@doist/doistbot-cli 1.0.8 → 1.0.10

package/sandbox/dist/tasks/issue-triage/routing.js

@@ -15,21 +15,51 @@ function hasProductLabel(labels) {
 function stripMarkdownHeading(line) {
     return line.replace(/^#{1,6}\s+/, '').trim();
 }
-export function summarizeTriageReasoning(triageOutput, maxChars = 220) {
-    const summaryHeading = triageOutput.match(/^## Summary\s*$/im);
-    const summaryBody = summaryHeading?.index === undefined
-        ? triageOutput
-        : triageOutput.slice(summaryHeading.index + summaryHeading[0].length);
-    const nextHeading = summaryBody.search(/^##\s/m);
-    const candidate = nextHeading >= 0 ? summaryBody.slice(0, nextHeading) : summaryBody;
+function stripListMarker(line) {
+    return line.replace(/^\s*(?:\d+[.)]|[-*])\s+/, '').trim();
+}
+// `undefined` means the heading is absent; `''` means it is present but empty.
+// Callers depend on the difference — an empty `## Summary` has its own fallback.
+function sectionBody(triageOutput, heading) {
+    const match = triageOutput.match(new RegExp(`^## ${heading}\\s*$`, 'im'));
+    if (match?.index === undefined) {
+        return undefined;
+    }
+    const rest = triageOutput.slice(match.index + match[0].length);
+    const nextHeading = rest.search(/^##\s/m);
+    return (nextHeading >= 0 ? rest.slice(0, nextHeading) : rest).trim();
+}
+// The single sentence-boundary heuristic. Returns whole sentences or nothing —
+// a half-sentence of context is what made the old routing comment worthless, so
+// detail that doesn't fit is dropped, not clipped. Any complete sentence is worth
+// keeping, however short: the boundary check already rules out fragments.
+function sentencesWithin(text, maxChars) {
+    if (text.length <= maxChars) {
+        return text;
+    }
+    const clipped = text.slice(0, maxChars);
+    const sentenceEnd = Math.max(clipped.lastIndexOf('. '), clipped.lastIndexOf('! '), clipped.lastIndexOf('? '));
+    return sentenceEnd > 0 ? clipped.slice(0, sentenceEnd + 1) : undefined;
+}
+// As above, but for text we would rather shorten than lose. Falls back to a word
+// boundary + ellipsis when the budget holds no complete sentence at all.
+function trimToSentence(text, maxChars) {
+    const whole = sentencesWithin(text, maxChars);
+    if (whole !== undefined) {
+        return whole;
+    }
+    const capped = text.slice(0, maxChars - 3);
+    const wordEnd = capped.lastIndexOf(' ');
+    return `${(wordEnd > 0 ? capped.slice(0, wordEnd) : capped).trimEnd()}...`;
+}
+export function summarizeTriageReasoning(triageOutput, maxChars = 320) {
+    const candidate = sectionBody(triageOutput, 'Summary') ?? triageOutput;
     const firstParagraph = candidate
         .trim()
         .split(/\n\s*\n/)
         .map((paragraph) => paragraph
         .split('\n')
-        .map((line) => stripMarkdownHeading(line)
-        .replace(/^[-*]\s+/, '')
-        .trim())
+        .map((line) => stripListMarker(stripMarkdownHeading(line)))
         .filter(Boolean)
         .join(' '))
         .find(Boolean) ?? '';
@@ -37,12 +67,89 @@ export function summarizeTriageReasoning(triageOutput, maxChars = 220) {
     if (!normalized) {
         return 'See the triage summary above for details.';
     }
-    if (normalized.length <= maxChars) {
-        return normalized;
+    return trimToSentence(normalized, maxChars);
+}
+const NEXT_STEP_MAX_CHARS = 300;
+// Anchored on purpose: the verdict is the leading word of the line ("Medium — ...").
+// A loose substring search matches "low" inside "low-volume", "workflow" and "flow",
+// which turned ~25% of agreeing issues into false mismatch alerts.
+const SEVERITY_VERDICT = /^\W*(?:priority[:\s]+)?(critical|high|medium|low)\b/i;
+function severityRank(value) {
+    return value.match(SEVERITY_VERDICT)?.[1]?.toLowerCase();
+}
+function labelSeverity(labels) {
+    const label = labels.find((l) => l.trim().toLowerCase().startsWith('severity:'));
+    return label ? severityRank(label.split(':')[1] ?? '') : undefined;
+}
+function firstLineOf(triageOutput, heading) {
+    return sectionBody(triageOutput, heading)
+        ?.split('\n')
+        .map((line) => stripListMarker(line))
+        .find(Boolean);
+}
+// Triage agrees with the `Severity:*` label ~93% of the time, so restating the
+// assessed priority is noise. The 7% where it disagrees is the part worth reading
+// — and it skews toward triage rating the issue *worse* than the label.
+function buildSeverityMismatchLine(triageOutput, labels, maxChars) {
+    const priority = firstLineOf(triageOutput, 'Issue Priority Assessment');
+    const assessed = priority ? severityRank(priority) : undefined;
+    const labelled = labelSeverity(labels);
+    if (!priority || !assessed || !labelled || assessed === labelled) {
+        return undefined;
+    }
+    const alert = `Labelled \`Severity:${labelled}\`, triage assessed ${assessed}`;
+    const rationale = sentencesWithin(priority.replace(/^\W*\w+\s*[—–-]\s*/, '').trim(), maxChars);
+    return rationale ? `- ⚠️ **${alert}** — ${rationale}` : `- ⚠️ **${alert}.**`;
+}
+// The ping only fires when the fix pipeline declined the issue, but the team can't
+// see that from the outside. Naming the state plus why no PR exists saves them from
+// re-deriving it, and turns `insufficient-context` into an explicit ask.
+function buildStatusLine(assessment, maxChars) {
+    if (!assessment) {
+        return undefined;
     }
-    return `${normalized.slice(0, maxChars - 3).trimEnd()}...`;
+    const state = assessment.classification === 'insufficient-context'
+        ? 'Blocked, no PR opened'
+        : assessment.classification === 'needs-human'
+            ? 'Needs a human, no PR opened'
+            : `No PR opened (${assessment.confidence} confidence)`;
+    const scope = sentencesWithin(assessment.scope.replace(/\s+/g, ' ').trim(), maxChars);
+    return scope ? `- **${state}** — ${scope}` : `- **${state}.**`;
 }
-export function buildHeroRoutingComment({ labels, primaryRepo, triageOutput, }) {
+// ~1 in 6 triages concludes the code lives outside the repos the labels resolved to.
+// Without this the pinged team opens the issue, decides it isn't theirs, and bounces it.
+function buildRepoMismatchLine(assessment, resolvedRepos) {
+    const targetRepo = assessment?.targetRepo?.trim();
+    if (!targetRepo || resolvedRepos.length === 0) {
+        return undefined;
+    }
+    const matches = resolvedRepos.some((repo) => repo.toLowerCase() === targetRepo.toLowerCase());
+    if (matches) {
+        return undefined;
+    }
+    return `- ⚠️ **Likely in \`${targetRepo}\`**, not ${resolvedRepos.map((repo) => `\`${repo}\``).join(' / ')} — may belong to another team.`;
+}
+/**
+ * Compact recap for the hero routing ping. The full triage comment sits directly
+ * above it with `## Summary` visible, so this deliberately carries only what that
+ * summary doesn't: the exceptions worth alerting on, and the next concrete action.
+ */
+export function buildTriageRecap({ triageOutput, assessment, labels = [], resolvedRepos = [], maxChars = 220, }) {
+    const nextStep = firstLineOf(triageOutput, 'Suggested Triage Action');
+    const lines = [
+        buildSeverityMismatchLine(triageOutput, labels, maxChars),
+        buildRepoMismatchLine(assessment, resolvedRepos),
+        buildStatusLine(assessment, maxChars),
+        // The action is the payload, so it gets a wider budget than supporting
+        // detail and keeps ellipsis as a last resort rather than being dropped.
+        nextStep && `- **Next step:** ${trimToSentence(nextStep, NEXT_STEP_MAX_CHARS)}`,
+    ].filter(Boolean);
+    if (lines.length === 0) {
+        return summarizeTriageReasoning(triageOutput);
+    }
+    return lines.join('\n');
+}
+export function buildHeroRoutingComment({ labels, primaryRepo, triageOutput, assessment, resolvedRepos, }) {
     if (!hasProductLabel(labels)) {
         return CX_NO_PRODUCT_LABEL_COMMENT;
     }
@@ -52,7 +159,9 @@ export function buildHeroRoutingComment({ labels, primaryRepo, triageOutput, })
     }
     return `${HERO_ROUTING_MARKER}
 
-Routing to ${heroGroup.mention} for triage. ${summarizeTriageReasoning(triageOutput)}`;
+Routing to ${heroGroup.mention} for triage.
+
+${buildTriageRecap({ triageOutput, assessment, labels, resolvedRepos })}`;
 }
 export function shouldPostHeroRoutingComment(fixAssessment) {
     if (!fixAssessment) {

package/sandbox/dist/tasks/issue-triage/triage.js

@@ -289,6 +289,8 @@ async function runIssueTriagePipeline(ctx, mode) {
                         ? resolution.repositories[0]
                         : undefined,
                     triageOutput: validated.output,
+                    assessment: fixAssessment,
+                    resolvedRepos: resolution.repositories.map((repo) => repo.fullName),
                 });
                 await postTriageRouting({
                     octokit: writeOctokit,

package/sandbox/dist/tasks/review/engines/multi-focus.js

@@ -10,22 +10,30 @@ import { buildGeminiFlashReviewSummaryModel } from '../summary-model.js';
 import { dedupeMultiFocusFindings } from './dedupe.js';
 import { buildGeminiFlashCredential, getApiKeys, getAvailableModels, getEnabledReviewModels, getOpenRouterReviewCredential, getProviderFailures, requestModelReview, } from './shared.js';
 const DEFAULT_REVIEW_MODELS_BY_FOCUS = {
-    general: ['deepseek', 'zai', 'openai'],
-    quality: ['deepseek', 'zai', 'openai'],
-    efficiency: ['deepseek', 'zai', 'openai'],
-    tests: ['deepseek', 'zai'],
-    standards: ['deepseek', 'zai'],
+    general: ['deepseek', 'grok', 'openai'],
+    security: ['deepseek', 'grok', 'openai'],
+    reuse: ['deepseek', 'grok', 'openai'],
+    quality: ['deepseek', 'grok', 'openai'],
+    efficiency: ['deepseek', 'grok', 'openai'],
+    tests: ['deepseek', 'grok'],
+    standards: ['deepseek', 'grok'],
 };
 // Default Gemini-flash pass on failure. tests/standards keep no fallback unless
 // a caller opts into a wider fallback set.
 const DEFAULT_GEMINI_FLASH_FALLBACK_FOCUSES = [
     'general',
+    'security',
+    'reuse',
     'quality',
     'efficiency',
 ];
 // Open models whose failed pass triggers that fallback.
-const GEMINI_FLASH_FALLBACK_MODELS = ['deepseek', 'zai'];
-const LOW_PRIORITY_SUMMARY_NOTICE = 'I also included a few optional follow-up notes in the details below.';
+const GEMINI_FLASH_FALLBACK_MODELS = ['deepseek', 'grok'];
+function lowPrioritySummaryNotice(count) {
+    return count === 1
+        ? 'I also left one optional follow-up note in the details below.'
+        : 'I also included a few optional follow-up notes in the details below.';
+}
 // Workflow phase orchestrating the council fan-out; one model span per pass.
 const MULTI_FOCUS_PHASE = 'review.multi-focus';
 const FOCUS_PASS_PHASE = 'review.focus-pass';
@@ -203,18 +211,18 @@ function appendSummaryOnlyFindings(summary, summaryFindings) {
         return summary;
     }
     const title = `Optional follow-up note${notes.length === 1 ? '' : 's'} (${notes.length})`;
-    return `${summary.trim()}\n\n${LOW_PRIORITY_SUMMARY_NOTICE}\n\n<details>\n<summary>${title}</summary>\n\n${notes.join('\n')}\n\n</details>`;
+    return `${summary.trim()}\n\n${lowPrioritySummaryNotice(notes.length)}\n\n<details>\n<summary>${title}</summary>\n\n${notes.join('\n')}\n\n</details>`;
 }
-function buildSummarySynthesisFindings({ inlineFindings, summaryFindings, }) {
-    return Array.from(new Set([
-        ...inlineFindings.map((finding) => finding.body.trim()),
-        ...summaryFindings.map((finding) => formatSummaryOnlyFinding(finding)),
-    ].filter(Boolean)));
+// Summary-only (P3) findings are deliberately excluded: they already render in
+// the "Optional follow-up note" details block, and feeding them here made the
+// summary restate a single P3 nit under "Few things worth tightening:".
+function buildSummarySynthesisFindings(inlineFindings) {
+    return Array.from(new Set(inlineFindings.map((finding) => finding.body.trim()).filter(Boolean)));
 }
 async function buildMultiFocusReviewOutput({ inlineFindings, summaryFindings, lineLookup, prTitle, prBody, reviewObservations, summaryModel, summaryFallbackModels, workspaceDir, }) {
     const comments = convertFindingsToComments(inlineFindings, lineLookup);
     const inlineCommentFindings = Array.from(new Set(comments.map((comment) => comment.body.trim()).filter(Boolean)));
-    const findings = buildSummarySynthesisFindings({ inlineFindings, summaryFindings });
+    const findings = buildSummarySynthesisFindings(inlineFindings);
     const hasReviewFindings = inlineFindings.length > 0 || summaryFindings.length > 0;
     if (!hasReviewFindings) {
         logger.info('Review summary synthesis skipped for clean review', {
@@ -339,7 +347,7 @@ export async function runMultiFocusReviewEngine({ ctx, prompt, lineLookup, prTit
         const routedFindings = routeLowPriorityFindings(dedupedFindings, lowPriorityFindingPlacement);
         const geminiSummaryFallback = buildGeminiFlashReviewSummaryModel(apiKeys.gemini);
         const summaryModel = ctx.reviewSummaryModel ??
-            (apiKeys.zai ? { provider: 'zai', credential: apiKeys.zai } : undefined);
+            (apiKeys.grok ? { provider: 'grok', credential: apiKeys.grok } : undefined);
         const review = await buildMultiFocusReviewOutput({
             inlineFindings: routedFindings.inlineFindings,
             summaryFindings: routedFindings.summaryFindings,

package/sandbox/dist/tasks/review/engines/shared.js

@@ -22,6 +22,8 @@ export function getAvailableModels(apiKeys) {
         models.push('openai');
     if (hasReviewProviderCredential(apiKeys.deepseek))
         models.push('deepseek');
+    if (hasReviewProviderCredential(apiKeys.grok))
+        models.push('grok');
     if (hasReviewProviderCredential(apiKeys.zai))
         models.push('zai');
     return models;

package/sandbox/dist/tasks/review/multi-focus-prompt.js

@@ -14,6 +14,8 @@ const MULTI_FOCUS_BASE_PROMPT_FILE = `${REVIEW_PROMPTS_DIR}/review-multi-focus-b
 const REVIEW_FOCUS_PROMPTS_DIR = `${REVIEW_PROMPTS_DIR}/review-focus-prompts`;
 export const REVIEW_FOCUS_DEFINITIONS = [
     { name: 'general', title: 'General', fileName: 'general.md' },
+    { name: 'security', title: 'Security', fileName: 'security.md' },
+    { name: 'reuse', title: 'Reuse', fileName: 'reuse.md' },
     { name: 'quality', title: 'Quality', fileName: 'quality.md' },
     { name: 'efficiency', title: 'Efficiency', fileName: 'efficiency.md' },
     { name: 'tests', title: 'Tests', fileName: 'tests.md' },

package/sandbox/dist/tasks/review/review-summary.js

@@ -39,7 +39,7 @@ ${findingsBlock}
 ${observationsBlock}
 </review_observations>
 
-Write the complete GitHub review summary body. Synthesize the findings into grouped themes—do not mirror every inline comment one-for-one or reference the reviewers, models, focus passes, or review observations. Use review observations only as context for the overview. Do not mention issues not present in the <review-finding> blocks. If there are no <review-finding> blocks, state that no issues were flagged.
+Write the complete GitHub review summary body. Synthesize the findings into grouped themes—do not mirror every inline comment one-for-one or reference the reviewers, models, focus passes, or review observations. Use review observations only as context for the overview. Do not mention issues not present in the <review-finding> blocks. If there are no <review-finding> blocks, state that no inline issues were flagged.
 
 Structure:
 - Start with one short sentence giving a concise overview of what changed.
@@ -51,7 +51,7 @@ Keep it human and compact: the opener should be 1 sentence, or 2 only if the PR
 Respond with only valid JSON in this exact shape:
 {"summary":"<your summary>","comments":[]}`;
 }
-export async function summarizeReviewFindings({ credential, provider = 'zai', modelName, fallbackModels = [], prTitle, prBody, findings, reviewObservations, workspaceDir, requestReview = requestReviewSummary, }) {
+export async function summarizeReviewFindings({ credential, provider = 'grok', modelName, fallbackModels = [], prTitle, prBody, findings, reviewObservations, workspaceDir, requestReview = requestReviewSummary, }) {
     const prompt = buildReviewSummaryPrompt({
         prTitle,
         prBody,

package/sandbox/dist/tasks/review/review.js

@@ -15,7 +15,8 @@ import { REVIEW_FOCUS_DEFINITIONS } from './multi-focus-prompt.js';
 import { appendReviewFeedbackLink } from './review-feedback.js';
 import { runSharedReview } from './runner.js';
 const REVIEW_EVENT = 'COMMENT';
-const REQUIRED_PRODUCTION_REVIEW_MODELS = getDefaultReviewModels();
+const DEFAULT_PRODUCTION_REVIEW_MODELS = getDefaultReviewModels();
+const SUPPORTED_PRODUCTION_REVIEW_MODELS = ['openai', 'deepseek', 'grok', 'zai'];
 const REVIEW_FOCUSES = REVIEW_FOCUS_DEFINITIONS.map((focus) => focus.name);
 function optionalEnv(name) {
     const value = process.env[name]?.trim();
@@ -43,8 +44,8 @@ function parseContext() {
 export function parseReviewModelsEnv(value) {
     return parseCsvAllowList({
         value,
-        allowedValues: REQUIRED_PRODUCTION_REVIEW_MODELS,
-        fallback: REQUIRED_PRODUCTION_REVIEW_MODELS,
+        allowedValues: SUPPORTED_PRODUCTION_REVIEW_MODELS,
+        fallback: DEFAULT_PRODUCTION_REVIEW_MODELS,
         envName: 'REVIEW_MODELS_ENABLED',
     });
 }
@@ -253,6 +254,7 @@ async function main() {
             }
             if (ctx.openRouterApiKey) {
                 reviewApiKeys.deepseek = createOpenRouterCredential(ctx.openRouterApiKey);
+                reviewApiKeys.grok = createOpenRouterCredential(ctx.openRouterApiKey);
                 reviewApiKeys.zai = createOpenRouterCredential(ctx.openRouterApiKey);
             }
             if (ctx.openaiApiKey) {

package/sandbox/dist/tasks/review/summary-model.js

@@ -6,7 +6,7 @@ import { genAiProviderName, withPhase } from '../../core/tracing.js';
 import { resolveReviewProviderCredential } from '../../providers/credentials.js';
 import { extractJsonCandidatesFromResponse } from '../../providers/helpers.js';
 const REVIEW_SUMMARY_TIMEOUT_MS = 5 * 60 * 1000;
-const REVIEW_SUMMARY_THINKING_LEVEL = 'medium';
+const DEFAULT_REVIEW_SUMMARY_THINKING_LEVEL = 'medium';
 // One model span per summary attempt; a fallback chain emits one span per model.
 const SUMMARY_PHASE = 'review.summary';
 const DIFF_SIDE = {
@@ -56,7 +56,7 @@ function getRequestFailureReason(result) {
         ? 'no_assistant_response'
         : 'request_failed';
 }
-export async function requestReviewSummary({ credential, provider = 'zai', modelName, prompt, workspaceDir, }) {
+export async function requestReviewSummary({ credential, provider = 'grok', modelName, prompt, workspaceDir, }) {
     const resolvedCredential = resolveReviewProviderCredential(credential);
     if (!resolvedCredential) {
         return failedReviewSummaryRequest({ reason: 'missing_credential' });
@@ -93,12 +93,13 @@ export async function requestReviewSummary({ credential, provider = 'zai', model
 }
 async function runReviewSummaryRequest({ provider, resolvedCredential, resolvedModelName, logModelName, prompt, workspaceDir, }) {
     try {
+        const thinkingLevel = provider === 'grok' ? 'high' : DEFAULT_REVIEW_SUMMARY_THINKING_LEVEL;
         logger.info('Review summary model request', {
             provider,
             modelName: logModelName,
             promptLength: prompt.length,
             timeoutMs: REVIEW_SUMMARY_TIMEOUT_MS,
-            thinkingLevel: REVIEW_SUMMARY_THINKING_LEVEL,
+            thinkingLevel,
         });
         const result = await invokePiPrompt({
             provider,
@@ -111,7 +112,7 @@ async function runReviewSummaryRequest({ provider, resolvedCredential, resolvedM
             cwd: workspaceDir ?? getWorkspaceDir(),
             timeoutMs: REVIEW_SUMMARY_TIMEOUT_MS,
             capabilityPreset: 'read-only',
-            thinkingLevel: REVIEW_SUMMARY_THINKING_LEVEL,
+            thinkingLevel,
         });
         if (!result.ok) {
             const reason = getRequestFailureReason(result);

package/sandbox/dist/tasks/review/thinking-level.js

@@ -4,6 +4,8 @@ export function getReviewThinkingLevel(model) {
             return 'high';
         case 'openai':
             return 'high';
+        case 'grok':
+            return 'high';
         case 'zai':
             return 'high';
         case 'gemini':

package/sandbox/dist/tasks/review-slice-plan/context.js

@@ -0,0 +1,69 @@
+import { ReviewHygieneDecisionKind, ReviewHygieneExclusionReason, ReviewHygieneReasonCode, } from 'doistbot-async-routing-contracts';
+import { optionalPositiveInt, parseBaseContext, requiredEnv, } from '../../core/shared.js';
+export function parseReviewSlicePlanContext(env = process.env) {
+    const dryRun = env.DRY_RUN?.toLowerCase() === 'true';
+    const commentId = optionalPositiveInt(env.COMMENT_ID);
+    if (!dryRun && commentId === undefined) {
+        throw new Error('Missing required environment variable: COMMENT_ID');
+    }
+    return {
+        ...parseBaseContext(env),
+        baseBranch: requiredEnv('BASE_BRANCH', env),
+        headSha: requiredEnv('HEAD_SHA', env),
+        commentId,
+        runId: requiredEnv('SLICE_PLAN_RUN_ID', env),
+        reviewHygiene: parseReviewHygieneDecision(requiredEnv('REVIEW_HYGIENE_DECISION', env)),
+        geminiApiKey: requiredEnv('GEMINI_API_KEY', env),
+        dryRun,
+    };
+}
+function parseReviewHygieneDecision(serialized) {
+    let value;
+    try {
+        value = JSON.parse(serialized);
+    }
+    catch (error) {
+        throw new Error(`Invalid REVIEW_HYGIENE_DECISION JSON: ${error instanceof Error ? error.message : String(error)}`);
+    }
+    if (!isRecord(value)) {
+        throw new Error('Invalid REVIEW_HYGIENE_DECISION: expected an object');
+    }
+    if (value.kind !== ReviewHygieneDecisionKind.Warn &&
+        value.kind !== ReviewHygieneDecisionKind.Block) {
+        throw new Error('Invalid REVIEW_HYGIENE_DECISION: expected warn or block');
+    }
+    if (!isReviewHygieneStats(value.stats) ||
+        !Array.isArray(value.reasons) ||
+        !value.reasons.every(isReviewHygieneReason)) {
+        throw new Error('Invalid REVIEW_HYGIENE_DECISION: missing stats or reasons');
+    }
+    if (value.exclusionSummary !== undefined &&
+        !isReviewHygieneExclusionSummary(value.exclusionSummary)) {
+        throw new Error('Invalid REVIEW_HYGIENE_DECISION: invalid exclusion summary');
+    }
+    return value;
+}
+function isReviewHygieneReason(value) {
+    return (isRecord(value) &&
+        Object.values(ReviewHygieneReasonCode).includes(value.code) &&
+        isFiniteNumber(value.actual) &&
+        isFiniteNumber(value.limit));
+}
+function isReviewHygieneExclusionSummary(value) {
+    return (isRecord(value) &&
+        ['files', 'additions', 'deletions', 'changedLines', 'reviewLoadLines'].every((key) => isFiniteNumber(value[key])) &&
+        Array.isArray(value.reasons) &&
+        value.reasons.every((reason) => Object.values(ReviewHygieneExclusionReason).includes(reason)));
+}
+function isFiniteNumber(value) {
+    return typeof value === 'number' && Number.isFinite(value);
+}
+function isReviewHygieneStats(value) {
+    if (!isRecord(value)) {
+        return false;
+    }
+    return ['additions', 'deletions', 'changedFiles', 'changedLines', 'reviewLoadLines'].every((key) => isFiniteNumber(value[key]));
+}
+function isRecord(value) {
+    return typeof value === 'object' && value !== null && !Array.isArray(value);
+}

package/sandbox/dist/tasks/review-slice-plan/plan.js

@@ -0,0 +1,209 @@
+import { Ajv } from 'ajv';
+import { extractJsonCandidatesFromResponse } from '../../providers/helpers.js';
+export const ReviewSliceStrategy = {
+    Stack: 'stack',
+    Independent: 'independent',
+    Hybrid: 'hybrid',
+    NoSafeSplit: 'no_safe_split',
+};
+const REVIEW_SLICE_PLAN_SCHEMA = {
+    type: 'object',
+    additionalProperties: false,
+    properties: {
+        strategy: {
+            type: 'string',
+            enum: Object.values(ReviewSliceStrategy),
+        },
+        summary: { type: 'string', minLength: 1, maxLength: 1_000 },
+        slices: {
+            type: 'array',
+            maxItems: 12,
+            items: {
+                type: 'object',
+                additionalProperties: false,
+                properties: {
+                    title: { type: 'string', minLength: 1, maxLength: 160 },
+                    purpose: { type: 'string', minLength: 1, maxLength: 800 },
+                    files: {
+                        type: 'array',
+                        minItems: 1,
+                        items: { type: 'string', minLength: 1, maxLength: 500 },
+                    },
+                    changes: {
+                        type: 'array',
+                        minItems: 1,
+                        maxItems: 8,
+                        items: { type: 'string', minLength: 1, maxLength: 800 },
+                    },
+                    validation: {
+                        type: 'array',
+                        minItems: 1,
+                        maxItems: 8,
+                        items: { type: 'string', minLength: 1, maxLength: 500 },
+                    },
+                    dependsOn: {
+                        anyOf: [{ type: 'integer', minimum: 1, maximum: 12 }, { type: 'null' }],
+                    },
+                },
+                required: ['title', 'purpose', 'files', 'changes', 'validation', 'dependsOn'],
+            },
+        },
+        caveats: {
+            type: 'array',
+            maxItems: 8,
+            items: { type: 'string', minLength: 1, maxLength: 800 },
+        },
+    },
+    required: ['strategy', 'summary', 'slices', 'caveats'],
+};
+const ajv = new Ajv();
+const validateReviewSlicePlanSchema = ajv.compile(REVIEW_SLICE_PLAN_SCHEMA);
+const AGENT_SPLIT_SUGGESTION = 'You can use your agent of choice (Codex/Claude etc) to help you split this PR 😊 Just copy the link to this comment and ask them `Can you please create a PR stack based on the suggestions in this comment`';
+export function parseReviewSlicePlanOutput(output, changedFiles) {
+    const candidates = extractJsonCandidatesFromResponse(output);
+    let lastError = 'Model response did not contain a JSON object';
+    for (const candidate of candidates) {
+        try {
+            const parsed = JSON.parse(candidate);
+            if (!validateReviewSlicePlanSchema(parsed)) {
+                lastError = `Model response did not match the slice-plan schema: ${ajv.errorsText(validateReviewSlicePlanSchema.errors)}`;
+                continue;
+            }
+            const plan = parsed;
+            validateReviewSlicePlanSemantics(plan, changedFiles);
+            return plan;
+        }
+        catch (error) {
+            lastError = error instanceof Error ? error.message : String(error);
+        }
+    }
+    throw new Error(lastError);
+}
+export function validateReviewSlicePlanSemantics(plan, changedFiles) {
+    if (plan.strategy === ReviewSliceStrategy.NoSafeSplit) {
+        if (plan.slices.length !== 0) {
+            throw new Error('A no_safe_split plan must not include slices');
+        }
+        if (plan.caveats.length === 0) {
+            throw new Error('A no_safe_split plan must explain the blocking coupling in caveats');
+        }
+        return;
+    }
+    if (plan.slices.length < 2) {
+        throw new Error('A slicing plan must contain at least two slices');
+    }
+    const changedPaths = new Set(changedFiles.map((file) => file.filename));
+    const coveredPaths = new Set();
+    const titles = new Set();
+    for (const [index, slice] of plan.slices.entries()) {
+        const sliceNumber = index + 1;
+        const normalizedTitle = slice.title.trim().toLowerCase();
+        if (titles.has(normalizedTitle)) {
+            throw new Error(`Slice ${sliceNumber} duplicates another slice title`);
+        }
+        titles.add(normalizedTitle);
+        const uniqueSlicePaths = new Set(slice.files);
+        if (uniqueSlicePaths.size !== slice.files.length) {
+            throw new Error(`Slice ${sliceNumber} contains duplicate file paths`);
+        }
+        for (const file of slice.files) {
+            if (!changedPaths.has(file)) {
+                throw new Error(`Slice ${sliceNumber} references an unchanged file: ${file}`);
+            }
+            coveredPaths.add(file);
+        }
+        if (slice.dependsOn !== null && slice.dependsOn >= sliceNumber) {
+            throw new Error(`Slice ${sliceNumber} must only depend on an earlier slice`);
+        }
+    }
+    const uncoveredPaths = [...changedPaths].filter((path) => !coveredPaths.has(path));
+    if (uncoveredPaths.length > 0) {
+        throw new Error(`Slice plan does not cover ${uncoveredPaths.length} changed file(s): ${uncoveredPaths
+            .slice(0, 10)
+            .join(', ')}`);
+    }
+    if (plan.strategy === ReviewSliceStrategy.Independent &&
+        plan.slices.some((slice) => slice.dependsOn !== null)) {
+        throw new Error('An independent plan cannot contain slice dependencies');
+    }
+    if (plan.strategy === ReviewSliceStrategy.Stack) {
+        for (const [index, slice] of plan.slices.entries()) {
+            const expectedDependency = index === 0 ? null : index;
+            if (slice.dependsOn !== expectedDependency) {
+                throw new Error('A stack plan must make each slice depend on its immediate predecessor');
+            }
+        }
+    }
+}
+export function renderReviewSlicePlan(plan, options) {
+    const compact = options.compact === true;
+    if (plan.strategy === ReviewSliceStrategy.NoSafeSplit) {
+        const explanation = [plan.summary, ...plan.caveats].join(' ');
+        return [
+            '<details>',
+            '<summary><h3>🪄 Suggested slicing plan 👇</h3></summary>',
+            '',
+            sanitizeMarkdown(explanation, compact ? 1_200 : 8_000),
+            '',
+            'A human familiar with the change should identify an intermediate, buildable boundary before splitting this PR.',
+            '',
+            '</details>',
+            '',
+            AGENT_SPLIT_SUGGESTION,
+        ].join('\n');
+    }
+    const branchNames = plan.slices.map((slice, index) => buildSuggestedBranchName(index + 1, slice.title));
+    const lines = [
+        '<details>',
+        '<summary><h3>🪄 Suggested slicing plan 👇</h3></summary>',
+        '',
+        sanitizeMarkdown(plan.summary, compact ? 600 : 1_000),
+        '',
+        '**PR order**',
+        '',
+        ...plan.slices.map((slice, index) => {
+            const base = slice.dependsOn === null ? options.baseBranch : branchNames[slice.dependsOn - 1];
+            return `${index + 1}. <code>${escapeHtml(branchNames[index])}</code> → base <code>${escapeHtml(base)}</code>`;
+        }),
+        '',
+    ];
+    for (const [index, slice] of plan.slices.entries()) {
+        lines.push(`#### PR ${index + 1} ${sanitizeMarkdown(slice.title, 160)}`, '', sanitizeMarkdown(slice.purpose, compact ? 300 : 800), '', `**Files (${slice.files.length}):**`, '', ...formatFileList(slice.files, compact ? 8 : slice.files.length), '');
+    }
+    lines.push('> This plan is based on the current PR head. Keep each slice buildable and move tests with the behavior they cover.', '', '</details>', '', AGENT_SPLIT_SUGGESTION);
+    return lines.join('\n').trim();
+}
+function formatFileList(files, limit) {
+    const visibleFiles = files.slice(0, limit);
+    const lines = visibleFiles.map((file) => `- <code>${escapeHtml(truncate(file, 240))}</code>`);
+    if (files.length > visibleFiles.length) {
+        lines.push(`- …and ${files.length - visibleFiles.length} more files from this group`);
+    }
+    return lines;
+}
+function buildSuggestedBranchName(sliceNumber, title) {
+    const slug = title
+        .normalize('NFKD')
+        .replace(/[^a-zA-Z0-9]+/g, '-')
+        .replace(/^-+|-+$/g, '')
+        .toLowerCase()
+        .slice(0, 40);
+    return `slice-${sliceNumber}-${slug || 'change'}`;
+}
+function sanitizeMarkdown(value, maxLength) {
+    const truncated = truncate(value.replace(/\s+/g, ' ').trim(), maxLength);
+    return escapeHtml(truncated.replace(/([\\`*_{}[\]()#+.!|>])/g, '\\$1'));
+}
+function truncate(value, maxLength) {
+    if (value.length <= maxLength) {
+        return value;
+    }
+    return `${value.slice(0, Math.max(0, maxLength - 1)).trimEnd()}…`;
+}
+function escapeHtml(value) {
+    return value
+        .replace(/&/g, '&amp;')
+        .replace(/</g, '&lt;')
+        .replace(/>/g, '&gt;')
+        .replace(/@/g, '&#64;');
+}