mixdog 0.9.153 → 0.9.155

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "0.9.153",
+  "version": "0.9.155",
   "private": false,
   "type": "module",
   "description": "Standalone mixdog coding-agent CLI/TUI workspace.",
@@ -31,6 +31,7 @@
     "!scripts/bench/",
     "!scripts/recall-bench-*.txt",
     "!scripts/*-test.mjs",
+    "!scripts/*.test.mjs",
     "!scripts/*-smoke.mjs",
     "!scripts/smoke-loop*.mjs",
     "!scripts/*-bench.mjs",
@@ -40,6 +41,8 @@
     "!scripts/*.jsx",
     "src/",
     "!src/tui/dev",
+    "!src/**/*.test.mjs",
+    "!src/**/*.test.jsx",
     "vendor/"
   ],
   "scripts": {
@@ -68,7 +71,8 @@
     "test:spec-advisory": "node scripts/advisory-spec-test.mjs && npm run test:spec-advisory --prefix apps/desktop",
     "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: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: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 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/rules/shared/60-verification.md

@@ -9,6 +9,8 @@
   buildability invariants. Treat mutable behavior, UX, exact text, snapshots,
   and implementation shape as advisory specifications; update them when the
   requested behavior changes instead of preserving obsolete behavior.
+- A new test asserts observable behavior, never source text or implementation
+  shape, and never re-asserts a contract another test already owns.
 - A check runs at the strictness the task requires; never raise a tool's own
   severity beyond it.
 - If verification fails, collect all failures, leave Verification, complete all

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/agent-loop.mjs

@@ -27,7 +27,6 @@ import {
     _intraTurnSig,
 } from './loop/tool-classify.mjs';
 import { preDispatchDenyForSession } from './loop/pre-dispatch-deny.mjs';
-import { runRecallFastTrackCompact } from './loop/recall-fasttrack.mjs';
 import { executeTool, _scopedCacheOutcomeForCall, resolveLiveToolCwd } from './loop/tool-exec.mjs';
 
 // classifyResultKind is imported from result-classification.mjs at the top of

package/src/runtime/agent/orchestrator/session/compact/handoff.mjs

@@ -239,28 +239,6 @@ export function collectWorkingFileGroups(messages, cap = Number.POSITIVE_INFINIT
     }, prior, limit, cwd);
 }
 
-export function collectWorkingFiles(messages, cap = Number.POSITIVE_INFINITY, options = {}) {
-    const groups = collectWorkingFileGroups(messages, cap, options);
-    return [...groups.modified, ...groups.referenced].map((entry) => entry.path);
-}
-
-export function stripWorkingFileSections(text) {
-    const out = [];
-    let skipping = false;
-    for (const raw of String(text || '').split('\n')) {
-        const line = raw.trim();
-        if (line === '## Working files') {
-            skipping = true;
-            continue;
-        }
-        if (skipping && (/^##\s+/.test(line) || /^<\/?prior-compacted-context>$/.test(line))) {
-            skipping = false;
-        }
-        if (!skipping) out.push(raw);
-    }
-    return out.join('\n').replace(/\n{3,}/g, '\n\n');
-}
-
 function indexToolResults(messages) {
     const map = new Map();
     for (const m of messages || []) {

package/src/runtime/agent/orchestrator/session/compact/messages.mjs

@@ -43,8 +43,8 @@ export function isProtectedContextUserMessage(m) {
 // content prefix (the `<skill>` envelope from buildSkillResultEnvelope) so the
 // check survives even if the synthetic `meta` field is dropped during a tail
 // rebuild. It is NOT the human's latest prompt and must be excluded from
-// "latest human request" selection (deriveCurrentRequest /
-// buildRecallFastTrackQuery). The `meta:'skill'` marker is also honoured.
+// "latest human request" selection (deriveCurrentRequest). The `meta:'skill'`
+// marker is also honoured.
 export function isInjectedSkillBodyMessage(m) {
     if (m?.role !== 'user') return false;
     if (m.meta === 'skill') return true;

package/src/runtime/agent/orchestrator/session/compact/runner.mjs

@@ -717,13 +717,6 @@ function truncateTailToCap(messages, cap) {
     return estimateMessagesTokens(sanitized) <= cap ? sanitized : [];
 }
 
-function splitRecallFitInputs(recallText) {
-    return {
-        recall: String(recallText || '').trim(),
-        prior: '',
-    };
-}
-
 function recallTailStartIndex(live, tail) {
     if (!tail.length) return live.length;
     const first = tail[0];
@@ -837,9 +830,9 @@ function _recallFastTrackCompactMessages(messages, budgetTokens, opts = {}) {
         head: recallHead,
         lastTurnStartIdx = recallHead.length,
     } = selectRecallPreservedTail(live, recallTailOpts);
-    const recallFit = splitRecallFitInputs(opts.recallText);
+    const recallText = String(opts.recallText || '').trim();
     if (recallHead.length === 0
-        && !(recallFit.recall || recallFit.prior || opts.allowEmptyRecall === true)) {
+        && !(recallText || opts.allowEmptyRecall === true)) {
         throw new Error('recallFastTrackCompactMessages: no compactable prior history before preserved tail');
     }
 
@@ -851,7 +844,7 @@ function _recallFastTrackCompactMessages(messages, budgetTokens, opts = {}) {
     }
     const budgetRaisedBy = Math.max(0, budget - originalBudget);
 
-    if (!recallFit.recall && !recallFit.prior && opts.allowEmptyRecall !== true) {
+    if (!recallText && opts.allowEmptyRecall !== true) {
         throw new Error('recallFastTrackCompactMessages: recall text is empty');
     }
     const oldHistory = recallHead;
@@ -871,14 +864,14 @@ function _recallFastTrackCompactMessages(messages, budgetTokens, opts = {}) {
         ? Math.min(recallRoomUncapped, Math.max(512, recallTokenCap - tailTokens))
         : recallRoomUncapped;
     const toolLines = collectToolOutcomeLines(recallHead)
-        .filter((line) => !recallFit.recall.includes(String(line || '').trim()));
+        .filter((line) => !recallText.includes(String(line || '').trim()));
     const workingFiles = collectWorkingFileGroups(recallHead, undefined, {
         cwd: opts.cwd,
         previousSummary: '',
         now: Date.now(),
     });
     const conversationLines = excludeTailFromConversation(
-        conversationLinesFromMemoryText(recallFit.recall),
+        conversationLinesFromMemoryText(recallText),
         recallTail,
     );
     const composedRecall = composeRecallHandoff({
@@ -894,7 +887,6 @@ function _recallFastTrackCompactMessages(messages, budgetTokens, opts = {}) {
         fittedRecall,
         recallRoom,
         recallMeta,
-        '',
     );
     if (!summaryMessage) {
         throw new Error(`recallFastTrackCompactMessages: summary cannot fit remaining budget=${recallRoom}`);
@@ -936,20 +928,15 @@ function _recallFastTrackCompactMessages(messages, budgetTokens, opts = {}) {
         remainingTokens: budget - mandatoryCost,
         recallTokenCap: (Number.isFinite(recallTokenCap) && recallTokenCap > 0) ? recallTokenCap : null,
         recallRoom,
-        recallChars: recallFit.recall.length,
-        recallBytes: textByteLength(recallFit.recall),
-        priorChars: 0,
-        priorBytes: 0,
+        recallChars: recallText.length,
+        recallBytes: textByteLength(recallText),
         summaryMessageChars: summaryContent.length,
         summaryMessageBytes: textByteLength(summaryContent),
-        recallEmpty: !recallFit.recall,
-        priorEmpty: !recallFit.prior,
-        recallTruncatedInSummary: !!recallFit.recall && !summaryContent.includes(recallFit.recall),
-        priorTruncatedInSummary: false,
+        recallEmpty: !recallText,
+        recallTruncatedInSummary: !!recallText && !summaryContent.includes(recallText),
         tailTruncated: recallTail.some((m) => messageContentHasMarker(m, RECALL_TAIL_TRUNCATION_MARKER) || messageContentHasMarker(m, RECALL_TAIL_SHORT_TRUNCATION_MARKER)),
         fileReattached: reattach.reattached,
         tailOptions: recallTailOpts,
-        previousSummary: false,
         durationMs: Date.now() - startedAt,
     };
     compactDebugLog('recall-fasttrack result', diagnostics);