fix-sentry

name	fix-sentry
description	Auto-fix high-frequency Sentry issues: fetch issues > N occurrences, analyze stack traces, fix code, create GitHub issues, and submit PRs. Supports user feedback issues (event.type "default") with attachment analysis (logs, screenshots) when include_feedback=true. Use when: (1) User says "/fix-sentry", (2) User asks to fix Sentry issues.

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."

Operating Modes

Batch Mode (default)

Invocation: /fix-sentry or /fix-sentry threshold=50

• Uses the specified threshold (default 100) to fetch issues
• Runs full Phase 1 → Phase 2 → Phase 3
• Fixes all qualifying issues

Daemon Mode

Invocation: /fix-sentry limit=1 (from daemon script)

• Phase 1 uses adaptive threshold descent: starts at 100, lowers progressively until a fixable issue is found
• Fixes only 1 issue (controlled by limit parameter), then exits
• If no fixable issue exists at any threshold → outputs [NO_FIXABLE_ISSUES] and exits

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 1.1b: Load Skip List (Daemon Mode Only)

In daemon mode (limit > 0), load the skip list to avoid re-analyzing issues that were already triaged in previous sessions. The skip list is stored at:

~/.wayland-fix-sentry/skip-list.json

Format:

{
  "ELECTRON-6X": { "reason": "already_fixed", "expires": "2026-03-28T03:00:00Z", "summary": "PR #1758 merged" },
  "ELECTRON-A7": { "reason": "system_level", "expires": "2026-04-03T03:00:00Z", "summary": "EPIPE in net.Socket" }
}

On load:

1. Read the file (if it doesn't exist, start with an empty skip list)
2. Remove all entries where expires < current time (expired entries get re-analyzed)
3. Keep the remaining entries as the active skip list

During Phase 1.2 (Fetch Issues):

When iterating through fetched issues, if an issue's short ID (e.g., ELECTRON-6X) is in the active skip list, skip it immediately without calling get_issue_details or doing any analysis. Log the skip: Skipping ELECTRON-6X (cached: already_fixed - PR #1758 merged)

In batch mode (limit=0): skip list is ignored - always analyze everything fresh.

Step 1.2: Fetch Unresolved Issues

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

Batch mode (no limit parameter or limit=0)

Use the specified threshold parameter (default 100) directly:

mcp__sentry__list_issues(
  projectSlugOrId="<project>",
  query="times_seen:><threshold> is:unresolved",
  sort="freq",
  limit=25
)

Daemon mode (limit > 0): Adaptive Threshold Descent

When limit is set, use adaptive threshold descent to find fixable issues. Start high and lower progressively - this ensures the most impactful issues are fixed first.

Threshold sequence: 100 → 80 → 60 → 40 → 20 → 10

For each threshold in the sequence:

1. Fetch issues: mcp__sentry__list_issues(query="times_seen:><threshold> is:unresolved", sort="freq", limit=25)
2. Run Steps 1.3–1.6 (filter, deduplicate, triage)
3. If any "Needs fix" issues are found → proceed to Phase 2 with the top limit issues
4. If all issues are skipped (already fixed, system-level, unfixable) → log and try the next lower threshold

If all thresholds are exhausted with no fixable issues, enter Deep Analysis Mode (Step 1.2b).

Step 1.2b: Deep Analysis Mode - Issues Without Stack Traces

When no fixable issues remain at any standard threshold, search for issues that lack stack traces but may still be fixable through code analysis:

mcp__sentry__list_issues(
  projectSlugOrId="<project>",
  query="!has:stacktrace is:unresolved",
  sort="freq",
  limit=10
)

For these issues, apply Step C (Defensive fix) logic from Step 1.6:

• Extract distinctive patterns from the error message (file names, paths, keywords)
• Search the codebase for matching code paths
• If a matching code path is found → classify as "Defensive fix" and proceed to Phase 2

If deep analysis also yields no fixable issues, output the following exact text and exit:

[NO_FIXABLE_ISSUES] All thresholds exhausted, no actionable issues found.

This marker is machine-readable - the daemon script uses it to determine backoff timing.

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
• Event type (event.type - error, default, etc.)
• Stack trace (file paths, line numbers, function names)
• First/last seen timestamps
• Release version(s) affected
• Frequency and affected users count

Step 1.5b: Attachment Analysis

For issues where event.type is default (user feedback) or stack traces are absent, check for attachments that may contain diagnostic information.

When to run: After Step 1.5, for any issue group that meets either condition:

• event.type is default (user feedback / bug report)
• No usable stack trace was extracted in Step 1.5

Prerequisite: Only runs when include_feedback=true is set for event.type: default issues. Issues without stack traces are always eligible regardless of this flag.

Procedure:

1. Get events for the issue (if not already fetched):

   mcp__sentry__list_issue_events(
     issueUrl="<sentry-url>",
     limit=3
   )
   

2. List attachments for each event:

   mcp__sentry__get_event_attachment(
     projectSlug="<project>",
     eventId="<eventId>"
   )
   

3. Download and analyze attachments by type:

   Attachment type	Name / MIME pattern	Analysis method
   Logs	logs.gz, *.log, *.txt	Download → decompress → search for error patterns, stack traces, warnings
   Screenshots	*.png, *.jpg, screenshot*	Download → use vision to identify UI state, error dialogs, frozen screens
   Config / state	*.json, *.xml	Download → check for misconfigurations

   mcp__sentry__get_event_attachment(
     projectSlug="<project>",
     eventId="<eventId>",
     attachmentId="<attachmentId>"
   )
   

4. Extract diagnostic signals from logs:

   • Error / Warning lines (error, warn, fatal, crash, EPIPE, etc.)
   • Stack traces embedded in log output
   • Performance indicators (memory usage, CPU spikes, event loop delays)
   • Repeated failure patterns around the reported issue time

5. Extract signals from screenshots:

   • Visible error messages or dialogs
   • UI freeze indicators (loading spinners stuck, unresponsive elements)
   • Memory / resource warnings visible in UI

6. Combine user description + attachment signals to form a diagnosis hypothesis. Pass the combined evidence to Step 1.6 for triage classification.

Rate-limit note: Same as Step 1.5 - call attachment APIs sequentially, never in parallel.

Step 1.6: Triage - Can We Fix It?

Classify each issue group using the detailed decision flow in references/triage-rules.md.

Quick reference - seven categories:

Category	Action
Direct fix	Stack trace → our code → fix
Defensive fix	No trace, but pattern matches our code → fix with guards
Feedback fix	User feedback + attachment analysis → identifiable code path → fix
Pending merge	Open PR exists → skip or improve
Already fixed	Merged PR / resolved → skip
System-level	EPIPE, ENOSPC, EIO, uv, Chromium → skip
Unfixable	No trace, no matching code, no diagnostic attachments → skip

Output a triage report (see references/report-template.md for format), 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 all checks in order. Every check must pass before proceeding to Step 2.6.

# 1. Lint + auto-fix
bun run lint:fix

# 2. Format + auto-fix
bun run format

# 3. Type check - MUST pass
bunx tsc --noEmit

# 4. Tests - MUST pass
bun run test

i18n check (run if any src/renderer/, locales/, or src/common/config/i18n files were modified):

bun run i18n:types
node scripts/check-i18n.js

• i18n:types must run before check-i18n.js
• If check-i18n.js exits with errors → fix them before proceeding
• If check-i18n.js exits with warnings only → may proceed

Final CI verification - replicate the exact CI check locally:

prek run --from-ref origin/main --to-ref HEAD

• If prek reports issues → fix them (run bun run lint:fix and bun run format again), then re-run prek
• prek uses check-only commands (lint, format:check) - it will catch anything the auto-fix missed

Gate rules:

Check	Result	Action
Type check	FAIL	Fix type errors and re-run. Max 3 attempts, then abandon issue.
Tests	FAIL	Adjust fix/test and re-run. Max 3 attempts, then abandon issue.
i18n	FAIL	Fix missing keys and re-run.
prek	FAIL	Fix reported issues and re-run.

Step 2.5: Verify Fix

Verification strategy depends on which process the error originates from:

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: unit tests from Step 2.3 are sufficient. Mark as verified if tests pass.
• Renderer: use CDP for live verification. See references/cdp-verification.md for full flow.

Step 2.6: Commit & Create PR

Do NOT invoke /commit or /oss-pr - they may prompt for confirmation, which blocks the daemon flow. Instead, commit and create the PR directly using the commands below.

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 directly:

   git add <changed-files>
   git commit -m "<type>(<scope>): <subject>"
   

   Follow project commit conventions: <type>(<scope>): <subject> in English. No AI signatures. Reference the Sentry issue IDs in the commit body if needed.

2. Push branch and create PR as Draft:

   git push -u origin fix/sentry-<shortId>
   
   gh pr create --draft --title "<type>(<scope>): <subject>" --body "$(cat <<'EOF'
   ## Summary
   
   <1-3 bullet points describing the fix>
   
   ## Sentry Issues
   
   - <Sentry issue ID> - <error message> (<occurrence count> occurrences)
   
   ## Test Plan
   
   - [x] Unit tests pass
   - [x] Type check passes
   - [x] Lint and format pass
   EOF
   )"
   

   PR title: under 70 characters, <type>(<scope>): <description> format. NEVER add AI-generated signatures, Generated with, or Co-Authored-By lines.

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"
   

Step 2.7: Return to Main

git checkout main

Proceed to the next group.

Phase 3: Summary Report

After all groups are processed, output a summary report. See references/report-template.md for the exact format.

Step 3.1: Update Skip List (Daemon Mode Only)

In daemon mode (limit > 0), after the summary report, update ~/.wayland-fix-sentry/skip-list.json with all issues that were skipped in this session. This prevents the next session from re-analyzing the same issues.

TTL by classification:

Classification	TTL	Reason
system_level	7 days	These never change (EPIPE, ENOSPC, EIO, uv, etc.)
already_fixed	48 hours	Re-check in case of regression
feedback_no_clue	48 hours	User feedback with no diagnostic clues in attachments
unfixable	24 hours	Might become fixable with new code changes
fix_pending_merge	12 hours	PR might get merged, issue might resolve

Write rules:

1. Read the existing file first (preserve entries from previous sessions that haven't expired)
2. For each skipped issue in this session, add or update its entry with the appropriate TTL
3. For issues that were fixed in this session (PR created), do NOT add to skip list - they should be detected as "already fixed" by the next session's normal triage
4. Write the merged result back to the file

Example output:

{
  "ELECTRON-A7": { "reason": "system_level", "expires": "2026-04-03T03:00:00Z", "summary": "EPIPE in net.Socket" },
  "ELECTRON-6X": { "reason": "already_fixed", "expires": "2026-03-29T03:00:00Z", "summary": "PR #1758 merged" },
  "ELECTRON-16": { "reason": "system_level", "expires": "2026-04-03T03:00:00Z", "summary": "Electron SingletonCookie" },
  "ELECTRON-X": { "reason": "fix_pending_merge", "expires": "2026-03-27T15:00:00Z", "summary": "PR #1498 open" }
}

Configuration

Default parameters (can be overridden via skill args):

Parameter	Default	Description
threshold	100	Minimum occurrence count (batch mode only)
project	electron	Sentry project slug
sort	freq	Sort order for issues
limit	0	Max issues to fix per invocation (0 = unlimited, >0 = daemon mode)
include_feedback	false	Include event.type: default (user feedback) issues

Override examples:

• Batch mode: /fix-sentry threshold=50 project=electron
• Daemon mode: /fix-sentry limit=1 project=electron
• Include user feedback: /fix-sentry include_feedback=true

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.