Andrey

curious about building things for humans

[Andrey @onsails]

right-agent 0.3.0 is out!

Most of it went into the parts you don't see until something breaks. Skill learning, dashboards, secrets, gateway recovery. The kind of thing that only matters when an agent runs for days. Showcase below.

[Andrey @onsails]

@SarahLakzit @SemiAnalysis_ ... yet

[Andrey @onsails]

Future work: classifier recalls hindsight memory and/or chat history to determine whether this has been done before.

[Andrey @onsails]

right-agent 0.3.0 is out!

Most of it went into the parts you don't see until something breaks. Skill learning, dashboards, secrets, gateway recovery. The kind of thing that only matters when an agent runs for days. Showcase below.

[Andrey @onsails]

@williewonder_ Small Haiku classifier at the end of every foreground turn. Inputs: user msg, assistant reply, skills already cited, this turn's cost/elapsed vs the agent's P50/P90/P99 baseline. Output: skip / patch existing / create new. Probe-writer only forks on patch or create.

[Andrey @onsails]

Smaller fixes worth flagging:
- OpenShell gateway down at startup used to kill the bot. Stays up and recovers now.
- Silent background failures now emit a diagnostic.
- Idle 1M sessions bloating context. Auto-compact after 2h over 40%.

https://t.co/pig6btsFKs

[Andrey @onsails]

Prompt cache was leaking. Per-turn chat context lived in the system prompt, and Hindsight recall changed it every turn, so the prefix was never byte-stable. Moved that block to a volatile stdin prefix. System prompt now stays stable per agent, cache hits across turns.

[Andrey @onsails]

@shl Hermes → Right Agent (same self improvement loop, natively lives in Telegram)

[Andrey @onsails]

Resume across sessions is just reading the doc. Pick up at first non-done stage.

Repo: https://t.co/bcNRnHlLS8

[Andrey @onsails]

Everyone fights Codex vs Claude Code. I use both for different parts of a big milestone.

Claude Opus 1M context holds the whole thing. Specs, plans, decisions, open questions. All in one place. Including the review.

The actual coding goes to Codex. Headless, in an isolated worktree.

So I shipped codex-sprint to wire it into one sprint flow.

[Andrey @onsails]

Per stage: brainstorm a spec. Write a plan. Isolate a worktree. Codex implements. Headless claude /code-review --fix loops the diff until clean. Verify. Commit. Land. Update the sprint doc. Next stage.

Main context only holds the doc. No diffs, no build logs.

[Andrey @onsails]

Per stage: brainstorm a spec. Write a plan. Isolate a worktree. Codex implements with --write. Headless /code-review --fix loops. Verify. Commit. Land. Update the sprint doc. Next stage.

Main context only holds the doc. No diffs, no build logs.

[Andrey @onsails]

@0xIlyy Zellij is only cool when run locally. On ssh scroll is very flaky.

[Andrey @onsails]

Bigger thing is the pattern. AI infra is rebuilding crypto primitives. Opaque blobs, signatures, stateful protocols, replay protection. All shipped without 30 years of attack literature.
The two fields are converging in production. Fast.

[Andrey @onsails]

Last week a cryptographer poked at an AI feature and found a leak. Normal news for crypto. Fresh news for the AI stack. He didn't break any math. He just looked at how much the model thought.

[Andrey @onsails]

Second: the blobs replay. Same encrypted reasoning works across different sessions and different accounts. That's a textbook crypto bug shipped on top of a 2026 model.
https://t.co/7Hlr8oZ7H8