Native harness audit: accepted reliability and instruction fixes

app/src/main/kotlin/com/youcoded/app/runtime/SessionService.kt

@@ -4234,6 +4234,9 @@ class SessionService : Service() {
             // native:set-binding below — has a msg.id, so this replies
             // not-implemented-on-mobile rather than no-op'ing.
             "native:queue-remove",
+            // "Send now" on a waiting message — request/response like
+            // native:queue-remove, so it replies not-implemented-on-mobile.
+            "native:queue-send-now",
             "native:interrupt",
             // Stalled-turn Retry. Fire-and-forget (no msg.id) exactly like
             // native:send / native:interrupt, so this is a correct no-op here:

desktop/line-budgets.json

@@ -7,22 +7,22 @@
     "renderer/dev/workbench/fixtures/marketplace/registry.ts": "generated sample of the marketplace registries (regenerated by a script, not hand-edited)"
   },
   "budgets": {
-    "main/ipc-handlers.ts": 5728,
-    "main/harness/native-session-host.ts": 5296,
-    "renderer/App.tsx": 4966,
-    "main/remote-server.ts": 4217,
+    "main/ipc-handlers.ts": 5731,
+    "main/harness/native-session-host.ts": 5311,
+    "renderer/App.tsx": 4981,
+    "main/remote-server.ts": 4223,
     "main/harness/harness-session.ts": 4229,
     "renderer/components/SettingsPanel.tsx": 3256,
-    "renderer/remote-shim.ts": 3176,
+    "renderer/remote-shim.ts": 3177,
     "renderer/state/chat-reducer.ts": 3397,
     "renderer/styles/globals.css": 2960,
     "renderer/components/SessionStrip.tsx": 2784,
     "main/main.ts": 2604,
-    "shared/types.ts": 2391,
+    "shared/types.ts": 2394,
     "renderer/components/SyncPanel.tsx": 2060,
     "main/engine/engine-manager.ts": 2035,
     "renderer/components/ResumeBrowser.tsx": 2069,
-    "main/preload.ts": 1961,
+    "main/preload.ts": 1963,
     "renderer/components/StatusBar.tsx": 1809,
     "renderer/components/ToolCard.tsx": 1636
   }

desktop/package.json

@@ -17,7 +17,7 @@
     "lint": "oxlint --type-aware",
     "postinstall": "node scripts/patch-node-pty.js && node scripts/patch-xterm-webgl-mipmap.js",
     "typecheck": "tsgo --noEmit -p tsconfig.json && tsgo --noEmit -p tsconfig.tests.json",
-    "lint:design": "oxlint -c .oxlintrc.design.json --format default --max-warnings 542"
+    "lint:design": "oxlint -c .oxlintrc.design.json --format default --max-warnings 537"
   },
   "keywords": [],
   "author": "itsdestin",

desktop/src/main/harness/busy-message-boundary.ts

@@ -0,0 +1,66 @@
+import type { ModelMessage } from 'ai';
+import { finalizeRemainingCalls } from './tool-group-finalization';
+
+export interface BusyMessage {
+  id: string;
+  text: string;
+  attachments: string[];
+  restore?: () => void;
+}
+
+/** WHY: the claim and acceptance are one synchronous boundary. An exception
+ * returns the unaccepted head to its host, rather than losing a queued send. */
+export function claimBusyMessage(
+  take: (() => BusyMessage | undefined) | undefined,
+  accept: (item: BusyMessage) => void,
+  beforeAccept?: () => void,
+): boolean {
+  const item = take?.();
+  if (!item) return false;
+  try {
+    beforeAccept?.();
+    accept(item);
+  } catch (err) {
+    item.restore?.();
+    throw err;
+  }
+  return true;
+}
+
+/** WHY: the emitted event is the authority for both live and rebuilt history;
+ * no synthetic steer or app-generated marker is used for human input. */
+export function appendUserHistory(
+  text: string, attachments: string[], emit: () => string, appGenerated: boolean,
+  imageParts: (paths: string[]) => Array<{ type: 'file'; mediaType: string; data: Buffer }>,
+  markAppGenerated: (message: ModelMessage) => ModelMessage,
+  history: ModelMessage[], origins: Array<string[] | null>, record: (uuid: string) => void,
+): void {
+  const images = imageParts(attachments);
+  const uuid = emit();
+  const message = (images.length
+    ? { role: 'user', content: [{ type: 'text', text }, ...images] } as ModelMessage
+    : { role: 'user', content: text } as ModelMessage);
+  history.push(appGenerated ? markAppGenerated(message) : message);
+  origins.push([uuid]);
+  record(uuid);
+}
+
+/** Pair every unstarted call before adding new human input. Already completed
+ * results stay at the front of the same tool message. */
+export function supersedeToolGroup<TCall extends { toolCallId: string; toolName: string }, TPart>(
+  calls: readonly TCall[], from: number, completed: TPart[], completedOrigins: string[],
+  emit: (data: { toolUseId: string; toolName: string; toolResult: string; isError: true }) => string,
+  part: (call: TCall, text: string) => TPart,
+  record: (uuid: string) => void, commit: (parts: TPart[], origins: string[]) => void,
+  injectCompleted: (calls: TCall[]) => void,
+  reason = 'Not run: new user input arrived before this action started.',
+  afterCompleted?: () => void,
+): void {
+  const remaining = finalizeRemainingCalls(calls, from,
+    () => reason,
+    (call, text) => emit({ toolUseId: call.toolCallId, toolName: call.toolName, toolResult: text, isError: true }), part);
+  remaining.origins.forEach(record);
+  commit([...completed, ...remaining.parts], [...completedOrigins, ...remaining.origins]);
+  injectCompleted(calls.slice(0, from));
+  afterCompleted?.();
+}

desktop/src/main/harness/eval/run-case.ts

@@ -9,6 +9,7 @@ import { ASSISTANT_PRESET } from '../../../shared/harness-manifest';
 import { CORE_TOOLS } from '../tools';
 import { seedFixtureWorkspace } from './fixture-workspace';
 import { assembleSystemPrompt } from '../prompt-assembly';
+import { prepareProjectInstructions } from '../injection/project-instructions';
 import { resolvePreset } from '../preset-registry';
 import { BATTERY_PROMPT } from './battery';
 import type { TranscriptEvent } from '../../../shared/types';
@@ -411,6 +412,9 @@ export async function runCase(opts: RunCaseOpts): Promise<CaseRun> {
   const wrapUpPrompt = opts.wrapUpPrompt ?? WRAP_UP_PROMPT;
 
   const fixtureRoot = seedFixtureWorkspace(opts.instructions);
+  // WHY: evaluator uses the production async inventory, but fixture experiments
+  // cannot silently inherit real instructions above their disposable root.
+  const projectInstructionFiles = await prepareProjectInstructions(fixtureRoot, 20_000, fixtureRoot);
   const events: TranscriptEvent[] = [];
   let toolCalls = 0;
   let asks = 0;
@@ -486,6 +490,7 @@ export async function runCase(opts: RunCaseOpts): Promise<CaseRun> {
         presetBody: resolvePreset('assistant').body,
         cwd: fixtureRoot,
         appVersion: 'eval',
+        projectInstructionFiles,
       }),
       // Load-bearing (2026-08-11): omitted, HarnessSession takes fitToContext's
       // own 32_768 default, which against this harness's output ceiling left a