Open Agent Hackathon 2026

Reverse-engineer any codebase.
Know why every line exists.

siddarth-ba-72/Interview-Prep-Pilot-AIreading the code…

ai-service/app/llm.py

23def get_client() -> AsyncOpenAI:
24 global _client
25 if _client is None:
26 if settings.llm_provider == "gemini":→ Gemini
27 _client = AsyncOpenAI(
28 api_key=_required(settings.gemini_api_key, ...),
29 base_url=settings.gemini_base_url,
30 timeout=settings.llm_timeout_seconds,
31 )
32 else:
33 _client = AsyncOpenAI(
34 api_key=_required(settings.openai_api_key, ...),
35 timeout=settings.llm_timeout_seconds,
36 )
37 return _client
application-dev.yml:5llm.py:26llm.py:107Gatewayspring cloud gatewayAI serviceai-service · fastapiGemini APImodel providerOpenAI APImodel provider

0%

of a developer's day goes to understanding code.

Xia et al., IEEE TSE 2018

Commit68a5228

Add ky.retry() to force retries from afterResponse hooks

Pull request#779

Add ky.retry() to force retries

Issue#778

Simplify forcing a retry in afterResponse hook?

Discussion#778

4 comments: sholladay, sindresorhus, MrSquaare

source/core/constants.ts

263export const retry = (options) => new RetryMarker(options);

?

The reason exists. It's just scattered.

How it works

01Paste a repo

Any public GitHub repository.

02Agents take it apart

Readers work in parallel; a Verifier checks every quote.

03A map you can trust

Every arrow cites the line that proves it.

04Ask why

Click an arrow. Get the commit, PR and issue behind it.

05See what breaks

Pick a part. Everything that depends on it lights up.

06Follow it, keep it true

Play a request end to end. Every push is re-checked.

The proof

Built on Zetaris, Meterless and NVIDIA.

01 / 04

Zetaris

Build a USL · query it · add a DQ rule

Every line, commit, pull request and issue in one semantic layer.

1 · Nine tables, one provenance layer

2 · One query from a line to its reason

3 · A rule that scores explainability

02 / 04

Meterless H-MEM

An engine in a working feature

Cartograph's memory: it keeps approved answers, and knows when they go stale.

1 · An approved answer becomes a memory

2 · The next question starts from it

3 · Code changes, the memory is set aside

03 / 04

Meterless Runtime

Primary AI provider

Paste Nightshift's connect link and every agent runs through it. Without it, NVIDIA answers directly.

1 · One paste connects it

2 · Agent → Nightshift → Nemotron

3 · Runtime offline, NVIDIA steps in

04 / 04

NVIDIA Nemotron

Nemotron in every agent

Ultra plans and judges. Lightning reads and writes. Every call is checked.

1 · Every agent, its model

2 · Two tiers, every call logged

3 · Output is validated, never trusted

Draw your codebase.

owner/repo or a github.com link · a map in a few minutes