crashlytics-crash-report-android

name	crashlytics-crash-report-android
description	Analyze Android Crashlytics logs with mandatory git blame analysis and code-level fixes. Multi-agent architecture: classifier-android → firebase-fetcher → forensics-android → validate-report. Use when the user invokes /crash-report-android.

Converted from Claude Code command /crash-report-android. Review and adapt: hooks and MCP tool IDs may need manual mapping for Codex.

Android Crash Analysis - Multi-Agent Edition

Analyze crash errors from Firebase Crashlytics using specialized agents.

Configuration

Before starting, check if a config file exists at .claude/crashlytics.local.md. If it exists, read and use these settings:

• language — output language for body content (default: English). Section headers stay English.
• default_branch — local branch name; analysis runs on origin/<default_branch> (default: master)
• forensics_model — model for forensics agent (default: opus)
• output_format — both / detailed_only / jira_only (default: both)
• firebase_project_id — pre-configured project ID (skip auto-detection if set)
• firebase_app_id_android — pre-configured app ID (skip auto-detection if set)

If config doesn't exist, auto-create it with defaults and continue:

Bash: mkdir -p .claude && cat <<'EOF' > .claude/crashlytics.local.md
---
language: en
default_branch: master   # local branch name; plugin uses origin/<this> for git blame
default_platform: android
forensics_model: opus
output_format: both
firebase_project_id: ""
firebase_app_id_android: ""
firebase_app_id_ios: ""
---

# Crashlytics Plugin Config

Auto-created with defaults. Run `/crash-config` to customize.
EOF

Multi-Agent Architecture

classifier(Haiku) → fetcher(Haiku) → forensics(Opus) → validate-report.py → output

Workflow

Prerequisites Check (cached)

Skip if .claude/crashlytics-prereqs-ok exists. Otherwise run and cache on success.

Bash: test -f .claude/crashlytics-prereqs-ok && echo "CACHED_OK"

If not cached:
  Bash: plugins/crashlytics/scripts/check-prerequisites.sh
  Parse: OK → pass, MISSING → show fix instruction
  If ALL OK → Bash: touch .claude/crashlytics-prereqs-ok
  If any MISSING → show checklist, degrade gracefully (do NOT cache)

Fix instructions: node → brew install node, firebase → npm install -g firebase-tools, firebase-auth → firebase login in terminal, python3 → brew install python3.

STEP 0: Firebase Data

NEVER use mcp__plugin_crashlytics_firebase__firebase_login — broken.

Path A: direct MCP fast-path (preferred)

If firebase_project_id and firebase_app_id_android are set in config AND mcp__plugin_crashlytics_firebase__* tools are available:

1. Call: mcp__plugin_crashlytics_firebase__crashlytics_get_issue
     appId: {firebase_app_id_android}
     issueId: {ISSUE_ID}
2. Call: mcp__plugin_crashlytics_firebase__crashlytics_batch_get_events
     appId: {firebase_app_id_android}
     names: [sampleEvent from step 1]

Skip Path B if both calls succeed.

Path B: firebase-fetcher agent (discovery / fallback)

Task(
  subagent_type="firebase-fetcher",
  model="haiku",
  prompt="Fetch crash data:
    Issue ID: {ISSUE_ID}
    Platform: android
    Config project_id: {firebase_project_id or ''}
    Config app_id: {firebase_app_id_android or ''}"
)

Parse result: extract issue data, events, stack traces, console_url. If fallback → ask user for stack trace manually with Console URL link.

STEP 1: Get data

If user provided a Firebase Issue ID — first try loading via firebase-fetcher (Step 0). Only ask for stack trace if auto-loading failed.

If no Issue ID — ask to provide:

• Stack trace (required)
• Context: crash count, % users, device, app version

STEP 2: Call crash-classifier-android

Task(
  subagent_type="crash-classifier-android",
  model="haiku",
  prompt="Classify this Android crash:
    Stack trace: {stack_trace}
    Context: Events={event_count}, Users={user_count}%, Version={app_version}, Device={device}"
)

STEP 3: Call crash-forensics-android

Use forensics_model from config (default: opus).

Task(
  subagent_type="crash-forensics-android",
  model="{forensics_model}",
  prompt="Analyze this Android crash with git blame:
    Classification: {classifier_output}
    Firebase data: {firebase_output}
    Stack trace: {stack_trace}
    console_url: {console_url}
    branch_ref: origin/{default_branch from config, default master}"
)

STEP 3.5: Quality Gate

Pass the full forensics output through a tmp file — echo "..." mangles markdown with backticks/$/code fences.

1. Write tool: /tmp/crashlytics-forensics-{ISSUE_ID}.md
   contents: <full forensics_output verbatim, no truncation>

2. Bash: python3 plugins/crashlytics/scripts/validate-report.py \
           --console-url "{console_url}" \
           < /tmp/crashlytics-forensics-{ISSUE_ID}.md

Parse the YAML result:

• pass: true → output report as-is.
• pass: false:
  • Если все элементы missing[] относятся к одной секции — fill in those fields using the existing forensics output.
  • Если score >= 8/14 и language config != en — высоко вероятен false-positive validator'а на переведённых заголовках. Проверить вручную каждое поле из missing[] по факту наличия в полном отчёте; если все на месте, считать pass: true и вывести.
  • Иначе — fill missing fields from previous steps. Do NOT re-call forensics. Do NOT re-validate against an abbreviated summary — это даёт false 1/14.
• pass: null (has needs_review items) → evaluate those fields yourself and decide pass/fail.

STEP 4: Output results

Format 1: Detailed Analysis — basic info, stack trace analysis, git blame, root cause, fix (before/after), assignee, context.

Format 2: JIRA Brief — name, key stack trace lines, root cause, code fix, component, assignee, reproduction steps.

Output based on output_format from config (default: both).

Fallback mode (if agents unavailable)

Perform analysis directly: classify → Glob/Grep files → git blame → assignee → fix → output both formats.

Pre-submit checklist

• Classification completed (component, trigger)
• Files found via Glob/Grep or reason explained
• git blame executed on configured branch
• Assignee determined with source (git blame line X)
• Report formats match config (both by default)
• Reviewer passed or missing fields filled in
• console_url included in JIRA Brief

Git blame + code search = MANDATORY
"TBD" = "I analyzed and ownership is unclear", NOT "I didn't check"