Startup Builds Git-Backed Documentation System After Losing Confluence Data

·2 views

A startup founder lost years of system architecture documentation when a Confluence space was silently deleted, prompting a fundamental rethink of how technical knowledge is stored and owned. The team adopted a 'Documentation-as-Contract' philosophy, requiring engineers to write structured markdown files directly within the code repository before any business logic is developed. These git-backed documents serve as the sole, high-fidelity context fed to AI tools, which the team uses strictly as an architectural advisor rather than a code generator. The pipeline is automated through GitLab CI/CD and Google Workspace to create a managed RAG (Retrieval-Augmented Generation) workflow with minimal overhead. The approach is designed to keep human intent central to engineering decisions while preventing AI from producing architecturally incoherent, low-quality code.

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)

Fastify test suite had passing tests that could never fail, exposing a coverage blind spot

A developer discovered that several tests in the Fastify Node.js project contained assertions like `assert.ok(true)` that would pass regardless of whether the code under test actually worked. One test claimed to verify a server had started, but deleting the server logic entirely left the test green, since the only assertion was unconditionally true. The issue highlights a broader problem where code coverage metrics count executed lines without checking whether those lines would catch real regressions. The developer opened a pull request on the Fastify repository to replace the hollow assertions with meaningful checks, such as verifying the server is actively listening. To help teams detect similar issues, they also released a command-line scanner called `margyn-scan` that flags tests whose assertions can never fail.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Free AI Tokens Still Carry Hidden Ops Costs for Batch Workloads

Platform teams using free AI model endpoints often overlook operational costs that emerge even when token pricing is zero, according to a technical cost drill published by MonkeyCode. When token fees disappear, the real bill shifts to engineer time, retry overhead, and queue age — all of which can quietly erode a project deadline. The drill uses a minimal single-threaded Python worker to process a batch queue and track metrics like queue age ratio and deadline slack in a running CSV ledger. Two key alert thresholds are proposed: a queue age ratio above 10% and a deadline slack turning negative, either of which should trigger a switch to a paid, SLA-backed endpoint. The approach is not suited for regulated data or jobs with strict latency SLOs, where paid infrastructure remains the appropriate choice.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Five Defensive Coding Gates to Catch Silent Failures in Free AI Server Setups

A developer testing MonkeyCode's free log-summarizing API deliberately crashed their own server to expose where silent failures occur in free AI infrastructure. During the experiment, the server process died with no restart policy, no health probe, and no alerts, yet the AI model remained online for four hours, masking the outage entirely. To address this, the author outlines five progressive safeguards: a file-based kill switch, a self-managed daily token budget counter, a hard request timeout, an API response shape validator, and a lightweight health endpoint monitored by an external uptime checker. Each gate is designed to make failures loud and explicit rather than allowing them to propagate silently to end users. The article notes it was prepared as part of MonkeyCode's product outreach.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Knowledge Graphs Turn Simple AI Agents Into Persistent AI Employees

Most AI agents follow a basic think-act-observe loop suited only for short-lived tasks, but building a true AI employee requires persistent memory of ownership, history, and pending actions across days or weeks. A knowledge graph architecture addresses this by storing state, relationships, and context that the agent reads each time it wakes up. Using a minimal TypeScript implementation, the article demonstrates how nodes and edges can represent an AI employee like 'Maya' who owns sales leads and responds to events such as email replies. When a lead's status changes, an event identifies the responsible employee, triggers the agent loop, and writes results back to the graph before the agent sleeps again. Scheduling future events — such as a follow-up reminder — allows the agent to remain dormant until needed, making the system practical for real-world, long-running workflows.

0 comments Read more at DEV Community