Fix Sentry Skill

Automated workflow: Sentry issues → analyze → fix → GitHub Issue → PR.

Announce at start: "I'm using fix-sentry skill to find and fix high-frequency Sentry issues."

Prerequisites

• Sentry MCP must be configured (global or project scope) with mcp__sentry__* tools available
• gh CLI must be authenticated
• Working directory must be clean (git status shows no uncommitted changes)

Workflow

Phase 1: Collect & Filter Issues

Step 1.1: Verify Environment

git status --porcelain   # must be clean
git branch --show-current

If working directory is dirty, STOP and ask user to commit or stash first.

Step 1.2: Fetch High-Frequency Unresolved Issues

Always include is:unresolved to exclude issues already marked as resolved in Sentry.

mcp__sentry__list_issues(
  projectSlugOrId="electron",
  query="times_seen:>100 is:unresolved",
  sort="freq",
  limit=25
)

Step 1.3: Evidence-Based Filtering

Determine whether each issue has already been addressed. Only skip issues with concrete evidence of a fix — version distribution alone is NOT sufficient to conclude an issue is fixed (the latest release may simply have fewer users).

1. Get the latest release version:

   gh release list --repo <org>/<repo> --limit 3
   

2. Search for existing fixes (concrete evidence required):

   gh release view <latest-tag> --repo <org>/<repo>
   git log --oneline --since="<release-date>" --grep="<keyword-from-error>"
   

3. Cross-reference with Sentry issue metadata:

   • If the issue has a GitHub annotation linking to a merged PR, skip it
   • If the issue status is resolved with inRelease, skip it
   • If release notes explicitly mention a fix for this error, skip it

4. Check for existing OPEN PRs:

   gh pr list --repo <org>/<repo> --state open --search "<error-keyword>" --json number,title,state
   

   • If an OPEN PR already addresses this issue, do NOT create a duplicate
   • Classify as "fix pending merge" — the issue is still occurring because the fix hasn't been deployed yet
   • If the OPEN PR has quality issues (e.g., missing tests), note it for improvement

Important: version distribution is supplementary info, NOT a skip criterion. "Only seen on v1.8.30, not on v1.8.31" does NOT mean the issue is fixed — the latest version may have too few users to trigger the error. Include version info in the triage report for context, but never use it as the sole reason to skip an issue.

Classification criteria (three states):

Condition	Classification	Action
Has merged PR / mentioned in release notes	Already fixed	Skip
Resolved with inRelease in Sentry	Already fixed	Skip
Has OPEN PR addressing the root cause	Fix pending merge	Skip (or improve existing PR)
No concrete fix evidence found	Needs fix	Fix it

Step 1.4: Deduplicate by Root Cause

Sentry creates separate issues for the same error across different releases or slight variations. Group issues by their root cause (same function + same error type):

Example: ELECTRON-5, ELECTRON-6X, ELECTRON-1A are all fetchModelList + "Missing credentials" → Treat as one fix group, reference all Sentry IDs in the PR.

Step 1.5: Get Stack Traces (Rate-Limit Aware)

For each unique issue group, get details one at a time:

mcp__sentry__get_issue_details(issueUrl="<sentry-url>")

Important: Sentry API rate limit is 5 requests/second. Call get_issue_details sequentially, never in parallel. If you hit a 429, wait a moment and retry.

Extract:

• Error message and type
• Stack trace (file paths, line numbers, function names)
• First/last seen timestamps
• Release version(s) affected
• Frequency and affected users count

Step 1.6: Triage — Can We Fix It?

Classify each issue group into one of four categories using the decision flow below.

Step A: Skip — System-level or framework-internal errors

These errors originate outside our codebase and cannot be fixed by application code changes. Skip immediately without further analysis.

Error pattern	Source	Action
write EPIPE / broken pipe	OS pipe closed	Skip
ENOSPC: no space left on device	Disk full	Skip
write EIO (no app code in stack)	I/O hardware/driver	Skip
uv__loop_interrupt	libuv internal	Skip
SingletonCookie / SingletonLock	Chromium internal	Skip
ERR_INTERNET_DISCONNECTED	Network offline	Skip

Step B: Direct fix — Stack trace points to our code

When a stack trace is available and points to our codebase:

Criteria	Result
Stack trace points to src/ files in our repo	Fix
Error cause is clear from trace	Fix
Fix is straightforward (null check, try-catch, type guard)	Fix
Stack trace points to third-party lib only (no app code)	Skip
Fix requires architectural redesign	Skip

Note on file paths: Sentry stack traces reference build output paths (e.g., src/common/chatLib.ts). After refactoring, files may have moved (e.g., → src/common/chat/chatLib.ts). Use Glob to locate the actual file in the current codebase.

Step C: Defensive fix — No stack trace, but error pattern is identifiable

Some errors (especially native Node.js fs, net errors) are reported without stack traces. These should NOT be automatically skipped — the error message itself often contains enough information to locate the responsible code.

Approach: Extract distinctive patterns from the error message (file name fragments, path structures, keywords), then search the codebase for code that produces or consumes matching patterns. If a matching code path is found, trace its error handling and apply a defensive fix (guards, try-catch, existence checks) even without 100% certainty it's the exact source.

Scenario	Result
Error pattern matches a code path in our codebase	Defensive fix
Error is purely user-specific with no matching code	Skip
Error references app-internal files (config, resources)	Defensive fix

Step D: Skip filters (apply to all categories)

Condition	Action
Has merged PR / mentioned in release notes	Skip (already fixed)
Resolved with inRelease in Sentry	Skip (already fixed)
Has OPEN PR addressing the root cause	Skip (or improve existing PR)

Classification summary

Each issue ends up in one of these categories:

Category	Criteria	Action
Direct fix	Stack trace → our code, clear cause	Fix with targeted code change
Defensive fix	No stack trace, but error path matches our code	Fix with defensive guards
Pending merge	Existing OPEN PR addresses the root cause	Skip or improve existing PR
Already fixed	Merged PR / resolved in Sentry	Skip
System-level	EPIPE, ENOSPC, EIO, uv, Chromium internal	Skip
Unfixable	No stack trace, no matching code path, third-party	Skip

Output a triage report to the user before proceeding:

=== Sentry Issue Triage ===

Will fix — direct (N groups):
  1. [ELECTRON-XX] Error description (N events)
     → file:line — root cause summary

Will fix — defensive (N groups):
  1. [ELECTRON-YY] Error description (N events)
     → Pattern: "batch-export-*.zip" matches createZip in fsBridge.ts
     → Defensive fix: ensure parent directory exists before write

Fix pending merge (P groups):
  1. [ELECTRON-ZZ] Error description (N events)
     → PR #1234 (OPEN) — fix submitted but not yet merged/deployed

Skipped (M issues):
  1. [ELECTRON-AA] EPIPE (N events) → System-level: OS pipe closed
  2. [ELECTRON-BB] SingletonCookie (N events) → Chromium internal
  3. [ELECTRON-CC] Error (N events) → Already fixed: PR #456 merged

Output the triage report for transparency, then proceed immediately — do not wait for user confirmation.

Phase 2: Fix Issues (Serial, One Group at a Time)

Phase 2 handles two types of work:

• New fixes: issues with no existing PR → full flow (Steps 2.1–2.7)
• Pending-merge fixes: issues with an OPEN PR that needs improvement (e.g., missing tests) → checkout existing branch, add tests, push update (Steps 2.1b–2.5, then 2.7)

Process all groups serially: pending-merge groups first (quick improvement), then new fixes.

Step 2.1: Create Branch (New Fix)

For issues with no existing PR:

git checkout main
git pull origin main
git checkout -b fix/sentry-<primary-issue-shortId>

Branch naming: fix/sentry-<shortId> using the highest-frequency issue in the group (e.g., fix/sentry-ELECTRON-6X).

Step 2.1b: Checkout Existing Branch (Pending-Merge Fix)

For issues with an existing OPEN PR that needs improvement (e.g., missing tests):

# Get the branch name from the PR
gh pr view <pr-number> --repo <org>/<repo> --json headRefName --jq '.headRefName'
# Checkout and sync
git checkout <branch-name>
git pull origin <branch-name>

Then skip Step 2.2 (code fix already exists) and go directly to Step 2.3 (Write Tests).

Step 2.2: Locate and Fix Code

1. Use Glob to find the actual file path (may differ from Sentry stack trace due to refactoring)
2. Read the file(s) identified in the stack trace
3. Understand the surrounding context (read neighboring code, types, callers)
4. Implement the minimal fix:
   • Add null/undefined guards
   • Add try-catch for unhandled exceptions
   • Fix incorrect type assertions
   • Add missing error handling
   • Fix race conditions with proper async handling
5. Do NOT refactor surrounding code — fix only the reported issue

Step 2.3: Write Tests for the Fix

Every bug fix MUST have a corresponding unit test. This is enforced by the commit skill and the testing skill — do not skip it.

1. Check if a test file already exists for the modified module (e.g., utils.test.ts for utils.ts)
2. If no test file exists, create one following the testing skill conventions
3. Write test(s) that:
   • Reproduce the bug: a test that would have failed before the fix
   • Verify the fix: the same test now passes with the fix applied
   • Cover at least one failure path (e.g., null input, missing key, invalid URL)
4. Run bun run test to confirm the new tests pass
5. If the fix is in code that's hard to unit test (e.g., deep Electron API dependency), document why in a code comment and add the closest possible test

Examples of good fix tests:

• Fix: added null check for apiKey → Test: call function with undefined apiKey, assert graceful error
• Fix: wrapped fs.readdir in try-catch → Test: mock fs.readdir to throw EPERM, assert no crash
• Fix: validated URL before new URL() → Test: pass invalid URL string, assert error response

Step 2.4: Quality Checks

Run quality checks with fallback commands. Some projects use bun run scripts, others need direct npx/bunx invocation. Try the script first, fall back to direct invocation.

# Lint — try script first, fall back to npx
bun run lint:fix 2>/dev/null || npx oxlint --fix

# Format — try script first, fall back to npx
bun run format 2>/dev/null || npx oxfmt

# Type check — always works
bunx tsc --noEmit

# Tests — run if available, warn if test script is missing
bun run test 2>/dev/null || echo "Warning: no test script found, skipping tests"

Type check must pass. Lint and format are best-effort with fallback. If tests fail due to the fix, adjust the fix. If tests fail for unrelated reasons, note it in the PR.

Step 2.5: Verify Fix

Verification strategy depends on which process the error originates from.

Determine process type from the Sentry stack trace:

Culprit path / error origin	Process	Verification method
src/process/, src/index.ts	main	Unit tests only
src/process/worker/	worker	Unit tests only
src/renderer/, src/common/ (IPC)	renderer	CDP + unit tests

Main / Worker process errors → Unit tests only

Most high-frequency Sentry errors originate from the main process (fs, net, cron, IPC bridge providers). CDP (Chrome DevTools Protocol) connects to the renderer process and cannot inspect main or worker process errors.

For these fixes:

1. Unit tests from Step 2.3 are the primary and sufficient verification
2. Quality checks from Step 2.4 must pass
3. No CDP verification needed — do not attempt it
4. Mark as verified if unit tests pass

Renderer process errors → CDP verification

Only use CDP when the error originates from renderer-side code (React components, UI hooks, renderer-side IPC calls). These are errors visible in the browser DevTools console.

Prerequisites:

• mcp__chrome-devtools__* tools must be available
• CDP is enabled by default in dev mode on port 9230
• Start the app if not running: bun run start & Wait ~20s, then poll mcp__chrome-devtools__list_pages until pages appear.

CRITICAL — MCP session rules:

1. NEVER run claude mcp remove/add mid-session — tools become permanently unavailable.
2. The chrome-devtools MCP server connects to CDP lazily — app can be started during workflow.
3. If MCP tools return "No such tool", classify as skipped and rely on unit tests.

See docs/cdp.md for CDP configuration details.

CDP verification flow:

1. Navigate to the relevant page using mcp__chrome-devtools__navigate_page
2. Reproduce the error scenario via click, fill, press_key, evaluate_script
3. Check for errors: list_console_messages, take_screenshot, list_network_requests
4. Pass: error no longer occurs. Fail: error still occurs or new error introduced.

On failure — retry loop (max 3 attempts):

Adjust the fix → re-run tests → re-run quality checks → re-verify. After 3 failures, proceed to commit & PR but mark verification as FAILED.

On success — collect evidence (screenshots, console logs) for the PR.

Step 2.6: Commit & Create PR

Delegate to existing skills — do not manually construct commit messages or PR bodies.

Pre-flight duplicate check (safety net, supplements triage-phase filtering):

gh pr list --repo <org>/<repo> --state open --search "<error-keyword-or-file>" --json number,title
gh issue list --repo <org>/<repo> --state open --search "<error-keyword>" --json number,title

If an existing OPEN PR/issue addresses the same root cause, STOP — do not create a duplicate. Instead, report to the user and suggest updating the existing PR if needed.

1. Commit: Invoke the commit skill (/commit). The commit skill will analyze changes, run quality checks, format the commit message, and handle all conventions (no AI signatures, no --no-verify, etc.). Provide context: this is a Sentry bug fix, reference the Sentry issue IDs.

2. Create PR as Draft: Invoke the PR skill (/pr). The PR skill will create a GitHub issue if needed, push the branch, and create the PR with proper formatting and issue linkage. Always create as Draft (gh pr create --draft) — PR starts in WIP state. Provide context: include Sentry issue IDs, occurrence counts, error details, and verification results (screenshots, console logs, pass/fail status) so the PR skill can incorporate them into the issue and PR body.

3. Mark PR Ready based on verification result:

   Process	Verification Result	PR Action
   main	Unit tests pass	gh pr ready <pr-number> — mark as Ready for Review
   main	Unit tests fail / not writable	Keep as Draft, add needs-manual-review label
   renderer	CDP pass	gh pr ready <pr-number> — mark as Ready for Review
   renderer	CDP fail (3 attempts exhausted)	Keep as Draft, add needs-manual-review label

   # On pass (unit tests pass for main, or CDP pass for renderer):
   gh pr ready <pr-number>
   
   # On fail:
   gh pr edit <pr-number> --add-label "needs-manual-review"
   

This ensures all commits and PRs follow the project's established conventions without duplicating rules across skills.

Step 2.7: Wait for CI & Auto-Merge

After the PR is created and marked as Ready for Review, wait for all CI checks to pass, then automatically merge the PR.

Only auto-merge when ALL of these conditions are met:

1. PR is marked as Ready for Review (not Draft)
2. All required CI checks pass (see list below)
3. No check is in failure state

If the PR is Draft (verification failed/skipped → needs-manual-review), skip auto-merge and proceed to the next group.

Polling flow:

# Poll CI status (max 15 minutes, check every 30 seconds)
gh pr checks <pr-number> --repo <org>/<repo> --watch --fail-fast

If gh pr checks --watch is not available, use a manual polling loop:

max_wait = 900  # 15 minutes
interval = 30   # seconds
elapsed = 0

while elapsed < max_wait:
    checks = gh pr checks <pr-number>
    if all checks passed:
        break
    if any check failed:
        report failure, skip merge
        break
    sleep interval
    elapsed += interval

if elapsed >= max_wait:
    report timeout, skip merge

CI checks to monitor (fast checks only — ignore slow Build Test jobs):

Check	Monitor
Code Quality	Yes
Unit Tests (all platforms)	Yes
Coverage Test	Yes
I18n Check	Yes
Release Script Test	Yes
Build Test (all platforms)	Skip
CodeQL / Analyze	Skip

Only wait for the "Yes" checks above. Build and CodeQL jobs are slow and non-blocking for bug-fix PRs — do not wait for them.

When polling, check only the monitored jobs. If all monitored checks pass, proceed to merge even if Build Test / CodeQL are still pending or skipped.

On all monitored checks passed — merge:

gh pr merge <pr-number> --repo <org>/<repo> --squash --delete-branch

Use --squash to keep commit history clean. --delete-branch cleans up the remote branch.

On any check failed:

1. Do NOT merge
2. Add ci-failed label to the PR
3. Report the failed check(s) in the summary
4. Proceed to the next group

On timeout (15 minutes):

1. Do NOT merge
2. Add ci-timeout label
3. Report in summary that CI did not complete in time

Step 2.8: Return to Main

git checkout main

Proceed to the next group.

Phase 3: Summary Report

After all groups are processed, output:

=== Fix Sentry Results ===

Fixed & Merged (N groups, covering X Sentry issues):
  1. [ELECTRON-5, ELECTRON-6X, ELECTRON-1A] Missing credentials in fetchModelList
     PR: <pr-url> (merged ✓)
     Issue: #<number>
     Verification: PASS — screenshot attached, no console errors
     CI: all checks passed, auto-merged via squash

  2. ...

Fixed, Pending Manual Review (P groups):
  1. [ELECTRON-YY] Worker process error
     PR: <pr-url> (draft)
     Verification: skipped — worker process, not verifiable via chrome-devtools
     → Requires manual review and merge

Fixed, CI Failed (F groups):
  1. [ELECTRON-ZZ] Error description
     PR: <pr-url> (ci-failed)
     → Failed check: Build Test (windows-x64)

Already fixed (M issues):
  1. [ELECTRON-6, ELECTRON-6Y] Unsupported message type 'finished'
     → Evidence: PR #456 merged in v1.8.31

Skipped (K issues):
  1. [ELECTRON-J] write EPIPE
     → Reason: System-level error, no application code

Total: N fixed (A auto-merged, B pending review, C ci-failed), M already fixed, K skipped

Configuration

Default parameters (can be overridden via skill args):

Parameter	Default	Description
threshold	100	Minimum occurrence count
project	electron	Sentry project slug
sort	freq	Sort order for issues

Override example: /fix-sentry threshold=50 project=electron

Mandatory Rules

No AI Signature

NEVER add any AI-related signatures to commits, PRs, or issues.

Minimal Fix Only

Fix the reported error. Do NOT refactor, add features, or "improve" surrounding code.

No Blocking Questions

The entire workflow runs end-to-end without stopping for user confirmation. Output the triage report for transparency, then proceed immediately. The goal is uninterrupted automation — questions block the flow.

No Duplicate PRs

Before creating a new PR/issue, always check for existing OPEN PRs addressing the same root cause. If found, improve the existing PR (e.g., add missing tests) instead of creating a duplicate.

One Root Cause = One Branch = One PR

Group duplicate Sentry issues by root cause. Each unique root cause gets one branch, one GitHub issue, and one PR.

Rate Limit Awareness

Sentry API has a rate limit of ~5 requests/second. Always call get_issue_details sequentially, never in parallel.

Skill Changes Stay Separate

Do NOT include changes to .claude/skills/ in bug-fix branches. Skill updates should go through their own branch and PR.