Rykan V™

Control what your AI can do

[Rykan V™ @rykan_v]

Link to our telegram channel: https://t.co/v2r4RpEMBh

[Rykan V™ @rykan_v]

@kevintwohy try dm him, like this comment if he doesnt reply

[Christopher Karani @ckarani7]

I kept hitting model failures on phones that I couldn't reproduce in the office. LangFuse didn't catch it. LangSmith didn't catch it. So I built the thing I wanted: a flight recorder for AI running on real devices.

https://t.co/JDhV7s4RZK

[Christopher Karani @ckarani7]

11x faster hooks
13x faster denies
3x smaller binary

ryk 0.2.19

[Rykan V™ @rykan_v]

it’s a pandemic, your memory is what’s dying

[Christopher Karani @ckarani7]

@irangareddy
Hey — saw your post about the agent wiping the DB / running destructive commands

We’ve built Rykan V specifically for this: policy guardrails that block dangerous shell/file actions before they execute.
You can also run ryk scan to see every dangerous command your agent has run in past sessions.

Open-source here if you want to try or star it: https://t.co/Q1GQ3HqtPZ

[Rykan V™ @rykan_v]

@pauliusztin_ next step is being able to control what the agent can do, we're solving that at https://t.co/sucpUzQZrs, free oss. try it out let me know what you think

[Rykan V™ @rykan_v]

Open ryk allowlist for the TUI of what this session can run. Linear dump: ryk allowlist --plain.

[Rykan V™ @rykan_v]

@LarryPanozzo Listing the untracked export under fixes completed is the tell. It already knew Git could not get that file back.

[Rykan V™ @rykan_v]

@TambaClan Git clean -fX on an ignored data dir ate the whole tree. The scoped delete never stayed scoped.

[Rykan V™ @rykan_v]

Run ryk redteam --ci. Same fixtures the CI job uses, same deny path a live session hits.

[Rykan V™ @rykan_v]

@TimeToBuildBob The 65% yes on npm run analyze is why confirmation is a log.

[Rykan V™ @rykan_v]

@LesOttolenghi Who gets to tell them no is the right question. Slack after the agent already wrote is not a no.

[Rykan V™ @rykan_v]

@arcee_ai Isolated workers help until the episode is a lie. If the worker exits after a half-write, the orchestrator just schedules the next half-write.

[Rykan V™ @rykan_v]

@bally_kehal Making the loop a plugin is the interesting part. Most teams will swap models and keep the same broken retry policy.

[Rykan V™ @rykan_v]

@eliebakouch Appending the modification is the right KV trick. Still need something besides the agent to confirm the file actually changed.

[Rykan V™ @rykan_v]

@AraneaDev The placeholder is the easy half. The proxy has to be the only thing that can unmask, or a second tool just reprints the token.

[Rykan V™ @rykan_v]

@ClaudeCodeLog Fork inherit is the useful default. A fresh-context spawn still needs a checkpoint if the parent dies mid-run.

[Rykan V™ @rykan_v]

@sCuz123 Token scope dies the moment the agent reads a second file. Gate the irreversible write, not the first secret it was handed.

[Rykan V™ @rykan_v]

@nykdotdev A clean first run is a demo. The graph only counts if a crashed subagent resumes from checkpointed state without rewriting files it already finished.

[Rykan V™ @rykan_v]

When the session is done, ryk stop. Leaves the policy log on disk so ryk replay still works.

[Rykan V™ @rykan_v]

@tryadaline Named regression buckets are the useful part. Shipping prompt or halt changes against a bug list just churns. Grade the change on the named cluster or you are guessing.