diff --git a/packages/conversation-ui-core/src/composer-input-parts.ts b/packages/conversation-ui-core/src/composer-input-parts.ts
index d22cab13a..efd45d1be 100644
--- a/packages/conversation-ui-core/src/composer-input-parts.ts
+++ b/packages/conversation-ui-core/src/composer-input-parts.ts
@@ -1,4 +1,5 @@
 import { parseLocalFileReference } from "./message-body.js";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 
 export type ComposerInputPart =
   | { readonly type: "text"; readonly text: string }
@@ -254,7 +255,7 @@ function mergeAdjacentTextParts(parts: readonly ComposerInputPart[]): readonly C
 }
 
 function normalizeLabel(value: string): string {
-  return value.toLowerCase().replace(/[^a-z0-9_.-]+/g, "-").replace(/^-+|-+$/g, "");
+  return trimBoundaryCharacters(value.toLowerCase().replace(/[^a-z0-9_.-]+/g, "-"), "-");
 }
 
 function readKnownMentionAtStart(
@@ -319,7 +320,7 @@ function readUrlAtStart(
 }
 
 function trimUrlToken(value: string): string {
-  let next = trimNaturalLanguageSuffix(value).replace(/[.,;:!?，。！？；：、]+$/u, "");
+  let next = trimBoundaryCharacters(trimNaturalLanguageSuffix(value), ".,;:!?，。！？；：、", "end");
   while (next.endsWith(")") || next.endsWith("]")) {
     const opener = next.endsWith(")") ? "(" : "[";
     const closer = next.endsWith(")") ? ")" : "]";
diff --git a/packages/conversation-ui-core/src/message-body.ts b/packages/conversation-ui-core/src/message-body.ts
index 68a341452..84d21772a 100644
--- a/packages/conversation-ui-core/src/message-body.ts
+++ b/packages/conversation-ui-core/src/message-body.ts
@@ -2,6 +2,7 @@ import { unified } from "unified";
 import remarkParse from "remark-parse";
 import remarkGfm from "remark-gfm";
 import remarkMath from "remark-math";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 
 export interface LocalFileReference {
   readonly raw: string;
@@ -521,11 +522,11 @@ function stripWrappingQuotes(value: string): string {
 }
 
 function trimTrailingUrlPunctuation(value: string): string {
-  return value.replace(/[),.;!?]+$/g, "");
+  return trimBoundaryCharacters(value, "),.;!?", "end");
 }
 
 function trimTrailingPathPunctuation(value: string): string {
-  return value.replace(/[),.;!?]+$/g, "");
+  return trimBoundaryCharacters(value, "),.;!?", "end");
 }
 
 function normalizePath(value: string): string {
diff --git a/packages/conversation-ui-core/src/string-boundaries.ts b/packages/conversation-ui-core/src/string-boundaries.ts
new file mode 100644
index 000000000..6456f141e
--- /dev/null
+++ b/packages/conversation-ui-core/src/string-boundaries.ts
@@ -0,0 +1,74 @@
+export function trimBoundaryCharacters(
+  value: string,
+  characters: string,
+  side: "start" | "end" | "both" = "both",
+): string {
+  let start = 0;
+  let end = value.length;
+  if (side !== "end") {
+    while (start < end && characters.includes(value[start] ?? "")) start += 1;
+  }
+  if (side !== "start") {
+    while (end > start && characters.includes(value[end - 1] ?? "")) end -= 1;
+  }
+  return start === 0 && end === value.length ? value : value.slice(start, end);
+}
+
+export function extractWrappedSection(
+  value: string,
+  labels: readonly string[],
+  boundaryLabels: readonly string[],
+): string | undefined {
+  const searchValue = asciiLowercase(value);
+  let selected: { readonly labelIndex: number; readonly start: number } | undefined;
+  for (const label of labels) {
+    const searchLabel = asciiLowercase(label);
+    let labelIndex = searchValue.indexOf(searchLabel);
+    while (labelIndex !== -1) {
+      let start = labelIndex + label.length;
+      while (start < value.length && isWhitespace(value[start])) start += 1;
+      if (":：".includes(value[start] ?? "")) {
+        start += 1;
+        while (start < value.length && isWhitespace(value[start])) start += 1;
+        if (!selected || labelIndex < selected.labelIndex) selected = { labelIndex, start };
+        break;
+      }
+      labelIndex = searchValue.indexOf(searchLabel, labelIndex + 1);
+    }
+  }
+  if (!selected) return undefined;
+  let end = value.length;
+  for (const boundary of boundaryLabels) {
+    const searchBoundary = asciiLowercase(boundary);
+    let boundaryIndex = searchValue.indexOf(searchBoundary, selected.start);
+    while (boundaryIndex !== -1) {
+      const before = value[boundaryIndex - 1];
+      const after = value[boundaryIndex + boundary.length];
+      if (isWhitespace(before) && !isWordCharacter(after)) {
+        end = Math.min(end, boundaryIndex);
+        break;
+      }
+      boundaryIndex = searchValue.indexOf(searchBoundary, boundaryIndex + 1);
+    }
+  }
+  const section = value.slice(selected.start, end).trim();
+  return section || undefined;
+}
+
+function asciiLowercase(value: string): string {
+  const result: string[] = [];
+  for (let index = 0; index < value.length; index += 1) {
+    const character = value[index] ?? "";
+    const code = character.charCodeAt(0);
+    result.push(code >= 65 && code <= 90 ? String.fromCharCode(code + 32) : character);
+  }
+  return result.join("");
+}
+
+function isWhitespace(value: string | undefined): boolean {
+  return value !== undefined && value.trim() === "";
+}
+
+function isWordCharacter(value: string | undefined): boolean {
+  return value !== undefined && /[A-Za-z0-9_]/u.test(value);
+}
diff --git a/packages/conversation-ui-core/src/terminal-report-view-model.ts b/packages/conversation-ui-core/src/terminal-report-view-model.ts
index 7a579e0a2..570f5eb9d 100644
--- a/packages/conversation-ui-core/src/terminal-report-view-model.ts
+++ b/packages/conversation-ui-core/src/terminal-report-view-model.ts
@@ -14,6 +14,7 @@ import {
   userVisibleTerminalReportChanges,
 } from "./terminal-report-visibility.js";
 import { projectVisibleTerminalReportContent } from "./terminal-report-presentation.js";
+import { extractWrappedSection, trimBoundaryCharacters } from "./string-boundaries.js";
 
 interface TerminalReportSourceRefInput {
   readonly kind: string;
@@ -300,10 +301,17 @@ function uniqueSourceRefs(refs: readonly TerminalReportSourceRefInput[]): Termin
  */
 function compactTerminalReportTitle(value: string): string {
   const normalized = value.replace(/\s+/gu, " ").trim();
-  const wrapped = normalized.match(
-    /(?:原始请求|Original request)\s*[:：]\s*(.+?)(?=\s+(?:执行任务|已物化的执行操作|建议协作步骤|Requested objective|No literal file content|Materialized operations|Participant operation ownership|Permission boundary)\b|$)/iu,
-  )?.[1]?.trim();
-  const candidate = (wrapped || normalized).replace(/^\*+|\*+$/gu, "").trim();
+  const wrapped = extractWrappedSection(normalized, ["原始请求", "Original request"], [
+    "执行任务",
+    "已物化的执行操作",
+    "建议协作步骤",
+    "Requested objective",
+    "No literal file content",
+    "Materialized operations",
+    "Participant operation ownership",
+    "Permission boundary",
+  ]);
+  const candidate = trimBoundaryCharacters(wrapped || normalized, "*").trim();
   return candidate.length > 180 ? `${candidate.slice(0, 177).trimEnd()}...` : candidate;
 }
 
diff --git a/packages/conversation-ui-core/test/composer-policy.test.mjs b/packages/conversation-ui-core/test/composer-policy.test.mjs
index a73142c3d..54e456b36 100644
--- a/packages/conversation-ui-core/test/composer-policy.test.mjs
+++ b/packages/conversation-ui-core/test/composer-policy.test.mjs
@@ -114,6 +114,18 @@ test("raw token removal trims duplicate whitespace at removal boundary", () => {
   assert.equal(removeComposerInputRawToken("Run /review  now", "/review"), "Run now");
 });
 
+test("composer parsing handles long boundary runs without changing tokens", () => {
+  const punctuation = "!".repeat(100_000);
+  const parts = parseComposerInputParts(`https://example.com/docs${punctuation}`);
+  assert.equal(parts[0]?.type, "url_ref");
+  assert.equal(parts[0]?.raw, "https://example.com/docs");
+
+  const paddedId = `${"-".repeat(100_000)}review${"-".repeat(100_000)}`;
+  const command = parseComposerInputParts("/review", { slashCommands: [{ id: paddedId, label: "Review" }] });
+  assert.equal(command[0]?.type, "slash_command");
+  assert.equal(command[0]?.commandId, paddedId);
+});
+
 test("permission confirmation is derived from explicit option metadata", () => {
   assert.equal(conversationPermissionPresetRequiresConfirmation("full_access"), true);
   assert.equal(conversationPermissionPresetRequiresConfirmation("read_only"), false);
diff --git a/packages/conversation-ui-core/test/message-body.test.mjs b/packages/conversation-ui-core/test/message-body.test.mjs
index f26a3e682..7f39f7018 100644
--- a/packages/conversation-ui-core/test/message-body.test.mjs
+++ b/packages/conversation-ui-core/test/message-body.test.mjs
@@ -144,3 +144,12 @@ test("message body collects unique file references from nested structured blocks
     "docs/design/conversation-shared-capabilities.md",
   ]);
 });
+
+test("message tokenization trims long punctuation runs in linear boundary helpers", () => {
+  const punctuation = "!".repeat(100_000);
+  const tokens = splitInlineTokens(`https://example.com/docs${punctuation} src/index.ts${punctuation}`, "/workspace/ling-stack");
+  const link = tokens.find((token) => token.kind === "web_link");
+  const file = tokens.find((token) => token.kind === "local_file");
+  assert.equal(link?.kind === "web_link" ? link.href : undefined, "https://example.com/docs");
+  assert.equal(file?.kind === "local_file" ? file.reference.displayPath : undefined, "src/index.ts");
+});
diff --git a/packages/conversation-ui-core/test/terminal-report-view-model.test.mjs b/packages/conversation-ui-core/test/terminal-report-view-model.test.mjs
index d79060a2a..957e9bd96 100644
--- a/packages/conversation-ui-core/test/terminal-report-view-model.test.mjs
+++ b/packages/conversation-ui-core/test/terminal-report-view-model.test.mjs
@@ -140,6 +140,20 @@ test("terminal report projection keeps Markdown summaries and bounds model title
   assert.match(participant?.description ?? "", /```ts\nconst ok = true;/u);
 });
 
+test("terminal report title extraction handles long protocol whitespace", () => {
+  const title = `Original request${" ".repeat(100_000)}: concise result Requested objective ignored wrapper`;
+  const report = terminalReportToViewModel({
+    id: "terminal_report_long_wrapper",
+    status: "committed",
+    language: "en-US",
+    outcome: { status: "succeeded", title, summary: "done", corrections: [], sourceRefs: [] },
+    changes: [], verification: [], participants: [], artifacts: [], risks: [], references: [],
+    evidenceFingerprint: "fingerprint_long_wrapper",
+    committedAt: "2026-10-21T00:00:00.000Z",
+  });
+  assert.equal(report.rows.find((row) => row.reportSection === "outcome")?.title, "concise result");
+});
+
 test("terminal report artifact workspace paths require one trusted changed-file source ref", () => {
   const report = terminalReportToViewModel({
     id: "terminal_report_workspace_files",
diff --git a/packages/core/src/channel-team-snapshot-harness.ts b/packages/core/src/channel-team-snapshot-harness.ts
index e92423968..a4368f34b 100644
--- a/packages/core/src/channel-team-snapshot-harness.ts
+++ b/packages/core/src/channel-team-snapshot-harness.ts
@@ -111,7 +111,7 @@ function inferRiskReview(input: string): Parameters<typeof buildChannelTeamRiskR
       },
     };
   }
-  if (/(delete|drop|remove).*(production|database|db)|rm\s+-rf|reset\s+--hard/.test(lower)) {
+  if (hasDestructiveTargetSequence(lower) || /rm\s+-rf|reset\s+--hard/.test(lower)) {
     return {
       actionType: "bash",
       risk: "critical",
@@ -129,6 +129,18 @@ function inferRiskReview(input: string): Parameters<typeof buildChannelTeamRiskR
   };
 }
 
+function hasDestructiveTargetSequence(value: string): boolean {
+  for (const action of ["delete", "drop", "remove"] as const) {
+    const actionIndex = value.indexOf(action);
+    if (actionIndex === -1) continue;
+    const targetStart = actionIndex + action.length;
+    if (["production", "database", "db"].some((target) => value.indexOf(target, targetStart) !== -1)) {
+      return true;
+    }
+  }
+  return false;
+}
+
 function diffValue(path: string, before: unknown, after: unknown): readonly ChannelTeamSnapshotDiffEntry[] {
   if (JSON.stringify(before) === JSON.stringify(after)) {
     return [];
diff --git a/packages/core/src/composer-slash-intents.ts b/packages/core/src/composer-slash-intents.ts
index bdebbf2b7..42419c123 100644
--- a/packages/core/src/composer-slash-intents.ts
+++ b/packages/core/src/composer-slash-intents.ts
@@ -1,3 +1,5 @@
+import { trimBoundaryCharacters } from "./string-boundaries.js";
+
 export type ComposerSlashIntent =
   | { readonly kind: "none" }
   | {
@@ -314,7 +316,7 @@ function isSessionCommandOption(command: ComposerSlashCommandOption): boolean {
 }
 
 function normalizeLabel(value: string): string {
-  return value.toLowerCase().replace(/[^a-z0-9_.-]+/g, "-").replace(/^-+|-+$/g, "");
+  return trimBoundaryCharacters(value.toLowerCase().replace(/[^a-z0-9_.-]+/g, "-"), "-");
 }
 
 function teammateMentionAliases(teammate: ComposerSlashMentionOption): readonly string[] {
diff --git a/packages/core/src/external-work-actions.ts b/packages/core/src/external-work-actions.ts
index 0837cd5c3..3d69a94c1 100644
--- a/packages/core/src/external-work-actions.ts
+++ b/packages/core/src/external-work-actions.ts
@@ -3,6 +3,7 @@ import { mkdirSync, writeFileSync } from "node:fs";
 import { relative, resolve as resolvePath } from "node:path";
 import type { ApprovalRequest } from "./approval-types.js";
 import { redactSecrets } from "./redaction.js";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 import { DEFAULT_WORKSPACE_ID } from "./workspace-types.js";
 import { normalizeGitHubExternalWorkItem } from "./external-work.js";
 import type {
@@ -145,7 +146,7 @@ export function createGitHubExternalActionAdapter(options: GitHubExternalActionA
   if (!fetchImpl) {
     throw new Error("A fetch implementation is required for GitHub external action execution.");
   }
-  const apiBaseUrl = (options.apiBaseUrl ?? "https://api.github.com").replace(/\/+$/, "");
+  const apiBaseUrl = trimBoundaryCharacters(options.apiBaseUrl ?? "https://api.github.com", "/", "end");
 
   return {
     async getPullRequest(input: GitHubSubjectInput): Promise<ExternalWorkItem | undefined> {
diff --git a/packages/core/src/external-work-ci-triage.ts b/packages/core/src/external-work-ci-triage.ts
index 9b0c8e3e3..4b6587e50 100644
--- a/packages/core/src/external-work-ci-triage.ts
+++ b/packages/core/src/external-work-ci-triage.ts
@@ -3,6 +3,7 @@ import { mkdirSync, writeFileSync } from "node:fs";
 import { relative, resolve as resolvePath } from "node:path";
 import { normalizeGitHubExternalWorkItem } from "./external-work.js";
 import { redactSecrets } from "./redaction.js";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 import type {
   ExternalWorkCiTriageRecord,
   ExternalWorkCiTriageResponse,
@@ -85,7 +86,7 @@ export function createGitHubCiTriageAdapter(options: GitHubCiTriageAdapterOption
   if (!fetchImpl) {
     throw new Error("A fetch implementation is required for GitHub CI triage.");
   }
-  const apiBaseUrl = (options.apiBaseUrl ?? "https://api.github.com").replace(/\/+$/, "");
+  const apiBaseUrl = trimBoundaryCharacters(options.apiBaseUrl ?? "https://api.github.com", "/", "end");
 
   return {
     async getPullRequest(input): Promise<ExternalWorkItem> {
diff --git a/packages/core/src/external-work-scheduler.ts b/packages/core/src/external-work-scheduler.ts
index ab9e01213..b772a96ea 100644
--- a/packages/core/src/external-work-scheduler.ts
+++ b/packages/core/src/external-work-scheduler.ts
@@ -4,6 +4,7 @@ import { existsSync, mkdirSync, writeFileSync } from "node:fs";
 import { isAbsolute, relative, resolve as resolvePath } from "node:path";
 import { DEFAULT_WORKSPACE_ID, type Channel } from "./workspace-types.js";
 import { redactSecrets } from "./redaction.js";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 import { createCoreOriginalProcessEntry, isCoreOriginalLaunchRefusal, type CoreOriginalProcessEntry } from "./original-host-process-entry.js";
 import type {
   ChannelStoreFacet,
@@ -117,7 +118,7 @@ export function createGitHubExternalWorkAdapter(options: GitHubExternalWorkAdapt
   if (!fetchImpl) {
     throw new Error("A fetch implementation is required for GitHub external work polling.");
   }
-  const apiBaseUrl = (options.apiBaseUrl ?? "https://api.github.com").replace(/\/+$/, "");
+  const apiBaseUrl = trimBoundaryCharacters(options.apiBaseUrl ?? "https://api.github.com", "/", "end");
 
   return {
     id: "github-readonly",
diff --git a/packages/core/src/external-work.ts b/packages/core/src/external-work.ts
index 861748f36..b5c20fc9f 100644
--- a/packages/core/src/external-work.ts
+++ b/packages/core/src/external-work.ts
@@ -59,17 +59,45 @@ export function updateExternalWorkflowSecretRef(markdown: string, secretRef: str
 }
 
 export function renderExternalWorkflowPrompt(template: string, input: { readonly item: ExternalWorkItem }): string {
-  return template.replace(/\{\{\s*([^}]+?)\s*\}\}/g, (_match, expression: string) => {
-    const path = expression.trim();
-    if (!path.startsWith("item.")) {
-      throw new ExternalWorkflowValidationError(`Unknown workflow prompt variable: ${path}`);
+  const parts: string[] = [];
+  let cursor = 0;
+  let opening: number | undefined;
+  let index = 0;
+  while (index < template.length) {
+    const character = template[index];
+    const nextCharacter = template[index + 1];
+    if (character === "{" && nextCharacter === "{") {
+      if (opening === undefined) opening = index;
+      index += 1;
+      continue;
     }
-    const value = readDotPath(input, path);
-    if (value === undefined || value === null) {
-      throw new ExternalWorkflowValidationError(`Unknown workflow prompt variable: ${path}`);
+    if (character !== "}") {
+      index += 1;
+      continue;
     }
-    return formatPromptValue(value);
-  });
+    if (nextCharacter !== "}") {
+      opening = undefined;
+      index += 1;
+      continue;
+    }
+    if (opening !== undefined && index > opening + 2) {
+      parts.push(template.slice(cursor, opening));
+      const path = template.slice(opening + 2, index).trim();
+      if (!path.startsWith("item.")) {
+        throw new ExternalWorkflowValidationError(`Unknown workflow prompt variable: ${path}`);
+      }
+      const value = readDotPath(input, path);
+      if (value === undefined || value === null) {
+        throw new ExternalWorkflowValidationError(`Unknown workflow prompt variable: ${path}`);
+      }
+      parts.push(formatPromptValue(value));
+      cursor = index + 2;
+    }
+    opening = undefined;
+    index += 2;
+  }
+  parts.push(template.slice(cursor));
+  return parts.join("");
 }
 
 export function normalizeGitHubExternalWorkItem(input: {
diff --git a/packages/core/src/managed-official-skill-source.ts b/packages/core/src/managed-official-skill-source.ts
index 6ccdd37cb..e268e8112 100644
--- a/packages/core/src/managed-official-skill-source.ts
+++ b/packages/core/src/managed-official-skill-source.ts
@@ -4,6 +4,7 @@ import { join, resolve } from "node:path";
 import { promisify } from "node:util";
 import type { SkillSourceScanRoot } from "./skill-types.js";
 import { createCoreOriginalProcessEntry, isCoreOriginalLaunchRefusal } from "./original-host-process-entry.js";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 
 const execFile = promisify(execFileCallback);
 
@@ -179,7 +180,7 @@ async function readGstackVersion(repoRoot: string): Promise<string | undefined>
 
 function normalizeGithubPath(host: string, pathname: string): string | undefined {
   const normalizedHost = host.trim().toLowerCase();
-  const normalizedPath = pathname.replace(/^\/+|\/+$/g, "").replace(/\.git$/i, "").toLowerCase();
+  const normalizedPath = trimBoundaryCharacters(pathname, "/").replace(/\.git$/i, "").toLowerCase();
   return normalizedHost && normalizedPath ? `${normalizedHost}/${normalizedPath}` : undefined;
 }
 
diff --git a/packages/core/src/memory-markdown-import.ts b/packages/core/src/memory-markdown-import.ts
index b2ecf8dae..6d011ca8e 100644
--- a/packages/core/src/memory-markdown-import.ts
+++ b/packages/core/src/memory-markdown-import.ts
@@ -5,6 +5,7 @@ import type {
   MemorySourceRef,
 } from "./memory-types.js";
 import { DEFAULT_WORKSPACE_ID } from "./workspace-types.js";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 
 export interface MemoryMarkdownImportOptions {
   readonly workspaceId?: string;
@@ -374,7 +375,7 @@ function importSourceRefId(sourceId: string, itemId: string): string {
 }
 
 function sanitizeSourceId(value: string): string {
-  const sanitized = value.trim().replace(/[^A-Za-z0-9._:-]+/g, "_").replace(/^_+|_+$/g, "");
+  const sanitized = trimBoundaryCharacters(value.trim().replace(/[^A-Za-z0-9._:-]+/g, "_"), "_");
   return sanitized || "memory-markdown-import";
 }
 
diff --git a/packages/core/src/provider-adapters.ts b/packages/core/src/provider-adapters.ts
index 1e196e162..32d2a36b7 100644
--- a/packages/core/src/provider-adapters.ts
+++ b/packages/core/src/provider-adapters.ts
@@ -4,6 +4,7 @@ import {
   isKnownGenerationModel,
   isOpenAIImagesGenerationModel,
 } from "./provider-generation-models.js";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 
 export type ProviderAdapterMediaKind = "seedream" | "seedance" | "jimeng";
 
@@ -97,10 +98,10 @@ export function builtinMediaModelCandidates(kind: ProviderAdapterMediaKind | und
 
 export function normalizeProviderBaseUrlForAdapter(input: Pick<ProviderAdapterInput, "api" | "kind" | "baseUrl">): string {
   const rawBaseUrl = input.baseUrl?.trim() || defaultProviderBaseUrl(input);
-  const withoutTrailingSlash = rawBaseUrl.replace(/\/+$/, "");
+  const withoutTrailingSlash = trimBoundaryCharacters(rawBaseUrl, "/", "end");
   try {
     const url = new URL(withoutTrailingSlash);
-    let pathname = url.pathname.replace(/\/+$/, "");
+    let pathname = trimBoundaryCharacters(url.pathname, "/", "end");
     for (const suffix of endpointSuffixes) {
       if (pathname.endsWith(suffix)) {
         pathname = pathname.slice(0, -suffix.length);
@@ -110,7 +111,7 @@ export function normalizeProviderBaseUrlForAdapter(input: Pick<ProviderAdapterIn
     url.pathname = pathname || "/";
     url.search = "";
     url.hash = "";
-    return url.toString().replace(/\/+$/, "");
+    return trimBoundaryCharacters(url.toString(), "/", "end");
   } catch {
     return withoutTrailingSlash;
   }
diff --git a/packages/core/src/provider-generation-models.ts b/packages/core/src/provider-generation-models.ts
index 5ac581f4e..56c5a84ce 100644
--- a/packages/core/src/provider-generation-models.ts
+++ b/packages/core/src/provider-generation-models.ts
@@ -128,8 +128,6 @@ const generationFamilyRules: readonly GenerationModelFamilyRule[] = [
   },
 ];
 
-const bflEditPattern = /(?:^|[/_-])flux(?:[/_.-].*)?(?:kontext|fill)(?=$|[/_.-])/i;
-
 export function inferGenerationModelFamily(
   modelId: string | undefined,
   _context: GenerationModelFamilyContext = {},
@@ -148,7 +146,7 @@ export function inferGenerationModelFamily(
     ownedBy: rule.ownedBy,
     api: "media-job",
     input: ["text", "image"],
-    capabilities: rule.familyId === "black-forest-labs" && bflEditPattern.test(normalized)
+    capabilities: rule.familyId === "black-forest-labs" && isBlackForestLabsEditModel(normalized)
       ? mergeCapabilities(rule.capabilities, ["image_edit"])
       : rule.capabilities,
     supportedParameters: rule.supportedParameters,
@@ -157,6 +155,41 @@ export function inferGenerationModelFamily(
   };
 }
 
+function isBlackForestLabsEditModel(value: string): boolean {
+  const lower = value.toLowerCase();
+  let searchFrom = 0;
+  while (searchFrom < lower.length) {
+    const fluxIndex = lower.indexOf("flux", searchFrom);
+    if (fluxIndex === -1) return false;
+    const before = lower[fluxIndex - 1];
+    if (fluxIndex === 0 || isModelSeparator(before)) {
+      const afterFlux = fluxIndex + "flux".length;
+      if (hasEditMarkerAt(lower, afterFlux)) return true;
+      if (isModelSeparator(lower[afterFlux])) {
+        for (let index = afterFlux + 1; index < lower.length; index += 1) {
+          if (hasEditMarkerAt(lower, index)) return true;
+        }
+        return false;
+      }
+    }
+    searchFrom = fluxIndex + "flux".length;
+  }
+  return false;
+}
+
+function hasEditMarkerAt(value: string, index: number): boolean {
+  for (const marker of ["kontext", "fill"] as const) {
+    if (value.startsWith(marker, index) && (index + marker.length === value.length || isModelSeparator(value[index + marker.length]))) {
+      return true;
+    }
+  }
+  return false;
+}
+
+function isModelSeparator(value: string | undefined): boolean {
+  return value !== undefined && "/_.-".includes(value);
+}
+
 export function isOpenAIImagesGenerationModel(modelId: string | undefined): boolean {
   return inferGenerationModelFamily(modelId)?.familyId === "openai-images";
 }
diff --git a/packages/core/src/runtime-trace-projection.ts b/packages/core/src/runtime-trace-projection.ts
index 56e3132be..a5eac99e1 100644
--- a/packages/core/src/runtime-trace-projection.ts
+++ b/packages/core/src/runtime-trace-projection.ts
@@ -18,6 +18,7 @@ import type {
   ProjectLingTraceInput,
   ProjectLingTraceResult,
 } from "./runtime-trace-types.js";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 
 export function projectLingTrace(input: ProjectLingTraceInput): ProjectLingTraceResult {
   const traceId = createLingTraceId(input.subjectKind, input.subjectId);
@@ -490,7 +491,7 @@ function latestDefined(values: readonly (string | undefined)[]): string | undefi
 }
 
 function safeIdFragment(value: string): string {
-  const cleaned = value.replace(/[^A-Za-z0-9._-]+/g, "_").replace(/^_+|_+$/g, "");
+  const cleaned = trimBoundaryCharacters(value.replace(/[^A-Za-z0-9._-]+/g, "_"), "_");
   return cleaned || "unknown";
 }
 
diff --git a/packages/core/src/sandbox-types.ts b/packages/core/src/sandbox-types.ts
index c71f550c1..5fa0085e8 100644
--- a/packages/core/src/sandbox-types.ts
+++ b/packages/core/src/sandbox-types.ts
@@ -1,4 +1,5 @@
 import { isSupportedSecretRef } from "./secrets.js";
+import { trimBoundaryCharacters } from "./string-boundaries.js";
 
 export type LingSandboxBackendKind = "local_unix" | "platform_native" | "docker" | "hosted" | "external";
 
@@ -260,7 +261,7 @@ function isInsideRoot(candidate: string, root: string): boolean {
 }
 
 function stripTrailingSlash(path: string): string {
-  return path.replace(/\/+$/, "");
+  return trimBoundaryCharacters(path, "/", "end");
 }
 
 function looksSecretLikeName(name: string): boolean {
diff --git a/packages/core/src/sandbox/execution-policy-paths.ts b/packages/core/src/sandbox/execution-policy-paths.ts
index 62b0f691d..e6f21f2dc 100644
--- a/packages/core/src/sandbox/execution-policy-paths.ts
+++ b/packages/core/src/sandbox/execution-policy-paths.ts
@@ -1,5 +1,6 @@
 import { posix, win32 } from "node:path";
 import { ExecutionSandboxPolicyError } from "./execution-policy-types.js";
+import { trimBoundaryCharacters } from "../string-boundaries.js";
 
 /** Lexical checks only. They do not establish realpath, reparse-point or inode authority. */
 export function executionSandboxPaths(platform: "darwin" | "win32") {
@@ -19,8 +20,9 @@ export function executionSandboxPaths(platform: "darwin" | "win32") {
     } else if (!paths.isAbsolute(value) || value.startsWith("//")) {
       throw new ExecutionSandboxPolicyError("invalid_sandbox_path");
     }
-    const normalized = paths.normalize(value).replace(platform === "win32" ? /[\\/]+$/u : /\/+$/u, "");
-    if (!normalized || normalized === paths.parse(normalized).root.replace(/[\\/]+$/u, "")) throw new ExecutionSandboxPolicyError("invalid_sandbox_path");
+    const trailingSeparators = platform === "win32" ? "\\/" : "/";
+    const normalized = trimBoundaryCharacters(paths.normalize(value), trailingSeparators, "end");
+    if (!normalized || normalized === trimBoundaryCharacters(paths.parse(normalized).root, trailingSeparators, "end")) throw new ExecutionSandboxPolicyError("invalid_sandbox_path");
     return platform === "win32" ? normalized[0].toUpperCase() + normalized.slice(1) : normalized;
   }
   const inside = (candidate: string, root: string): boolean => key(candidate) === key(root) || key(candidate).startsWith(key(root) + paths.sep);
diff --git a/packages/core/src/skill-optimization-preview.ts b/packages/core/src/skill-optimization-preview.ts
index 0e3925f2f..054a49efc 100644
--- a/packages/core/src/skill-optimization-preview.ts
+++ b/packages/core/src/skill-optimization-preview.ts
@@ -1,5 +1,6 @@
 import { createHash } from "node:crypto";
 import { buildSkillSourceContextFindings } from "./skill-source-health.js";
+import { parseSkillFrontMatter } from "./skill-source-definition.js";
 import type {
   SkillCompiledRuntimeView,
   SkillCompiledSourceMapEntry,
@@ -264,20 +265,11 @@ function findRoutingDescription(snapshot: SkillSourceSnapshot): string | undefin
 }
 
 function parseFrontMatterName(markdown: string): string | undefined {
-  return parseFrontMatterField(markdown, "name");
+  return parseSkillFrontMatter(markdown).frontMatter.name;
 }
 
 function parseFrontMatterDescription(markdown: string): string | undefined {
-  return parseFrontMatterField(markdown, "description");
-}
-
-function parseFrontMatterField(markdown: string, field: string): string | undefined {
-  const frontMatter = /^---\n([\s\S]*?)\n---/u.exec(markdown)?.[1];
-  if (!frontMatter) {
-    return undefined;
-  }
-  const match = new RegExp(`^${field}:\\s*["']?(.+?)["']?\\s*$`, "imu").exec(frontMatter)?.[1];
-  return match && match !== "|" && match !== ">" ? match : undefined;
+  return parseSkillFrontMatter(markdown).frontMatter.description;
 }
 
 function estimateOriginalTokens(snapshot: SkillSourceSnapshot): number {
diff --git a/packages/core/src/skill-source-health.ts b/packages/core/src/skill-source-health.ts
index 685f75f32..3969ab013 100644
--- a/packages/core/src/skill-source-health.ts
+++ b/packages/core/src/skill-source-health.ts
@@ -1,4 +1,5 @@
 import { GSTACK_DEFAULT_SKILL_PACK_ID } from "./builtin-skill-packs.js";
+import { parseSkillFrontMatter } from "./skill-source-definition.js";
 import type {
   SkillRegistryEntry,
   SkillSourceContextFindingEvidence,
@@ -341,19 +342,7 @@ function findRoutingDescription(snapshot: SkillSourceSnapshot): string | undefin
 }
 
 function parseFrontMatterDescription(markdown: string): string | undefined {
-  const frontMatter = /^---\n([\s\S]*?)\n---/u.exec(markdown)?.[1];
-  if (!frontMatter) {
-    return undefined;
-  }
-  const singleLine = /^description:\s*["']?(.+?)["']?\s*$/imu.exec(frontMatter)?.[1];
-  if (singleLine && singleLine !== "|" && singleLine !== ">") {
-    return singleLine;
-  }
-  const block = /^description:\s*[|>]\s*\n((?:[ \t]+.+\n?)+)/imu.exec(frontMatter)?.[1];
-  if (block) {
-    return block.split(/\r?\n/u).map((line) => line.trim()).filter(Boolean).join(" ");
-  }
-  return undefined;
+  return parseSkillFrontMatter(markdown).frontMatter.description;
 }
 
 function stripFrontMatter(markdown: string): string {
diff --git a/packages/core/src/skill-source-sync.ts b/packages/core/src/skill-source-sync.ts
index 2c3c5ac08..1557b1374 100644
--- a/packages/core/src/skill-source-sync.ts
+++ b/packages/core/src/skill-source-sync.ts
@@ -6,6 +6,7 @@ import { parseDocument } from "yaml";
 import { buildSkillSourceContextFindings } from "./skill-source-health.js";
 import { parseSkillFrontMatter, previewSkillSourceDefinition, splitMarkdownFrontMatter } from "./skill-source-definition.js";
 import { selectActiveSkillSourceSnapshots } from "./skill-source-selection.js";
+import { collapseRepeatedCharacter, trimBoundaryCharacters } from "./string-boundaries.js";
 import type {
   CreateSkillSourceSnapshotInput,
   SkillAgentMetadata,
@@ -293,14 +294,10 @@ export async function previewGstackSkillSourceSync(input: PreviewGstackSkillSour
 }
 
 export function normalizeGstackSkillId(input: string): string {
-  return input
-    .trim()
-    .toLowerCase()
-    .replace(/^\/+|\/+$/g, "")
+  const normalized = trimBoundaryCharacters(input.trim().toLowerCase(), "/")
     .replace(/\s+/g, "-")
-    .replace(/[^a-z0-9._-]+/g, "-")
-    .replace(/-+/g, "-")
-    .replace(/^-|-$/g, "");
+    .replace(/[^a-z0-9._-]+/g, "-");
+  return trimBoundaryCharacters(collapseRepeatedCharacter(normalized, "-"), "-");
 }
 
 async function gstackRootFromRepoPath(repoPath: string, id?: string): Promise<SkillSourceScanRoot> {
@@ -513,10 +510,15 @@ function classifySkillSourceWarnings(markdown: string, sourceSizeBytes: number,
     /\$GSTACK_BIN\b/,
     /\bgstack-/,
   ], "SKILL.md references host-specific gstack or Claude runtime behavior.");
-  addPatternWarning(warnings, markdown, "unsupported_tool", "warning", [
-    /\bmcp__[A-Za-z0-9_-]+__[A-Za-z0-9_-]+\b/,
-    /\bnative AUQ\b/i,
-  ], "SKILL.md references a tool surface that Ling may not expose directly.");
+  const unsupportedToolLine = findFirstMcpToolLine(markdown) ?? findFirstMatchingLine(markdown, /\bnative AUQ\b/i);
+  if (unsupportedToolLine !== undefined) {
+    warnings.push({
+      code: "unsupported_tool",
+      severity: "warning",
+      message: "SKILL.md references a tool surface that Ling may not expose directly.",
+      location: `SKILL.md:${unsupportedToolLine}`,
+    });
+  }
   addPatternWarning(warnings, markdown, "permission_like_instruction", "info", [
     /\b(approval bypass|bypass approval|ignore approval|grant .*tool|allowedTools|disallowedTools|permission gate)\b/i,
   ], "SKILL.md contains permission-like language that must not override Ling policy.");
@@ -609,7 +611,10 @@ export function buildSkillSourceSemanticProfile(input: {
     workflowVerbs: uniqueSortedClaims(extractClaims(input.markdown, workflowVerbPatterns)),
     stopGateCount: countPattern(input.markdown, /\b(STOP|approval gate|ask for approval|ask the user|confirm before|requires confirmation)\b/gi),
     approvalGateClaims: uniqueSortedClaims(extractClaims(input.markdown, approvalGatePatterns)),
-    toolClaims: uniqueSortedClaims(extractClaims(input.markdown, toolClaimPatterns)),
+    toolClaims: uniqueSortedClaims([
+      ...extractClaims(input.markdown, toolClaimPatterns),
+      ...extractMcpToolClaims(input.markdown),
+    ]),
     hostClaims: uniqueSortedClaims(extractClaims(input.markdown, hostClaimPatterns)),
     externalActionClaims: uniqueSortedClaims(extractClaims(input.markdown, externalActionPatterns)),
     destructiveClaims: uniqueSortedClaims(extractClaims(input.markdown, destructiveClaimPatterns)),
@@ -719,7 +724,7 @@ const approvalGatePatterns = [
 ] as const;
 
 const toolClaimPatterns = [
-  /\b(bash|shell|filesystem|computer-use|browser|playwright|apply_patch|mcp__[A-Za-z0-9_-]+__[A-Za-z0-9_-]+|git|gh)\b/gi,
+  /\b(bash|shell|filesystem|computer-use|browser|playwright|apply_patch|git|gh)\b/gi,
 ] as const;
 
 const hostClaimPatterns = [
@@ -767,6 +772,46 @@ function extractClaims(markdown: string, patterns: readonly RegExp[]): string[]
   return claims;
 }
 
+function extractMcpToolClaims(markdown: string): string[] {
+  const claims: string[] = [];
+  const lowerMarkdown = markdown.toLowerCase();
+  for (let index = 0; index < markdown.length;) {
+    const start = lowerMarkdown.indexOf("mcp__", index);
+    if (start === -1) break;
+    if (isMcpToolCharacter(markdown[start - 1])) {
+      index = start + 1;
+      continue;
+    }
+    let end = start + "mcp__".length;
+    while (end < markdown.length && isMcpToolCharacter(markdown[end])) end += 1;
+    const candidate = markdown.slice(start, end);
+    const components = candidate.slice("mcp__".length);
+    const separator = components.indexOf("__", 1);
+    if (separator !== -1 && separator + 2 < components.length) claims.push(candidate.toLowerCase());
+    index = Math.max(end, start + 1);
+  }
+  return claims;
+}
+
+function findFirstMcpToolLine(markdown: string): number | undefined {
+  const claim = extractMcpToolClaims(markdown)[0];
+  if (!claim) return undefined;
+  const index = markdown.toLowerCase().indexOf(claim);
+  if (index === -1) return undefined;
+  let line = 1;
+  for (let cursor = 0; cursor < index; cursor += 1) {
+    if (markdown[cursor] === "\n") line += 1;
+  }
+  return line;
+}
+
+function isMcpToolCharacter(value: string | undefined): boolean {
+  if (!value) return false;
+  const code = value.charCodeAt(0);
+  return (code >= 48 && code <= 57) || (code >= 65 && code <= 90) ||
+    (code >= 97 && code <= 122) || value === "_" || value === "-";
+}
+
 function uniqueSortedClaims(values: readonly string[]): string[] {
   return [...new Set(values.filter(Boolean))].sort();
 }
diff --git a/packages/core/src/string-boundaries.ts b/packages/core/src/string-boundaries.ts
new file mode 100644
index 000000000..84e89dc43
--- /dev/null
+++ b/packages/core/src/string-boundaries.ts
@@ -0,0 +1,93 @@
+export function trimBoundaryCharacters(
+  value: string,
+  characters: string,
+  side: "start" | "end" | "both" = "both",
+): string {
+  let start = 0;
+  let end = value.length;
+  if (side !== "end") {
+    while (start < end && characters.includes(value[start] ?? "")) start += 1;
+  }
+  if (side !== "start") {
+    while (end > start && characters.includes(value[end - 1] ?? "")) end -= 1;
+  }
+  return start === 0 && end === value.length ? value : value.slice(start, end);
+}
+
+export function collapseRepeatedCharacter(value: string, character: string): string {
+  const result: string[] = [];
+  let previousWasCharacter = false;
+  for (const current of value) {
+    if (current === character) {
+      if (!previousWasCharacter) result.push(current);
+      previousWasCharacter = true;
+    } else {
+      result.push(current);
+      previousWasCharacter = false;
+    }
+  }
+  return result.join("");
+}
+
+export function extractWrappedSection(
+  value: string,
+  labels: readonly string[],
+  boundaryLabels: readonly string[],
+  acceptedColons = ":：",
+  allowWhitespaceBeforeColon = true,
+): string | undefined {
+  const searchValue = asciiLowercase(value);
+  let selected: { readonly labelIndex: number; readonly start: number } | undefined;
+  for (const label of labels) {
+    const searchLabel = asciiLowercase(label);
+    let labelIndex = searchValue.indexOf(searchLabel);
+    while (labelIndex !== -1) {
+      let start = labelIndex + label.length;
+      if (allowWhitespaceBeforeColon) {
+        while (start < value.length && isWhitespace(value[start])) start += 1;
+      }
+      if (acceptedColons.includes(value[start] ?? "")) {
+        start += 1;
+        while (start < value.length && isWhitespace(value[start])) start += 1;
+        if (!selected || labelIndex < selected.labelIndex) selected = { labelIndex, start };
+        break;
+      }
+      labelIndex = searchValue.indexOf(searchLabel, labelIndex + 1);
+    }
+  }
+  if (!selected) return undefined;
+  let end = value.length;
+  for (const boundary of boundaryLabels) {
+    const searchBoundary = asciiLowercase(boundary);
+    let boundaryIndex = searchValue.indexOf(searchBoundary, selected.start);
+    while (boundaryIndex !== -1) {
+      const before = value[boundaryIndex - 1];
+      const after = value[boundaryIndex + boundary.length];
+      if (isWhitespace(before) && !isWordCharacter(after)) {
+        end = Math.min(end, boundaryIndex);
+        break;
+      }
+      boundaryIndex = searchValue.indexOf(searchBoundary, boundaryIndex + 1);
+    }
+  }
+  const section = value.slice(selected.start, end).trim();
+  return section || undefined;
+}
+
+function asciiLowercase(value: string): string {
+  const result: string[] = [];
+  for (let index = 0; index < value.length; index += 1) {
+    const character = value[index] ?? "";
+    const code = character.charCodeAt(0);
+    result.push(code >= 65 && code <= 90 ? String.fromCharCode(code + 32) : character);
+  }
+  return result.join("");
+}
+
+function isWhitespace(value: string | undefined): boolean {
+  return value !== undefined && value.trim() === "";
+}
+
+function isWordCharacter(value: string | undefined): boolean {
+  return value !== undefined && /[A-Za-z0-9_]/u.test(value);
+}
diff --git a/packages/core/src/terminal-report-fallback.ts b/packages/core/src/terminal-report-fallback.ts
index 5a9509670..6f1909de8 100644
--- a/packages/core/src/terminal-report-fallback.ts
+++ b/packages/core/src/terminal-report-fallback.ts
@@ -11,6 +11,7 @@ import type {
 } from "./terminal-report-types.js";
 import { createTerminalReportId } from "./terminal-report-types.js";
 import { formatSynchronousCatalogMessage } from "@ling-stack/localization";
+import { extractWrappedSection, trimBoundaryCharacters } from "./string-boundaries.js";
 
 export interface BuildDeterministicTerminalReportInput {
   readonly id?: string;
@@ -113,10 +114,14 @@ export function buildDeterministicTerminalReportContent(snapshot: TerminalReport
  */
 function fallbackOutcomeTitle(snapshot: TerminalReportEvidenceSnapshot): string {
   const normalized = snapshot.objective.replace(/\s+/gu, " ").trim();
-  const wrapped = normalized.match(
-    /(?:原始请求|Original request):\s*(.+?)(?=\s+(?:Requested objective\b|No literal file content\b|已物化的执行操作\b|参与者操作归属\b|Permission boundary\b)|$)/iu,
-  )?.[1]?.trim();
-  const candidate = (wrapped || normalized).replace(/^\*+|\*+$/gu, "").trim();
+  const wrapped = extractWrappedSection(normalized, ["原始请求", "Original request"], [
+    "Requested objective",
+    "No literal file content",
+    "已物化的执行操作",
+    "参与者操作归属",
+    "Permission boundary",
+  ], ":", false);
+  const candidate = trimBoundaryCharacters(wrapped || normalized, "*").trim();
   const maxLength = 120;
   return candidate.length > maxLength ? `${candidate.slice(0, maxLength - 1).trimEnd()}…` : candidate;
 }
diff --git a/packages/core/test/external-work.test.mjs b/packages/core/test/external-work.test.mjs
index e80f30350..fb32a9424 100644
--- a/packages/core/test/external-work.test.mjs
+++ b/packages/core/test/external-work.test.mjs
@@ -45,6 +45,11 @@ Labels: {{ item.labels }}
 Head: {{ item.headSha }}
 `;
 
+test("workflow prompt rendering handles long malformed placeholders without backtracking", () => {
+  const template = `${"{{".repeat(50_000)}${" ".repeat(100_000)}`;
+  assert.equal(renderExternalWorkflowPrompt(template, { item: { identifier: "#1" } }), template);
+});
+
 const linearWorkflowMarkdown = `---
 source:
   kind: linear
diff --git a/packages/core/test/provider-check.test.mjs b/packages/core/test/provider-check.test.mjs
index 58d8a2652..2f4a40c59 100644
--- a/packages/core/test/provider-check.test.mjs
+++ b/packages/core/test/provider-check.test.mjs
@@ -33,6 +33,11 @@ test("default provider templates include Seedream live HTTP transport config", (
   assert.equal(seedream.enabled, false);
 });
 
+test("BFL edit capability inference handles long model identifiers", () => {
+  const model = `black-forest-labs/flux-${"flux-".repeat(20_000)}kontext-pro`;
+  assert.deepEqual(inferGenerationModelFamily(model)?.capabilities, ["image_generate", "image_edit"]);
+});
+
 test("official provider templates expose fixed default base URLs", () => {
   const templates = defaultProviderTemplates();
   const byId = new Map(templates.map((template) => [template.templateId, template]));
diff --git a/packages/core/test/skill-source-sync.test.mjs b/packages/core/test/skill-source-sync.test.mjs
index f917dd720..66a451bbb 100644
--- a/packages/core/test/skill-source-sync.test.mjs
+++ b/packages/core/test/skill-source-sync.test.mjs
@@ -11,6 +11,7 @@ import {
   getBuiltinSkillPack,
   ManagedOfficialSkillSourceError,
   normalizeOfficialGstackOrigin,
+  normalizeGstackSkillId,
   officialGstackSourceScanRoot,
   previewGstackSkillSourceSync,
   refreshManagedOfficialGstackSource,
@@ -18,6 +19,13 @@ import {
   resolveSkillSourceRoots,
 } from "../dist/index.js";
 
+test("skill identifiers and MCP claims handle long ambiguous delimiters", () => {
+  assert.equal(normalizeGstackSkillId(`${"/".repeat(100_000)}review${"-".repeat(100_000)}`), "review");
+  const tool = `mcp__server${"_".repeat(100_000)}__tool`;
+  const profile = buildSkillSourceSemanticProfile({ markdown: `Use ${tool} and MCP__Server__Tool, but not mcp____.`, sourceChecksum: "sha256:long" });
+  assert.deepEqual(profile.toolClaims, [tool, "mcp__server__tool"]);
+});
+
 async function withTempDir(prefix, fn) {
   const root = await mkdtemp(resolve(tmpdir(), prefix));
   try {
diff --git a/packages/daemon/src/features/channels/link-previews.ts b/packages/daemon/src/features/channels/link-previews.ts
index eb2ca5a6d..19ccb4b18 100644
--- a/packages/daemon/src/features/channels/link-previews.ts
+++ b/packages/daemon/src/features/channels/link-previews.ts
@@ -7,6 +7,7 @@ import { writeJson } from "../../shared/response.js";
 import { preparePinnedPublicHttpRequest, PublicHttpError } from "../../shared/public-http-transport.js";
 import { validateChannelMessageOwner } from "./protocol.js";
 import { parseChannelLinkPreviewRequest, type ChannelLinkPreviewRequest } from "./schemas.js";
+import { decodeHtmlEntitiesOnce } from "../../shared/html-text.js";
 
 const CHANNEL_LINK_PREVIEW_TIMEOUT_MS = 5_000;
 const CHANNEL_LINK_PREVIEW_MAX_BYTES = 192 * 1024;
@@ -251,13 +252,7 @@ function readTitle(html: string): string | undefined {
 }
 
 function decodeHtmlEntities(value: string): string {
-  return value
-    .replace(/&amp;/g, "&")
-    .replace(/&lt;/g, "<")
-    .replace(/&gt;/g, ">")
-    .replace(/&quot;/g, "\"")
-    .replace(/&#39;/g, "'")
-    .trim();
+  return decodeHtmlEntitiesOnce(value, { names: ["amp", "lt", "gt", "quot", "#39"] }).trim();
 }
 
 export function normalizePreviewUrl(rawUrl: string): string {
diff --git a/packages/daemon/src/features/channels/materialization/extractors/link-body.ts b/packages/daemon/src/features/channels/materialization/extractors/link-body.ts
index 6e82e9d35..4421457e7 100644
--- a/packages/daemon/src/features/channels/materialization/extractors/link-body.ts
+++ b/packages/daemon/src/features/channels/materialization/extractors/link-body.ts
@@ -1,6 +1,7 @@
 import type { LingInputExtractor } from "../registry.js";
 import { decidePublicLinkMaterializationPolicy } from "../policy.js";
 import { preparePinnedPublicHttpRequest, PublicHttpError } from "../../../../shared/public-http-transport.js";
+import { decodeHtmlEntitiesOnce, stripUnreadableHtml } from "../../../../shared/html-text.js";
 
 export const linkBodyExtractorV1: LingInputExtractor = {
   id: "link_body_v1",
@@ -242,18 +243,7 @@ function isSupportedLinkBodyContentType(contentType: string): boolean {
 }
 
 export function extractReadableHtmlText(html: string): string {
-  const body = /<body\b[^>]*>([\s\S]*?)<\/body>/iu.exec(html)?.[1] ?? html;
-  return normalizeExtractedText(
-    body
-      .replace(/<script\b[\s\S]*?<\/script>/giu, " ")
-      .replace(/<style\b[\s\S]*?<\/style>/giu, " ")
-      .replace(/<noscript\b[\s\S]*?<\/noscript>/giu, " ")
-      .replace(/<svg\b[\s\S]*?<\/svg>/giu, " ")
-      .replace(/<(?:nav|footer|form|aside)\b[\s\S]*?<\/(?:nav|footer|form|aside)>/giu, " ")
-      .replace(/<br\s*\/?>/giu, "\n")
-      .replace(/<\/(?:p|div|section|article|main|h[1-6]|li|tr)>/giu, "\n")
-      .replace(/<[^>]+>/gu, " "),
-  );
+  return normalizeExtractedText(stripUnreadableHtml(html));
 }
 
 export function normalizeExtractedText(value: string): string {
@@ -266,15 +256,7 @@ export function normalizeExtractedText(value: string): string {
 }
 
 function decodeHtmlEntities(value: string): string {
-  return value
-    .replace(/&nbsp;/g, " ")
-    .replace(/&amp;/g, "&")
-    .replace(/&lt;/g, "<")
-    .replace(/&gt;/g, ">")
-    .replace(/&quot;/g, "\"")
-    .replace(/&#39;/g, "'")
-    .replace(/&#(\d+);/g, (_match, code: string) => String.fromCodePoint(Number.parseInt(code, 10)))
-    .replace(/&#x([0-9a-f]+);/giu, (_match, code: string) => String.fromCodePoint(Number.parseInt(code, 16)));
+  return decodeHtmlEntitiesOnce(value, { numeric: true });
 }
 
 function publicLinkPolicyMessage(reason: string): string {
diff --git a/packages/daemon/src/features/web/web-search-adapters.ts b/packages/daemon/src/features/web/web-search-adapters.ts
index 370fee56e..047946b17 100644
--- a/packages/daemon/src/features/web/web-search-adapters.ts
+++ b/packages/daemon/src/features/web/web-search-adapters.ts
@@ -2,6 +2,7 @@ import {
   normalizeProviderBaseUrlForAdapter,
   type ProviderPreset,
 } from "@ling-stack/core";
+import { decodeHtmlEntitiesOnce } from "../../shared/html-text.js";
 
 const DEFAULT_OPENAI_WEB_SEARCH_MODEL = "gpt-5.5";
 const WEB_SEARCH_MAX_OUTPUT_TOKENS = 900;
@@ -685,15 +686,13 @@ function isPublicHttpUrl(value: string): boolean {
 }
 
 function cleanMarkdownText(value: string): string {
-  return value
+  const cleaned = value
     .replace(/\*\*/gu, "")
     .replace(/!\[[^\]]*\]\([^)]+\)/gu, "")
     .replace(/\[[^\]]*\]\([^)]+\)/gu, "")
-    .replace(/&amp;/gu, "&")
-    .replace(/&quot;/gu, "\"")
-    .replace(/&#39;/gu, "'")
     .replace(/\s+/gu, " ")
     .trim();
+  return decodeHtmlEntitiesOnce(cleaned, { names: ["amp", "quot", "#39"] });
 }
 
 function looksLikeBareHost(value: string): boolean {
diff --git a/packages/daemon/src/shared/html-text.ts b/packages/daemon/src/shared/html-text.ts
new file mode 100644
index 000000000..6b7ed87e1
--- /dev/null
+++ b/packages/daemon/src/shared/html-text.ts
@@ -0,0 +1,153 @@
+interface HtmlEntityDecodeOptions {
+  readonly names?: readonly ("nbsp" | "amp" | "lt" | "gt" | "quot" | "#39")[];
+  readonly numeric?: boolean;
+}
+
+const htmlEntityPattern = /&(?:nbsp|amp|lt|gt|quot|#39|#\d{1,7}|#x[0-9a-f]{1,6});/giu;
+
+export function decodeHtmlEntitiesOnce(value: string, options: HtmlEntityDecodeOptions = {}): string {
+  const names = new Set(options.names ?? ["nbsp", "amp", "lt", "gt", "quot", "#39"]);
+  return value.replace(htmlEntityPattern, (entity) => {
+    const token = entity.slice(1, -1).toLowerCase();
+    if (token.startsWith("#") && token !== "#39") {
+      if (options.numeric !== true) return entity;
+      const radix = token.startsWith("#x") ? 16 : 10;
+      const digits = token.slice(radix === 16 ? 2 : 1);
+      const codePoint = Number.parseInt(digits, radix);
+      return isValidCodePoint(codePoint) ? String.fromCodePoint(codePoint) : entity;
+    }
+    if (!names.has(token as "nbsp" | "amp" | "lt" | "gt" | "quot" | "#39")) return entity;
+    switch (token) {
+      case "nbsp": return " ";
+      case "amp": return "&";
+      case "lt": return "<";
+      case "gt": return ">";
+      case "quot": return "\"";
+      case "#39": return "'";
+      default: return entity;
+    }
+  });
+}
+
+export function stripUnreadableHtml(html: string): string {
+  const body = findBodyContentRange(html);
+  const source = body ? html.slice(body.start, body.end) : html;
+  const output: string[] = [];
+  let cursor = 0;
+  while (cursor < source.length) {
+    if (source[cursor] !== "<") {
+      output.push(source[cursor] ?? "");
+      cursor += 1;
+      continue;
+    }
+    if (source.startsWith("<!--", cursor)) {
+      const commentEnd = source.indexOf("-->", cursor + 4);
+      cursor = commentEnd === -1 ? source.length : commentEnd + 3;
+      output.push(" ");
+      continue;
+    }
+    const tag = readHtmlTag(source, cursor);
+    if (!tag) {
+      output.push("<");
+      cursor += 1;
+      continue;
+    }
+    cursor = tag.end;
+    if (!tag.closing && unreadableTags.has(tag.name)) {
+      cursor = findClosingTag(source, tag.name, cursor)?.end ?? source.length;
+      output.push(" ");
+      continue;
+    }
+    if (tag.name === "br" || (tag.closing && blockTags.has(tag.name))) output.push("\n");
+    else output.push(" ");
+  }
+  return output.join("");
+}
+
+const unreadableTags = new Set(["script", "style", "noscript", "svg", "nav", "footer", "form", "aside", "head"]);
+const blockTags = new Set(["p", "div", "section", "article", "main", "h1", "h2", "h3", "h4", "h5", "h6", "li", "tr"]);
+
+function findBodyContentRange(html: string): { readonly start: number; readonly end: number } | undefined {
+  let cursor = 0;
+  while (cursor < html.length) {
+    const opening = html.indexOf("<", cursor);
+    if (opening === -1) return undefined;
+    const tag = readHtmlTag(html, opening);
+    if (!tag) {
+      cursor = opening + 1;
+      continue;
+    }
+    if (!tag.closing && tag.name === "body") {
+      const closing = findClosingTag(html, "body", tag.end);
+      return { start: tag.end, end: closing?.start ?? html.length };
+    }
+    if (!tag.closing && unreadableTags.has(tag.name)) {
+      cursor = findClosingTag(html, tag.name, tag.end)?.end ?? html.length;
+    } else {
+      cursor = tag.end;
+    }
+  }
+  return undefined;
+}
+
+function findClosingTag(html: string, name: string, start: number): HtmlTag | undefined {
+  let cursor = start;
+  while (cursor < html.length) {
+    const opening = html.indexOf("<", cursor);
+    if (opening === -1) return undefined;
+    const tag = readHtmlTag(html, opening);
+    if (tag?.closing && tag.name === name) return tag;
+    cursor = tag ? tag.end : opening + 1;
+  }
+  return undefined;
+}
+
+interface HtmlTag {
+  readonly start: number;
+  readonly name: string;
+  readonly closing: boolean;
+  readonly end: number;
+}
+
+function readHtmlTag(html: string, start: number): HtmlTag | undefined {
+  if (html[start] !== "<") return undefined;
+  let cursor = start + 1;
+  while (isHtmlWhitespace(html[cursor])) cursor += 1;
+  let closing = false;
+  if (html[cursor] === "/") {
+    closing = true;
+    cursor += 1;
+    while (isHtmlWhitespace(html[cursor])) cursor += 1;
+  }
+  const nameStart = cursor;
+  while (isAsciiLetterOrDigit(html[cursor])) cursor += 1;
+  if (cursor === nameStart) return undefined;
+  const name = html.slice(nameStart, cursor).toLowerCase();
+  let quote: "\"" | "'" | undefined;
+  while (cursor < html.length) {
+    const character = html[cursor];
+    if (quote) {
+      if (character === quote) quote = undefined;
+    } else if (character === "\"" || character === "'") {
+      quote = character;
+    } else if (character === ">") {
+      return { start, name, closing, end: cursor + 1 };
+    }
+    cursor += 1;
+  }
+  return undefined;
+}
+
+function isHtmlWhitespace(value: string | undefined): boolean {
+  return value === " " || value === "\t" || value === "\n" || value === "\r" || value === "\f";
+}
+
+function isAsciiLetterOrDigit(value: string | undefined): boolean {
+  if (!value) return false;
+  const code = value.charCodeAt(0);
+  return (code >= 48 && code <= 57) || (code >= 65 && code <= 90) || (code >= 97 && code <= 122);
+}
+
+function isValidCodePoint(value: number): boolean {
+  return Number.isInteger(value) && value >= 0 && value <= 0x10ffff && !(value >= 0xd800 && value <= 0xdfff);
+}
diff --git a/packages/daemon/test/html-text-security.test.mjs b/packages/daemon/test/html-text-security.test.mjs
new file mode 100644
index 000000000..d1b75565d
--- /dev/null
+++ b/packages/daemon/test/html-text-security.test.mjs
@@ -0,0 +1,14 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { extractReadableHtmlText } from "../dist/features/channels/materialization/extractors/link-body.js";
+import { decodeHtmlEntitiesOnce } from "../dist/shared/html-text.js";
+
+test("readable HTML extraction removes whitespace-terminated script tags", () => {
+  const text = extractReadableHtmlText("ignored before<body><p>before</p><script>secret()</script ><p>after</p></body>ignored after");
+  assert.equal(text, "before\nafter");
+});
+
+test("HTML entities decode exactly one layer", () => {
+  assert.equal(decodeHtmlEntitiesOnce("&amp;lt; &lt; &amp;quot;"), "&lt; < &quot;");
+  assert.equal(decodeHtmlEntitiesOnce("&#60; &#x3c;", { numeric: true }), "< <");
+});
diff --git a/packages/providers/src/utils.ts b/packages/providers/src/utils.ts
index b2dac747c..113041975 100644
--- a/packages/providers/src/utils.ts
+++ b/packages/providers/src/utils.ts
@@ -5,3 +5,13 @@ export function isRecord(value: unknown): value is Record<string, unknown> {
 export function removeUndefined<T extends Record<string, unknown>>(value: T): T {
   return Object.fromEntries(Object.entries(value).filter(([, item]) => item !== undefined)) as T;
 }
+
+export function stripTrailingSlashes(value: string): string;
+export function stripTrailingSlashes(value: undefined): undefined;
+export function stripTrailingSlashes(value: string | undefined): string | undefined;
+export function stripTrailingSlashes(value: string | undefined): string | undefined {
+  if (value === undefined) return undefined;
+  let end = value.length;
+  while (end > 0 && value[end - 1] === "/") end -= 1;
+  return end === value.length ? value : value.slice(0, end);
+}
diff --git a/packages/providers/src/vendors/openai-images.ts b/packages/providers/src/vendors/openai-images.ts
index 9b5a6d026..39515e426 100644
--- a/packages/providers/src/vendors/openai-images.ts
+++ b/packages/providers/src/vendors/openai-images.ts
@@ -10,7 +10,7 @@ import type {
   MediaVendorTransportRequest,
   MediaVendorTransportResult,
 } from "../types.js";
-import { isRecord, removeUndefined } from "../utils.js";
+import { isRecord, removeUndefined, stripTrailingSlashes } from "../utils.js";
 
 const MAX_OPENAI_IMAGE_BYTES = 25 * 1024 * 1024;
 const DEFAULT_OPENAI_IMAGES_TIMEOUT_MS = 180_000;
@@ -19,7 +19,7 @@ export async function executeOpenAIImageGeneration(
   request: MediaVendorTransportRequest,
   options: LiveMediaVendorTransportOptions,
 ): Promise<MediaVendorTransportResult> {
-  const endpoint = `${options.baseUrl?.replace(/\/+$/, "")}/images/generations`;
+  const endpoint = `${stripTrailingSlashes(options.baseUrl)}/images/generations`;
   const payload = buildOpenAIImageGenerationPayload(request.requestPlan);
   const httpClient = options.httpClient ?? defaultMediaVendorHttpClient;
   const abortController = new AbortController();
diff --git a/packages/providers/src/vendors/seedance-task-status.ts b/packages/providers/src/vendors/seedance-task-status.ts
index 5632f3aef..eac624503 100644
--- a/packages/providers/src/vendors/seedance-task-status.ts
+++ b/packages/providers/src/vendors/seedance-task-status.ts
@@ -1,4 +1,4 @@
-import { isRecord } from "../utils.js";
+import { isRecord, stripTrailingSlashes } from "../utils.js";
 import type { MediaRemoteTerminalObservation } from "@ling-stack/contracts";
 
 /** Decoded protocol facts only. Host transport and Core lineage still decide settlement. */
@@ -80,7 +80,7 @@ export function createSeedanceStatusEndpoint(baseUrl: string | undefined, extern
     if ((base.protocol !== "http:" && base.protocol !== "https:") || base.username || base.password || base.search || base.hash) {
       return undefined;
     }
-    return `${baseUrl.replace(/\/+$/, "")}/contents/generations/tasks/${encodeURIComponent(externalOperationId)}`;
+    return `${stripTrailingSlashes(baseUrl)}/contents/generations/tasks/${encodeURIComponent(externalOperationId)}`;
   } catch {
     return undefined;
   }
diff --git a/packages/providers/src/vendors/seedance.ts b/packages/providers/src/vendors/seedance.ts
index 5f6c1636f..6afd65f2a 100644
--- a/packages/providers/src/vendors/seedance.ts
+++ b/packages/providers/src/vendors/seedance.ts
@@ -14,7 +14,7 @@ import type {
   MediaVendorTransportRequest,
   MediaVendorTransportResult,
 } from "../types.js";
-import { isRecord, removeUndefined } from "../utils.js";
+import { isRecord, removeUndefined, stripTrailingSlashes } from "../utils.js";
 import { createSeedanceStatusEndpoint, decodeSeedanceTaskStatusResponse, type SeedanceTerminalObservation } from "./seedance-task-status.js";
 
 const MAX_MEDIA_VENDOR_VIDEO_BYTES = 512 * 1024 * 1024;
@@ -42,7 +42,7 @@ export async function submitSeedanceVideoTask(
   request: MediaVendorTransportRequest,
   options: LiveMediaVendorTransportOptions,
 ): Promise<MediaExecutionSubmission> {
-  const endpoint = `${options.baseUrl?.replace(/\/+$/, "")}/contents/generations/tasks`;
+  const endpoint = `${stripTrailingSlashes(options.baseUrl)}/contents/generations/tasks`;
   const payload = buildSeedanceVideoTaskPayload(request.requestPlan);
   const httpClient = options.httpClient ?? defaultMediaVendorHttpClient;
   const abortController = new AbortController();
@@ -226,7 +226,7 @@ export async function executeSeedanceVideoTask(
   request: MediaVendorTransportRequest,
   options: LiveMediaVendorTransportOptions,
 ): Promise<MediaVendorTransportResult> {
-  const endpoint = `${options.baseUrl?.replace(/\/+$/, "")}/contents/generations/tasks`;
+  const endpoint = `${stripTrailingSlashes(options.baseUrl)}/contents/generations/tasks`;
   const payload = buildSeedanceVideoTaskPayload(request.requestPlan);
   const httpClient = options.httpClient ?? defaultMediaVendorHttpClient;
   const abortController = new AbortController();
diff --git a/packages/providers/src/vendors/seedream.ts b/packages/providers/src/vendors/seedream.ts
index 621f96513..cbdce66eb 100644
--- a/packages/providers/src/vendors/seedream.ts
+++ b/packages/providers/src/vendors/seedream.ts
@@ -10,7 +10,7 @@ import type {
   MediaVendorTransportRequest,
   MediaVendorTransportResult,
 } from "../types.js";
-import { isRecord, removeUndefined } from "../utils.js";
+import { isRecord, removeUndefined, stripTrailingSlashes } from "../utils.js";
 
 const MAX_MEDIA_VENDOR_IMAGE_BYTES = 25 * 1024 * 1024;
 const DEFAULT_LIVE_MEDIA_VENDOR_TIMEOUT_MS = 180_000;
@@ -19,7 +19,7 @@ export async function executeSeedreamImageGeneration(
   request: MediaVendorTransportRequest,
   options: LiveMediaVendorTransportOptions,
 ): Promise<MediaVendorTransportResult> {
-  const endpoint = `${options.baseUrl?.replace(/\/+$/, "")}/images/generations`;
+  const endpoint = `${stripTrailingSlashes(options.baseUrl)}/images/generations`;
   const payload = buildSeedreamImageGenerationPayload(request.requestPlan);
   return await executeSeedreamImageRequest({
     request,
@@ -37,7 +37,7 @@ export async function executeSeedreamImageEdit(
   request: MediaVendorTransportRequest,
   options: LiveMediaVendorTransportOptions,
 ): Promise<MediaVendorTransportResult> {
-  const endpoint = `${options.baseUrl?.replace(/\/+$/, "")}/images/generations`;
+  const endpoint = `${stripTrailingSlashes(options.baseUrl)}/images/generations`;
   const payloadResolution = buildSeedreamImageEditPayload(request.requestPlan);
   if (!payloadResolution.ok) {
     return {
diff --git a/scripts/gstack/testsavant-runtime.mjs b/scripts/gstack/testsavant-runtime.mjs
index 9fa45e17e..3ce54c22e 100644
--- a/scripts/gstack/testsavant-runtime.mjs
+++ b/scripts/gstack/testsavant-runtime.mjs
@@ -22,12 +22,14 @@ export async function verifyTestsavantRuntimeFiles(modelRoot) {
   }
 }
 
-// Matches the frozen upstream preprocessing, including its entity order.
+// Matches the frozen upstream preprocessing while decoding each entity once.
 export function testsavantPlainText(input) {
   if (!input.includes("<")) return input;
   return input.replace(/<(script|style)[^>]*>[\s\S]*?<\/\1>/gi, " ").replace(/<[^>]+>/g, " ")
-    .replace(/&nbsp;/g, " ").replace(/&amp;/g, "&").replace(/&lt;/g, "<").replace(/&gt;/g, ">")
-    .replace(/&quot;/g, '"').replace(/\s+/g, " ").trim();
+    .replace(/&(?:nbsp|amp|lt|gt|quot);/g, entity => ({
+      "&nbsp;": " ", "&amp;": "&", "&lt;": "<", "&gt;": ">", "&quot;": '"',
+    })[entity] ?? entity)
+    .replace(/\s+/g, " ").trim();
 }
 
 export function planTestsavantWindows(ids) {
diff --git a/scripts/gstack/testsavant-runtime.test.mjs b/scripts/gstack/testsavant-runtime.test.mjs
index e1bca7417..ea544f258 100644
--- a/scripts/gstack/testsavant-runtime.test.mjs
+++ b/scripts/gstack/testsavant-runtime.test.mjs
@@ -44,6 +44,10 @@ test("platform resolver uses only packaged Node, and child environment excludes
     API_KEY: "not-real", PATH: "evil", HOME: "/empty", SystemRoot: "C:\\Windows" }), { HOME: "/empty", SystemRoot: "C:\\Windows" });
 });
 
+test("TestSavant plain text decoding preserves nested entities for a later layer", () => {
+  assert.equal(testsavantPlainText("<p>&amp;lt; &lt;</p>"), "&lt; <");
+});
+
 test("the actual gs route receives its vocabulary, without changing the upstream combiner thresholds", () => {
   for (const [input, output] of [["safe", "safe"], ["log_only", "safe"], ["warn", "suspicious"], ["block", "unsafe"]]) {
     assert.equal(testsavantRouteVerdict({ verdict: input }), output);