The self-healing loop

The self-healing support loop

The Dev Server plan connects TicketMan AI to agentic coding models running on a separate, isolated dev server. It turns a crash into a shipped fix while keeping a human wherever you want one. The loop has six stages.


1. Error captured

The Diagnostics SDK catches an unhandled crash or a qualifying HTTP error in your application and forwards the full context: exception type and stack, breadcrumbs, environment, and runtime state. Nothing sensitive leaves unmasked; secrets are redacted before a report is ever written.

2. Grouped and filed

TicketMan AI receives the report, computes a fingerprint, and groups repeat occurrences into a single incident. The incident is filed as a structured, pre-categorized ticket with priority and context already attached, so no one triages the same crash twice.

3. AI root-cause triage

The platform's LLM layer reads the trace, the grouped history and the available context and proposes a likely root cause and an area of code to look at. This uses the same bring-your-own-model subsystem as the rest of the platform, so it can run on a local model if you require it. See AI and models.

4. Agentic dev server

The triaged ticket is handed to an agentic coding model, such as Claude Code or Codex, running on a dedicated dev server with access to the relevant repository. The agent reproduces the issue and works the fix in isolation from your production environment.

5. Fix, pull request, review

The agent commits a fix and opens a pull request. From here you choose the level of autonomy per repository:

• Operator approval required: TicketMan AI asks for a go-ahead before the agent acts, and again before the PR is opened.
• Human review: the PR waits for an engineer to review and merge, exactly like any other contribution.
• Auto-approve: trusted, well-scoped changes can merge automatically.

If the agent needs more information, it loops back for another pass rather than guessing.

6. CI/CD ships it

Your existing pipeline builds, tests and delivers the merged change. The incident is marked resolved, and the crash that started the loop is gone. The loop closes.

You decide how far it runs

Every gate in the loop is configurable. You can run the whole thing on autopilot for a low-risk service, and require operator approval at every step for a critical one. The default is conservative: a human approves before any code is written.

What you need

• TicketMan AI self-hosted, with the Diagnostics SDK reporting from your app.
• A dev server the agentic model can run on, with access to the target repository.
• A CI/CD pipeline that ships merged pull requests.

Want to see the loop in action? Book a demo from the home page.