Custos — Development skill for Claude Code
Continuous, hook-driven session capture so an agent's history survives Claude Code's own context compaction and /clear -- Ed25519-signed checkpoints, verified live.
How to install Custos
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open RARS-oss/custos and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Custos does
Continuous, hook-driven session capture so an agent's history survives Claude Code's own context compaction and /clear -- Ed25519-signed checkpoints, verified live.
Alternatives in Development
- Claudeception — A Claude Code skill for autonomous skill extraction and continuous learning 2.2k ★
- Continuous Claude — Run Claude Code in a continuous loop, autonomously creating PRs, waiting for checks, and merging 1.3k ★
- /autopilot — Autonomous hunt loop with deterministic scope safety and configurable checkpoints 1.2k ★
README
custos
**Continuous, hook-driven session capture so an agent's history survives Claude Code's own context compaction and `/clear`.** Every turn is appended to a hash-chained ledger the moment it happens — not because the agent remembered to save it, but because a Claude Code hook did. Lifecycle boundaries (`PreCompact`, `SessionEnd`) get an Ed25519-signed checkpoint, ported from [`bulla`](https://github.com/RARS-oss/bulla)'s receipt shape, over everything captured since the last one.
Why
[`tabularium`](https://github.com/RARS-oss/tabularium) already solves durable, checkable memory for whatever an agent *chooses* to write. But "chooses to write" is the weak link: it depends on remembering to call it, turn after turn, for the whole session. This project exists because that weak link showed up live — a real session hit Claude Code's `/clear to save N tokens` prompt and the obvious question followed: what happens to everything not already saved? `custos` is a client of tabularium, not a competing memory engine: it makes capture automatic, and leaves curation (what's actually worth remembering) to `custos consolidate` and, ultimately, tabularium.
Scope — grounded in Claude Code's real hook surface, not assumed capability
Checked before writing a line of code (`docs/DESIGN.md` §3):
PreCompact / PostCompact |
Exist, but notification-only — cannot block or steer compaction. Used as a checkpoint trigger anyway: it's the only "something's about to happen" signal there is. |
SessionEnd |
Gives a why (clear|resume|logout|prompt_input_exit|other) — the best lifecycle signal for "about to lose this session." Also a checkpoint trigger. |
Stop / UserPromptSubmit / PostToolUse |
The actual capture surface — custos never parses Claude Code's internal transcript file (async, unsupported, versioned); these three hooks are the ground truth instead. |
| Token / context-usage API | **Verified not to exist anywhere in t |
Related Skills
Claude Compaction Viewer
View and analyze Claude Code conversation history and compaction events with a TUI and CLI tool for clear sess
Auth Licensing
Skill: Auth & Licensing (Ed25519 signed keys, no user accounts in Y1)
Compact
Prepare session state for context compaction — capture learnings, save state, then compact
Post Compact Reminder
Claude Code hook that detects context compaction and injects a reminder to re-read AGENTS.md, preventing post-
Skill Evaluation Graph
Description Evidence-driven evaluation, benchmarking, and verified repair for Agent Skills across Codex, Claud
Bench Run All
Agent-driven bench round for one size — drive the 2 variant folders + judge + capture, hands-off
Related Agents
Tp Session Restorer
PROACTIVELY use this as the FIRST step after /clear, compaction, or a fresh window when a recent session_snaps
Continuous Learner
Use at session end (auto-triggered by SessionEnd hook) or via /learn command. Extracts repeatable patterns, de
Grow Agent
Session orchestrator for one domain. Owns Rule 20a session-start gate + Rule 4 goal-driven loop + Rule 22 deci