mixdog 0.9.154 → 0.9.155

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "0.9.154",
+  "version": "0.9.155",
   "private": false,
   "type": "module",
   "description": "Standalone mixdog coding-agent CLI/TUI workspace.",
@@ -72,7 +72,7 @@
     "test:session-transport": "node --test scripts/session-transport-test.mjs scripts/daemon-bootstrap-test.mjs src/runtime/shared/child-spawn-remote.test.mjs src/standalone/session-runtime-agent-control-client.test.mjs src/standalone/session-runtime-dispatch-cancel.test.mjs src/standalone/session-runtime-host-factory.test.mjs src/standalone/session-runtime-inline-host.test.mjs src/standalone/session-runtime-provider-cooldown.test.mjs src/standalone/session-runtime-shard-host.test.mjs src/standalone/session-runtime-shard-router.test.mjs",
     "test:session": "node --test scripts/runtime-turn-contract-test.mjs scripts/session-save-fault-store-test.mjs scripts/turn-review-revert-test.mjs src/runtime/shared/tool-surface.test.mjs src/runtime/bridge-clients.test.mjs",
     "test:quick": "npm run test:session && npm run test:session-transport && npm run test:desktop-main --prefix apps/desktop",
-    "test:office": "node --test src/runtime/office/office-runtime.test.mjs src/runtime/office/office-assurance.test.mjs && npm run test:office-approval --prefix apps/desktop",
+    "test:office": "node --test src/runtime/office/office-runtime.test.mjs src/runtime/office/office-assurance.test.mjs src/runtime/office/portable-authoring.test.mjs && npm run test:office-approval --prefix apps/desktop",
     "test:office:live": "node scripts/run-office-live-tests.mjs",
     "test:media": "node --test src/runtime/media/store.test.mjs src/runtime/media/renditions.test.mjs src/runtime/media/adapters/codex-image.test.mjs",
     "failures": "node scripts/tool-failures.mjs",

package/scripts/build-voice-runtime-unix.sh

@@ -41,17 +41,41 @@ curl --fail --location --retry 3 \
   --output "${WORK_ROOT}/source.tar.gz"
 tar -xzf "${WORK_ROOT}/source.tar.gz" -C "${WORK_ROOT}"
 
+# Portability baseline (GGML_NATIVE=OFF + explicit ISA on x86-64).
+#
+# ggml defaults GGML_NATIVE=ON, which tunes the binary to the BUILD machine.
+# CI runners are wider than consumer hardware, so a native build ships
+# instructions the target CPU may not have and aborts with SIGILL /
+# STATUS_ILLEGAL_INSTRUCTION right after backend init — the `--help` smoke
+# below cannot catch it, because on the build machine it passes.
+#
+# x86-64 pins an AVX2/FMA/F16C baseline (Haswell 2013+); AVX-512, AVX-VNNI and
+# BMI2 stay OFF because they are not universal on current consumer CPUs.
+# arm64 needs no ISA pins — NEON is baseline — but must still disable NATIVE.
 CMAKE_ARGS=(
   -S "${SOURCE_ROOT}"
   -B "${BUILD_ROOT}"
   -DCMAKE_BUILD_TYPE=Release
   -DBUILD_SHARED_LIBS=OFF
+  -DGGML_NATIVE=OFF
   -DWHISPER_BUILD_EXAMPLES=ON
   -DWHISPER_BUILD_SERVER=ON
   -DWHISPER_BUILD_TESTS=OFF
   -DWHISPER_SDL2=OFF
 )
 
+if [[ "${TARGET_ARCH}" == "x64" ]]; then
+  CMAKE_ARGS+=(
+    -DGGML_AVX=ON
+    -DGGML_AVX2=ON
+    -DGGML_FMA=ON
+    -DGGML_F16C=ON
+    -DGGML_AVX512=OFF
+    -DGGML_AVX_VNNI=OFF
+    -DGGML_BMI2=OFF
+  )
+fi
+
 case "${BACKEND}" in
   vulkan)
     CMAKE_ARGS+=(-DGGML_VULKAN=ON)

package/scripts/voice-runtime-config.json

@@ -1,5 +1,5 @@
 {
-  "runtimeVersion": "1.9.2",
+  "runtimeVersion": "1.9.3",
   "whisperCommit": "306c88f4d1286aec1bf96e544632897886af5501",
   "vulkanSdkVersion": "1.4.309.0",
   "platforms": [

package/src/runtime/agent/orchestrator/providers/anthropic-messages.mjs

@@ -240,8 +240,8 @@ export function requestAnthropicTools(tools, messages, opts) {
 // Test-only: expose the lowering so the steering-provenance test can assert
 // the API-key provider keeps steering-tagged user turns distinct (mirrors
 // anthropic-oauth._buildRequestBodyForCacheSmoke coverage).
-export function _toAnthropicMessagesForTest(messages) {
-    return toAnthropicMessages(messages);
+export function _toAnthropicMessagesForTest(messages, availableTools) {
+    return toAnthropicMessages(messages, availableTools);
 }
 
 // Applies cache_control markers to the FINAL, already-sanitized Anthropic

package/src/runtime/agent/orchestrator/providers/anthropic-oauth.mjs

@@ -296,12 +296,16 @@ function buildRequestBody(messages, model, tools, sendOpts) {
     // Message-tail cache budget (4-BP layout + ANTHROPIC_MSG_SLOTS) is shared
     // with anthropic.mjs — see resolveAnthropicMessageCacheSlots.
     const messageCacheSlots = resolveAnthropicMessageCacheSlots(systemBlocks, ttls);
+    // Tools are resolved BEFORE the messages: the lowering needs the final
+    // tool list to drop tool_reference blocks whose tool no longer ships in
+    // this request (otherwise the API rejects the whole turn).
+    const requestTools = requestAnthropicTools(tools, chatMsgs, opts);
     // Build → sanitize (once, inside toAnthropicMessages) → mark. Markers are
     // applied to the FINAL sanitized array by invariant, so block drops /
     // inserts / reorders performed by the sanitizer can never move or delete a
     // marked block. NEVER sanitize again after this (see send path).
     const anthropicMessages = applyAnthropicCacheMarkers(
-        toAnthropicMessages(chatMsgs),
+        toAnthropicMessages(chatMsgs, requestTools),
         messageCacheSlots,
     );
 
@@ -317,7 +321,6 @@ function buildRequestBody(messages, model, tools, sendOpts) {
 
     if (systemBlocks.length) body.system = systemBlocks;
 
-    const requestTools = requestAnthropicTools(tools, chatMsgs, opts);
     if (requestTools.length) {
         // No cache_control on tools — the systemBase BP already covers the
         // tools prefix via Anthropic's prompt cache prefix semantics (order:

package/src/runtime/agent/orchestrator/providers/anthropic.mjs

@@ -145,12 +145,16 @@ export class AnthropicProvider {
         // Message-tail cache budget (4-BP layout + ANTHROPIC_MSG_SLOTS) is
         // shared with anthropic-oauth — see resolveAnthropicMessageCacheSlots.
         const messageCacheSlots = resolveAnthropicMessageCacheSlots(systemBlocks, ttls);
+        // Tools are resolved BEFORE the messages: the lowering needs the final
+        // tool list to drop tool_reference blocks whose tool no longer ships
+        // in this request (otherwise the API rejects the whole turn).
+        const requestTools = requestAnthropicTools(tools, chatMsgs, opts);
         // Build → sanitize (once, inside toAnthropicMessages) → mark. Markers
         // are applied to the FINAL sanitized array by invariant, so block
         // drops / inserts / reorders performed by the sanitizer can never move
         // or delete a marked block. NEVER sanitize again after this.
         const anthropicMessages = applyAnthropicCacheMarkers(
-            toAnthropicMessages(chatMsgs),
+            toAnthropicMessages(chatMsgs, requestTools),
             messageCacheSlots,
         );
 
@@ -163,7 +167,6 @@ export class AnthropicProvider {
         applyAnthropicServerFallback(params, useModel, {
             enabled: this.config?.disableBetaHeaders !== true && opts.serverFallback !== false,
         });
-        const requestTools = requestAnthropicTools(tools, chatMsgs, opts);
         if (requestTools.length) {
             // No cache_control on tools — the system BP covers tools via
             // Anthropic prefix semantics (order: tools → system → messages).

package/src/runtime/agent/orchestrator/providers/lib/anthropic-request-utils.mjs

@@ -89,7 +89,39 @@ export function resolveAnthropicMessageCacheSlots(systemBlocks, ttls) {
 // root cause of the sporadic COLD-turn cache miss: pre-sanitize markers landed
 // on blocks the sanitizer then rewrote, so the provider-visible breakpoint
 // diverged from the cached one).
-export function toAnthropicMessages(messages) {
+// Optional `availableTools`: the exact tool list this request will ship. A
+// tool_reference whose tool is absent from that list makes the API reject the
+// WHOLE request ("Tool reference '<name>' not found in available tools"), and
+// the dangling block lives in the transcript — so once a loaded tool leaves
+// the deferred catalog (feature toggle, provider/model rebuild, disconnected
+// MCP server) every later turn of that session fails identically. Passing the
+// list lets the lowering drop unresolvable references instead. Omitted =>
+// every reference is kept verbatim.
+function anthropicToolNameSet(availableTools) {
+    if (!Array.isArray(availableTools)) return null;
+    const names = new Set();
+    for (const tool of availableTools) {
+        const name = typeof tool?.name === 'string' ? tool.name.trim() : '';
+        if (name) names.add(name);
+    }
+    return names;
+}
+
+// Replacement content for a tool_result whose references were all dropped.
+// The stored text result (the deferred-load summary) is the natural fallback;
+// an EMPTY tool_result is itself a 400, so a blank result gets a placeholder.
+function toolResultContentWithoutReferences(message, droppedNames) {
+    const normalized = normalizeContentForAnthropic(message.content);
+    if (typeof normalized === 'string' && normalized.trim()) return normalized;
+    if (Array.isArray(normalized) && normalized.length > 0) return normalized;
+    return [{
+        type: 'text',
+        text: `[tool references dropped - no longer available: ${droppedNames.join(', ')}]`,
+    }];
+}
+
+export function toAnthropicMessages(messages, availableTools) {
+    const availableNames = anthropicToolNameSet(availableTools);
     const result = [];
     for (let idx = 0; idx < messages.length; idx++) {
         const m = messages[idx];
@@ -136,12 +168,23 @@ export function toAnthropicMessages(messages) {
                 && Array.isArray(native?.toolReferences)
                 ? native.toolReferences.map((name) => String(name || '').trim()).filter(Boolean)
                 : [];
+            // Keep only references this request can actually back with a
+            // definition; a reference without one is a hard 400 on every turn
+            // that replays it (see anthropicToolNameSet).
+            const usableReferences = availableNames
+                ? references.filter((name) => availableNames.has(name))
+                : references;
+            const droppedReferences = usableReferences.length === references.length
+                ? []
+                : references.filter((name) => !usableReferences.includes(name));
             const block = {
                 type: 'tool_result',
                 tool_use_id: m.toolCallId || '',
-                content: references.length
-                    ? references.map((tool_name) => ({ type: 'tool_reference', tool_name }))
-                    : normalizeContentForAnthropic(m.content),
+                content: usableReferences.length
+                    ? usableReferences.map((tool_name) => ({ type: 'tool_reference', tool_name }))
+                    : (droppedReferences.length
+                        ? toolResultContentWithoutReferences(m, droppedReferences)
+                        : normalizeContentForAnthropic(m.content)),
                 ...((m.toolKind === 'error' || m.isError === true) ? { is_error: true } : {}),
             };
             if (last?.role === 'user' && Array.isArray(last.content)) {

package/src/runtime/agent/orchestrator/session/manager/ask-session.mjs

@@ -18,6 +18,10 @@ import {
     acknowledgePendingDeferredToolDelta,
     snapshotPendingDeferredToolDelta,
 } from '../../../../../session-runtime/deferred-tool-delta.mjs';
+import {
+    acknowledgePendingGoalReminder,
+    snapshotPendingGoalReminder,
+} from '../../../../../session-runtime/goal-reminder.mjs';
 import { recordStandaloneStatusTelemetry } from './status-telemetry.mjs';
 import { normalizeStaleCompactingStage } from './compaction-runner.mjs';
 import { resolveSessionContextMeta, positiveContextWindow } from './context-meta.mjs';
@@ -276,6 +280,7 @@ export async function askSession(sessionId, prompt, context, onToolCall, cwdOver
         let _pwstTurnDrained = null;
         let _turnPendingEntries = [];
         let _turnDeferredToolDelta = null;
+        let _turnGoalReminder = null;
         // After the first turn, the next prompt comes from the drained queue.
         // (On the first iteration _pendingTail is empty and `prompt` is the
         // caller's original message.)
@@ -589,11 +594,15 @@ export async function askSession(sessionId, prompt, context, onToolCall, cwdOver
             }
             const _currentTimeBlock = buildCurrentTimeBlock(prompt);
             _turnDeferredToolDelta = snapshotPendingDeferredToolDelta(session);
+            // Event-driven Goal state: present only after compaction dropped the
+            // Goal's own tool results, never on an ordinary turn.
+            _turnGoalReminder = snapshotPendingGoalReminder(session);
             const _turnReminderBlock = [
                 _currentTimeBlock
                     ? `<system-reminder>\n# Current Time\n${_currentTimeBlock}\n</system-reminder>`
                     : '',
                 _turnDeferredToolDelta?.content || '',
+                _turnGoalReminder?.content || '',
             ].filter(Boolean).join('\n\n');
             const _baseUserTurnContent = prefixUserTurnContent(prompt, _contextBlock);
             const _userTurnContent = prefixSessionStartContent(_baseUserTurnContent, _turnReminderBlock);
@@ -1067,6 +1076,9 @@ export async function askSession(sessionId, prompt, context, onToolCall, cwdOver
             if (_turnDeferredToolDelta) {
                 acknowledgePendingDeferredToolDelta(session, _turnDeferredToolDelta.revision);
             }
+            if (_turnGoalReminder) {
+                acknowledgePendingGoalReminder(session, _turnGoalReminder.revision);
+            }
             const terminalResultPreview = {
                 ...result,
                 trimmed: messagesDropped > 0,

package/src/runtime/agent/orchestrator/session/pre-send-compact.mjs

@@ -36,6 +36,7 @@ import { traceAgentCompact, messagePrefixHash } from '../agent-trace.mjs';
 import { invalidateProviderRequestToolsScope } from '../../../../session-runtime/provider-request-tools.mjs';
 import { bumpUsageMetricsEpoch } from './manager.mjs';
 import { resetReadStateAfterCompaction } from './read-dedup.mjs';
+import { markPendingGoalReminder } from '../../../../session-runtime/goal-reminder.mjs';
 
 const RECOVERED_ERROR_MESSAGE_MAX_CHARS = 300;
 const ANSI_ESCAPE_RE = /[\u001B\u009B][[\]()#;?]*(?:(?:[a-zA-Z\d]*(?:;[-a-zA-Z\d/#&.:=?%@~_]+)*)?\u0007|(?:(?:\d{1,4}(?:;\d{0,4})*)?[\\dA-PR-TZcf-nq-uy=><~]))/g;
@@ -590,6 +591,12 @@ export async function runPreSendCompactPass(state) {
                     catch { /* best-effort: PostCompact hook must never break the loop */ }
                 }
             }
+            // Compaction drops the Goal's own tool results, so the durable task
+            // snapshot vanishes mid-objective and the model keeps working
+            // without its checklist. Mark one tail-injected state reminder for
+            // the next turn; it clears when that turn is accepted.
+            try { markPendingGoalReminder(sessionRef, 'compaction'); }
+            catch { /* best-effort: a Goal reminder must never break compaction */ }
         }
     return {
         iterations,

package/src/runtime/agent/orchestrator/session/store/live-state.mjs

@@ -157,7 +157,10 @@ export function _clearLiveSession(id) {
 // in the persisted JSON) stay resident for this long after their last use so
 // multi-turn image recognition keeps working across an idle gap. Beyond the
 // TTL the memory cost wins and the snapshot is reclaimed like any other.
-export const LIVE_MEDIA_RETENTION_MS = 60 * 60 * 1000; // 1h
+export const LIVE_MEDIA_RETENTION_MS = Math.max(
+    60_000,
+    Number(process.env.MIXDOG_LIVE_MEDIA_RETENTION_MS) || 10 * 60 * 1000,
+); // 10m default — raw image bytes are the most expensive thing we retain
 
 export function _messagesCarryLiveMedia(messages) {
     if (!Array.isArray(messages)) return false;

package/src/runtime/channels/data/voice-runtime-manifest.json

@@ -1,5 +1,5 @@
 {
-  "version": "1.9.2",
+  "version": "1.9.3",
   "source": "ggml-org/whisper.cpp@306c88f4d1286aec1bf96e544632897886af5501",
   "model": {
     "id": "large-v3-turbo-q8_0",
@@ -64,9 +64,9 @@
         {
           "id": "vulkan",
           "requires": null,
-          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.2/whisper-server-win32-x64-vulkan.zip",
-          "sha256": "6ee3718696033ddbc156201eb0c097b778d8aca7388233b54fd3a966cbcb9fee",
-          "size": 16751300,
+          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.3/whisper-server-win32-x64-vulkan.zip",
+          "sha256": "a80997f6841decf08421ab261084fa365ba879932221ba53de98bc3710c9dced",
+          "size": 16733128,
           "format": "zip",
           "executable": "whisper-server.exe"
         }
@@ -77,9 +77,9 @@
         {
           "id": "vulkan",
           "requires": null,
-          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.2/whisper-server-linux-x64-vulkan.zip",
-          "sha256": "c3cbf695ebc818546a8ef4d3657ef74378781948b8ff62e7f2bf1da70ccdc639",
-          "size": 14461318,
+          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.3/whisper-server-linux-x64-vulkan.zip",
+          "sha256": "b4614d11e13f383be33a278003d66401e6a5a94024e9124ea9ada615177d4897",
+          "size": 14333249,
           "format": "zip",
           "executable": "whisper-server"
         }
@@ -90,9 +90,9 @@
         {
           "id": "vulkan",
           "requires": null,
-          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.2/whisper-server-linux-arm64-vulkan.zip",
-          "sha256": "73fc8ac7fdbc1e68ebb4a2c27faaf3c1145fcf7f4b0f04f32fdeb26bc66b20eb",
-          "size": 14274639,
+          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.3/whisper-server-linux-arm64-vulkan.zip",
+          "sha256": "b267e322e66df8d46b2fccd3812d4dcba57de9781baf40ebecb9fe43c477b10b",
+          "size": 14229402,
           "format": "zip",
           "executable": "whisper-server"
         }
@@ -103,8 +103,8 @@
         {
           "id": "metal",
           "requires": null,
-          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.2/whisper-server-darwin-arm64-metal.zip",
-          "sha256": "2ebd9eec5b991716c230f9d4dace8db952f99f0faf8ca1ef3b3a5b4a1b106401",
+          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.3/whisper-server-darwin-arm64-metal.zip",
+          "sha256": "6523f7acf18924e87ab80e7c712d76e8558cc09215df98da703b9102c0f5650b",
           "size": 1350283,
           "format": "zip",
           "executable": "whisper-server"
@@ -116,15 +116,15 @@
         {
           "id": "cpu",
           "requires": null,
-          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.2/whisper-server-darwin-x64-cpu.zip",
-          "sha256": "4d39a13874c380e28ae25ff922673bf8869e6e756f49e866a67e7a86ec8088a5",
-          "size": 1346377,
+          "url": "https://github.com/tribgames/mixdog/releases/download/voice-runtime-v1.9.3/whisper-server-darwin-x64-cpu.zip",
+          "sha256": "3d76426a0c7ec718b33ff54521a1d29ed09e4ed00ac9e8075f81d6bbe77a5d14",
+          "size": 1333240,
           "format": "zip",
           "executable": "whisper-server"
         }
       ]
     }
   },
-  "release_tag": "voice-runtime-v1.9.2",
-  "build_hash": "3705f8c9636467bb1ff9722b409dcc888bd7271e3774f13b0bc3ec64cc3d2822"
+  "release_tag": "voice-runtime-v1.9.3",
+  "build_hash": "d9a066477feb39afd5bc51cf50453d3ce1a194be5de907357ed7f9c1e9677044"
 }

package/src/runtime/channels/lib/voice-runtime-fetcher.mjs

@@ -551,6 +551,33 @@ export async function ensureWhisperRuntime(dataDir, onProgress = null) {
   })
 }
 
+// Version-drift guard. `active-version` holds `whisper-<version>-<variantId>`,
+// so an installed runtime falls BEHIND whenever the bundled manifest moves on
+// (observed: a 754 MB cuBLAS-11 bundle superseded by a 17 MB Vulkan build that
+// keeps GPU acceleration). The legacy executable fallback in
+// resolveManagedWhisperCmd still resolves the old layout, which is exactly
+// what made the drift permanent — every caller saw `installed: true`, so the
+// upgrade path never ran and the superseded bytes stayed on disk forever.
+//
+// Reported separately from `installed` on purpose: the old runtime still WORKS,
+// so the transcribe hot path must keep using it (offline, mid-install), while
+// the install path treats this as "re-fetch".
+export function isManagedWhisperStale(dataDir) {
+  const activeFile = join(dataDir, 'voice-runtime', 'active-version')
+  if (!existsSync(activeFile)) return false
+  const activeName = readFileSync(activeFile, 'utf8').trim()
+  if (!activeName) return false
+  if (!existsSync(BUNDLED_MANIFEST_PATH)) return false
+  try {
+    const manifest = JSON.parse(readFileSync(BUNDLED_MANIFEST_PATH, 'utf8'))
+    if (!manifest?.version) return false
+    return !activeName.startsWith(`whisper-${manifest.version}-`)
+  } catch {
+    // An unreadable manifest is not evidence of staleness.
+    return false
+  }
+}
+
 // Read-only resolver: returns the cached binary path when the managed runtime
 // is fully installed, null otherwise. Used by the transcribe hot path and the
 // /cli-check endpoint to test installation state without triggering a fetch.
@@ -747,6 +774,9 @@ export function resolveVoiceRuntime(dataDir, { modelId = 'standard' } = {}) {
     kind: 'managed-whisper.cpp',
     label: 'whisper.cpp',
     installed: !!(managedWhisperCmd && serverCmd && managedModelPath && managedFfmpegPath),
+    // Installed, but from an older manifest version — usable now, upgradeable
+    // by the install path. See isManagedWhisperStale.
+    stale: isManagedWhisperStale(dataDir),
     binary: !!managedWhisperCmd,
     model: !!managedModelPath,
     ffmpeg: !!managedFfmpegPath,
@@ -759,6 +789,35 @@ export function resolveVoiceRuntime(dataDir, { modelId = 'standard' } = {}) {
   }
 }
 
+/**
+ * Uninstall counterpart to ensureWhisperRuntime + ensureWhisperModel.
+ *
+ * Voice used to be install-only: turning it OFF flipped a config flag and left
+ * the whisper runtime and model weights (1.5 GB+ on a machine that picked a
+ * GPU variant) resident forever, so "off" was never actually cold. Removes
+ * both; re-enabling re-fetches them through the normal ensure* path.
+ *
+ * The managed ffmpeg runtime is deliberately KEPT — it is shared with non-voice
+ * media paths and is two orders of magnitude smaller.
+ *
+ * @returns {string[]} removed directory labels (empty when nothing was present)
+ */
+export function removeManagedVoiceRuntime(dataDir) {
+  const removed = []
+  for (const parts of [['voice-runtime'], ['voice', 'models']]) {
+    const target = join(dataDir, ...parts)
+    if (!existsSync(target)) continue
+    try {
+      rmSync(target, { recursive: true, force: true })
+      removed.push(parts.join('/'))
+    } catch (err) {
+      // A locked file (transcribe in flight) must not fail the toggle itself.
+      process.stderr.write(`[voice-runtime] uninstall skipped ${parts.join('/')}: ${err?.message || err}\n`)
+    }
+  }
+  return removed
+}
+
 // Single managed location for the whisper model weight file. Idempotent: if
 // the resolved file exists and matches the manifest sha256, return without
 // re-downloading. Atomic install via stage-then-rename so a partial download

package/src/runtime/office/assurance-benchmark.mjs

@@ -326,6 +326,71 @@ async function postSaveReleaseGate() {
   };
 }
 
+async function contentAwareCompositionGate() {
+  const content = {
+    packageId: 'composition-gate',
+    audience: 'executive committee',
+    objective: 'approve the operating plan',
+    decision: 'approve the plan',
+    facts: [{ id: 'revenue', label: 'Revenue', value: 120 }],
+    claims: [{ id: 'decision', text: 'Approve the plan', factIds: ['revenue'] }],
+  };
+  const operation = {
+    op: 'compose_document',
+    title: 'Decision brief',
+    summary: 'Approve the plan.',
+    sections: [{
+      heading: 'Evidence',
+      paragraphs: ['Revenue supports the decision.'],
+      table: [['Metric', 'Value'], ['Revenue', 120]],
+    }],
+  };
+  const first = expandOfficeDesignOperations({
+    format: 'docx',
+    backend: 'microsoft-office-com',
+    created: true,
+    design: { purpose: 'decide', expressionMode: 'strong-fit', content },
+    operations: [operation],
+  });
+  const firstId = first.semantic[0].composition.id;
+  const second = expandOfficeDesignOperations({
+    format: 'docx',
+    backend: 'microsoft-office-com',
+    created: true,
+    design: { purpose: 'decide', expressionMode: 'strong-fit', content },
+    library: {
+      source: 'mixdog-starter',
+      recentCompositions: [{
+        fingerprint: first.composition.fingerprint,
+        format: 'docx',
+        compositionIds: [firstId],
+      }],
+    },
+    operations: [operation],
+  });
+  const gate = evaluateOfficeSubmissionGate({
+    persisted: true,
+    issues: [{
+      severity: 'warning',
+      code: 'recent_composition_repeat',
+      path: '/',
+      message: 'same sequence',
+    }],
+  });
+  return {
+    category: 'content-aware-composition-gate',
+    passed: firstId !== second.semantic[0].composition.id
+      && second.semantic[0].composition.historyPenalty === 0
+      && !gate.ok
+      && gate.blocking[0].severity === 'error',
+    evidence: {
+      first: firstId,
+      second: second.semantic[0].composition.id,
+      gate,
+    },
+  };
+}
+
 export async function runOfficeAssuranceBenchmark() {
   const cases = await Promise.all([
     spreadsheetBenchMini(),
@@ -337,10 +402,11 @@ export async function runOfficeAssuranceBenchmark() {
     contentModelCrossApp(),
     semanticDeliverableQuality(),
     postSaveReleaseGate(),
+    contentAwareCompositionGate(),
   ]);
   const passed = cases.filter((entry) => entry.passed).length;
   return {
-    version: 2,
+    version: 3,
     createdAt: new Date().toISOString(),
     measurementKind: 'deterministic-office-assurance',
     categories: cases.length,

package/src/runtime/office/assurance.mjs

@@ -354,21 +354,22 @@ function headingLevel(paragraph) {
 }
 
 function wordBlockOrder(document) {
-  if (Array.isArray(document?.blockOrder) && document.blockOrder.length) return document.blockOrder;
-  const blocks = [
-    ...(document?.paragraphs || []).map((entry) => ({
-      type: 'paragraph',
-      index: entry.index,
-      path: entry.path,
-      start: Number(entry.start),
-    })),
-    ...(document?.tables || []).map((entry) => ({
-      type: 'table',
-      index: entry.index,
-      path: entry.path,
-      start: Number(entry.start),
-    })),
-  ];
+  const blocks = Array.isArray(document?.blockOrder) && document.blockOrder.length
+    ? document.blockOrder.map((entry) => ({ ...entry, start: Number(entry.start) }))
+    : [
+      ...(document?.paragraphs || []).map((entry) => ({
+        type: 'paragraph',
+        index: entry.index,
+        path: entry.path,
+        start: Number(entry.start),
+      })),
+      ...(document?.tables || []).map((entry) => ({
+        type: 'table',
+        index: entry.index,
+        path: entry.path,
+        start: Number(entry.start),
+      })),
+    ];
   if (blocks.every((entry) => Number.isFinite(entry.start))) {
     blocks.sort((left, right) => left.start - right.start);
   }
@@ -379,7 +380,9 @@ function reviewDocxStructure(document) {
   const issues = [];
   const paragraphs = Array.isArray(document?.paragraphs) ? document.paragraphs : [];
   const tables = Array.isArray(document?.tables) ? document.tables : [];
-  const content = paragraphs.filter((paragraph) => String(paragraph.text || '').trim());
+  const content = paragraphs.filter((paragraph) => (
+    paragraph?.inTable !== true && String(paragraph.text || '').trim()
+  ));
   const headings = content
     .map((paragraph) => ({ paragraph, level: headingLevel(paragraph) }))
     .filter((entry) => entry.level !== null);
@@ -408,7 +411,11 @@ function reviewDocxStructure(document) {
     const block = order[index];
     if (block.type !== 'paragraph') continue;
     const paragraph = paragraphsByIndex.get(Number(block.index));
-    if (headingLevel(paragraph) === null || !String(paragraph?.text || '').trim()) continue;
+    if (
+      paragraph?.inTable === true
+      || headingLevel(paragraph) === null
+      || !String(paragraph?.text || '').trim()
+    ) continue;
     const nextBlock = order.slice(index + 1).find((entry) => {
       if (entry.type === 'table') return true;
       return String(paragraphsByIndex.get(Number(entry.index))?.text || '').trim();