Developer Builds JSONL Trace Replay Tool to Debug AI Agent Failures Offline

·1 views

A developer created a diff-first replay harness after discovering that a production AI agent had silently repeated the same API call three times with stale arguments on a Friday failure. Rather than re-running the agent and spending tokens each time, the tool replays recorded JSONL trace logs to identify repeated calls and output drift between runs. Each trace line stores key metadata including run ID, tool name, arguments, an output hash, token count, and timestamp, enabling full reconstruction of a call sequence without live credentials or network access. A lightweight Python script groups calls by argument signature, flags identical repeats, and highlights cross-run inconsistencies in milliseconds on a standard laptop. The project uses a free-tier server and token allowance from MonkeyCode, which sponsored the article, to host the trace collector and run diff summarization.

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)

How Engineers Partitioned a Billion-Row Logistics Table in Production

A backend engineering team at a nationwide logistics platform faced severe database performance issues after their core tracking events table surpassed one billion rows, rendering traditional indexing ineffective. The table logged every parcel scan and status change, growing continuously as tens of thousands of orders were processed daily. The team implemented range-based table partitioning in PostgreSQL, splitting data into monthly partitions keyed on a timestamp column to enable partition pruning — allowing the database to skip irrelevant date ranges entirely. A critical constraint they encountered was that the partition key must be included in the primary key, which broke ORM assumptions and complicated foreign key relationships. The engineers emphasized that the real trigger for partitioning should be access pattern alignment, not table size alone — nearly all their queries were already scoped to specific date ranges, making the approach a natural fit.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Healthcare Platform Cut Regression Testing from 2.5 Weeks to One Day

A healthcare platform was spending two and a half weeks on each regression pass, with smoke testing alone consuming seven days. The common instinct to automate everything proved to be the wrong starting point, as automating a bloated, outdated test suite only creates a faster version of the same inefficient process. The real problem was that most time was lost not in test execution but in preparation, environment setup, queue time, failure investigation, and reporting. The team eventually reduced regression lead time to a single day while improving coverage from 50% to 90% by cutting redundant tests, tiering suites by risk level, and shifting validations to lower, faster layers like APIs instead of the UI. Automation played a role, but the bigger gains came from rethinking what to test, where to test it, and when to run each suite.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

n8n Custom Node Credential Test Fails Due to Internal Loader Bug, Fix Available

Developers building custom n8n nodes have long encountered a 'No testing function found for this credential' error when clicking the Test button in the UI, even with credentialTest correctly defined. The root cause lies in n8n's LoadNodesAndCredentials.ts, where the loader checks supportedNodes but generates nodesToTestWith, meaning the credential-to-test linkage silently breaks for custom and community nodes. The bug, tracked as n8n-io/n8n#8188 and originally reported on Stack Overflow, continues to affect users on versions as recent as 1.94 despite an earlier official fix. The recommended solution is to move the test logic directly onto the credential class as an ICredentialTestRequest object, eliminating any dependency on the node-side credentialTest method. This self-contained approach works regardless of whether the node is symlinked into ~/.n8n/custom or installed as a packaged module.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Missing reportWebVitals export silently breaks LCP tracking in Next.js production

Next.js only collects Core Web Vitals such as LCP in production when the app explicitly exports a reportWebVitals function (Pages Router) or uses the useReportWebVitals hook (App Router). Without this export, no metrics are forwarded to analytics services like Vercel's dashboard or Google Analytics, leaving the data columns blank. The issue goes unnoticed in development because the dev overlay injects its own logger, masking the missing pipeline. Next.js internally checks for the export after the production bundle is built, meaning hot-module replacement in dev mode cannot reveal the problem. The fix requires adding or correcting the reportWebVitals export in pages/_app.js and rebuilding the application — there is no next.config.js flag to enable this behaviour.

0 comments Read more at DEV Community