pi-smart-compact 8.0.7 → 9.0.0

package/dist/provider-scenario-eval.js

@@ -82,6 +82,35 @@ function assertSuccessful(message) {
   }
   return message;
 }
+async function withProviderDeadline(opts, invoke) {
+  if (opts.signal?.aborted)
+    throw new Error("LLM request aborted before dispatch");
+  const controller = new AbortController;
+  const watchdogMs = resolveCodexWatchdogMs(opts.maxTokens, opts.codexWatchdogMs);
+  const abort = Promise.withResolvers();
+  const abortFromCaller = () => {
+    controller.abort(opts.signal?.reason);
+    abort.reject(new Error("LLM request aborted by caller"));
+  };
+  opts.signal?.addEventListener("abort", abortFromCaller, { once: true });
+  const timeout = Promise.withResolvers();
+  const timer = setTimeout(() => {
+    controller.abort("provider-watchdog");
+    timeout.reject(new Error("Provider watchdog stopped generation after " + watchdogMs + "ms"));
+  }, watchdogMs);
+  if (typeof timer === "object" && "unref" in timer)
+    timer.unref();
+  try {
+    return await Promise.race([
+      invoke({ ...opts, signal: controller.signal }),
+      abort.promise,
+      timeout.promise
+    ]);
+  } finally {
+    clearTimeout(timer);
+    opts.signal?.removeEventListener("abort", abortFromCaller);
+  }
+}
 async function completeChatGptCodex(model, body, opts) {
   const controller = new AbortController;
   const watchdogMs = resolveCodexWatchdogMs(opts.maxTokens, opts.codexWatchdogMs);
@@ -125,10 +154,12 @@ async function completeChatGptCodex(model, body, opts) {
 var rawLlmClient = {
   complete: async (model, body, originalOpts) => {
     const opts = withCodexWireLimit(model, originalOpts);
-    if (isChatGptCodex(model))
-      return completeChatGptCodex(model, body, opts);
-    const response = opts.reasoning === undefined ? await (await resolveComplete())(model, body, opts) : await (await resolveCompleteSimple())(model, body, opts);
-    return assertSuccessful(response);
+    return withProviderDeadline(opts, async (bounded) => {
+      if (isChatGptCodex(model))
+        return completeChatGptCodex(model, body, bounded);
+      const response = bounded.reasoning === undefined ? await (await resolveComplete())(model, body, bounded) : await (await resolveCompleteSimple())(model, body, bounded);
+      return assertSuccessful(response);
+    });
   }
 };
 var defaultLlmClient = rawLlmClient;
@@ -138,7 +169,7 @@ function getLlmClient() {
 }
 
 // src/constants.ts
-var VERSION = "8.0.7";
+var VERSION = "9.0.0";
 var COMPACT_SYSTEM_PREFIX = "You are an expert conversation summarizer for a coding agent. " + "Produce structured markdown summaries. " + "Follow output format exactly. " + "Use EXACT names \u2014 never paraphrase code identifiers. " + "Trust deterministic extraction data over intuition.";
 var PROFILES = {
   light: {
@@ -180,7 +211,7 @@ var DEFAULT_CONFIG = {
   backupEnabled: true,
   backupDir: "",
   minContextPercent: 60,
-  requireApproval: false,
+  requireApproval: true,
   scrubSecrets: true,
   scrubPii: false,
   maxLlmCalls: 8,
@@ -220,6 +251,8 @@ Output EXACTLY this format:
 - **[Decision]**: [Rationale]
 ` + `## Files Modified
 - [Verified list from deterministic extraction]
+` + `## Files Deleted
+- [Verified deleted paths from deterministic extraction]
 ` + `## Files Read
 - [Verified list from deterministic extraction]
 ` + `## Next Steps
@@ -243,6 +276,7 @@ Rules for Accuracy:
 ` + `**Summary**: [2-4 sentences: what happened, errors, code changes with paths]
 ` + `**Decisions**: [comma-separated, or "None"]
 ` + `**Modified**: [comma-separated paths, or "None"]
+` + `**Deleted**: [comma-separated paths, or "None"]
 ` + `**Read**: [comma-separated paths, or "None"]
 `;
 var ASSEMBLY_PROMPT_PREFIX = `Merge these topic summaries into ONE coherent summary.
@@ -274,6 +308,8 @@ var ASSEMBLY_PROMPT_PREFIX = `Merge these topic summaries into ONE coherent summ
 - **[Decision]**: [Rationale]
 ` + `## Files Modified
 - [Verified deterministic list]
+` + `## Files Deleted
+- [Verified deterministic deleted paths]
 ` + `## Files Read
 - [Verified deterministic list]
 ` + `## Next Steps
@@ -328,6 +364,7 @@ var TRUNC = {
   BACKUP_PREVIEW_LINES: 5,
   FINGERPRINT_SEG: 2
 };
+var LIKELY_ERROR_RE = /(?:command not found|no such file|permission denied|syntax error|cannot find|module not found|compilation error|build failed|test failed|^FAIL\b|ERROR:)/i;
 var METRICS_BUFFER_MAX = 200;
 var RUNTIME_LOG_MAX_BYTES = 5 * 1024 * 1024;
 var EXPLORER_SYSTEM_PROMPT = `You are a conversation analyst. You have deterministic extraction data and can query the raw conversation using tools.
@@ -406,6 +443,9 @@ function calibrationKey(provider, model) {
 }
 
 // src/utils/type-guards.ts
+function isTextBlock(c) {
+  return typeof c === "object" && c !== null && c.type === "text" && typeof c.text === "string";
+}
 var KNOWN_METHODS = new Set(["eesv", "single-pass", "heuristic"]);
 var KNOWN_PROFILES = new Set(["light", "balanced", "aggressive"]);
 var KNOWN_MODES = new Set(["balanced", "aggressive", "fast", "thorough"]);
@@ -425,37 +465,9 @@ var GENERIC_BASENAMES = new Set([
   "lib.rs",
   "__init__.py"
 ]);
-var MIN_BARE_BASENAME_LEN = 5;
 function normalizePath(filePath) {
   return filePath.replace(/\\/g, "/").replace(/^\.\//, "").toLowerCase();
 }
-function buildPathNeedles(filePath) {
-  const parts = normalizePath(filePath).split("/").filter(Boolean);
-  if (parts.length === 0)
-    return [];
-  const needles = [];
-  const basename = parts[parts.length - 1];
-  if (!GENERIC_BASENAMES.has(basename) && basename.length >= MIN_BARE_BASENAME_LEN) {
-    needles.push(basename);
-  }
-  for (let j = parts.length - 2;j >= 0; j--) {
-    needles.push(parts.slice(j).join("/"));
-  }
-  return needles;
-}
-function buildUniquePathNeedles(filePath, allPaths) {
-  const normalized = allPaths.map(normalizePath);
-  return buildPathNeedles(filePath).filter((needle) => {
-    let owners = 0;
-    for (const candidate of normalized) {
-      if (candidate === needle || candidate.endsWith("/" + needle))
-        owners++;
-      if (owners > 1)
-        return false;
-    }
-    return owners === 1;
-  });
-}
 function isKnownPathReference(ref, knownPaths) {
   const normalizedRef = normalizePath(ref);
   return knownPaths.some((path) => {
@@ -499,6 +511,8 @@ function classifyHeading(raw) {
     return "files-modified";
   if (text.includes("file") && (text.includes("read") || text.includes("viewed")))
     return "files-read";
+  if (text.includes("file") && (text.includes("delet") || text.includes("remov")))
+    return "files-deleted";
   if (text.includes("next step") || text === "next actions")
     return "next-steps";
   if (text.includes("critical context") || text === "important context")
@@ -569,12 +583,25 @@ function findSection(summary, kind) {
 }
 
 // src/utils/extraction.ts
+function extractText(content) {
+  if (typeof content === "string")
+    return content;
+  if (!Array.isArray(content))
+    return "";
+  return content.map((b) => {
+    if (typeof b === "string")
+      return b;
+    if (isTextBlock(b))
+      return b.text;
+    return "";
+  }).join("");
+}
 var CONSTRAINT_PATTERNS = [
   { re: /\b(?:must|need|require|has to|important)\b.*\b(?:be|use|have|include|support)\b/i, cat: "requirement", conf: TUNING.CONFIDENCE_HIGH },
   { re: /\b(?:don't|never|avoid|shouldn't|must not|do not|no\s+(?:need|want))\b/i, cat: "prohibition", conf: TUNING.CONFIDENCE_MEDIUM },
   { re: /\b(?:prefer|like|want|would rather|should)\b.*\b(?:use|be|have|with)\b/i, cat: "preference", conf: TUNING.CONFIDENCE_LOW },
-  { re: /(?<![A-Za-z0-9_])(?:kritik|kritikal|\u00F6nemli|onemli|\u015Fart|sart|zorunlu|\u015Fart ko\u015Ful|\u00F6nemli \u015Fart|kesinlikle|kesinlikle \u015Fart|asla|sak\u0131n|sak\u0131nha|bunu yapma|b\u00F6yle olsun|b\u00F6yle yap\u0131n|\u015F\u00F6yle olsun|\u015F\u00F6yle yap\u0131n)(?![A-Za-z0-9_])/iu, cat: "requirement", conf: TUNING.CONFIDENCE_MEDIUM },
-  { re: /(?<![A-Za-z0-9_])(?:yapma|kullanma|sak\u0131n|asla\s+(?:kullanma|yapma|getirme))(?![A-Za-z0-9_])/iu, cat: "prohibition", conf: TUNING.CONFIDENCE_MEDIUM },
+  { re: /(?<![A-Za-z0-9_])(?:yapma|kullanma|sak\u0131n|sak\u0131nha|asla(?:\s+(?:kullanma|yapma|getirme))?|bunu yapma)(?![A-Za-z0-9_])/iu, cat: "prohibition", conf: TUNING.CONFIDENCE_MEDIUM },
+  { re: /(?<![A-Za-z0-9_])(?:kritik|kritikal|\u00F6nemli|onemli|\u015Fart|sart|zorunlu|\u015Fart ko\u015Ful|\u00F6nemli \u015Fart|kesinlikle|kesinlikle \u015Fart|b\u00F6yle olsun|b\u00F6yle yap\u0131n|\u015F\u00F6yle olsun|\u015F\u00F6yle yap\u0131n)(?![A-Za-z0-9_])/iu, cat: "requirement", conf: TUNING.CONFIDENCE_MEDIUM },
   { re: /(?<![A-Za-z0-9_])(?:tercih|isterim|olsun|kullanal\u0131m|yapal\u0131m|istiyorum)(?![A-Za-z0-9_])/iu, cat: "preference", conf: TUNING.CONFIDENCE_LOW }
 ];
 function isDiagnosticConstraintText(text) {
@@ -641,7 +668,6 @@ function readJsonlTail(target, limit, maxBytes = 512 * 1024) {
   }
   return values.slice(-limit);
 }
-var scheduledTailTrims = new Set;
 
 // src/infra/clock.ts
 var systemClock = {
@@ -655,14 +681,23 @@ import crypto from "crypto";
 var SECRET_PATTERNS = [
   { kind: "private-key", regex: /-----BEGIN [A-Z ]*PRIVATE KEY-----[\s\S]*?-----END [A-Z ]*PRIVATE KEY-----/g },
   { kind: "aws-access-key", regex: /\bAKIA[0-9A-Z]{16}\b/g },
-  { kind: "github-token", regex: /\bgh[pousr]_[A-Za-z0-9]{30,}\b/g },
+  { kind: "google-api-key", regex: /\bAIza[0-9A-Za-z_-]{30,}\b/g },
+  { kind: "stripe-key", regex: /\b[rs]k_(?:live|test)_[0-9A-Za-z]{16,}\b/g },
+  { kind: "gitlab-token", regex: /\bglpat-[0-9A-Za-z_-]{20,}\b/g },
+  { kind: "npm-token", regex: /\bnpm_[0-9A-Za-z]{30,}\b/g },
+  { kind: "github-token", regex: /\bgh[pousr]_[A-Za-z0-9]{20,}\b/g },
   { kind: "api-key", regex: /\bsk-(?:ant-)?[A-Za-z0-9_-]{20,}\b/g },
-  { kind: "slack-token", regex: /\bxox[baprs]-[A-Za-z0-9-]{20,}\b/g },
+  { kind: "slack-token", regex: /\bxox[baprs]-[A-Za-z0-9-]{10,}\b/g },
   { kind: "jwt", regex: /\beyJ[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,}\b/g },
-  { kind: "bearer-token", regex: /\bBearer\s+[A-Za-z0-9._~+\/-]{24,}/gi, replacement: () => "Bearer [REDACTED:bearer-token]" },
+  { kind: "bearer-token", regex: /\bBearer\s+[A-Za-z0-9._~+\/-]{12,}=*/gi, replacement: () => "Bearer [REDACTED:bearer-token]" },
+  {
+    kind: "connection-password",
+    regex: /\b([a-z][a-z0-9+.-]*:\/\/[^:\s/@]+:)[^@\s/]+(@)/gi,
+    replacement: (prefix, suffix) => prefix + "[REDACTED:password]" + suffix
+  },
   {
     kind: "credential",
-    regex: /\b(api[_-]?key|access[_-]?token|auth[_-]?token|token|password|passwd|secret)\s*([:=])\s*["']?([A-Za-z0-9._~+\/-]{16,})["']?/gi,
+    regex: /\b((?:[A-Za-z0-9]+[_-])*(?:api[_-]?key|access[_-]?token|auth[_-]?token|token|password|passwd|secret(?:[_-]?(?:access)?[_-]?key)?|client[_-]?secret)(?:[_-][A-Za-z0-9]+)*)\s*([:=])\s*["']?([^\s"']{16,})["']?/gi,
     replacement: (name, separator) => name + separator + "[REDACTED:credential]"
   }
 ];
@@ -690,6 +725,31 @@ function mergeFindings(target, findings) {
   for (const finding of findings)
     target.set(finding.kind, (target.get(finding.kind) ?? 0) + finding.count);
 }
+var SECRET_KEY_NAMES = new Set([
+  "api_key",
+  "apikey",
+  "access_token",
+  "auth_token",
+  "authorization",
+  "password",
+  "passwd",
+  "secret",
+  "secret_key",
+  "secret_access_key",
+  "client_secret",
+  "private_key",
+  "database_url",
+  "connection_string"
+]);
+function normalizeObjectKey(key) {
+  return key.replace(/([a-z0-9])([A-Z])/g, "$1_$2").toLowerCase().replace(/[^a-z0-9]+/g, "_").replace(/^_+|_+$/g, "");
+}
+function isSecretBearingKey(key) {
+  const normalized = normalizeObjectKey(key);
+  if (SECRET_KEY_NAMES.has(normalized))
+    return true;
+  return /(?:^|_)(?:api_key|access_token|auth_token|password|passwd|secret_access_key|client_secret|private_key)(?:_|$)/.test(normalized);
+}
 
 class SecretScrubber {
   secretsEnabled;
@@ -719,6 +779,10 @@ class SecretScrubber {
   scrubValue(input) {
     const findings = new Map;
     const seen = new WeakMap;
+    const recordCredential = () => {
+      findings.set("credential", (findings.get("credential") ?? 0) + 1);
+      this.total++;
+    };
     const visit = (value2) => {
       if (typeof value2 === "string") {
         const result = this.scrubText(value2);
@@ -739,8 +803,14 @@ class SecretScrubber {
       }
       const output = {};
       seen.set(value2, output);
-      for (const [key, item] of Object.entries(value2))
-        output[key] = visit(item);
+      for (const [key, item] of Object.entries(value2)) {
+        if (this.secretsEnabled && isSecretBearingKey(key) && typeof item === "string" && item.length > 0) {
+          output[key] = "[REDACTED:credential]";
+          recordCredential();
+        } else {
+          output[key] = visit(item);
+        }
+      }
       return output;
     };
     const value = visit(input);
@@ -1028,6 +1098,52 @@ function readMetricsLog(limit = 100) {
 }
 
 // src/phases/verify.ts
+var HIGH_RISK_OUTCOME_RE = /(?:\ball\s+tests?\s+(?:pass|passed|passing)\b|\btests?\s+(?:pass|passed|passing)\b|\b(?:build|deployment|migration)\s+(?:completed|succeeded|passed|successful)\b|\b(?:deployed|published|released)\b|\b(?:bug|issue|error)\s+(?:fixed|resolved)\b|\bno\s+(?:errors?|failures?)\b|\bcompleted successfully\b|\btestler?\s+(?:ge\u00E7ti|ba\u015Far\u0131l\u0131)\b|\bba\u015Far\u0131yla\s+(?:tamamland\u0131|da\u011F\u0131t\u0131ld\u0131|yay\u0131nland\u0131)\b|\b(?:deploy edildi|yay\u0131nland\u0131|hata yok)\b)/iu;
+var NEGATED_OUTCOME_RE = /\b(?:not|never|pending|failed|failing|unresolved|hen\u00FCz|de\u011Fil|ba\u015Far\u0131s\u0131z)\b/iu;
+function outcomeClaims(summary) {
+  return Array.from(new Set(summary.split(/\r?\n/).map((line) => line.replace(/^\s*(?:[-*+]|\d+[.)])\s+/, "").replace(/^\[[ x]\]\s+/i, "").trim()).filter((line) => line.length > 0 && !line.startsWith("#")).filter((line) => HIGH_RISK_OUTCOME_RE.test(line)).filter((line) => /\bno\s+(?:errors?|failures?)\b/i.test(line) || !NEGATED_OUTCOME_RE.test(line)))).slice(0, 12);
+}
+function classifyOutcomeClaim(claim) {
+  const lower = claim.toLowerCase();
+  if (/\btests?\b|\btestler?\b/.test(lower))
+    return "test";
+  if (/\bbuild\b|\bcompil(?:e|ed|ation)\b|\btypecheck\b/.test(lower))
+    return "build";
+  if (/\bdeploy(?:ed|ment)?\b|\bpublish(?:ed)?\b|\breleas(?:e|ed)\b/.test(lower))
+    return "release";
+  if (/\bbug\b|\bissue\b|\berror\b|\bfail(?:ed|ure)?\b|\bhata\b/.test(lower))
+    return "error";
+  if (/\bfile\b|\bdosya\b/.test(lower))
+    return "file";
+  return "generic";
+}
+function successfulToolSupportsClaim(claim, messages, extraction) {
+  const shape = semanticShape(claim);
+  const category = classifyOutcomeClaim(claim);
+  if (category === "error" && extraction.errors.some((error) => error.resolved && hasSemanticEvidence(claim, error.message)))
+    return true;
+  if (category === "file" && extraction.modifiedFiles.some((file) => claim.toLowerCase().includes(file.path.toLowerCase())))
+    return true;
+  for (const message of messages) {
+    if (message.role !== "toolResult" || message.isError)
+      continue;
+    const bounded = extractText(message.content).slice(0, 8000);
+    if (!bounded.trim() || LIKELY_ERROR_RE.test(bounded))
+      continue;
+    if (hasSemanticEvidence(claim, bounded))
+      return true;
+    const lower = bounded.toLowerCase();
+    if (category === "test" && /\b\d+\s+(?:tests?\s+)?pass(?:ed)?\b/.test(lower) && !/\b(?:fail(?:ed|ures?)?|errors?)\s*[:=]?\s*[1-9]\d*\b/.test(lower))
+      return true;
+    if (category === "build" && /\b(?:build|compile|typecheck)\b/.test(lower) && /\b(?:succeeded|successful|passed|exit(?:ed)?\s+(?:code\s+)?0)\b/.test(lower))
+      return true;
+    if (category === "release" && /\b(?:publish|release|deploy)\b/.test(lower) && /\b(?:succeeded|successful|completed|published|deployed|released)\b/.test(lower))
+      return true;
+    if (category === "error" && shape.concepts.length > 0 && /\b(?:fixed|resolved|passed|succeeded|successful)\b/.test(lower) && hasSemanticEvidence(claim, bounded))
+      return true;
+  }
+  return false;
+}
 var NEGATION_MARKERS = new Set([
   "no",
   "not",
@@ -1186,7 +1302,7 @@ function hasSemanticContradiction(source, target) {
     return conditional && !negative && !tokens.some((token) => CONDITION_MARKERS.has(token));
   });
 }
-function verifySummary(summary, extraction, continuity = null) {
+function verifySummary(summary, extraction, continuity = null, evidence = {}) {
   const parsed = parseSummary(summary);
   const gaps = [];
   const lower = summary.toLowerCase().replace(/\\/g, "/");
@@ -1206,9 +1322,18 @@ function verifySummary(summary, extraction, continuity = null) {
     ...extraction.errors.filter((error) => !error.resolved).map((error) => ({ message: error.message })),
     ...(continuity?.unresolvedErrors ?? []).map((error) => ({ message: error.message }))
   ], (item) => item.message);
+  const resolvedEvidence = uniqueByText([
+    ...extraction.errors.filter((error) => error.resolved).map((error) => ({ message: error.message })),
+    ...(continuity?.resolvedErrors ?? []).map((error) => ({ message: error.message }))
+  ], (item) => item.message).slice(-5);
+  const steeringConstraints = [
+    evidence.steering?.focus ? { text: "Preserve detail about: " + evidence.steering.focus } : null,
+    evidence.steering?.note ? { text: evidence.steering.note } : null
+  ].filter((item) => Boolean(item?.text.trim()));
   const constraintEvidence = uniqueByText([
     ...extraction.constraints.filter((item) => item.confidence >= 0.8).map((item) => ({ text: item.text })),
-    ...(continuity?.constraints ?? []).filter((item) => item.confidence >= 0.8).map((item) => ({ text: item.text }))
+    ...(continuity?.constraints ?? []).filter((item) => item.confidence >= 0.8).map((item) => ({ text: item.text })),
+    ...steeringConstraints
   ], (item) => item.text).filter((item) => !isDiagnosticConstraintText(item.text));
   const decisionEvidence = uniqueByText([
     ...extraction.decisions.filter((item) => item.type === "explicit").map((item) => ({ summary: item.summary })),
@@ -1226,14 +1351,29 @@ function verifySummary(summary, extraction, continuity = null) {
       score -= req.penalty;
     }
   }
+  const listedPaths = (kind) => new Set((findSection(parsed, kind)?.body ?? "").split(`
+`).map((line) => line.replace(/^\s*(?:[-*+]|\d+[.)])\s+/, "").replace(/^\[[ x]\]\s+/i, "").trim()).map((line) => line.startsWith("`") && line.endsWith("`") ? line.slice(1, -1) : line).filter((line) => line.length > 0 && !/^none(?: recorded)?[.!]?$/i.test(line)).map(normalizePath));
   const modifiedPaths = extraction.modifiedFiles.map((file) => file.path);
-  for (const file of extraction.modifiedFiles) {
-    const needles = buildUniquePathNeedles(file.path, modifiedPaths);
-    if (!needles.some((needle) => lower.includes(needle))) {
-      gaps.push({ kind: "missing-file", path: file.path });
-      score -= 5;
-    }
+  const modifiedListed = listedPaths("files-modified");
+  const readListed = listedPaths("files-read");
+  const deletedListed = listedPaths("files-deleted");
+  for (const file of modifiedPaths) {
+    if (!modifiedListed.has(normalizePath(file)))
+      gaps.push({ kind: "missing-file", path: file });
+  }
+  for (const file of extraction.readFiles) {
+    if (!readListed.has(normalizePath(file)))
+      gaps.push({ kind: "missing-read-file", path: file });
+  }
+  const deletedEvidence = Array.from(new Set([
+    ...extraction.deletedFiles,
+    ...continuity?.deletedFiles ?? []
+  ]));
+  for (const file of deletedEvidence) {
+    if (!deletedListed.has(normalizePath(file)))
+      gaps.push({ kind: "missing-deleted-file", path: file });
   }
+  score -= gaps.filter((gap) => gap.kind === "missing-file" || gap.kind === "missing-read-file" || gap.kind === "missing-deleted-file").length * 5;
   for (const error of unresolvedEvidence) {
     const snippet = summaryEvidenceLine(error.message, TRUNC.ERROR_SNIPPET).toLowerCase();
     if (snippet.length > 5 && !normalizedSummary.includes(snippet)) {
@@ -1241,6 +1381,13 @@ function verifySummary(summary, extraction, continuity = null) {
       score -= 5;
     }
   }
+  for (const error of resolvedEvidence) {
+    const snippet = summaryEvidenceLine(error.message, TRUNC.ERROR_SNIPPET).toLowerCase();
+    if (snippet.length > 5 && !normalizedSummary.includes(snippet)) {
+      gaps.push({ kind: "missing-error", message: error.message, resolved: true });
+      score -= 2;
+    }
+  }
   const constraintTarget = [
     findSection(parsed, "constraints")?.body ?? "",
     findSection(parsed, "critical-context")?.body ?? ""
@@ -1331,6 +1478,14 @@ function verifySummary(summary, extraction, continuity = null) {
     gaps.push({ kind: "missing-open-loops", unresolvedCount });
     score -= 5;
   }
+  if (evidence.sourceMessages) {
+    for (const claim of outcomeClaims(summary)) {
+      if (!successfulToolSupportsClaim(claim, evidence.sourceMessages, extraction)) {
+        gaps.push({ kind: "unsupported-claim", claim });
+        score -= 20;
+      }
+    }
+  }
   const finalScore = Math.max(0, score);
   return { ok: gaps.length === 0 && finalScore >= 85, gaps, score: finalScore };
 }