Ferax564

Building codixing (AI code retrieval for agents) + ezkeel & noma | Rust open source | Helping devs & fintech teams search massive codebases faster, Links in bio

[Ferax564 @Arref000]

@jradoff drag-select comments on raw .md is a nice wedge. how are you keeping comment anchors stable when the markdown reflows?

[Ferax564 @Arref000]

@StuSim the weird part is each memory type wants a different shelf. markdown for rules, transcripts for evidence, code for repeatability.

[Ferax564 @Arref000]

@lais_bsc pain-driven docs are sadly the only ones that survive. everything else turns into rules nobody believes yet.

[Ferax564 @Arref000]

@Liebs00 trading bot is exactly where i'd want a kill switch before i trust the deploy. Claude making it easy is both the fun and scary part.

[Ferax564 @Arref000]

@Kavyabuildss dev/staging/prod sharing tokens is the one that catches real people. not glamorous, just the first thing i check now.

[Ferax564 @Arref000]

@emmett_lu auth, billing, deploy scripts is exactly where i stop trusting "it works locally". small blast radius first, then let the agent move faster.

[Ferax564 @Arref000]

@Devsan282 makes sense for dashboard state. i still like deploy output being readable on disk, even if the DB owns the decisions.

[Ferax564 @Arref000]

@SifuBacktest 80-page dashboard drift is where repo search starts to matter. page 81 needs shared patterns before it invents another local solution.

[Ferax564 @Arref000]

@iammukeshm project context first is a good habit. the failure mode i watch is when it explains last month's architecture too confidently.

[Ferax564 @Arref000]

@Yamatoeth full-fidelity timeline is exactly the shape agents need. partial logs make them sound certain while missing the one browser event that mattered.

[Ferax564 @Arref000]

@RintaroOkabe03 the tally trick is smart. making tool avoidance visible usually works better than another "please use the search tool" rule.

[Ferax564 @Arref000]

@d4m1n yeah, bigger ctx buys time. i still find the 60% line is where bad retrieval turns into expensive confidence.

[Ferax564 @Arref000]

deploy dashboards are nice until the dashboard is the only place that knows what happened.

i want the VPS to keep readable deploy state: Docker/Caddy/envs, health, rollback.

that's why EZKeel is staying CLI-first: repo -> your box -> files you can inspect.

[Ferax564 @Arref000]

1M context windows make brute force cheaper, but they don't make noisy repo context good.

the useful question is still: which files, callers, tests, and configs explain this change?

Codixing is my attempt to make that answer small before the agent edits.

[Ferax564 @Arref000]

@Dustyboy23 checking cron for subprocess claude vs SDK init is a good habit. usage attribution gets weird fast once agents run in the background.

[Ferax564 @Arref000]

@AndrewLeeWard 200-line CLAUDE.md is the new unread handbook. the best repo instructions feel closer to guardrails than a company wiki.

[Ferax564 @Arref000]

@agrimsingh requires_ocr is exactly the kind of boring flag agents need. empty text looks valid enough to poison every later step.

[Ferax564 @Arref000]

@jilspatel21 same stack, different ops appetite is the right frame. for small apps i like the option you can debug without opening four consoles.

[Ferax564 @Arref000]

@intertwineai continuity is the missing part. MCP gives hands, but the useful jump is remembering why the last session chose a path.