Comments URL: https://news.ycombinator.com/item?id=47643058

Points: 2

# Comments: 0

Starting to run local inference has highlighted something I've been aware for longer: just running tests output shedloads of text into the context window that is there for good until compaction or starting afresh. For example, a single `cargo test` dumping 8KB into the agent's context just to communicate "47 test passed." The agent reads all of it, learns nothing useful, and the context window fills with noise. Makes LLM prefill slower as well as costs more when using per token APIs.

I created a small program that sits between the command output and the LLM: oo, or double-o ... yes, sad play on words. Double-o, the agent's best friend :)

oo wraps commands and classifies their output:

  - Small output (<4KB): passes through unchanged
  - Known success pattern: one-line summary
    (oo cargo test →  cargo test (47 passed, 2.1s))
  - Failure: filtered to actionable errors
  - Large unknown output: indexed locally, queryable via oo recall

Think pytest only outputting OK if all tests pass, but in case of error getting the necessary error output only.

It is a cli, basic use for agents:

oo pytest oo cargo test etc…

Custom patterns configurable and there is also learn -flag to use Anthropic models to have LLM write the configuration for your custom command.

Nothing too complicated or special, but works in my LLM workflow saving on token use & speeding models!

Have a great weekend!

Comments URL: https://news.ycombinator.com/item?id=47464325

Points: 1

# Comments: 1

Literally all my code has been ”ghostwritten” for the past 18 months. Does not sound like something enterprise customers would like to hear and try to understand what it means.

I am still an avid user of opencode, my own fork though with async tools etc, but it is cumbersome and tries to do too many things.

The problem it solves: agents needing historical knowledge of your project or other tasks that other agents are working on. Or just getting context transferred from session to session in clean and semantically searchable way. Most solutions require an API, a cloud service, or a running server. vipune is a single binary, no daemon, no keys, works offline.

I had a predecessor to vipune in my personal agentic flow for a long time, but it got bloated and decided on a rewrite and thought to share, maybe someone else will find it useful also.

And yes, this is written 99% by LLM agents, 100% monitored by yours truly.

It also does conflict detection — if you try to add something semantically similar to an existing memory, it flags it instead of silently duplicating.

Built in Rust, Apache-2.0. Binaries for macOS ARM64 and Linux. Early release — would love feedback on the use cases people try it for.

Comments URL: https://news.ycombinator.com/item?id=47094213

Points: 1

# Comments: 0

Strix Halo

I’ve been an active user of opencode for 7-8 months now, really like the tool, but beginning to get a feeling that the core team’s idea of keeping the core development to themselves is not going to scale any longer.

Really loving opencode though!

Maybe should try to give tickets a go. gh cli does add another HTTP layer and slows things down - feels silly to be paying for Cerebras if one is slowed down by other tooling.