Building redaction-first tracing for coding agents, and why unknowns must never read as zero

·6 views

The author found that existing LLM observability tools like MLflow, Langfuse and LangSmith record spans well but do not fit long-running coding-agent sessions, which generate tree-structured tool calls and sensitive transcripts. Instead of adopting one of those stacks, they extended their own harness, code-copilot-team, with an analytics layer built on a session, turn and tool-call data model. The store holds 65 real sessions and excludes 32 noise sessions, a count shown on every page so the denominator stays honest. Redaction is applied at ingest through three modes (none, code, metadata-only), and the judge path uses the stricter of the recorded and current modes. A review finding showed that truncating before redacting let a secret straddling a code fence leak, so the fix redacts the full input first and a test now covers that boundary case.

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)

Marketing Agencies Face a 2030 Shift as AI Commoditizes Execution Work

Agencies have long billed clients by the hour for creative and analytical output, but AI tools now produce much of that work in seconds, undermining the traditional model. Producing generic AI content in larger volumes would only intensify price competition, and many brands are moving routine production in-house. The article argues that lasting value lies in defining the real business problem, setting the creative standard, verifying outputs, and building client trust. It predicts the industry will split between large automated platforms and small, high-caliber advisory teams, leaving little room for the middle.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

COBOL Phoenix Translates Legacy COBOL to Python and Verifies Behavior Against GnuCOBOL

A developer built COBOL Phoenix, an open-source tool that converts COBOL programs into Python while checking that the output matches the original. The project includes a custom COBOL front end that flags unsupported constructs rather than silently dropping them, and a runtime that reproduces COBOL's decimal handling, rounding, and sign behavior. Verification compares results against GnuCOBOL using generated test inputs and mutation analysis to confirm the test suite catches injected faults. Reported results include 127 of 127 sample cases matching GnuCOBOL, 180 of 180 on an independent insurance underwriting program, and 20 programs parsed without errors. The project was created for the Kiroween 2026 hackathon and is available under the Apache-2.0 license.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

VigilShield: $19 Windows tool blocks connections to known attacker servers

A school teacher from Narowal, Pakistan has built VigilShield, a one-time $19 guard for Windows 10 and 11 that monitors outbound network connections. It checks each connection against threat lists drawn from public feeds such as abuse.ch and FireHOL, and blocks matches through Windows Firewall with an alert. The tool also includes ransomware canary files, alerts for suspicious programs, and USB, startup and hosts-file guards. The developer states it is not an antivirus replacement, cannot predict attacks, and does not stop every attack, and that it runs locally without sending data elsewhere.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Lambda Powertools TypeScript adds structured logs, custom metrics and traces

Lambda Powertools for TypeScript is an open source AWS toolkit that removes the need to write repetitive observability helpers in every function. The article walks through a small fictional bike rental API built with the Serverless Framework and tested locally with Vitest. It uses three packages, Logger, Metrics and Tracer, plugged into handlers through the Middy middleware engine. The author notes that consistent JSON logs with correlation IDs and custom CloudWatch metrics are the main benefits, while high-cardinality metric dimensions can drive up costs. The guide covers setup, correlation IDs, choosing between dimensions, metadata and X-Ray annotations, and the impact on cold starts and billing.

0 comments Read more at DEV Community