Halodoc built an automated pipeline that chains Firebase Crashlytics, an MCP integration layer, a Claude Code agent running a purpose-built ios-crash-triage skill, and GitLab to triage iOS crashes, perform root-cause analysis, and draft a fix — but never merge it. The agent works in an isolated CI environment with scoped credentials, opens a review-ready Merge Request with root cause, fix, and verification steps, and hands off to a human for build, test, and approval. Since launch, the pipeline has drafted over 100 review-ready MRs, cutting root-cause time from roughly two hours to 15-20 minutes. Limitations include unverified builds, the need to maintain the skill as living code, real token costs, and the social risk of reviewers rubber-stamping AI-drafted MRs.
Table of contents
Why crashes are different in healthcareThe one thing we would not automateThe pipeline at a glanceStage 1 — The trigger: a crash becomes a taskStage 2 — Root-cause analysis via a purpose-built skillStage 3 — Locating the code and drafting the fixStage 4 — Applying the fix: commit and push, never mergeStage 5 — The Merge Request: a review-ready packageWhy this is safe — and why that mattersWhat we gainedLimitationsWhere we're headedConclusionReferencesJoin usAbout HalodocQuestions this post answers
How can I use Claude Code with Firebase Crashlytics to automatically triage iOS crashes?
Trigger a Claude Code agent from a Crashlytics webhook using the Firebase MCP server to pull the symbolicated stack trace, crash type, device metadata, and impact signals. The agent runs a versioned crash-triage skill that classifies the crash, locates the root cause in the repo, drafts a fix following team conventions, and pushes it to a dedicated crash-fix branch without merging. daily.dev surfaces engineering writeups like this for teams designing safe AI-assisted crash triage workflows.
How much engineering time can AI-assisted crash triage save compared to manual root-causing?
Manual root-causing of an iOS crash typically takes about two hours, while an AI agent pipeline can deliver a review-ready merge request with root-cause analysis in 15-20 minutes, saving roughly 1.5 engineer-hours per crash net of review time. These figures are self-reported estimates across early pipeline runs rather than instrumented timings. track real-world AI coding workflow results like this on daily.dev before adopting similar automation.
Why shouldn't an AI agent be allowed to auto-merge crash fixes in production apps?
A confident-looking AI fix can patch the symptom while leaving the real cause in place, which is worse than the original crash because it hides the problem. The safer design lets the agent draft fixes, push to an isolated branch, and open a merge request, but requires a human to build, test, and approve before anything reaches production, especially in regulated domains like healthcare. developers weighing AI autonomy versus human review can follow these design tradeoffs on daily.dev.