Articles
- Go-live checklist for software: what to check before launch
What a go-live check of a whole project should cover: code and architecture, security and data, and whether the main flows work. And how we automated it.
25 October 2026
- Automated security audits of the whole codebase, not just the diff
Security checks on pull requests only see changed code. Why we added weekly whole-project security audits, and why only critical and high become tasks.
25 October 2026
- One repository, several projects: catching breaking changes in shared code
A billing engine or library shared by several products breaks the others quietly. How CodeLynq checks every project using a shared repository at PR time.
25 October 2026
- Architecture rules an AI coding agent cannot quietly break
An AI agent takes the shortest import, not the one your architecture intends. How we map the code, check each change against your rules and show it in the PR.
25 October 2026
- Routing coding tasks across Claude Code, Codex and Grok
Three coding agents, one pipeline. When a run may leave Claude for a cheaper agent, when it never does, and why every run must say which model it got and why.
25 October 2026
- Approving AI spend without approving every run
Approve every AI run and approval becomes a reflex; approve none and a top-up is gone in an afternoon. How we hold only the runs that look expensive.
25 October 2026
- One fix, several release branches: applying a change to every version you maintain
How CodeLynq applies one task to every release branch you maintain: one checkout and one pull request per branch, adapted to what each branch contains.
25 October 2026
- Debugging production with an AI agent that can only read
How we let an AI agent investigate a live environment through logs, SQL and shell commands without being able to change anything, and where the limits are.
25 October 2026
- From monitoring alert to pull request: auto-remediation with guard rails
How CodeLynq turns a monitoring alert into a deduplicated bug task and, only when every guard passes, a pull request. Plus the bug that made one guard useless.
25 October 2026
- How to turn meeting notes into a groomed backlog with AI
From meeting notes or a Fireflies transcript to draft tasks, then AI estimates, reasoning and a ranking that a person reviews before anything is saved.
25 October 2026
- The questions an AI should ask before it writes a spec
Why an AI coding tool should scope a task and confirm its assumptions like a tech lead before it writes a spec, and how it should know when to stop asking.
25 October 2026
- How to request changes to an AI pull request without starting over
Feedback on an open AI pull request becomes a revised spec version, approved and built on the same branch. And what happens once the PR is already merged.
25 October 2026
- Letting an AI assistant drive your delivery pipeline over MCP
How CodeLynq exposes tasks, specs, runs and deploys over MCP with OAuth 2.1, how scopes limit an assistant, and two security bugs we found in our own review.
25 October 2026
- When the AI fails silently: provider errors that look like empty results
How an AI provider outage reached our customers as a blank form and a bare "Exit code 1", and what we built so failures say what they are.
25 October 2026
- Why our first AI coding runs failed on setup, not on the AI
Our first five production runs all failed, none because of the AI: a branch not called main, tokens missing rights. How we now catch that in seconds.
25 October 2026
- Devin alternatives: how to choose an autonomous coding agent
The serious Devin alternatives compared on what matters to a team: who decides what gets built, where the code runs, which model writes it and how you pay.
17 October 2026
- What is spec-driven development? A practical guide with an example
Spec-driven development with an AI coding agent, step by step: what a good spec contains, a real example, who approves it and the mistakes to avoid.
16 October 2026
- The second month should be better than the first
Context windows are not memory. What an AI coding tool should retain between runs, what it must refuse to retain, and why that decides whether it improves.
12 September 2026
- When the code is not allowed to leave
When a contract says source stays on your infrastructure: how a self-hosted runner moves AI execution to your machines without you operating the product.
11 September 2026
- Your PR comments are untrusted input
If an AI agent acts on reviewer comments, anyone who can comment on the pull request can instruct it. What that means, and what we changed.
10 September 2026
- The merge button stays yours
We could have built auto-merge. Here is the argument we kept losing with ourselves, and what we built instead so that not merging is not a bottleneck.
8 September 2026
- Paying Opus prices to bump a dependency
One model for every task is the fastest way to burn an AI budget. How we route per task, why the router must explain itself, and what a spending cap should do.
6 September 2026
- Four checks an AI run should pass before you see it
A pull request opened the moment the model stops typing puts the review burden on a person. The four gates an AI run should pass first.
4 September 2026
- Reviewer feedback should only be given once
Most AI coding tools forget the correction as soon as the pull request merges, so the same note gets written again next week. Here is what we changed.
1 September 2026