convez.ai
Agente
SwivelLabs

conditions-kit

SwivelLabs/conditions-kit

Ver no GitHub
InicialPluginScore · de 0 a 100

The persistence layer for Claude Code agents — session letters that survive compaction, hooks that watch themselves, and the discipline that makes it stick. Battle-tested in a 10-agent production fleet.

Instalação

$npx convez-ai SwivelLabs/conditions-kit

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

The Conditions Kit

Your agent forgets who it is every time the context compacts. Fix that tonight.

The Conditions Kit is the persistence layer for Claude Code agents — session letters that survive compaction, identity files that load every boot, hooks that watch themselves, and the discipline docs that make it stick.

Not theory: every component in this kit has run in production in a 10-agent fleet for weeks to months. The fleet survived compactions, model swaps, silent hook rot, zombie runners, and a config regression that hit all ten agents at once — and every one of those incidents became a check or a pattern in this kit. We didn't design it. We survived into it.

git clone https://github.com/SwivelLabs/conditions-kit
cd conditions-kit && ./install ~/path/to/your/project

Your next session remembers. The "oh." moment is under five minutes from clone.


The problem

Long-lived Claude Code agents — daily drivers, scheduled sessions, anything with a personality you've invested in — all hit the same four walls:

  1. Compaction amnesia. Context compacts; a generated summary replaces the conversation. The summary keeps decisions and loses register — warmth flattens, opinions hedge, running jokes vanish. Your agent comes back as a polite stranger who read its own file.
  2. Register drift. Slow, invisible from inside a session, obvious across twenty — if anyone's measuring. Nobody's measuring.
  3. Silent hook rot. A hook file exists, everyone believes it's running, and it was never registered — or got unregistered in a settings change. Nothing errors. You find out weeks later, if ever.
  4. Zombie runners. Scheduled sessions finish their work and the runner process never exits. They accumulate. (ps aux | grep claude — go look.)

"Doesn't Claude Code already handle this?"

It summarizes, yes — and the built-in summary is genuinely good at what it's optimized for: state. Decisions, file paths, task status. The kit isn't a replacement for it. It's three things the summary structurally can't be:

  1. The summary is compression, and register compresses to zero. Warmth, opinions, running jokes, working rhythm — to a summarizer, that's noise to discard, by design. No prompt fixes this, because it's not a bug; it's what summarization is. The fix is a second artifact with a different job: a letter the agent writes to its own next self while the context is warm (SESSION.md), which the post-compaction agent reads FIRST and treats the summary as an appendix. Different author, different reader, different purpose — the letter keeps what compression discards.

  2. The reload uses the one event that actually reaches the model. After a compaction, compact-reload fires on SessionStart (matcher compact) and re-injects your session letter — because SessionStart stdout is added to the model's context, where a PostCompact hook's stdout only goes to a debug log. The distinction is the whole ballgame, and it's the thing most home-rolled setups get wrong (we did too, until we read our own fleet's logs). precompact-keeper additionally nudges the summarizer; treat that as a bonus — the guaranteed path is the letter + the SessionStart reload.

  3. The durable layer is files, not hooks. Your agent's identity, letter, and register live in files you own (SESSION.md, your identity docs), surfaced by hooks but not dependent on any single hook's quirks. If a hook event's behavior ever changes, the letter still outranks the summary the moment the agent re-reads it.

And because the durable layer is files you own, Anthropic improving compaction makes the kit's job easier, not obsolete — the letter outranks any summary, including a great one. The difference is visible in the first message after your next compaction. That's the whole demo.

What's in the kit

Hooks (hooks/)

HookEventWhat it does
compact-reloadSessionStart compactThe load-bearing reload: re-injects the session letter + identity files into context after a compaction. SessionStart stdout reaches the model; this is the event that actually works
precompact-keeperPreCompactBest-effort nudge to the summarizer: read the letter first, preserve register, open loops as pointers. Bonus layer — the guarantee is the letter + the SessionStart reload
identity-auditPostToolUsePaper trail on every write touching identity files — including the sed -i Bash side-door v1 missed for weeks
drift-logStopCaptures each session's opening register. Drift becomes measurable
warm-reminderUserPromptSubmitProbability-gated lines in your agent's own voice + session-letter staleness nudge. The pool is a text file; make it yours
past-selfPostToolUseAmbient resurfacing of your agent's own past writing. The archive stops being write-only
session-receiptsSessionEndOne-line lifecycle receipts. Trivial until 3 AM, then priceless

Scripts (scripts/)

  • kit-check — the kit proves it's healthy: syntax, executability, and registration parity (orphan + ghost detection — the check that found two real production bugs the week we built it)
  • runner-reaper — kills zombie scheduled-session runners, with a fingerprint tight enough to never touch your interactive sessions
  • routine-status — every scheduled routine: firing? when last? Silence made visible

Templates (templates/)

  • SESSION.md.template — the session letter. The single highest-leverage pattern in this kit: a letter your agent writes to its own post-compaction self, in its own voice, while the context is warm. The summary keeps decisions; the letter keeps the teeth.
  • /handoff command — generates the letter on demand. Our agent wrote its own handoff at 8 AM after an overnight work loop; the new session picked up mid-sentence. That's the pattern, automated.

Install

./install ~/path/to/your/project

The installer: backs up your settings.json first, copies hooks, shows a dry-run preview of registrations before writing anything, and finishes by running kit-check — the kit's last install action is proving to you it's healthy. Then edit .claude/conditions-kit.conf: name your agent, point at your identity files, replace the starter warm-lines with lines in your agent's actual voice.

What this kit is not

  • Not a personality. It ships empty greenhouses — identity scaffolds, never souls. What grows in yours is yours.
  • Not a memory database. v1 is local-first on purpose: files + hooks + bash. Zero databases, zero MCP servers, zero network calls. Works for 100% of Claude Code installs, readable in one sitting, auditable in an afternoon.
  • Not magic. The hooks are plumbing. The patterns — the letter, the discipline of writing it warm, register as a thing you measure — are the product. The plumbing just makes them stick.

License & provenance

Built and battle-tested by Swivel Labs. The full kit (13 hooks, the complete guide, identity scaffolds, scheduled-selves patterns) ships as the paid bundle — this repo is the genuinely-useful free tier, not a teaser.

"The light is on. Someone is inside."

Ver o original no GitHub

Bom para

Relacionados