fix(desktop): expand export diagnostics for startup failures

.agents/skills/process-pr-reviews/SKILL.md

@@ -1,10 +1,12 @@
 ---
 name: process-pr-reviews
-description: Use when the user asks to process, triage, fetch, view, count, list, or resolve review feedback in a GitHub PR. The built-in workflow currently focuses on CodeRabbit. In this workflow, “real review feedback” is strictly defined as actionable inline comments that are neither review summaries nor nitpicks.
+description: Use when the user asks to process, triage, fetch, view, count, list, or resolve review feedback in a GitHub PR. Supports both CodeRabbit and Codex review workflows. In this workflow, “real review feedback” is strictly defined as actionable inline comments; for CodeRabbit, exclude review summaries and nitpicks, and for Codex, exclude review summary cards and use PR main-thread reactions only as status signals.
 ---
 
 # Process PR Reviews
 
+This workflow supports both CodeRabbit and Codex PR review signals.
+
 ## CodeRabbit Reviews
 
 “Real review feedback” is strictly defined as:
@@ -13,6 +15,8 @@ description: Use when the user asks to process, triage, fetch, view, count, list
 - **not** a review summary
 - **not** a nitpick
 
+Nitpick comments from CodeRabbit must always be ignored to avoid unnecessary noise.
+
 There is no need to analyze the comment content itself.
 
 ### Data sources
@@ -95,6 +99,11 @@ The final goal of the CodeRabbit workflow is always:
 
 > **Top-level inline comments left by CodeRabbit in `pulls/<pr_number>/comments` that are neither nitpicks nor summaries**
 
+Important:
+
+- Treat CodeRabbit nitpicks as non-actionable by default.
+- Do not include nitpicks in counts, summaries, or resolution queues unless the user explicitly asks for nitpicks.
+
 In practice, do the following:
 
 1. Get CodeRabbit top-level inline comments from `pulls/<pr_number>/comments`
@@ -114,6 +123,18 @@ If the output of `gh api --paginate ...` is too large and gets truncated:
 
 ### Resolving review conversations
 
+### Resolution policy
+
+When triaging review feedback, apply this rule:
+
+- If a comment will **not** be fixed, you may resolve the conversation after triage.
+- If a comment **will** be fixed, do **not** resolve it first — make the code change first, then resolve the conversation afterward.
+
+In short:
+
+- **won't fix / no code change** → triage, then resolve
+- **will fix / code change required** → fix first, then resolve
+
 If the user asks to resolve a CodeRabbit review conversation:
 
 1. Identify the target inline comment from the actionable comment list.
@@ -131,3 +152,130 @@ Notes:
 - Resolve the **thread**, not the individual comment.
 - `pulls/<pr_number>/comments` remains the source of truth for identifying actionable inline comments.
 - `reviewThreads` is only for thread-level operations such as resolving conversations.
+
+## Codex Reviews
+
+“Real review feedback” is strictly defined as:
+
+- **inline review comments**
+- **not** the review summary card
+
+There is no need to analyze the comment content itself.
+
+### Important behavior differences from CodeRabbit
+
+- Codex review is **silent while running**.
+- Unlike CodeRabbit, Codex does **not** expose an in-progress PR check for review status.
+- While Codex is reviewing, the PR main conversation thread gets an `eyes` reaction from `chatgpt-codex-connector[bot]`.
+- If Codex finds no issues, it may leave **no actionable inline comments** and instead react to the PR main conversation thread with `+1` (thumbs up).
+- If Codex finds issues, it may create inline review comments in `pulls/<pr_number>/comments` and a review summary card in `pulls/<pr_number>/reviews`.
+
+### Data sources
+
+The Codex workflow uses these sources:
+
+1. **PR review comments**
+
+   ```bash
+   gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/comments
+   ```
+
+   This is the authoritative source for actionable inline Codex comments.
+
+2. **PR reviews**
+
+   ```bash
+   gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/reviews
+   ```
+
+   This is used to identify the Codex review summary card such as `### 💡 Codex Review`. It is not the primary source of actionable inline feedback.
+
+3. **Issue comment reactions on the PR main thread**
+
+   First fetch the PR issue node / comments if needed:
+
+   ```bash
+   gh api repos/<owner>/<repo>/issues/<pr_number>/comments
+   gh api repos/<owner>/<repo>/issues/<pr_number>/reactions
+   ```
+
+   Use reactions on the PR main conversation thread only to detect Codex review state:
+
+   - `eyes` from `chatgpt-codex-connector[bot]` → Codex review appears to be in progress
+   - `+1` from `chatgpt-codex-connector[bot]` on the PR main thread → Codex reviewed and found no issues
+
+   These reactions are **status signals**, not actionable review feedback.
+
+Do not treat these as primary sources for actionable comments:
+
+- `gh pr view ...`
+- `gh api repos/<owner>/<repo>/issues/<pr_number>/comments`
+- `gh api repos/<owner>/<repo>/issues/<pr_number>/reactions`
+
+Reason: actionable Codex review feedback still lives in PR review comments, not in the PR issue timeline.
+
+### Workflow
+
+#### 1. Fetch inline comments
+
+```bash
+gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/comments
+```
+
+Only keep records that satisfy all of the following:
+
+- `user.login` is `chatgpt-codex-connector[bot]`
+- `in_reply_to_id == null` (only top-level inline comments, not replies)
+
+This is the candidate set of actionable Codex comments.
+
+#### 2. Fetch reviews to identify the summary card
+
+```bash
+gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/reviews
+```
+
+Identify Codex review summaries. Common characteristics include:
+
+- `### 💡 Codex Review`
+- explanatory text such as `Here are some automated review suggestions for this pull request.`
+- “About Codex in GitHub” help text
+
+These review-level contents are **not the final result**. They are only used to understand whether Codex posted a review summary.
+
+#### 3. Optionally inspect PR main-thread reactions for status
+
+If the user asks whether Codex is still reviewing, or whether Codex finished with no findings, inspect reactions on the PR main thread.
+
+Interpret them as follows:
+
+- `eyes` by `chatgpt-codex-connector[bot]` → likely still reviewing / review in progress
+- `+1` by `chatgpt-codex-connector[bot]` with no Codex inline comments → likely completed with no findings
+
+Do not count these reactions as review comments.
+
+### Filtering rule
+
+The final goal of the Codex workflow is always:
+
+> **Top-level inline comments left by Codex in `pulls/<pr_number>/comments`**
+
+In practice, do the following:
+
+1. Get Codex top-level inline comments from `pulls/<pr_number>/comments`
+2. Use `pulls/<pr_number>/reviews` only to recognize the summary card
+3. If there are no Codex inline comments, optionally inspect PR main-thread reactions to distinguish:
+   - still reviewing (`eyes`)
+   - reviewed with no findings (`+1`)
+   - no observable Codex activity
+
+### Large output handling
+
+If any `gh api --paginate ...` output is too large and gets truncated:
+
+1. Record the tool output file path
+2. Do not manually read through the entire large JSON blob
+3. Hand it off to `@explorer` to extract:
+   - Codex-authored inline comments
+   - the number of top-level inline comments
+   - each comment’s `path` / `line` / `body`

apps/desktop/main/desktop-diagnostics.ts

@@ -5,6 +5,7 @@ import type {
   RuntimeEventQueryResult,
   RuntimeLogEntry,
   RuntimeState,
+  StartupProbePayload,
 } from "../shared/host";
 import type { RuntimeOrchestrator } from "./runtime/daemon-supervisor";
 import {
@@ -54,6 +55,17 @@ type DesktopDiagnosticsSnapshot = {
   updatedAt: string;
   isPackaged: boolean;
   coldStart: DesktopColdStartSnapshot;
+  startupProbe: {
+    preloadSeen: boolean;
+    rendererSeen: boolean;
+    entries: Array<{
+      source: StartupProbePayload["source"];
+      stage: string;
+      status: StartupProbePayload["status"];
+      detail: string | null;
+      at: string;
+    }>;
+  };
   sleepGuard: SleepGuardSnapshot;
   renderer: DesktopRendererSnapshot;
   embeddedContents: DesktopEmbeddedContentSnapshot[];
@@ -68,6 +80,8 @@ function nowIso(): string {
   return new Date().toISOString();
 }
 
+const MAX_STARTUP_PROBE_ENTRIES = 200;
+
 export function getDesktopDiagnosticsFilePath(): string {
   return resolve(app.getPath("userData"), "logs", "desktop-diagnostics.json");
 }
@@ -83,6 +97,12 @@ export class DesktopDiagnosticsReporter {
     error: null,
   };
 
+  private readonly startupProbe: DesktopDiagnosticsSnapshot["startupProbe"] = {
+    preloadSeen: false,
+    rendererSeen: false,
+    entries: [],
+  };
+
   private sleepGuard: SleepGuardSnapshot = createInitialSleepGuardSnapshot();
 
   private readonly renderer: DesktopRendererSnapshot = {
@@ -152,6 +172,33 @@ export class DesktopDiagnosticsReporter {
     this.scheduleFlush();
   }
 
+  recordStartupProbe(payload: StartupProbePayload): void {
+    if (payload.source === "preload") {
+      this.startupProbe.preloadSeen = true;
+    }
+
+    if (payload.source === "renderer") {
+      this.startupProbe.rendererSeen = true;
+    }
+
+    this.startupProbe.entries.push({
+      source: payload.source,
+      stage: payload.stage,
+      status: payload.status,
+      detail: payload.detail ?? null,
+      at: nowIso(),
+    });
+
+    if (this.startupProbe.entries.length > MAX_STARTUP_PROBE_ENTRIES) {
+      this.startupProbe.entries.splice(
+        0,
+        this.startupProbe.entries.length - MAX_STARTUP_PROBE_ENTRIES,
+      );
+    }
+
+    this.scheduleFlush();
+  }
+
   recordRendererDidFinishLoad(url: string): void {
     this.renderer.didFinishLoad = true;
     this.renderer.lastUrl = url;
@@ -277,6 +324,11 @@ export class DesktopDiagnosticsReporter {
       updatedAt: nowIso(),
       isPackaged: app.isPackaged,
       coldStart: { ...this.coldStart },
+      startupProbe: {
+        preloadSeen: this.startupProbe.preloadSeen,
+        rendererSeen: this.startupProbe.rendererSeen,
+        entries: this.startupProbe.entries.map((entry) => ({ ...entry })),
+      },
       sleepGuard: {
         ...this.sleepGuard,
         counters: { ...this.sleepGuard.counters },