_line_number\`: null.${getProjectMemorySuggestionPromptRules(memoryEnabled, 'review') ? `
+Every finding must set \`file_path\` to a path in \`<review_scope>\` and set at least one line number identifying a changed diff line responsible for the issue. Do not use an unchanged context line merely because it is near the defect. Use an empty findings array when there are no in-scope genuine issues. The JSON must be valid and must not be wrapped in code fences. Escape all JSON string contents correctly. If a snippet would be awkward to escape, set \`line_content\` to null. For pure additions, set only \`new_line_number\` and use \`old_line_number\`: null. For pure deletions, set only \`old_line_number\` and use \`new_line_number\`: null.${getProjectMemorySuggestionPromptRules(memoryEnabled, 'review') ? `
 
 ${getProjectMemorySuggestionPromptRules(memoryEnabled, 'review')}` : ''}
 
@@ -268,13 +290,16 @@ const specCoverageReviewPrompt = ({ ticket, diff, reviewContext, specContext, ag
     });
     const diffPromptBlock = buildDiffPromptBlock(diff);
     const continuationBlock = buildReviewContinuationBlock(continuation, 'spec-coverage');
-    return prependSkillPromptBlock(`Verify the implementation against this specification. Focus on missing acceptance criteria, unhandled edge cases, and scope drift — not generic code-quality defects already covered by the blind review.
+    return prependSkillPromptBlock(`Verify the implementation against this specification. Focus on missing acceptance criteria, unhandled edge cases, and scope drift — not generic code-quality defects already covered by the blind review. The supplied diff is the authoritative review scope.
+
+You may inspect unchanged code to understand the implementation, but it is supporting evidence rather than an independent review target. Report only conformance gaps introduced by, exposed by, or expected to be implemented in this diff, and anchor every finding to the changed path and changed line responsible. Do not report unrelated missing functionality, pre-existing defects, or cleanup opportunities. When causality or patch responsibility is uncertain, omit the finding.
 
 Before writing findings, perform this investigation workflow:
 1. ${diffPromptBlock.investigationInstruction}
 2. Read the full contents of every changed file that matters to the spec, not just the diff hunk.
 3. Compare the implementation against the spec_context block: are all acceptance criteria satisfied, are listed edge cases handled, and does the change stay within scope?
 4. Only then report genuine conformance gaps that are supported by the code you inspected.
+5. Before returning the review, discard every candidate finding that cannot be traced directly to a changed line in the supplied diff.
 
 ${SUBAGENT_REVIEW_GUIDANCE}
 
@@ -314,7 +339,7 @@ After the markdown findings, append a machine-readable block using these exact m
 {"findings":[{"ordinal":1,"title":"...","severity":"high","file_path":"path/to/file.ts","old_line_number":12,"new_line_number":15,"line_content":"...","hunk_header":"@@ -12,3 +15,4 @@","what_text":"...","why_text":"...","fix_text":"..."}]}
 </archon-review-findings>
 
-Use an empty findings array when there are no genuine issues. The JSON must be valid and must not be wrapped in code fences. Escape all JSON string contents correctly. If a snippet would be awkward to escape, set \`line_content\` to null. For pure additions, set only \`new_line_number\` and use \`old_line_number\`: null. For pure deletions, set only \`old_line_number\` and use \`new_line_number\`: null.
+Every finding must set \`file_path\` to a path in \`<review_scope>\` and set at least one line number identifying a changed diff line responsible for the conformance gap. Do not use an unchanged context line merely because it is related. Use an empty findings array when there are no in-scope genuine issues. The JSON must be valid and must not be wrapped in code fences. Escape all JSON string contents correctly. If a snippet would be awkward to escape, set \`line_content\` to null. For pure additions, set only \`new_line_number\` and use \`old_line_number\`: null. For pure deletions, set only \`old_line_number\` and use \`new_line_number\`: null.
 
 Do not include planning text, implementation steps, or commentary about changing modes.`, ticket.skills);
 };
@@ -378,6 +403,9 @@ const runReviewCompletionAutomation = async (ticket, project, reviewRunId, parse
             return true;
         }
         case 'fix_iteration': {
+            if (!reviewRunId) {
+                throw new Error('Review fix cannot start without persisted review findings');
+            }
             const previousIteration = bundleState?.review_iteration_count ?? 0;
             const previousFixLoopActive = bundleState?.fix_loop_active ?? 0;
             upsertBundleAutomationState(projectId, bundleId, {
@@ -385,6 +413,11 @@ const runReviewCompletionAutomation = async (ticket, project, reviewRunId, parse
                 fix_loop_active: 1,
             });
             const instruction = buildAutomationFixInstruction(config.reviewSeverity);
+            const reviewFixWorkflowContext = validateWorkflowContext({
+                ...(initiation.workflowContext ?? {}),
+                review_finding_run_id: reviewRunId,
+                review_severity: config.reviewSeverity,
+            });
             let reviewFix = null;
             try {
                 reviewFix = progression
@@ -394,8 +427,10 @@ const runReviewCompletionAutomation = async (ticket, project, reviewRunId, parse
                         ticketId: ticket.id,
                         ticketTitle: ticket.title,
                         progressionRunId: progression.id,
-                        workflowContext: initiation.workflowContext,
+                        workflowContext: reviewFixWorkflowContext,
                         triggeringReviewRunId: executionContext.runId,
+                        reviewFindingRunId: reviewRunId,
+                        reviewSeverity: config.reviewSeverity,
                         broadcaster: asGenericEmitter(io),
                     })
                     : null;
@@ -416,7 +451,7 @@ const runReviewCompletionAutomation = async (ticket, project, reviewRunId, parse
                 initiation: {
                     initiator: 'automation',
                     causalParentRunId: reviewFix?.runId ?? executionContext.runId,
-                    workflowContext: initiation.workflowContext,
+                    workflowContext: reviewFixWorkflowContext,
                 },
             });
             if (!started.ok) {
@@ -613,6 +648,7 @@ export function startReview(ticket, io, options) {
             }
             let activityRegistered = false;
             let registeredReviewProcess = null;
+            let activeReviewRunId = null;
             let finalizeStoppedReviewAfterFailure = null;
             try {
                 registerBundleActivity(acceptedTicket, 'review');
@@ -626,7 +662,17 @@ export function startReview(ticket, io, options) {
                     throw new Error('Ticket project no longer exists');
                 }
                 const cwd = currentTicket.bundle?.kind === 'project_root' ? project.repo_path : resolveWorktreePath(branch, project);
-                const tool = resolveTicketTool(ticket, config.tool ?? 'opencode');
+                const tool = resolveTicketTool(currentTicket, config.tool ?? 'opencode');
+                const reviewerConfig = tool === 'opencode'
+                    ? resolveReviewerConfig(currentTicket.bundle, project, {
+                        model: currentTicket.model,
+                        variant: currentTicket.variant,
+                    })
+                    : {
+                        model: currentTicket.model,
+                        variant: currentTicket.model ? currentTicket.variant : null,
+                        source: 'fallback',
+                    };
                 const reviewProcessState = registerReviewProcess(acceptedTicket.id);
                 registeredReviewProcess = reviewProcessState;
                 const createdAgentRunIds = [];
@@ -679,6 +725,14 @@ export function startReview(ticket, io, options) {
                         if (messageId) {
                             bindTicketThinkingMessage(ticket.id, messageId);
                         }
+                        if (activeReviewRunId) {
+                            finalizeReviewRun({
+                                reviewRunId: activeReviewRunId,
+                                status: 'stopped',
+                                messageId,
+                                failureMessage: 'Review was stopped by the user',
+                            });
+                        }
                         db.prepare('UPDATE tickets SET agent_status = ?, agent_log = NULL, updated_at = CURRENT_TIMESTAMP WHERE id = ? AND project_id = ?')
                             .run('stopped', ticket.id, executionContext.projectId);
                         terminalizeAgentRunsForExecutionRun(executionContext.runId, 'stopped', null);
@@ -723,7 +777,6 @@ export function startReview(ticket, io, options) {
                     finalizeStoppedReview();
                     return true;
                 };
-                beginTicketThinkingRun(ticket.id);
                 withExecutionRunTransaction(transaction => {
                     db.prepare('UPDATE tickets SET agent_status = ?, agent_operation = ?, agent_log = ?, streaming_response = NULL, updated_at = CURRENT_TIMESTAMP WHERE id = ? AND project_id = ?').run('running', 'review', 'Preparing review baseline...', ticket.id, executionContext.projectId);
                     transaction.transitionExecutionRun({
@@ -733,19 +786,59 @@ export function startReview(ticket, io, options) {
                         domainStatus: 'running',
                         eventType: 'review_started',
                     });
+                    activeReviewRunId = createReviewRun({
+                        ticketId: currentTicket.id,
+                        projectId: currentTicket.project_id,
+                        worktreeBundleId: currentTicket.worktree_bundle_id,
+                        diffSignature: null,
+                        workspaceSignature: null,
+                        messageId: null,
+                        model: reviewerConfig.model,
+                        variant: reviewerConfig.variant,
+                        status: 'running',
+                    });
+                    transaction.addExecutionRunLink({
+                        projectId: executionContext.projectId,
+                        runId: executionContext.runId,
+                        link: {
+                            entityType: 'review_run',
+                            entityId: activeReviewRunId,
+                            relationship: 'result',
+                            expectedParent: { entityType: 'ticket', entityId: currentTicket.id },
+                        },
+                    });
                 }, asGenericEmitter(io));
+                beginTicketThinkingRun(ticket.id);
                 updateStatus(ticket.id, 'running', 'Preparing review baseline...', io, 'review');
                 await captureStageBaseline(ticket.id, 'review');
                 if (stopReviewBeforeAgentIfRequested()) {
                     return;
                 }
                 updateStatus(ticket.id, 'running', 'Computing review diff...', io);
-                const diffText = (await getReviewDiffDataAsync({ cwd, branch, project })).diff;
+                const trackedPullRequest = ticket.project_id && ticket.worktree_bundle_id
+                    ? getBundlePullRequestRecord(ticket.worktree_bundle_id, ticket.project_id)
+                    : null;
+                const diffData = await getReviewDiffDataAsync({
+                    cwd,
+                    branch,
+                    project,
+                    pullRequestBaseBranch: trackedPullRequest?.baseBranch ?? null,
+                });
+                const diffText = diffData.diff;
                 if (stopReviewBeforeAgentIfRequested()) {
                     return;
                 }
-                const reviewContext = buildReviewContext({ cwd, branch, project });
+                const reviewContext = buildReviewContext({
+                    cwd,
+                    branch,
+                    project,
+                    diffData,
+                    baseBranch: trackedPullRequest?.baseBranch ?? null,
+                });
                 const diffSignature = computeDiffSignature(diffText);
+                if (!activeReviewRunId || !updateReviewRunPreparation(activeReviewRunId, diffSignature, reviewContext.workspaceSignature)) {
+                    throw new Error('The running review record could not be updated after diff preparation');
+                }
                 const supplementalRequirements = ticket.project_id && ticket.worktree_bundle_id
                     ? formatBuildOpenQuestionAnswersPrompt(listAnsweredBuildOpenQuestionsForBundle(ticket.project_id, ticket.worktree_bundle_id))
                     : '';
@@ -794,20 +887,28 @@ export function startReview(ticket, io, options) {
                     continuation: options?.continuation,
                 });
                 const blindRun = async () => {
-                    return runAgent({
-                        ticketId: ticket.id,
-                        tool,
-                        prompt: blindPrompt,
-                        cwd,
-                        model: ticket.model,
-                        variant: ticket.variant,
-                        resumeSessionId: options?.continuation?.blindSessionId ?? null,
-                        emitSuccessUpdate: false,
-                        executionContext,
-                        io,
-                        runStream: blindRunStreamTarget,
-                        reviewProcessState,
-                    });
+                    try {
+                        return await runAgent({
+                            ticketId: ticket.id,
+                            tool,
+                            prompt: blindPrompt,
+                            cwd,
+                            model: reviewerConfig.model,
+                            variant: reviewerConfig.variant,
+                            resumeSessionId: options?.continuation?.blindSessionId ?? null,
+                            emitSuccessUpdate: false,
+                            executionContext,
+                            io,
+                            runStream: blindRunStreamTarget,
+                            reviewProcessState,
+                        });
+                    }
+                    catch (error) {
+                        if (tool === 'opencode' && reviewerConfig.model && isOpencodeModelResolutionError(error)) {
+                            throw getReviewerExecutionError(error, reviewerConfig.model, reviewerConfig.variant);
+                        }
+                        throw error;
+                    }
                 };
                 const specCoverageRun = async () => {
                     if (!specContext) {
@@ -827,8 +928,8 @@ export function startReview(ticket, io, options) {
                                 continuation: options?.continuation,
                             }),
                             cwd,
-                            model: ticket.model,
-                            variant: ticket.variant,
+                            model: reviewerConfig.model,
+                            variant: reviewerConfig.variant,
                             resumeSessionId: options?.continuation?.specCoverageSessionId ?? null,
                             emitSuccessUpdate: false,
                             executionContext,
@@ -852,8 +953,11 @@ export function startReview(ticket, io, options) {
                     if (blindSettled.status === 'rejected') {
                         throw blindSettled.reason;
                     }
+                    if (specCoverageSettled.status === 'rejected') {
+                        throw specCoverageSettled.reason;
+                    }
                     blindResult = blindSettled.value;
-                    specCoverageResult = specCoverageSettled.status === 'fulfilled' ? specCoverageSettled.value : null;
+                    specCoverageResult = specCoverageSettled.value;
                 }
                 else {
                     blindResult = await blindRun();
@@ -874,30 +978,32 @@ export function startReview(ticket, io, options) {
                     return;
                 }
                 const { normalizedParent: normalizedResponse, parsed: blindParsed, } = resolveStructuredReviewOutput(finalResponse, blindResult.transcriptRuns ?? []);
-                let reviewRunId = null;
+                const constrainedBlindParsed = blindParsed
+                    ? constrainReviewOutputToDiff(blindParsed, diffText)
+                    : null;
+                let reviewRunId = activeReviewRunId;
                 let parsedFindings = [];
                 let parsedStatus = null;
                 let persistedMessageId = null;
                 let reviewContent = null;
                 let findingsToPersist = [];
-                if (blindParsed?.cleanedMarkdown) {
-                    const specCoverageSpecOutput = typeof specCoverageResult?.response === 'string'
-                        ? stripSystemReminderBlocks(specCoverageResult.response)
-                        : '';
-                    const specParsed = specCoverageSpecOutput
-                        ? parseStructuredReviewOutput(specCoverageSpecOutput)
+                if (constrainedBlindParsed?.status === 'completed') {
+                    const specOutput = resolveStructuredReviewOutput(specCoverageResult?.response ?? null, specCoverageResult?.transcriptRuns ?? [], 'spec-coverage').parsed;
+                    const specParsed = specOutput
+                        ? constrainReviewOutputToDiff(specOutput, diffText)
                         : null;
                     const mergedFindings = specParsed && specParsed.status === 'completed'
-                        ? [...blindParsed.findings, ...specParsed.findings.map(finding => ({
+                        ? [...constrainedBlindParsed.findings, ...specParsed.findings.map(finding => ({
                                 ...finding,
-                                ordinal: finding.ordinal + blindParsed.findings.length,
+                                ordinal: finding.ordinal + constrainedBlindParsed.findings.length,
                             }))]
-                        : blindParsed.findings;
-                    const combinedMarkdown = specParsed && specParsed.cleanedMarkdown
-                        ? `${blindParsed.cleanedMarkdown}\n\n## Spec-Coverage Review\n${specParsed.cleanedMarkdown}`
-                        : blindParsed.cleanedMarkdown;
+                        : constrainedBlindParsed.findings;
+                    const blindMarkdown = formatParsedReviewMarkdown(constrainedBlindParsed);
+                    const combinedMarkdown = specParsed?.status === 'completed'
+                        ? `${blindMarkdown}\n\n## Spec-Coverage Review\n${formatParsedReviewMarkdown(specParsed)}`
+                        : blindMarkdown;
                     parsedFindings = mergedFindings;
-                    parsedStatus = blindParsed.status;
+                    parsedStatus = constrainedBlindParsed.status;
                     reviewContent = combinedMarkdown;
                     findingsToPersist = mergedFindings;
                 }
@@ -922,19 +1028,15 @@ export function startReview(ticket, io, options) {
                             transcriptRuns: transcriptRuns.length > 0 ? transcriptRuns : null,
                         })
                         : null;
-                    let createdReviewRunId = null;
-                    if (messageId && reviewContent && parsedStatus) {
-                        createdReviewRunId = createReviewRun({
-                            ticketId: ticket.id,
-                            projectId: ticket.project_id,
-                            worktreeBundleId: ticket.worktree_bundle_id,
-                            diffSignature,
-                            workspaceSignature: reviewContext.workspaceSignature,
+                    if (activeReviewRunId) {
+                        finalizeReviewRun({
+                            reviewRunId: activeReviewRunId,
+                            status: parsedStatus ?? 'parse_failed',
                             messageId,
-                            status: parsedStatus,
+                            failureMessage: parseFailed ? 'Review output could not be parsed' : null,
                         });
                         if (findingsToPersist.length > 0) {
-                            insertReviewFindings(createdReviewRunId, findingsToPersist);
+                            insertReviewFindings(activeReviewRunId, findingsToPersist);
                         }
                     }
                     if (messageId) {
@@ -949,18 +1051,6 @@ export function startReview(ticket, io, options) {
                             },
                         });
                     }
-                    if (createdReviewRunId) {
-                        transaction.addExecutionRunLink({
-                            projectId: executionContext.projectId,
-                            runId: executionContext.runId,
-                            link: {
-                                entityType: 'review_run',
-                                entityId: createdReviewRunId,
-                                relationship: 'result',
-                                expectedParent: { entityType: 'ticket', entityId: ticket.id },
-                            },
-                        });
-                    }
                     db.prepare('UPDATE tickets SET agent_status = ?, agent_log = NULL, updated_at = CURRENT_TIMESTAMP WHERE id = ? AND project_id = ?')
                         .run('done', ticket.id, executionContext.projectId);
                     terminalizeAgentRunsForExecutionRun(executionContext.runId, parseFailed ? 'error' : 'done', parseFailed ? 'Review output could not be parsed' : null);
@@ -972,7 +1062,7 @@ export function startReview(ticket, io, options) {
                         ...(parseFailed ? { failure: { code: 'parse_failed', message: 'Review output could not be parsed' } } : {}),
                         eventType: parseFailed ? 'review_parse_failed' : 'review_completed',
                     });
-                    return { messageId, reviewRunId: createdReviewRunId };
+                    return { messageId, reviewRunId: activeReviewRunId };
                 }, asGenericEmitter(io));
                 persistedMessageId = persisted.messageId;
                 reviewRunId = persisted.reviewRunId;
@@ -990,8 +1080,8 @@ export function startReview(ticket, io, options) {
                 if (tool === 'claude' && usageSnapshot) {
                     storeTicketUsageSnapshotDirect(ticket.id, usageSnapshot);
                 }
-                emitAgentRunsUpdated(asGenericEmitter(io), 'ticket', ticket.id);
                 updateStatus(ticket.id, 'done', null, io);
+                emitAgentRunsUpdated(asGenericEmitter(io), 'ticket', ticket.id);
                 const completedWithMessage = getTicket(ticket.id);
                 if (completedWithMessage) {
                     io.emit('ticket:updated', completedWithMessage);
@@ -1075,8 +1165,9 @@ export function startReview(ticket, io, options) {
                 }
                 try {
                     withExecutionRunTransaction(transaction => {
+                        let errorMessageId = null;
                         if (errorContent) {
-                            const messageId = createTicketMessage({
+                            errorMessageId = createTicketMessage({
                                 ticketId: acceptedTicket.id,
                                 role: 'assistant',
                                 kind: 'error',
@@ -1088,7 +1179,7 @@ export function startReview(ticket, io, options) {
                                 runId: executionContext.runId,
                                 link: {
                                     entityType: 'ticket_message',
-                                    entityId: messageId,
+                                    entityId: errorMessageId,
                                     relationship: 'result',
                                     expectedParent: { entityType: 'ticket', entityId: acceptedTicket.id },
                                 },
@@ -1097,6 +1188,14 @@ export function startReview(ticket, io, options) {
                         db.prepare('UPDATE tickets SET agent_status = ?, agent_log = ?, streaming_response = NULL, updated_at = CURRENT_TIMESTAMP WHERE id = ? AND project_id = ?')
                             .run('error', message, acceptedTicket.id, executionContext.projectId);
                         terminalizeAgentRunsForExecutionRun(executionContext.runId, 'error', message);
+                        if (activeReviewRunId) {
+                            finalizeReviewRun({
+                                reviewRunId: activeReviewRunId,
+                                status: 'error',
+                                messageId: errorMessageId,
+                                failureMessage: message,
+                            });
+                        }
                         transaction.transitionExecutionRun({
                             projectId: executionContext.projectId,
                             runId: executionContext.runId,
@@ -1123,6 +1222,13 @@ export function startReview(ticket, io, options) {
                     const currentExecutionRun = getExecutionRun(executionContext.projectId, executionContext.runId);
                     if (currentExecutionRun && ['queued', 'running', 'waiting'].includes(currentExecutionRun.status)) {
                         terminalizeAgentRunsForExecutionRun(executionContext.runId, 'error', message);
+                        if (activeReviewRunId) {
+                            finalizeReviewRun({
+                                reviewRunId: activeReviewRunId,
+                                status: 'error',
+                                failureMessage: message,
+                            });
+                        }
                         transitionExecutionRun({
                             projectId: executionContext.projectId,
                             runId: executionContext.runId,