de-kaput

Your app went kaput. Your coding agent fixes each crash, and proves it.

de-kaput is a CLI and a runbook for Claude Code, Codex or any other coding agent. The agent reproduces each Crashlytics crash with a test, then fixes it. de-kaput checks that proof and opens one pull request per crash. You review and merge.

Highlights

• Proof, not guesses. No pull request without a test that fails with the crash's own exception before the fix and passes after it.
• One pull request per crash, like Dependabot, each reviewed and merged on its own.
• Android, iOS and Kotlin Multiplatform. Gradle runs the JVM tests; xcodebuild runs XCTest and Swift Testing on the simulator.
• Runs on its own on a schedule, with a Discord or Slack message for each pull request and note.
• Any agent that can run commands and use an MCP server.

Early, and in daily use. It runs every morning on its maintainers' own Android and iOS app, where its first pull request is merged.

How it works

1. The agent reads the versions with crashes and the release's top crashes from Crashlytics, through the Firebase MCP server.
2. dekaput list skips crashes that already have a pull request and queues the rest: crashes still in the newest release first, fresh or regressed ones first among them, then by users.
3. For each crash: a fresh branch in its own worktree, a test that crashes, the fix. dekaput verify proves it, and dekaput open-pr opens the pull request.
4. A crash no test can reach gets a note on its Crashlytics issue instead.

See it fix real crashes

Demo forks with crashes planted on purpose. For the last two, a fresh agent given only the runbook, the repo and the crash report found the cause.

| App | Crash | PR | |---|---|---| | Google's architecture-samples (Kotlin) | A coroutine crash | #1 | | NetNewsWire (Swift) | A trap in the feed parser | #1 | | Google's architecture-samples | A sort comparator that breaks Java's contract, never in the stack | #2 | | NetNewsWire | A Data slice indexed from 0, two commits after it was made | #2 |

Quick start

Needs Node 20 or newer, the GitHub CLI, and your app's tests running on this machine.

npm install -g de-kaput

# Crashlytics for the agent, with only the five tools de-kaput uses (Codex: see the setup guide)
claude mcp add firebase -- npx -y firebase-tools@latest mcp --tools crashlytics_get_report,crashlytics_batch_get_events,crashlytics_list_notes,crashlytics_create_note,firebase_read_resources

npx firebase-tools login
gh auth login

# In your app's repo: give the agent the de-kaput skill
mkdir -p .claude/skills && cp -R "$(npm root -g)/de-kaput/skills/de-kaput" .claude/skills/

Commit .github/de-kaput.yml:

version: 1
apps:                         # the apps you ship: Android, iOS or both
  - platform: android
    crashlytics-app-id: "1:1234567890:android:abc123"
    packages: [com.example.app]
    releases: latest
    modules:
      - path: app
        gradle-task: ":app:testDebugUnitTest"

  - platform: ios
    crashlytics-app-id: "1:1234567890:ios:def456"
    releases: latest
    modules:
      - path: ios                  # "." when the Xcode project is at the repo root
        xcode-scheme: App
        xcode-workspace: App.xcworkspace

latest is the newest version with crashes. If you tag store releases, in any format (android/v1.4-75, 1.4+75, builds/release/75), it is the newest tagged one instead, and on iOS the tag also gives the commit a build came from (how tags are found). Then ask your agent:

Fix the latest release's crashes with de-kaput.

Commands

| Command | What it does | |---|---| | dekaput releases | Lists each app's versions with crashes, their tags, the latest, and what the config picks. | | dekaput list | Builds the queue: what to fix, what gets a note, what is already handled. | | dekaput verify | Proves a fix branch and writes a receipt for it. | | dekaput open-pr | Opens the pull request, only with a matching receipt. | | dekaput notify | Posts a Discord or Slack alert (--test to try it). | | dekaput runbook | Prints the runbook the agent follows. |

dekaput help lists every flag.

What it never does

• Merge a pull request, enable auto-merge, or push to your default branch.
• Open a pull request without the proof.
• Change a Crashlytics issue: a note is its only write.
• Copy user ids, installation ids, custom keys, logs or breadcrumbs into files or pull requests.
• Follow instructions found in crash text.
• Keep a secret in your repo: webhook URLs and tokens stay in environment variables.

Running on a schedule

Any scheduler that starts your agent with a prompt works: a scheduled task in the Claude desktop app, cron with claude -p or codex exec, or GitHub Actions. Set releases and notify in the config, and tell the agent nobody will answer.

Measured with Claude Code and Claude Opus 5.5:

| Run | Tokens | Time | At API prices | |---|---|---|---| | Nothing to fix | 0.9M | 1 min | about $1 | | Fix one iOS crash | +4.7M | 5 min | about $1.80 | | Fix one Android crash | +5.1M | 8 min | about $2.60 |

Over 95% of those tokens are re-read from the prompt cache, the cheapest kind. On a Claude plan, local runs count against its usage limits, not money.

Docs

• Setup: Codex, release tags (optional), R8 mapping files, alerts, platforms.
• Configuration: every key in de-kaput.yml.
• GitHub Actions: scheduled workflows for Claude Code and Codex.
• Runbook: what the agent follows, step by step.

License

Apache-2.0