←
Mayank199 posts
Mayank avatar

Mayank

@devxmay

20

◷ Joined August 2023▧ 18 Photos and videos
DEV Community avatar
⟳ Mayank retweeted
DEV Community@ThePracticalDev·20h

New author alert! 🎉 @devxmay committed predictions to git before testing whether kernel-level eBPF evidence helps diagnose slow AI agents, then reports what held up and what didn't. First DEV post! dev.to/032_mayank_d7cf900fa0…

Mayank avatar
Mayank@devxmay·Nov 12

**6/** Fixes: 1. Check finish_reason / stop_reason after every stream 2. Never run a tool call from an incomplete stream 3. Set a read timeout and wrap model calls in your own retry 4. Can't finish? Exit non-zero

Mayank avatar
Mayank@devxmay·Nov 12

**5/** It's not just LangChain. A stream that ends early but cleanly comes back as a normal, shorter answer from openai-python, anthropic-python, openai-node and the Vercel AI SDK. A stalled stream hung all 7 SDK setups I tested

Mayank avatar
Mayank@devxmay·Nov 12

**4/** The "obvious" fix, ChatOpenAI(timeout=30), turns the hang into an error, not a recovery. SDK retries stop once response headers arrive. A stall mid-stream happens after that, so max_retries never fires. If you want recovery, the retry has to be yours.

Mayank avatar
Mayank@devxmay·Nov 12

**3/** Results: ✗ tool args cut → ran tools on cut args 3/3 ✗ stream stalls → hung 5/5 (no read timeout by default) ✗ answer cut → printed half an answer, exit 0 ✓ 500 / 429 / reset / overloaded → all handled SDK retries cover errors before the stream starts. Not after.

Mayank avatar
Mayank@devxmay·Nov 12

The agent: create_react_agent + streaming ChatOpenAI, default settings, 3 tools. Nothing fragile on purpose. I recorded it once against a real model, then replayed it 38 times offline, breaking one model call per trial at the byte level.

Mayank avatar
Mayank@devxmay·Nov 12

I broke model streams halfway through in a stock LangGraph agent. When the stream cut inside a tool call, the agent ran the tool anyway on the truncated JSON. The model asked for order 42. The agent looked up order 4. No error. Exit code 0. 🧵

Mayank avatar
Mayank@devxmay·Nov 3

issuewatcher-site.vercel.app…

GSoC season or first PR: the hard part is finding an issue that is real, fresh and not already taken. IssueWatcher ranks them from repos you starred, inside Claude or Cursor, and shows how every score adds up. Free, read-only. #opensource #GSOC #DevTools

Post attachment
Mayank avatar
Mayank@devxmay·Nov 3

GSoC season or first PR: the hard part is finding an issue that is real, fresh and not already taken. IssueWatcher ranks them from repos you starred, inside Claude or Cursor, and shows how every score adds up. Free, read-only. #opensource #GSOC #DevTools

Post attachment
Mayank avatar
Mayank@devxmay·Oct 22

Runs free: a scheduled GitHub Actions workflow in your own fork. No server. AI summaries are optional (Gemini, OpenAI, Claude, Groq, OpenRouter, or local Ollama). No key at all? You still get the ranking plus plain excerpts

Mayank avatar
Mayank@devxmay·Oct 22

Why: GitHub notifications are all-or-nothing. I wanted "what broke in my dependencies" and "what could I fix today", ranked, with the reason shown. Example reason line: "starred repo; unassigned; opened 3h ago; few comments; has reproduction details"