A my na badania nad metodami diagnostycznymi nowotworów musimy zbierać na zrzutkach. W tym miesiącu udało nam się zebrać 450 zł, czyli jedną setną kwoty, jaką potrzebujemy na opłaty publikacyjne, o serwerach nie wspomnę.
https://t.co/0wgYhnVNga
https://t.co/1YW3htJyrz
Single bash file. curl | bash. Zero LLM calls on hot path. Three hooks: UserPromptSubmit, PostCompact, SessionStart{clear}.
MIT. v0.2.0 just out.
Long Claude Code sessions drift - agent forgets your ADRs, re-suggests approaches you already rejected.
Shipped context-revive v0.2.0: auto re-injects a deterministic project brief on cadence + after /compact + /clear.
Real demo↓
Test on a long-running session today.
51 prompts in, ran /compact:
[15:25:23] post-compact signal written
[15:25:23] post-compact: forcing emit (counter=51)
Brief landed in the agent immediately, bypassing cadence. Same flow on /clear hours later.
After /clear (full context wipe), asked the fresh agent: "what is this project?"
It answered with ADR 0025, ADR 0029, recent commit 61a691a, and the hard "success = DORMANT→ACTIVE, not MRR" constraint.
Full recall, prompt #1. That's the brief working.
Just shipped ruby_llm-contract 0.8.0 - Contracts + Evals for RubyLLM. New thinking/reasoning_effort DSL, adapter unification via Chat#with_thinking, sharper narrative. https://t.co/TM2RaXWGgu
Why pasting a directory tree into CLAUDE.md hurts:
You move one file → the tree is wrong → agent cites paths that don't exist.
Agent can `ls` itself. Paste what code WON'T tell it: ADR decisions, past incidents, "never X" rules.
2026 AGENTS.md research was blunt on one thing: don't auto-generate architecture overviews or directory trees.
Evidence: reduces agent success by 0.5–2%, costs 20%+ more.
https://t.co/7HFTzasG20
So revive injects only what the agent can't grep out of the code itself.
My laptop just told me I have 163 hours invested in a project I forgot existed. that's 163 hours of thinking. 163 hours of context. 163 hours I'll lose unless I reopen it.
Built armillary to scan my folders and tell me what's rotting. https://t.co/YmHmlsEytw
ruby_llm-contract 0.6.0: define quality rules, get told which model to use.
Step.recommend(candidates: [...]) — evaluates model + reasoning_effort combos, picks cheapest that passes, builds your retry chain.
Data, not vibes.
https://t.co/ZG4VvrvtVz
Project Boundary update: patched command substitution bypass + fixed quoted paths with spaces. No more sneaking past the guard. https://t.co/wJ2hxT9uRC
EccDNA miało historię trudną i poniekąd romantyczną. Zapraszam do przeczytania naszego artykułu, chętnie odpowiemy na wszystkie pytania!
https://t.co/qeergyotiV
@ComplexiaSC Totally agree - the image part especially. Once a bad image gets into context, every message after that just errors out and the session is done. Ended up writing a hook that intercepts image reads before they can poison it. https://t.co/OJbT3dffgp
Claude Code + images is still broken — crashes on transparency, large PNGs, and kills the whole session.
Last night I shipped a hook that proxies image reads through a Haiku subprocess. Zero image data touches your main context.
https://t.co/ocMZxptyoS