fix(compression): end-to-end audit — fixes across the whole compression flow (#4323)

* fix(compression): SLM worker resolves deps+worker file without import.meta.url (B-SLM) The Next.js standalone bundle (webpack) replaces createRequire(import.meta.url) with a stub that always throws MODULE_NOT_FOUND, and freezes import.meta.url to the build-machine path. So depsAvailable() was always false (the worker never spawned) and resolveWorkerFile() anchored on a path absent at runtime — the SLM silently fell back to the aggressive summarizer in production. Confirmed by inspecting dist/.build/next/server/chunks/26410.js (stub module 215743 + frozen file:// path). Replace both with filesystem probing from runtime anchors (process.cwd(), process.argv[1]) that survive the bundle. Necessary complement to #4286 (deps co-location) for the SLM to actually engage in prod; still fail-open without it. VPS live validation deferred (Rule #18); local resolver regression tests added. * fix(compression): ultra heuristic preserves code blocks / inline code / URLs (B-ULTRA-CODE) ultra.ts called pruneByScore on raw text with no tombstoning, so the token pruner dropped low-score code tokens (`b)`, `{`, `+`) inside fenced blocks while leaving the fence markers intact — output that looked like valid code but was syntactically destroyed. caveman + llmlingua both extract/restore preserved blocks first; ultra was the only pruning engine that didn't. Add pruneProseOnly(): extractPreservedBlocks tombstones fenced code, inline code, URLs, CONST_CASE, versions; only the prose between placeholders is pruned; preserved blocks are re-stitched verbatim. * fix(compression): GCF round-trips values containing the inline-array pattern [..]: (B-GCF-QUOTE) A value like `ERR[404]: Not Found` / `[Speaker 1]: Hello` nested one level deep was emitted bare and re-parsed by the decoder as an inline-array header → it threw `count_mismatch` (or silently decoded wrong), losing the whole block. headroomEngine .apply() ships such blobs in prod, so this was a reachable lossless violation. Two complementary fixes, both per SPEC §2.4: - encode: needsQuote() now quotes strings matching `[`…`]``:` (spec compliance / other decoders). - decode: the inline-array branch only fires when the bracket is in the KEY position (no `=` before it), so a quoted `note="ERR[404]: …"` value falls through to key=value. * fix(compression): aggressive fidelity — keep text blocks, compress Anthropic tool_result, don't corrupt JSON (B-AGG-*) Three fidelity fixes in the aggressive path (each TDD, aggressive-fidelity.test.ts): - B-AGG-TEXTDROP: replaceTextContent dropped 2nd+ text blocks unconditionally; now a trailing block is dropped only when its text is already subsumed by newText, else kept. - B-AGG-ANTHROPIC-TR: tool-result compression only fired for OpenAI role:tool messages; now Anthropic-shape tool_result content blocks (inside user messages) are compressed too, preserving tool_use_id + block structure. - B-AGG-JSONTAG: the [COMPRESSED:aging:*] prefix corrupted JSON/code payloads; pure JSON is now kept verbatim+untagged (stays parseable), fenced blocks get the tag on a preceding line. * fix(compression): accessibility collapse preserves [ref] anchors + fires on interleaved trees (B-MCPA11Y-*) - B-MCPA11Y-ANCHORS: collapseRepeated silently dropped the omitted middle siblings' [ref=eNN] anchors (the agent could no longer click them); now every omitted ref is kept alongside the collapse notice. Wires the previously-dead preserveRefPattern. Invariant: extractRefs(input) ⊆ extractRefs(output). - B-MCPA11Y-COLLAPSE: noise removal blanked lines (replace→""), and a blank line broke the sibling run so collapse never fired on realistic interleaved trees; noise lines are now deleted, and the sibling walk skips stray blanks. * fix(compression): rtk intensity scales the line budget (B-RTK-INTENSITY) The intensity knob only set smartTruncate's preserveHead/Tail (16↔24), which rarely fired because the matched filter capped lines first — so minimal/standard/aggressive produced byte-identical output on filter-matched tool output. effectiveMaxLines() now scales the effective line budget (minimal 1.5x, standard 1x, aggressive 0.5x) at both the per-filter and engine-level truncation sites. Both go through smartTruncate with priorityPatterns, so error/failure lines survive at every intensity (tested). * fix(compression): robust language detection + auto-detect honors the detected pack (B-LANG-*) - B-LANG-DETECTOR: detector was first-match-wins on a single keyword, and some hints are English-ambiguous ("configuration" in fr, "error" in es) → English text misclassified. Now score-based (count native-keyword hits, highest wins), and the two English-ambiguous words are removed from the hint lists, so a lone shared word never misclassifies while sparse-keyword languages (id) still detect on a single native word. - B-LANG-DORMANT: with autoDetectLanguage on but enabledPacks ["en"], detected non-English text fell back to the English pack, whose `articles` rule deletes foreign articles (pt-BR "a"/"o"). Auto-detect now uses the detected pack directly (it always has rules); enabledPacks still gates manual selection. * fix(compression): mode selection enables its engine + align stacked allowlist (B-MODE-ENGINE-DECOUPLE, B-PIPELINE-DIVERGENCE) - B-MODE-ENGINE-DECOUPLE: picking the standard/rtk MODE now runs caveman/rtk regardless of the per-engine enabled flag — the mode selection is the enable signal (the per-engine flag still gates stacked pipeline steps). Previously an operator who picked a mode but left the engine toggle off got silent 0% compression. - B-PIPELINE-DIVERGENCE: the global stackedPipeline normalizer stripped session-dedup/ccr/headroom/llmlingua (engines the combo path accepts via KNOWN_ENGINE_IDS). The allowlist now matches, so the global setting can use all registered engines. * docs(compression): correct SLM "stable" claim + document partial packs / stacked telemetry limits - The llmlingua `stable:true` comment claimed the bundle walk-up + deps-gate were "confirmed against the live install" — that was wrong (webpack froze import.meta.url and stubbed createRequire, so the worker never spawned in prod). Corrected to reflect B-SLM. - COMPRESSION_ENGINES.md: add a Known limitations section (SLM dep co-location requirement, partial de/fr/ja packs, no-op engines absent from engineBreakdown). * fix(compression): cast normalized engine id to CompressionPipelineStep['engine'] (typecheck)
2026-08-03 13:52:09 +03:00 · 2026-06-19 21:51:12 -03:00
parent 2c0fd04704
commit 5d89fa84e7
25 changed files with 1006 additions and 105 deletions
--- a/docs/compression/COMPRESSION_ENGINES.md
+++ b/docs/compression/COMPRESSION_ENGINES.md
@@ -284,6 +284,25 @@ Compression exposes five MCP tools:
 | `omniroute_list_compression_combos` | `read:compression`  | List compression combos          |
 | `omniroute_compression_combo_stats` | `read:compression`  | Read combo/engine analytics      |

+## Known limitations
+
+- **LLMLingua-2 (SLM) requires co-located optional deps.** The worker only runs in a
+  production build when `@atjsh/llmlingua-2` + peers are co-located into
+  `dist/node_modules` (see `scripts/build/colocateOptionals.mjs`, #4286). Without them the
+  engine fail-opens (returns the original text). Worker resolution no longer depends on
+  `import.meta.url` (it dies in the standalone bundle) — it anchors on the runtime
+  cwd / `argv[1]`.
+- **Caveman language packs `de` / `fr` / `ja` are partial.** They ship `context` +
+  `filler` + `structural` rules but no `dedup` / `ultra` packs, so `ultra` intensity is
+  no stronger than `full` for those languages (they use only their own rules — there is no
+  silent fall-back to the English `dedup`/`ultra` rules, which would mangle foreign text).
+  `en` / `es` / `id` / `pt-BR` are complete. Contributions of `dedup.json` + `ultra.json`
+  for the partial packs are welcome.
+- **Stacked telemetry only lists engines that compressed.** A stacked-pipeline step whose
+  engine ran but produced 0 % savings returns `stats:null` and so does not appear in
+  `engineBreakdown` — indistinguishable from a step that was skipped. Distinguishing
+  "ran, 0 %" from "skipped" would require a breakdown-model change and is deferred.
+
 ## Validation

 The focused gates for this area are:
--- a/open-sse/services/compression/aggressive.ts
+++ b/open-sse/services/compression/aggressive.ts
@@ -1,6 +1,10 @@
 import type { AggressiveConfig, CompressionStats, Summarizer } from "./types.ts";
 import { DEFAULT_AGGRESSIVE_CONFIG } from "./types.ts";
-import { compressToolResult } from "./toolResultCompressor.ts";
+import {
+  compressToolResult,
+  compressAnthropicToolResultBlock,
+  isAnthropicToolResultBlock,
+} from "./toolResultCompressor.ts";
 import { applyAging } from "./progressiveAging.ts";
 import { RuleBasedSummarizer } from "./summarizer.ts";
 import { cavemanCompress } from "./caveman.ts";
@@ -65,15 +69,35 @@ export function compressAggressive(
  try {
    const afterToolResult = currentMessages.map((msg) => {
      if (cfg.preserveSystemPrompt !== false && msg.role === "system") return msg;
-      if (msg.role !== "tool" && msg.role !== "function") return msg;
-      const text = extractTextContent(msg.content);
-      if (!text || COMPRESSED_MARKER_RE.test(text)) return msg;

-      const result = compressToolResult(text, cfg.toolStrategies);
-      if (result.strategy === "none" || result.saved <= 0) return msg;
+      // OpenAI-shape: a dedicated tool/function message whose content is the result text.
+      if (msg.role === "tool" || msg.role === "function") {
+        const text = extractTextContent(msg.content);
+        if (!text || COMPRESSED_MARKER_RE.test(text)) return msg;

-      toolResultSavings += result.saved;
-      return setContent(msg, result.compressed);
+        const result = compressToolResult(text, cfg.toolStrategies);
+        if (result.strategy === "none" || result.saved <= 0) return msg;
+
+        toolResultSavings += result.saved;
+        return setContent(msg, result.compressed);
+      }
+
+      // Anthropic-shape: `tool_result` content blocks live inside a (typically user)
+      // message's content array. Compress the text inside each block while preserving
+      // the tool_use_id and block structure exactly (B-AGG-ANTHROPIC-TR).
+      if (!Array.isArray(msg.content)) return msg;
+      if (!msg.content.some(isAnthropicToolResultBlock)) return msg;
+
+      let blockSavings = 0;
+      const nextContent = msg.content.map((part) => {
+        if (!isAnthropicToolResultBlock(part)) return part;
+        const { block, saved } = compressAnthropicToolResultBlock(part, cfg.toolStrategies);
+        blockSavings += saved;
+        return block;
+      });
+      if (blockSavings <= 0) return msg;
+      toolResultSavings += blockSavings;
+      return { ...msg, content: nextContent };
    });
    currentMessages = afterToolResult;
  } catch (err) {
--- a/open-sse/services/compression/caveman.ts
+++ b/open-sse/services/compression/caveman.ts
@@ -522,11 +522,17 @@ export function cavemanCompress(
        ? detectCompressionLanguage(textPart)
        : (config.language ?? "en");
      const enabledPacks = config.enabledLanguagePacks ?? ["en", detectedLanguage];
-      const language = enabledPacks.includes(detectedLanguage)
+      // When auto-detect is on, honor the detected language directly: the detector only
+      // returns languages that have a rule pack, and falling back to the English pack on
+      // non-English text mangles it (the EN `articles` rule deletes pt-BR "a"/"o").
+      // enabledPacks still gates MANUAL pack selection (auto-detect off). (B-LANG-DORMANT)
+      const language = config.autoDetectLanguage
        ? detectedLanguage
-        : enabledPacks.includes("en")
-          ? "en"
-          : detectedLanguage;
+        : enabledPacks.includes(detectedLanguage)
+          ? detectedLanguage
+          : enabledPacks.includes("en")
+            ? "en"
+            : detectedLanguage;
      const rules = getRulesForContext(msg.role, config.intensity, language).filter(
        (rule) => !config.skipRules.includes(rule.name)
      );
--- a/open-sse/services/compression/engines/headroom/gcf/decode_generic.ts
+++ b/open-sse/services/compression/engines/headroom/gcf/decode_generic.ts
@@ -139,10 +139,13 @@ function parseObjectBody(
      continue;
    }

-    // Inline array.
+    // Inline array. The bracket must be in the KEY position: an inline-array header is
+    // `name[...]: …`, never `key=value` whose value happens to contain `[..]:` (e.g. a
+    // quoted `note="ERR[404]: Not Found"`). Guard on no `=` before the bracket so such
+    // values fall through to the key=value branch instead of throwing (B-GCF-QUOTE).
    if (!content.startsWith("@") && !content.startsWith("##")) {
      const bracketIdx = content.indexOf("[");
-      if (bracketIdx > 0) {
+      if (bracketIdx > 0 && !content.slice(0, bracketIdx).includes("=")) {
        const rest = content.slice(bracketIdx);
        const closeIdx = rest.indexOf("]");
        if (closeIdx >= 0) {
--- a/open-sse/services/compression/engines/headroom/gcf/scalar.ts
+++ b/open-sse/services/compression/engines/headroom/gcf/scalar.ts
@@ -8,6 +8,10 @@

 const JSON_NUMBER_RE = /^-?(?:0|[1-9]\d*)(?:\.\d+)?(?:[eE][+-]?\d+)?$/;
 const NUMERIC_LIKE_RE = /^[+-]\.?\d|^\.\d|^0\d/;
+// SPEC §2.4: a bracket pair immediately followed by `:` (e.g. `ERR[404]: Not Found`,
+// `[Speaker 1]: Hello`). Bare, on a line-level key=value RHS, the decoder re-parses
+// this as an inline-array header → count_mismatch / wrong value (B-GCF-QUOTE).
+const INLINE_ARRAY_RE = /\[[^\]]*\]:/;

 /** Check if a string value must be quoted per Section 2.4. */
 export function needsQuote(s: string): boolean {
@@ -17,6 +21,7 @@ export function needsQuote(s: string): boolean {
  if (NUMERIC_LIKE_RE.test(s)) return true;
  if (s[0] === " " || s[s.length - 1] === " ") return true;
  if (s[0] === "#" || s[0] === "@" || s[0] === ".") return true;
+  if (INLINE_ARRAY_RE.test(s)) return true;
  for (let i = 0; i < s.length; i++) {
    const c = s.charCodeAt(i);
    if (
--- a/open-sse/services/compression/engines/llmlingua/index.ts
+++ b/open-sse/services/compression/engines/llmlingua/index.ts
@@ -354,9 +354,13 @@ export const llmlinguaEngine: CompressionEngine = {
    inputScope: "messages",
    targetLatencyMs: 200,
    supportsPreview: false,
-    // Promoted to stable after VPS validation (2026-06-16): the deployed worker
-    // compressed real prose (209→107 ch, ok=true), and the bundle's walk-up
-    // resolution + optional-deps gate were confirmed against the live install.
+    // Stable. The worker model itself was VPS-validated (real prose 209→107 ch, ok=true),
+    // but the EARLIER "walk-up + optional-deps gate confirmed in the bundle" claim was
+    // wrong: the Next standalone bundle (webpack) froze `import.meta.url` to the build path
+    // and stubbed `createRequire`, so in production the gate was always false and the worker
+    // never spawned (it silently fell back to the aggressive summarizer). Fixed in B-SLM —
+    // worker.ts now resolves deps + worker file from runtime anchors (cwd / argv[1]). The
+    // optional deps must also be co-located into dist/node_modules (#4286) to actually run.
    stable: true,
  },

--- a/open-sse/services/compression/engines/llmlingua/worker.ts
+++ b/open-sse/services/compression/engines/llmlingua/worker.ts
@@ -31,8 +31,6 @@
 */

 import { Worker } from "node:worker_threads";
-import { createRequire } from "node:module";
-import { fileURLToPath } from "node:url";
 import path from "node:path";
 import fs from "node:fs";

@@ -49,27 +47,83 @@ const FIRST_CALL_TIMEOUT_MS = 60000;
 /**
 * Gate probe: `@atjsh/llmlingua-2` is the entry package that declares the others
 * (`@huggingface/transformers`, `@tensorflow/tfjs`, `js-tiktoken`) as peers. We probe
- * ONLY it because the peers are ESM-only — e.g. `@huggingface/transformers@3.5.2`'s
- * `exports` has no `require`/`default` condition, so `require.resolve()` throws
- * `MODULE_NOT_FOUND` for it even when it is installed and `import()`-able (verified on
- * the VPS). Gating on all four would therefore always fail-open. The worker still
- * fail-opens if a peer is genuinely missing at `import()` time.
+ * ONLY it (by manifest existence) because the peers are ESM-only and `require.resolve`
+ * throws for them even when installed; the worker still fail-opens if a peer is
+ * genuinely missing at `import()` time.
+ *
+ * ⚠️ We do NOT use `createRequire(import.meta.url).resolve()` nor any other
+ * `import.meta.url`-based resolution: the Next.js standalone bundle (webpack) replaces
+ * `createRequire(import.meta.url)` with a stub module that ALWAYS throws
+ * `MODULE_NOT_FOUND` and freezes `import.meta.url` to the build-machine path, so such a
+ * gate is always false / mis-anchored in production (B-SLM). We probe the filesystem
+ * from runtime anchors that survive the bundle instead.
 */
-const GATE_DEP = "@atjsh/llmlingua-2";
+const GATE_DEP_REL = path.join("node_modules", "@atjsh", "llmlingua-2", "package.json");
+
+/** Relative path (from an install root) to the esbuild'd / source worker entry. */
+const WORKER_JS_REL = path.join(
+  "open-sse",
+  "services",
+  "compression",
+  "engines",
+  "llmlingua",
+  "onnxWorker.js"
+);
+const WORKER_TS_REL = path.join(
+  "open-sse",
+  "services",
+  "compression",
+  "engines",
+  "llmlingua",
+  "onnxWorker.ts"
+);
+
+const MAX_WALK_UP = 8;
+
+/**
+ * Walk up from each anchor directory (≤ MAX_WALK_UP levels) and return the first
+ * ancestor that actually contains `relPath`, or null. Pure + exported for tests.
+ *
+ * This deliberately avoids `import.meta.url`/`__dirname` (both dead in the standalone
+ * bundle) — see the GATE_DEP_REL comment.
+ */
+export function firstAncestorWith(anchors: string[], relPath: string): string | null {
+  for (const anchor of anchors) {
+    if (!anchor) continue;
+    let dir = path.resolve(anchor);
+    for (let i = 0; i <= MAX_WALK_UP; i++) {
+      if (fs.existsSync(path.join(dir, relPath))) return dir;
+      const parent = path.dirname(dir);
+      if (parent === dir) break;
+      dir = parent;
+    }
+  }
+  return null;
+}
+
+/**
+ * Runtime install-root anchors that SURVIVE the standalone bundle:
+ *  - `process.cwd()` — `dist/server.js` runs `process.chdir(__dirname)` → the dist root.
+ *  - `dirname(process.argv[1])` — the entry script (server.js / bin), walked up.
+ */
+function runtimeAnchors(): string[] {
+  const anchors = [process.cwd()];
+  const argv1 = process.argv[1];
+  if (typeof argv1 === "string" && argv1) anchors.push(path.dirname(argv1));
+  return anchors;
+}

 // ─── optional-deps gate (memoized) ──────────────────────────────────────────────

 let _depsAvailable: boolean | null = null;

-/** Lazily (and once) check whether the optional LLMLingua dependency stack is installed. */
-function depsAvailable(): boolean {
+/**
+ * Lazily (and once) check whether the optional LLMLingua dependency stack is installed,
+ * by probing `node_modules/@atjsh/llmlingua-2/package.json` from the runtime anchors.
+ */
+export function depsAvailable(): boolean {
  if (_depsAvailable !== null) return _depsAvailable;
-  try {
-    createRequire(import.meta.url).resolve(GATE_DEP);
-    _depsAvailable = true;
-  } catch {
-    _depsAvailable = false;
-  }
+  _depsAvailable = firstAncestorWith(runtimeAnchors(), GATE_DEP_REL) !== null;
  return _depsAvailable;
 }

@@ -106,55 +160,33 @@ const warmedModels = new Set<string>();
 let idleTimer: NodeJS.Timeout | null = null;

 /**
- * Resolve the worker entry file across dev and prod.
+ * Resolve the worker entry file across dev and prod WITHOUT `import.meta.url`.
 *
- * Dev: `onnxWorker.ts` sits next to this file and runs via the tsx loader.
+ * Prod: the worker is esbuild'd to `<distRoot>/open-sse/.../onnxWorker.js`
+ * (scripts/build/prepublish.ts) + kept by the pack-artifact allowlist. The install
+ * root is found by walking up the runtime anchors (cwd / argv[1] dir), since the
+ * bundled module location (`import.meta.url`) is frozen to the build machine.
 *
- * Prod: this module is collapsed into a `.next` chunk and the worker is esbuild'd to
- * `<appRoot>/open-sse/services/compression/engines/llmlingua/onnxWorker.js`
- * (scripts/build/prepublish.ts) + kept by the pack-artifact allowlist. The process
- * `cwd` is NOT the app root (pm2 starts it from `/root`), so cwd-relative resolution
- * is unreliable — we instead WALK UP from this module's location (`import.meta.url`,
- * which in the standalone bundle is a real `<appRoot>/.next/...` path) until we find
- * an ancestor that actually contains the worker at its known relative path. cwd
- * candidates remain as a last-resort fallback. First existing candidate wins; a `.ts`
- * choice gets the tsx loader, a `.js` choice runs natively.
+ * Dev (tsx): the same relative path resolves to the `.ts` source under the project
+ * root (cwd) and runs via the tsx loader.
+ *
+ * First existing candidate wins; a `.js` choice runs natively, a `.ts` choice gets the
+ * tsx loader. Exported for tests.
 */
-function resolveWorkerFile(): { workerFile: string; execArgv: string[] } {
-  const moduleDir = path.dirname(fileURLToPath(import.meta.url));
-  const rel = path.join(
-    "open-sse",
-    "services",
-    "compression",
-    "engines",
-    "llmlingua",
-    "onnxWorker.js"
-  );
+export function resolveWorkerFile(): { workerFile: string; execArgv: string[] } {
+  const anchors = runtimeAnchors();

-  // 1. Dev: sibling source/compiled file next to this module.
-  const devTs = path.join(moduleDir, "onnxWorker.ts");
-  if (fs.existsSync(devTs)) return { workerFile: devTs, execArgv: ["--import", "tsx/esm"] };
-  const devJs = path.join(moduleDir, "onnxWorker.js");
-  if (fs.existsSync(devJs)) return { workerFile: devJs, execArgv: [] };
+  // Prod first: the esbuild'd .js under the install root.
+  const jsRoot = firstAncestorWith(anchors, WORKER_JS_REL);
+  if (jsRoot) return { workerFile: path.join(jsRoot, WORKER_JS_REL), execArgv: [] };

-  // 2. Prod: walk up from the bundled module location, then cwd, looking for the
-  //    esbuild'd worker at <root>/open-sse/.../onnxWorker.js.
-  const roots: string[] = [];
-  let dir = moduleDir;
-  for (let i = 0; i < 12; i++) {
-    roots.push(dir);
-    const parent = path.dirname(dir);
-    if (parent === dir) break;
-    dir = parent;
-  }
-  roots.push(process.cwd(), path.join(process.cwd(), "app"));
-  for (const root of roots) {
-    const candidate = path.join(root, rel);
-    if (fs.existsSync(candidate)) return { workerFile: candidate, execArgv: [] };
-  }
+  // Dev: the .ts source (tsx loader).
+  const tsRoot = firstAncestorWith(anchors, WORKER_TS_REL);
+  if (tsRoot)
+    return { workerFile: path.join(tsRoot, WORKER_TS_REL), execArgv: ["--import", "tsx/esm"] };

-  // 3. Nothing found — return the sibling .js path; the spawn will fail-open.
-  return { workerFile: devJs, execArgv: [] };
+  // Nothing found — return a cwd-relative .js path; the spawn will fail-open.
+  return { workerFile: path.join(process.cwd(), WORKER_JS_REL), execArgv: [] };
 }

 /** Reset the idle eviction timer; terminates the worker after the idle window. */
--- a/open-sse/services/compression/engines/mcpAccessibility/collapseRepeated.ts
+++ b/open-sse/services/compression/engines/mcpAccessibility/collapseRepeated.ts
@@ -1,5 +1,28 @@
+import { MCP_ACCESSIBILITY_DEFAULTS } from "./constants.ts";
+
 const SIBLING_PATTERN = /^(\s*)-\s*([a-zA-Z]+)\b/;

+/**
+ * Extract every `[ref=eNN]` anchor from a blob, preserving order, de-duplicated. These refs are how
+ * an agent clicks elements, so they MUST survive collapse. Uses the shared `preserveRefPattern`.
+ */
+export function extractRefs(text: string): string[] {
+  const seen = new Set<string>();
+  const refs: string[] = [];
+  // Fresh regex per call: the shared pattern carries the global flag (stateful lastIndex).
+  const pattern = new RegExp(
+    MCP_ACCESSIBILITY_DEFAULTS.preserveRefPattern.source,
+    MCP_ACCESSIBILITY_DEFAULTS.preserveRefPattern.flags
+  );
+  for (const m of text.matchAll(pattern)) {
+    if (!seen.has(m[0])) {
+      seen.add(m[0]);
+      refs.push(m[0]);
+    }
+  }
+  return refs;
+}
+
 export function findNthSiblingEnd(
  lines: string[],
  start: number,
@@ -64,6 +87,11 @@ export function collapseRepeated(
        j++;
        continue;
      }
+      // A blank line (noise removal can leave these) must not break a sibling run.
+      if (ln.trim() === "") {
+        j++;
+        continue;
+      }
      break;
    }
    const groupLen = j - i;
@@ -74,6 +102,15 @@ export function collapseRepeated(
      out.push(
        `${indent}... [${groupLen - keepHead - keepTail} similar "${role}" items omitted by OmniRoute MCP filter]`
      );
+      // BUG A invariant: the omitted middle siblings carry [ref=eNN] anchors the agent needs to
+      // click. Extract every ref from the dropped lines and keep them alongside the notice so
+      // extractRefs(input) ⊆ extractRefs(output) always holds.
+      const omittedRefs = extractRefs(lines.slice(headEnd, tailStart).join("\n"));
+      if (omittedRefs.length > 0) {
+        out.push(
+          `${indent}  [refs of omitted "${role}" items (clickable): ${omittedRefs.join(" ")}]`
+        );
+      }
      for (let k = tailStart; k < j; k++) out.push(lines[k]);
    } else {
      for (let k = i; k < j; k++) out.push(lines[k]);
--- a/open-sse/services/compression/engines/mcpAccessibility/index.ts
+++ b/open-sse/services/compression/engines/mcpAccessibility/index.ts
@@ -1,16 +1,25 @@
 import { collapseRepeated } from "./collapseRepeated.ts";
 import { MCP_ACCESSIBILITY_TAIL_RESERVE, type McpAccessibilityConfig } from "./constants.ts";

-const NOISE_PATTERNS: RegExp[] = [/^\s*-\s*generic:?\s*$/gm, /^\s*-\s*text:\s*""\s*$/gm];
+// Per-line (non-global, anchored) noise matchers. Used to DELETE whole noise lines rather than
+// blank them: `replace(pattern, "")` would leave empty strings behind, and a blank line between two
+// sibling headers breaks collapseRepeated's sibling run (an empty line is neither a sibling header
+// nor an indented child), so collapse would never fire on realistic interleaved trees.
+const NOISE_LINE_PATTERNS: RegExp[] = [/^\s*-\s*generic:?\s*$/, /^\s*-\s*text:\s*""\s*$/];
+
+function isNoiseLine(line: string): boolean {
+  return NOISE_LINE_PATTERNS.some((p) => p.test(line));
+}

 export function smartFilterText(text: string, config: McpAccessibilityConfig): string {
  if (typeof text !== "string" || text.length < config.minLengthToProcess) {
    return text;
  }
-  let out = text;
-  for (const pattern of NOISE_PATTERNS) {
-    out = out.replace(pattern, "");
-  }
+  // Drop noise lines entirely (not blank them) so interleaved noise does not split sibling runs.
+  let out = text
+    .split("\n")
+    .filter((line) => !isNoiseLine(line))
+    .join("\n");
  out = collapseRepeated(
    out,
    config.collapseThreshold,
--- a/open-sse/services/compression/engines/rtk/index.ts
+++ b/open-sse/services/compression/engines/rtk/index.ts
@@ -289,7 +289,7 @@ export function processRtkText(
      if (config.enabledFilters.length === 0 || config.enabledFilters.includes(filter.id)) {
        const filtered = applyLineFilter(result, {
          ...filter,
-          maxLines: filter.maxLines || config.maxLinesPerResult,
+          maxLines: effectiveMaxLines(filter.maxLines || config.maxLinesPerResult, config.intensity),
        });
        result = filtered.text;
        if (filtered.appliedRules.length > 0) {
@@ -352,7 +352,7 @@ export function processRtkText(
    }
  });
  const truncated = smartTruncate(result, {
-    maxLines: config.maxLinesPerResult,
+    maxLines: effectiveMaxLines(config.maxLinesPerResult, config.intensity),
    maxChars: config.maxCharsPerResult,
    preserveHead: config.intensity === "aggressive" ? 16 : 24,
    preserveTail: config.intensity === "aggressive" ? 16 : 24,
@@ -458,6 +458,18 @@ function processRtkContent(
  };
 }

+/**
+ * Scale a line budget by intensity so minimal / standard / aggressive produce
+ * meaningfully different output on truncation-based filters (B-RTK-INTENSITY).
+ * Truncation always runs through smartTruncate with priorityPatterns, so error /
+ * failure lines survive at EVERY intensity. (Include/collapse filters like docker-logs
+ * compress by content, not line budget, so they are intensity-independent by nature.)
+ */
+export function effectiveMaxLines(base: number, intensity: string | undefined): number {
+  const factor = intensity === "aggressive" ? 0.5 : intensity === "minimal" ? 1.5 : 1;
+  return Math.max(1, Math.round(base * factor));
+}
+
 export function applyRtkCompression(
  body: Record<string, unknown>,
  options: { config?: Partial<RtkConfig>; stepConfig?: Record<string, unknown> } = {}
--- a/open-sse/services/compression/languageDetector.ts
+++ b/open-sse/services/compression/languageDetector.ts
@@ -1,17 +1,39 @@
 const LANGUAGE_HINTS: Record<string, RegExp[]> = {
  "pt-BR": [/\b(?:voce|você|preciso|arquivo|codigo|código|erro|falha|obrigado)\b/i],
-  es: [/\b(?:necesito|archivo|codigo|código|error|fallo|gracias|puedes)\b/i],
+  // NOTE: English-ambiguous words are intentionally excluded — "error" (es) and
+  // "configuration" (fr) are identical in English and would misclassify English text.
+  // Spanish/French keep their distinctive native spellings (fallo / erreur, etc).
+  es: [/\b(?:necesito|archivo|codigo|código|fallo|gracias|puedes)\b/i],
  de: [/\b(?:ich|datei|fehler|bitte|kannst|konfiguration|danke)\b/i],
-  fr: [/\b(?:fichier|erreur|merci|peux|configuration|besoin)\b/i],
+  fr: [/\b(?:fichier|erreur|merci|peux|besoin)\b/i],
  ja: [/[\u3040-\u30ff]/],
  id: [/\b(?:saya|kamu|anda|dengan|untuk|yang|tidak|bisa|terima\s+kasih|dari)\b/i],
 };

+/**
+ * Score each language by the NUMBER of native-keyword hits and pick the highest
+ * (English-ambiguous words are excluded from the hint lists, so a lone shared word
+ * never misclassifies English). Highest score wins; ties keep the earlier language;
+ * zero hits → English. (B-LANG-DETECTOR)
+ */
 export function detectCompressionLanguage(text: string): string {
+  let best = "en";
+  let bestScore = 0;
  for (const [language, patterns] of Object.entries(LANGUAGE_HINTS)) {
-    if (patterns.some((pattern) => pattern.test(text))) return language;
+    let score = 0;
+    for (const pattern of patterns) {
+      const global = pattern.flags.includes("g")
+        ? pattern
+        : new RegExp(pattern.source, pattern.flags + "g");
+      const matches = text.match(global);
+      if (matches) score += matches.length;
+    }
+    if (score > bestScore) {
+      bestScore = score;
+      best = language;
+    }
  }
-  return "en";
+  return best;
 }

 export function listSupportedCompressionLanguages(): string[] {
--- a/open-sse/services/compression/messageContent.ts
+++ b/open-sse/services/compression/messageContent.ts
@@ -63,12 +63,22 @@ export function replaceTextContent(msg: ChatMessageLike, newText: string): ChatM
    return { ...msg, content: newText };
  }

+  // The first text block receives `newText`. Trailing text blocks are normally
+  // already subsumed by `newText` (callers build it from the JOIN of all text
+  // blocks via extractTextContent), so we drop them to avoid duplicating content.
+  // But if a caller's `newText` does NOT contain a trailing block's text, dropping
+  // it would silently lose content the model can no longer see (B-AGG-TEXTDROP) —
+  // so in that case we keep the trailing block instead of returning [].
  let replaced = false;
  const content = msg.content.flatMap((part) => {
    if (!isTextBlock(part)) return [part];
-    if (replaced) return [];
-    replaced = true;
-    return [{ ...part, text: newText }];
+    if (!replaced) {
+      replaced = true;
+      return [{ ...part, text: newText }];
+    }
+    const partText = part.text ?? "";
+    if (partText && !newText.includes(partText)) return [part];
+    return [];
  });

  if (!replaced) {
--- a/open-sse/services/compression/progressiveAging.ts
+++ b/open-sse/services/compression/progressiveAging.ts
@@ -5,11 +5,54 @@ import { cavemanCompress } from "./caveman.ts";
 import { extractTextContent, replaceTextContent, type ChatMessageLike } from "./messageContent.ts";

 const COMPRESSED_MARKER_RE = /^\[COMPRESSED:/;
+const JSON_PREFIX_RE = /^\s*[{[]/;
+const FENCE_RE = /^\s*```/;

 function estimateTokens(text: string): number {
  return Math.ceil(text.length / 4);
 }

+/**
+ * Structured content that an inline `[COMPRESSED:...]` prefix would corrupt:
+ * a pure-JSON payload (parses as JSON) or a fenced code block (B-AGG-JSONTAG).
+ */
+type StructuredKind = "json" | "fenced" | null;
+
+function structuredKind(text: string): StructuredKind {
+  const trimmed = text.trim();
+  if (FENCE_RE.test(trimmed) && trimmed.endsWith("```")) return "fenced";
+  if (JSON_PREFIX_RE.test(trimmed)) {
+    try {
+      JSON.parse(trimmed);
+      return "json";
+    } catch {
+      return null;
+    }
+  }
+  return null;
+}
+
+/**
+ * Build the aged content for a message, keeping structured payloads intact:
+ * - pure JSON: leave verbatim and untagged (stays JSON.parse-able). Aging engines
+ *   (lite/caveman) do not shrink JSON anyway, so nothing is lost; tracked in stats
+ *   via the unchanged content. The recursion guard relies on structuredKind() on the
+ *   next pass, so re-running aging is a no-op (idempotent).
+ * - fenced code block: place the tag on its own line BEFORE the fence so the block
+ *   stays valid and the content still starts with `[COMPRESSED:` (recursion guard).
+ * - everything else: inline-prepend the tag as before.
+ */
+function tagAged(tier: string, originalText: string, compressed: string): string {
+  const kind = structuredKind(originalText);
+  if (kind === "json") {
+    return originalText;
+  }
+  if (kind === "fenced") {
+    return `[COMPRESSED:aging:${tier}]\n${originalText}`;
+  }
+  return `[COMPRESSED:aging:${tier}] ${compressed}`;
+}
+
 type ChatMessage = ChatMessageLike;

 type CompressedResult = {
@@ -62,7 +105,7 @@ export function applyAging(
          typeof compressed.body.messages[0].content === "string"
            ? compressed.body.messages[0].content
            : extractTextContent(compressed.body.messages[0].content);
-        const tagged = `[COMPRESSED:aging:light] ${newContent}`;
+        const tagged = tagAged("light", text, newContent);
        saved += estimateTokens(text) - estimateTokens(tagged);
        result.push(setContent(msg, tagged));
      } else {
@@ -75,7 +118,7 @@ export function applyAging(
          typeof compressed.body.messages[0].content === "string"
            ? compressed.body.messages[0].content
            : extractTextContent(compressed.body.messages[0].content);
-        const tagged = `[COMPRESSED:aging:moderate] ${newContent}`;
+        const tagged = tagAged("moderate", text, newContent);
        saved += estimateTokens(text) - estimateTokens(tagged);
        result.push(setContent(msg, tagged));
      } else {
@@ -84,12 +127,12 @@ export function applyAging(
    } else {
      if (msg.role === "assistant") {
        const summary = sum.summarize([msg]);
-        const tagged = `[COMPRESSED:aging:fullSummary] ${summary}`;
+        const tagged = tagAged("fullSummary", text, summary);
        saved += estimateTokens(text) - estimateTokens(tagged);
        result.push(setContent(msg, tagged));
      } else if (msg.role === "user") {
        const firstLine = text.split("\n")[0]?.slice(0, 120) ?? "";
-        const tagged = `[COMPRESSED:aging:fullSummary] ${firstLine}`;
+        const tagged = tagAged("fullSummary", text, firstLine);
        saved += estimateTokens(text) - estimateTokens(tagged);
        result.push(setContent(msg, tagged));
      } else {
--- a/open-sse/services/compression/strategySelector.ts
+++ b/open-sse/services/compression/strategySelector.ts
@@ -110,7 +110,9 @@ export function applyCompression(
  }
  if (mode === "rtk") {
    return applyRtkCompression(body, {
-      config: options?.config?.rtkConfig,
+      // Selecting the "rtk" mode IS the enable signal — run it even if the per-engine
+      // rtkConfig.enabled flag is off (that flag gates stacked steps). (B-MODE-ENGINE-DECOUPLE)
+      config: { ...(options?.config?.rtkConfig ?? {}), enabled: true },
    });
  }
  const adapter = adaptBodyForCompression(body);
@@ -147,6 +149,9 @@ export function applyCompression(
            ),
          }
        : {}),
+      // Selecting the "standard" mode runs caveman regardless of the per-engine
+      // cavemanConfig.enabled flag (that flag gates stacked steps). (B-MODE-ENGINE-DECOUPLE)
+      enabled: true,
    };
    const result = cavemanCompress(
      compressionBody as Parameters<typeof cavemanCompress>[0],
--- a/open-sse/services/compression/toolResultCompressor.ts
+++ b/open-sse/services/compression/toolResultCompressor.ts
@@ -9,6 +9,7 @@ export interface CompressionResult {
 const ANSI_RE = /\x1b\[[0-9;]*[a-zA-Z]/g;
 const SHELL_PROMPT_RE = /\$\s/;
 const JSON_PREFIX_RE = /^\s*[{[]/;
+const COMPRESSED_MARKER_RE = /^\[COMPRESSED:/;

 function isCodeLikeLine(rawLine: string): boolean {
  const line = rawLine.trimStart();
@@ -160,6 +161,68 @@ function estimateTokens(text: string): number {
  return Math.ceil(text.length / 4);
 }

+/** Minimal shape of an Anthropic `tool_result` content block. */
+export interface AnthropicToolResultBlock {
+  type: "tool_result";
+  tool_use_id?: string;
+  content?: string | Array<{ type?: string; text?: string; [key: string]: unknown }>;
+  [key: string]: unknown;
+}
+
+export function isAnthropicToolResultBlock(value: unknown): value is AnthropicToolResultBlock {
+  return (
+    !!value &&
+    typeof value === "object" &&
+    (value as { type?: unknown }).type === "tool_result"
+  );
+}
+
+/**
+ * Compress the text inside an Anthropic-shape `tool_result` content block,
+ * reusing the same per-type strategies as OpenAI-shape tool messages. The
+ * `tool_use_id` and block type are preserved exactly; only the inner text is
+ * compressed. Returns the (possibly unchanged) block plus tokens saved.
+ */
+export function compressAnthropicToolResultBlock(
+  block: AnthropicToolResultBlock,
+  opts: ToolStrategiesConfig
+): { block: AnthropicToolResultBlock; saved: number } {
+  const content = block.content;
+
+  if (typeof content === "string") {
+    if (!content || COMPRESSED_MARKER_RE.test(content)) return { block, saved: 0 };
+    const result = compressToolResult(content, opts);
+    if (result.strategy === "none" || result.saved <= 0) return { block, saved: 0 };
+    return { block: { ...block, content: result.compressed }, saved: result.saved };
+  }
+
+  if (Array.isArray(content)) {
+    let saved = 0;
+    let changed = false;
+    const nextContent = content.map((part) => {
+      if (
+        !part ||
+        typeof part !== "object" ||
+        part.type !== "text" ||
+        typeof part.text !== "string"
+      ) {
+        return part;
+      }
+      const text = part.text;
+      if (!text || COMPRESSED_MARKER_RE.test(text)) return part;
+      const result = compressToolResult(text, opts);
+      if (result.strategy === "none" || result.saved <= 0) return part;
+      saved += result.saved;
+      changed = true;
+      return { ...part, text: result.compressed };
+    });
+    if (!changed) return { block, saved: 0 };
+    return { block: { ...block, content: nextContent }, saved };
+  }
+
+  return { block, saved: 0 };
+}
+
 export function compressToolResult(content: string, opts: ToolStrategiesConfig): CompressionResult {
  if (opts.fileContent) {
    const result = compressFileContent(content);
--- a/open-sse/services/compression/ultra.ts
+++ b/open-sse/services/compression/ultra.ts
@@ -1,10 +1,39 @@
 import { pruneByScore } from "./ultraHeuristic.ts";
+import { extractPreservedBlocks } from "./preservation.ts";
 import { DEFAULT_ULTRA_CONFIG } from "./types.ts";
 import type { UltraConfig, CompressionStats, CompressionMode } from "./types.ts";
 import { extractTextContent, mapTextContent, type ChatMessageLike } from "./messageContent.ts";

 const COMPRESSED_PREFIX = "[COMPRESSED:";

+/**
+ * Prune PROSE only. Fenced code, inline code, URLs, CONST_CASE, versions, etc. are
+ * tombstoned by `extractPreservedBlocks` and re-stitched verbatim, so the heuristic
+ * NEVER mangles structured content (mirrors caveman.ts / llmlingua/index.ts).
+ *
+ * Without this, `pruneByScore` tokenizes the whole text and drops low-score tokens
+ * (`b)`, `{`, `+`, …) inside code blocks, corrupting them while leaving the fence
+ * markers intact — output that looks like valid code but isn't (B-ULTRA-CODE).
+ */
+function pruneProseOnly(text: string, rate: number, minScore: number): string {
+  const { text: withPlaceholders, blocks } = extractPreservedBlocks(text);
+  if (blocks.length === 0) return pruneByScore(text, rate, minScore);
+
+  const placeholderToContent = new Map(blocks.map((b) => [b.placeholder, b.content]));
+  const escaped = blocks.map((b) => b.placeholder.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"));
+  const splitRe = new RegExp(`(${escaped.join("|")})`, "g");
+
+  return withPlaceholders
+    .split(splitRe)
+    .map((part) => {
+      if (!part) return "";
+      const preserved = placeholderToContent.get(part);
+      if (preserved !== undefined) return preserved; // verbatim — never pruned
+      return pruneByScore(part, rate, minScore); // prose only
+    })
+    .join("");
+}
+
 export interface UltraCompressResult {
  messages: Array<{ role: string; content?: string | unknown[]; [key: string]: unknown }>;
  stats: CompressionStats;
@@ -40,7 +69,7 @@ export function ultraCompress(
    const next = mapTextContent(msg, (textPart) => {
      if (!textPart || textPart.startsWith(COMPRESSED_PREFIX)) return textPart;
      messageOriginalChars += textPart.length;
-      const pruned = pruneByScore(textPart, compressionRate, minScoreThreshold);
+      const pruned = pruneProseOnly(textPart, compressionRate, minScoreThreshold);
      messageCompressedChars += pruned.length;
      return pruned;
    }) as Message;
--- a/src/lib/db/compression.ts
+++ b/src/lib/db/compression.ts
@@ -228,23 +228,32 @@ function normalizeContextEditingConfig(value: unknown): ContextEditingConfig {
  };
 }

-function normalizeStackedPipeline(value: unknown): CompressionPipelineStep[] {
+// Engines allowed in the global stackedPipeline setting. MUST stay in sync with the
+// compression-combo KNOWN_ENGINE_IDS (src/lib/db/compressionCombos.ts) — otherwise the
+// global setting silently strips engines the combo path accepts (B-PIPELINE-DIVERGENCE).
+const STACKED_PIPELINE_ENGINE_IDS = new Set([
+  "lite",
+  "caveman",
+  "aggressive",
+  "ultra",
+  "rtk",
+  "headroom",
+  "session-dedup",
+  "ccr",
+  "llmlingua",
+]);
+
+export function normalizeStackedPipeline(value: unknown): CompressionPipelineStep[] {
  const source = Array.isArray(value) ? value : (DEFAULT_COMPRESSION_CONFIG.stackedPipeline ?? []);
  const pipeline: CompressionPipelineStep[] = [];
  for (const entry of source) {
    const record = toRecord(entry);
    const engine = record.engine;
-    if (
-      engine !== "lite" &&
-      engine !== "caveman" &&
-      engine !== "aggressive" &&
-      engine !== "ultra" &&
-      engine !== "rtk"
-    ) {
+    if (typeof engine !== "string" || !STACKED_PIPELINE_ENGINE_IDS.has(engine)) {
      continue;
    }
    pipeline.push({
-      engine,
+      engine: engine as CompressionPipelineStep["engine"],
      ...(typeof record.intensity === "string"
        ? { intensity: record.intensity as CompressionPipelineStep["intensity"] }
        : {}),
--- a/tests/unit/compression/aggressive-fidelity.test.ts
+++ b/tests/unit/compression/aggressive-fidelity.test.ts
@@ -0,0 +1,209 @@
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+import {
+  extractTextContent,
+  replaceTextContent,
+  type ChatMessageLike,
+} from "../../../open-sse/services/compression/messageContent.ts";
+import { applyAging } from "../../../open-sse/services/compression/progressiveAging.ts";
+import { compressAggressive } from "../../../open-sse/services/compression/aggressive.ts";
+import type { AgingThresholds } from "../../../open-sse/services/compression/types.ts";
+
+// ─── ISSUE 1 — B-AGG-TEXTDROP ────────────────────────────────────────────────
+// `replaceTextContent` previously dropped every text block after the first via
+// flatMap → []. With the standard call pattern (newText = compressed JOIN of all
+// text blocks), the joined original content must remain recoverable: nothing the
+// model can no longer see may be silently lost.
+describe("replaceTextContent — multi-text-block fidelity (B-AGG-TEXTDROP)", () => {
+  it("does not silently drop trailing text-block content absent from newText", () => {
+    const msg: ChatMessageLike = {
+      role: "user",
+      content: [
+        { type: "text", text: "FIRST block alpha" },
+        { type: "image", source: { foo: 1 } },
+        { type: "text", text: "SECOND block bravo" },
+        { type: "text", text: "THIRD block charlie" },
+      ],
+    };
+
+    // newText does NOT subsume the trailing blocks (worst case): a caller that
+    // only summarized the first block. The trailing blocks' content must not
+    // silently vanish.
+    const replaced = replaceTextContent(msg, "NEWTEXT-only-first");
+
+    const out = extractTextContent(replaced.content);
+    assert.ok(out.includes("NEWTEXT-only-first"), "replacement text missing");
+    assert.ok(out.includes("SECOND block bravo"), "second block silently dropped");
+    assert.ok(out.includes("THIRD block charlie"), "third block silently dropped");
+
+    // Non-text blocks (image) must survive unchanged.
+    const blocks = replaced.content as Array<{ type?: string }>;
+    assert.ok(
+      blocks.some((b) => b.type === "image"),
+      "non-text block must be preserved"
+    );
+  });
+
+  it("collapses trailing blocks when newText already subsumes them (no duplication)", () => {
+    const msg: ChatMessageLike = {
+      role: "user",
+      content: [
+        { type: "text", text: "alpha" },
+        { type: "text", text: "bravo" },
+      ],
+    };
+    // Standard call pattern: newText = compressed JOIN of all text blocks.
+    const joined = extractTextContent(msg.content); // "alpha\nbravo"
+    const replaced = replaceTextContent(msg, joined);
+    const blocks = replaced.content as Array<{ type?: string; text?: string }>;
+    const textBlocks = blocks.filter((b) => b.type === "text" || b.text !== undefined);
+    // Should collapse to a single text block — no duplicated "alpha"/"bravo".
+    assert.equal(textBlocks.length, 1, "subsumed trailing blocks should be collapsed");
+    assert.equal(extractTextContent(replaced.content), joined);
+  });
+
+  it("aging a multi-text-block message keeps all original text represented", () => {
+    // Build a long conversation so the first (multi-block) message ages out.
+    const first: ChatMessageLike = {
+      role: "user",
+      content: [
+        { type: "text", text: "alpha-marker request: fix login" },
+        { type: "text", text: "bravo-marker error: TypeError: boom" },
+      ],
+    };
+    const msgs: ChatMessageLike[] = [first];
+    for (let i = 1; i < 8; i++) {
+      msgs.push({ role: i % 2 ? "assistant" : "user", content: `filler ${i} ${"z".repeat(60)}` });
+    }
+    const result = applyAging(msgs, { fullSummary: 10, moderate: 10, light: 3, verbatim: 1 });
+    const out = extractTextContent(result.messages[0].content as ChatMessageLike["content"]);
+    // Light tier keeps content; both blocks' text must still be present (joined).
+    assert.ok(out.includes("alpha-marker"), "first text block lost during aging");
+    assert.ok(out.includes("bravo-marker"), "second text block silently dropped during aging");
+  });
+});
+
+// ─── ISSUE 2 — B-AGG-ANTHROPIC-TR ────────────────────────────────────────────
+// Anthropic-shape tool_result blocks (a {type:"tool_result"} content block inside
+// a user message) must be compressed too, preserving tool_use_id + block type.
+describe("aggressive — Anthropic tool_result compression (B-AGG-ANTHROPIC-TR)", () => {
+  it("compresses the text inside an Anthropic tool_result block", () => {
+    const bigJsonArray = JSON.stringify(
+      Array.from({ length: 200 }, (_, i) => ({ id: i, name: `item${i}`, data: "x".repeat(40) }))
+    );
+    const userMsg: ChatMessageLike = {
+      role: "user",
+      content: [
+        {
+          type: "tool_result",
+          tool_use_id: "toolu_ABC123",
+          content: [{ type: "text", text: bigJsonArray }],
+        },
+      ],
+    };
+
+    const result = compressAggressive([userMsg]);
+    const outMsg = result.messages[0];
+    const blocks = outMsg.content as Array<Record<string, unknown>>;
+    const tr = blocks.find((b) => b.type === "tool_result");
+
+    assert.ok(tr, "tool_result block must survive");
+    assert.equal(tr!.type, "tool_result", "block type must be unchanged");
+    assert.equal(tr!.tool_use_id, "toolu_ABC123", "tool_use_id must be preserved");
+
+    // The inner text must be smaller than the original (it was compressed).
+    const innerText =
+      typeof tr!.content === "string"
+        ? (tr!.content as string)
+        : ((tr!.content as Array<{ type?: string; text?: string }>) ?? [])
+            .filter((c) => c.type === "text")
+            .map((c) => c.text ?? "")
+            .join("\n");
+    assert.ok(
+      innerText.length < bigJsonArray.length,
+      `tool_result inner text was not compressed (orig ${bigJsonArray.length}, got ${innerText.length})`
+    );
+    assert.ok(result.stats.aggressive!.toolResultSavings > 0, "toolResultSavings must be > 0");
+  });
+
+  it("compresses a string-form Anthropic tool_result block", () => {
+    const errorOutput =
+      "TypeError: Cannot read property 'x' of undefined\n" +
+      Array.from({ length: 40 }, (_, i) => `    at fn${i} (file${i}.ts:${i + 1}:${i + 5})`).join(
+        "\n"
+      );
+    const userMsg: ChatMessageLike = {
+      role: "user",
+      content: [{ type: "tool_result", tool_use_id: "toolu_ERR", content: errorOutput }],
+    };
+    const result = compressAggressive([userMsg]);
+    const tr = (result.messages[0].content as Array<Record<string, unknown>>).find(
+      (b) => b.type === "tool_result"
+    );
+    assert.ok(tr, "tool_result block must survive");
+    assert.equal(tr!.tool_use_id, "toolu_ERR");
+    assert.ok((tr!.content as string).length < errorOutput.length, "string tool_result not compressed");
+  });
+});
+
+// ─── ISSUE 3 — B-AGG-JSONTAG ─────────────────────────────────────────────────
+// Aging must not corrupt structured content (JSON / fenced code block) with a
+// literal [COMPRESSED:...] inline prefix.
+describe("progressiveAging — structured-content tag safety (B-AGG-JSONTAG)", () => {
+  const thresholds: AgingThresholds = { fullSummary: 10, moderate: 10, light: 3, verbatim: 1 };
+
+  function agedFirstContent(content: string, t: AgingThresholds): string {
+    const msgs: ChatMessageLike[] = [
+      { role: "user", content },
+      { role: "assistant", content: "a " + "z".repeat(60) },
+      { role: "user", content: "b " + "z".repeat(60) },
+      { role: "assistant", content: "c" },
+    ];
+    const result = applyAging(msgs, t);
+    const c = result.messages[0].content;
+    return typeof c === "string" ? c : extractTextContent(c as ChatMessageLike["content"]);
+  }
+
+  it("keeps pure-JSON content JSON.parse-able after aging", () => {
+    const json = JSON.stringify({
+      status: "ok",
+      items: Array.from({ length: 5 }, (_, i) => ({ id: i, name: `n${i}` })),
+      meta: { a: 1, b: 2 },
+    });
+    const aged = agedFirstContent(json, thresholds);
+    // Must remain valid JSON (no inline tag corruption).
+    assert.doesNotThrow(() => JSON.parse(aged), `aged JSON not parseable: ${aged.slice(0, 80)}`);
+  });
+
+  it("keeps a fenced code block valid after aging (tag outside the fence)", () => {
+    const fenced = "```json\n{\n  \"a\": 1,\n  \"b\": [1, 2, 3]\n}\n```";
+    const aged = agedFirstContent(fenced, thresholds);
+    // The fenced block must still be present and intact.
+    assert.ok(aged.includes("```json"), "opening fence lost");
+    assert.ok(aged.trimEnd().endsWith("```"), "closing fence lost");
+    assert.ok(aged.includes('"a": 1'), "fenced payload corrupted");
+  });
+
+  it("does not re-compress structured content on a second aging pass (recursion guard)", () => {
+    const json = JSON.stringify({ status: "ok", items: [{ id: 0 }, { id: 1 }], meta: { a: 1 } });
+    const msgs: ChatMessageLike[] = [
+      { role: "user", content: json },
+      { role: "assistant", content: "a " + "z".repeat(60) },
+      { role: "user", content: "b " + "z".repeat(60) },
+      { role: "assistant", content: "c" },
+    ];
+    const first = applyAging(msgs, thresholds);
+    const second = applyAging(first.messages as ChatMessageLike[], thresholds);
+    const firstContent = JSON.stringify(
+      first.messages.map((m) => (m as ChatMessageLike).content)
+    );
+    const secondContent = JSON.stringify(
+      second.messages.map((m) => (m as ChatMessageLike).content)
+    );
+    assert.equal(secondContent, firstContent, "second aging pass changed structured content");
+    // And it must still be parseable.
+    const c0 = (second.messages[0] as ChatMessageLike).content;
+    const text0 = typeof c0 === "string" ? c0 : extractTextContent(c0 as ChatMessageLike["content"]);
+    assert.doesNotThrow(() => JSON.parse(text0), "JSON corrupted after two aging passes");
+  });
+});
--- a/tests/unit/compression/gcf-inline-array-quote.test.ts
+++ b/tests/unit/compression/gcf-inline-array-quote.test.ts
@@ -0,0 +1,34 @@
+/**
+ * Regression guard for B-GCF-QUOTE (lossless violation).
+ *
+ * SPEC §2.4 requires quoting a string that contains the inline-array pattern `[`…`]``:`
+ * (e.g. `ERR[404]: Not Found`, `[Speaker 1]: Hello`). needsQuote() lacked that rule, so
+ * such a value emitted bare on a line-level `key=value` RHS is re-parsed by the decoder
+ * (decode_generic.ts:142-160) as an inline-array header → throws `count_mismatch` (or
+ * silently decodes wrong). Reachable in prod: headroomEngine.apply() ships the blob.
+ */
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { needsQuote } from "@omniroute/open-sse/services/compression/engines/headroom/gcf/scalar.ts";
+import { encodeGeneric } from "@omniroute/open-sse/services/compression/engines/headroom/gcf/generic.ts";
+import { decodeGeneric } from "@omniroute/open-sse/services/compression/engines/headroom/gcf/decode_generic.ts";
+
+test("needsQuote flags the inline-array pattern [..]:  (SPEC §2.4)", () => {
+  assert.equal(needsQuote("ERR[404]: Not Found"), true);
+  assert.equal(needsQuote("[Speaker 1]: Hello"), true);
+  assert.equal(needsQuote("[1]: y"), true);
+  // Must not over-trigger on innocuous brackets without the colon.
+  assert.equal(needsQuote("arr[0] index"), false);
+  assert.equal(needsQuote("plain value here"), false);
+});
+
+test("GCF round-trips nested values containing [..]: losslessly (B-GCF-QUOTE)", () => {
+  const data = [
+    { id: 1, status: "x", code: 1, meta: { note: "ERR[404]: Not Found" } },
+    { id: 2, status: "y", code: 2, meta: { note: "[Speaker 1]: Hello world" } },
+    { id: 3, status: "z", code: 3, meta: { note: "plain note" } },
+  ];
+  const encoded = encodeGeneric(data);
+  const decoded = decodeGeneric(encoded);
+  assert.deepEqual(decoded, data);
+});
--- a/tests/unit/compression/language-detect-select.test.ts
+++ b/tests/unit/compression/language-detect-select.test.ts
@@ -0,0 +1,50 @@
+/**
+ * Guards for B-LANG-DETECTOR + B-LANG-DORMANT.
+ *
+ * B-LANG-DETECTOR: the detector was first-match-wins on a single keyword, and some hint
+ * words are English-ambiguous ("configuration" in fr, "error" in es) → English text
+ * misclassified as fr/es. Now it is score-based and needs ≥2 hits to leave English.
+ *
+ * B-LANG-DORMANT: with autoDetectLanguage on but enabledPacks ["en"], detected non-English
+ * text fell back to the English pack, whose `articles` rule deletes foreign articles
+ * (pt-BR "a"/"o"). Auto-detect must use the detected pack (it always has rules).
+ */
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { detectCompressionLanguage } from "@omniroute/open-sse/services/compression/languageDetector.ts";
+import { cavemanCompress } from "@omniroute/open-sse/services/compression/caveman.ts";
+
+test("detector ignores single English-ambiguous keywords (configuration/error)", () => {
+  assert.equal(
+    detectCompressionLanguage("Please update the configuration and fix the error in the file"),
+    "en"
+  );
+});
+
+test("detector still recognizes genuine non-English text (>=2 native keywords)", () => {
+  assert.equal(detectCompressionLanguage("Por favor preciso do arquivo com erro"), "pt-BR");
+  assert.equal(detectCompressionLanguage("これはテストですコードを確認"), "ja");
+});
+
+test("auto-detect uses the detected pt-BR pack, not the mangling English pack (B-LANG-DORMANT)", () => {
+  // pt-BR prose with an article the English `articles` rule would delete ("a configuração").
+  const text =
+    "Por favor, você poderia revisar a configuração do arquivo? Obrigado pela ajuda com isso.";
+  const res = cavemanCompress({ messages: [{ role: "user", content: text }] } as Record<
+    string,
+    unknown
+  >, {
+    enabled: true,
+    autoDetectLanguage: true,
+    enabledLanguagePacks: ["en"], // the production-default that used to force the English pack
+    intensity: "full",
+    compressRoles: ["user"],
+    minMessageLength: 0,
+  } as Record<string, unknown>);
+  const rules = res.stats?.rulesApplied ?? [];
+  // A pt-BR rule must have run (proves the pt-BR pack was selected, not English).
+  assert.ok(
+    rules.some((r) => r.startsWith("pt_")),
+    `expected a pt_* rule to apply, got: ${JSON.stringify(rules)}`
+  );
+});
--- a/tests/unit/compression/llmlingua-worker-resolution.test.ts
+++ b/tests/unit/compression/llmlingua-worker-resolution.test.ts
@@ -0,0 +1,76 @@
+/**
+ * Regression guard for B-SLM: the LLMLingua worker must resolve its deps + worker
+ * file WITHOUT relying on `import.meta.url`.
+ *
+ * Root cause (confirmed via dist/.build/next/server/chunks/26410.js): webpack
+ * replaces `createRequire(import.meta.url)` with a stub that throws MODULE_NOT_FOUND
+ * and freezes `import.meta.url` to the build-machine path. Both make the worker
+ * never spawn in the standalone bundle. The resolution must use runtime anchors
+ * (process.cwd() / process.argv[1]) instead.
+ */
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+
+import {
+  firstAncestorWith,
+  resolveWorkerFile,
+  depsAvailable,
+} from "@omniroute/open-sse/services/compression/engines/llmlingua/worker.ts";
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const WORKER_SRC = path.resolve(
+  here,
+  "../../../open-sse/services/compression/engines/llmlingua/worker.ts"
+);
+
+/** Strip // line and block comments so we scan CODE, not doc-comments that may mention the banned APIs. */
+function stripComments(src: string): string {
+  return src.replace(/\/\*[\s\S]*?\*\//g, "").replace(/(^|[^:])\/\/.*$/gm, "$1");
+}
+
+test("worker.ts CODE never uses import.meta.url / createRequire (both die in the standalone bundle)", () => {
+  const code = stripComments(fs.readFileSync(WORKER_SRC, "utf8"));
+  assert.ok(!code.includes("import.meta"), "worker.ts code must not reference import.meta");
+  assert.ok(
+    !code.includes('from "node:module"'),
+    "worker.ts must not import from node:module (createRequire)"
+  );
+  assert.ok(
+    !code.includes('from "node:url"'),
+    "worker.ts must not import from node:url (fileURLToPath)"
+  );
+});
+
+test("firstAncestorWith walks up from anchors to find a marker", () => {
+  const tmp = fs.mkdtempSync(path.join(os.tmpdir(), "slm-root-"));
+  try {
+    // Build <tmp>/dist/node_modules/@atjsh/llmlingua-2/package.json and an anchor deep inside.
+    const pkgDir = path.join(tmp, "dist", "node_modules", "@atjsh", "llmlingua-2");
+    fs.mkdirSync(pkgDir, { recursive: true });
+    fs.writeFileSync(path.join(pkgDir, "package.json"), "{}");
+    const anchor = path.join(tmp, "dist", ".build", "next", "server", "chunks");
+    fs.mkdirSync(anchor, { recursive: true });
+
+    const rel = path.join("node_modules", "@atjsh", "llmlingua-2", "package.json");
+    const found = firstAncestorWith([anchor], rel);
+    assert.equal(found, path.join(tmp, "dist"), "must find the dist root by walking up");
+    assert.equal(firstAncestorWith([anchor], path.join("node_modules", "nope")), null);
+  } finally {
+    fs.rmSync(tmp, { recursive: true, force: true });
+  }
+});
+
+test("resolveWorkerFile returns an existing onnxWorker file (no import.meta.url)", () => {
+  // In this test env cwd is the worktree root, which contains the real source tree.
+  const { workerFile } = resolveWorkerFile();
+  assert.ok(fs.existsSync(workerFile), `resolved worker file must exist: ${workerFile}`);
+  assert.ok(/onnxWorker\.(t|j)s$/.test(workerFile), `must point at onnxWorker: ${workerFile}`);
+});
+
+test("depsAvailable is true when @atjsh/llmlingua-2 is installed (symlinked node_modules)", () => {
+  assert.equal(depsAvailable(), true);
+});
--- a/tests/unit/compression/mcpAccessibility-anchors.test.ts
+++ b/tests/unit/compression/mcpAccessibility-anchors.test.ts
@@ -0,0 +1,57 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { smartFilterText } from "@omniroute/open-sse/services/compression/engines/mcpAccessibility/index.ts";
+import { DEFAULT_MCP_ACCESSIBILITY_CONFIG } from "@omniroute/open-sse/services/compression/engines/mcpAccessibility/constants.ts";
+
+/** Regex-extract every [ref=eNN] anchor from a blob, as a sorted unique list. */
+function extractRefs(s: string): string[] {
+  const refs = new Set<string>();
+  for (const m of s.matchAll(/\[ref=e\d+\]/g)) {
+    refs.add(m[0]);
+  }
+  return [...refs].sort();
+}
+
+/**
+ * Build a realistic accessibility snapshot: a `list` with 40 `listitem` siblings, each carrying a
+ * clickable child `link "..." [ref=eNN]`, with interleaved `- generic:` / `- text: ""` noise lines
+ * between siblings (exactly the kind of tree real MCP accessibility dumps produce).
+ */
+function buildSnapshot(): string {
+  const lines: string[] = ['- list "Results":'];
+  for (let i = 0; i < 40; i++) {
+    lines.push(`  - listitem:`);
+    lines.push(`    - generic:`);
+    lines.push(`    - link "Result ${i}" [ref=e${i}]`);
+    lines.push(`    - text: ""`);
+    // interleaved noise BETWEEN siblings — this is what breaks the sibling run for collapse
+    lines.push(`  - generic:`);
+    lines.push(`  - text: ""`);
+  }
+  // pad past minLengthToProcess so smartFilterText actually runs
+  return lines.join("\n").padEnd(3000, " ");
+}
+
+test("collapse fires on interleaved tree AND no [ref=eXX] anchor is lost", () => {
+  const input = buildSnapshot();
+  const out = smartFilterText(input, DEFAULT_MCP_ACCESSIBILITY_CONFIG);
+
+  // BUG B: collapse must actually fire despite interleaved noise lines.
+  assert.ok(out.length < input.length, "output shorter (compressed)");
+  assert.ok(
+    out.includes('items omitted by OmniRoute MCP filter'),
+    "collapse notice present (collapse fired)"
+  );
+
+  // BUG A: every [ref=eNN] in the input must survive in the output (agent can still click them).
+  const inRefs = extractRefs(input);
+  const outRefs = extractRefs(out);
+  assert.equal(inRefs.length, 40, "sanity: 40 refs in input");
+  for (const r of inRefs) {
+    assert.ok(outRefs.includes(r), `ref ${r} must survive collapse (extractRefs(input) ⊆ extractRefs(output))`);
+  }
+
+  // It still compresses meaningfully.
+  const savings = input.length - out.length;
+  assert.ok(savings > 0, "savings > 0");
+});
--- a/tests/unit/compression/mode-and-pipeline.test.ts
+++ b/tests/unit/compression/mode-and-pipeline.test.ts
@@ -0,0 +1,62 @@
+/**
+ * Guards for B-MODE-ENGINE-DECOUPLE and B-PIPELINE-DIVERGENCE.
+ *
+ * B-MODE-ENGINE-DECOUPLE: selecting a single MODE must run its engine even if the
+ * per-engine `enabled` flag is off — the mode selection IS the enable signal (the
+ * per-engine flag still gates STACKED pipeline steps). Previously standard/rtk silently
+ * no-op'd when cavemanConfig.enabled / rtkConfig.enabled was false.
+ *
+ * B-PIPELINE-DIVERGENCE: the global stackedPipeline normalizer stripped
+ * session-dedup/ccr/headroom/llmlingua (engines the combo path allows), so those engines
+ * could never run via the global setting. The allowlists must agree.
+ */
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { applyCompression } from "@omniroute/open-sse/services/compression/strategySelector.ts";
+import { normalizeStackedPipeline } from "../../../src/lib/db/compression.ts";
+
+test("standard mode compresses even when cavemanConfig.enabled is false (B-MODE-ENGINE-DECOUPLE)", () => {
+  const body = {
+    messages: [
+      {
+        role: "user",
+        content:
+          "Please could you kindly review the configuration of the file. ".repeat(8) +
+          "I would really appreciate it, thank you so much for your help here.",
+      },
+    ],
+  };
+  const res = applyCompression(body, "standard", {
+    config: {
+      cavemanConfig: { enabled: false, compressRoles: ["user"], intensity: "full", minMessageLength: 0 },
+    },
+  } as Record<string, unknown>);
+  assert.ok(res.compressed, "standard mode must run caveman regardless of cavemanConfig.enabled");
+});
+
+test("rtk mode compresses even when rtkConfig.enabled is false (B-MODE-ENGINE-DECOUPLE)", () => {
+  const content =
+    Array.from({ length: 60 }, (_, i) => `line ${String(i).padStart(3, "0")} routine output`).join(
+      "\n"
+    ) + "\nERROR: boom";
+  const res = applyCompression({ messages: [{ role: "tool", content }] }, "rtk", {
+    config: { rtkConfig: { enabled: false, intensity: "standard", applyToToolResults: true } },
+  } as Record<string, unknown>);
+  assert.ok(res.compressed, "rtk mode must run regardless of rtkConfig.enabled");
+});
+
+test("normalizeStackedPipeline keeps headroom/ccr/session-dedup/llmlingua (B-PIPELINE-DIVERGENCE)", () => {
+  const pipe = normalizeStackedPipeline([
+    { engine: "session-dedup" },
+    { engine: "ccr" },
+    { engine: "headroom" },
+    { engine: "llmlingua" },
+    { engine: "rtk", intensity: "standard" },
+    { engine: "bogus-engine" }, // unknown ids still dropped
+  ]);
+  const engines = pipe.map((s) => s.engine);
+  for (const e of ["session-dedup", "ccr", "headroom", "llmlingua", "rtk"]) {
+    assert.ok(engines.includes(e), `${e} must survive normalize`);
+  }
+  assert.ok(!engines.includes("bogus-engine"), "unknown engine ids are still dropped");
+});
--- a/tests/unit/compression/rtk-intensity.test.ts
+++ b/tests/unit/compression/rtk-intensity.test.ts
@@ -0,0 +1,49 @@
+/**
+ * Regression guard for B-RTK-INTENSITY: the intensity knob used to be nearly inert
+ * (it only set smartTruncate's preserveHead/Tail 16↔24). It must now scale the effective
+ * line budget so minimal / standard / aggressive differ on truncation-based filters,
+ * WITHOUT ever dropping error/failure lines (priorityPatterns protect them at every
+ * intensity). Include/collapse filters (e.g. docker-logs) compress by content and are
+ * intensity-independent by nature — so the deterministic proof is on effectiveMaxLines.
+ */
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import {
+  applyRtkCompression,
+  effectiveMaxLines,
+} from "@omniroute/open-sse/services/compression/engines/rtk/index.ts";
+
+test("effectiveMaxLines scales the line budget by intensity (minimal > standard > aggressive)", () => {
+  const min = effectiveMaxLines(120, "minimal");
+  const std = effectiveMaxLines(120, "standard");
+  const agg = effectiveMaxLines(120, "aggressive");
+  assert.equal(std, 120, "standard is the baseline");
+  assert.ok(min > std, `minimal (${min}) keeps more than standard (${std})`);
+  assert.ok(agg < std, `aggressive (${agg}) keeps fewer than standard (${std})`);
+  assert.ok(min > agg, "minimal keeps strictly more than aggressive");
+  assert.ok(effectiveMaxLines(1, "aggressive") >= 1, "never below 1 line");
+  assert.equal(effectiveMaxLines(120, undefined), 120, "unknown intensity = baseline");
+});
+
+test("rtk preserves error/failure lines at EVERY intensity", () => {
+  const lines: string[] = [];
+  for (let i = 0; i < 400; i++) {
+    if (i === 120) lines.push("ERROR: connection refused at step 120");
+    else if (i === 300) lines.push("FAILED: assertion mismatch at step 300");
+    else lines.push(`line ${String(i).padStart(4, "0")} routine output text here`);
+  }
+  const content = lines.join("\n");
+  for (const intensity of ["minimal", "standard", "aggressive"] as const) {
+    const res = applyRtkCompression({ messages: [{ role: "tool", content }] }, {
+      enabled: true,
+      intensity,
+      applyToToolResults: true,
+    } as Record<string, unknown>);
+    const out =
+      typeof res.body.messages?.[0]?.content === "string"
+        ? (res.body.messages[0].content as string)
+        : "";
+    assert.ok(out.includes("ERROR: connection refused"), `${intensity}: ERROR line must survive`);
+    assert.ok(out.includes("FAILED: assertion mismatch"), `${intensity}: FAILED line must survive`);
+  }
+});
--- a/tests/unit/compression/ultra-code-preservation.test.ts
+++ b/tests/unit/compression/ultra-code-preservation.test.ts
@@ -0,0 +1,32 @@
+/**
+ * Regression guard for B-ULTRA-CODE: the ultra heuristic must NOT corrupt fenced code
+ * blocks / inline code / URLs. It used to call pruneByScore on raw text (no tombstoning),
+ * so code tokens like `b)` / `{` / `+` scored < minScore and were pruned, turning
+ * `add(a, b) { return a + b; }` into `add(a, return`. caveman + llmlingua both
+ * extract/restore preserved blocks first; ultra must too.
+ */
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { ultraCompress } from "@omniroute/open-sse/services/compression/ultra.ts";
+
+test("ultraCompress preserves fenced code, inline code, and URLs byte-identical", () => {
+  const code = "```ts\nexport function add(a, b) {\n  return a + b;\n}\n```";
+  const inline = "`add(x, y)`";
+  const url = "https://example.com/api/v1/auth?id=42";
+  const filler =
+    "Here is a fairly long explanatory paragraph that should be pruned heavily because " +
+    "it contains lots of low information filler words and redundant phrasing repeated " +
+    "many times over and over again to ensure the heuristic actually triggers pruning. ";
+  // Realistic layout: fenced code block sits on its own line (markdown convention).
+  const text = `${filler}\n\n${code}\n\nThen call ${inline} and see ${url} for details.\n\n${filler}`;
+
+  const { messages } = ultraCompress([{ role: "user", content: text }], {
+    maxTokensPerMessage: 0,
+  });
+  const out = typeof messages[0].content === "string" ? messages[0].content : "";
+
+  assert.ok(out.includes(code), `fenced code block must survive byte-identical:\n${out}`);
+  assert.ok(out.includes(inline), `inline code must survive byte-identical:\n${out}`);
+  assert.ok(out.includes(url), `URL must survive byte-identical:\n${out}`);
+  assert.ok(out.length < text.length, "prose must still be compressed");
+});