📰 Dev.to · Mukunda Rao Katta

This is a submission for the Hermes Agent Challenge. I run agents on three different frameworks....

This is a submission for the Hermes Agent Challenge. I spent 20 minutes debugging a 400 error from...

A zero-dep Python lib that tracks cumulative LLM API cost and raises before you go over budget — with per-label breakdown and warn-at callbacks.

A zero-dep Python lib with composable stop conditions for LLM agent loops — turn limits, cost caps, response patterns, and custom predicates that you can OR/AND

A zero-dep Python lib that trims LLM conversation history to fit a token budget — and always keeps tool_use/tool_result pairs together.

A zero-dep Python lib that scrubs PII and provider keys from agent logs before they hit S3, the error tracker, or Slack.

A zero-dep Python lib for agent working memory — keyed notes, lists, counters, prompt injection, and JSONL logging in under 200 lines.

A zero-dep Python lib with a lightweight key-value state store for agent loops — snapshots, diffs, turn history, numeric helpers, and JSON persistence.

A zero-dep Python lib that enforces an allowlist of tool names in an LLM agent loop — raises immediately on blocked tools or filters them silently from the resp

A zero-dep Python lib that puts a hard turn cap on your LLM agent loop — with configurable warning thresholds before the limit hits.

A zero-dep Python lib that trims LLM message lists to a token budget — always keeps the system prompt and your last N turns, drops middle messages oldest-first.

A zero-dep Python lib that extracts JSON, bulleted lists, key-value pairs, and markdown sections from LLM response text — without writing ad-hoc regex every tim

A zero-dep Python lib that strips thinking/scratchpad tags from LLM responses — raw text and Anthropic content blocks — so you can use the reasoning internally

A zero-dep Python lib that hash-pins a system prompt and detects drift — so you know immediately when a prompt changes instead of debugging ghost behavior for h

A zero-dep Python lib that diffs two agent JSONL runs. See exactly what your prompt change moved: tool order, cost, step count, args.

A zero-dep Python lib that logs every LLM tool call and its result to a JSONL file — name, args, result, duration, error — for audit and replay.

A zero-dep Python lib that writes one JSONL record per agent turn so you can resume after a crash instead of starting from scratch.

A zero-dep Python lib that exports JSONL agent traces to CSV — auto-detects common field names for cost, tokens, duration, and error across different agent fram

A zero-dep Python lib that filters JSONL agent trace events by field value — full-text search, exact match, substring, regex, error presence, and invert — all A

A zero-dep Python lib that reads a JSONL agent trace and prints a compact summary — event count, errors, cost, tokens, duration, tools used, lanes.

A reflection on the boring-but-essential agent stack layer and why small single-purpose libraries ship faster and compose better than monoliths.

A pattern-based first line of defense for LLM apps. Five rules, zero runtime deps, deterministic, fast.

Wrap your Nemotron agent in one context manager, get one auditable RunReport that answers the six questions every production owner asks on day one.

An httpx transport plug-in that captures the exact JSON request body sent to Anthropic, OpenAI, or any httpx-based SDK. Credentials redacted by default.