SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Building an AI bug-fixing agent that runs inside your own GitHub Actions runner

0
·5 views

The author describes an AI agent that fixes production errors using a design where the coordinating service never has access to the codebase. The control plane is a GitHub App limited to metadata and workflow dispatch, so it can start a job in the user's repository but cannot read the code. The agent runs on the user's own GitHub Actions runner, which holds the code and the AI API key. A per-run, short-lived HMAC token scoped to one run lets the runner fetch error context and report results. The agent works through a small set of tools, including reading files, searching code, and opening a pull request, with no shell or network access.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

LedgerLens: Open-Source, Privacy-First AI Expense Auditor Built for Freelancers

A developer has released LedgerLens, an open-source expense auditing tool aimed at freelancers and solo business owners. The software automatically categorizes invoices and receipts, flags duplicate charges and calculation errors, and identifies potential tax compliance issues. It is built with FastAPI and runs Gemma 2 through Ollama, allowing local processing so sensitive client financial data need not go to third-party cloud services. The creator positions it as a low-cost alternative to subscription accounting platforms that charge $35 to $70 per month. The project is deployed using a Render Blueprint and is available on GitHub.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

A Practical Framework for Writing Email Sequences to Niche Audiences

Generic email sequences tend to underperform in niche markets because specialized audiences quickly notice when a sender lacks real understanding of their work. The framework starts by collecting 20 to 30 authentic phrases from customer replies, sales calls, competitor reviews and niche communities, then reusing that language in subject lines and openings. Each sequence should be built around one clear goal, such as a demo booking or completed setup, and a typical five-email welcome or nurture series runs about two weeks. The guide outlines a timed structure covering delivery and expectations, problem naming, a useful lesson, social proof and a final clear request. Keeping segments small and relevant, and only emailing people who have opted in, rounds out the approach.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

New UK Companies House API tool cleans addresses, flags directors and extracts balance sheet data

A developer has released a service that works with UK Companies House data to address common frustrations such as rate-limit errors and inconsistent records. The tool normalises raw addresses into standard formatting and attaches verified GPS coordinates. It also sorts board officers into active and resigned categories for know-your-business checks. Filing PDFs can be scanned to pull balance sheet figures into structured JSON, and a proxy allows official PDFs to be displayed in web portals. A live sandbox is available without an API key, and an open-source Laravel wrapper has been published on Packagist.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Shipment Reminders Need Durable Idempotency, Not Just 5-Minute Queue Deduplication

A developer argues that a shipment reminder system should assume at-least-once queue delivery and rely on a durable idempotency key before sending, since queue deduplication only covers a 5-minute window. Acknowledging messages only after a send is safely persisted prevents lost notifications, but it can cause duplicate emails or SMS messages if a worker crashes after the provider accepts a send and before the completion is recorded. The piece recommends an atomic claim on a logical key such as user_id, reminder_id, and scheduled_at, with duplicates acknowledged without resending. It also suggests alerting on mismatches between delivery attempts and unique send keys, and sizing the key store for the full retry horizon.

0 comments Read more at DEV Community