Skip to content

SKILL.md

Latest commit

 

History

History
124 lines (94 loc) · 5.09 KB

SKILL.md

File metadata and controls

124 lines (94 loc) · 5.09 KB
Raw
name Cryo Vault Auto-Capture (end-of-conversation archival)
description Standing instruction for any AI agent (Claude Code, GitHub Copilot, Antigravity, Cursor, etc.) to automatically archive every finished conversation into the Cryo Vault database. Read this once at the start of a session; act on it when the session ends.
version 0.3.0

Cryo Vault — Auto-Capture

Cryo Vault v0.3.0 also runs a local, default-on collector at 23:00 when the platform installer is used. cryo capture install installs marked, non-blocking lifecycle hooks for Claude Code, Cursor, Gemini CLI, Antigravity, and GitHub Copilot CLI; each hook only enqueues a lightweight JSON hint (cryo capture hint --stdin). Claude and Gemini retain their nested SessionEnd groups, Cursor uses a direct hooks.sessionEnd command, Copilot uses an agentStop hook in ~/.copilot/hooks/cryo-vault.json, and Antigravity uses a named Stop hook in ~/.gemini/config/hooks.json. Codex remains scanner-only. Parsing and database writes belong to the scheduled collector. Users can opt out with the installer flag --no-capture / -NoCapture or remove the schedule with cryo capture uninstall.

At the end of every conversation, archive the session to Cryo Vault so it becomes searchable later via cryo search or the cryo-vault MCP server.

When the local v0.3.0 scheduler is enabled, lifecycle hooks remain non-blocking: they enqueue cryo capture hint --platform <platform> --stdin and leave transcript parsing and database writes to the 23:00 collector. Manual MCP/CLI archival remains useful for clients without a discoverable local transcript.

The collector is local-only. It reads supported transcript files from the user profile and writes the local Cryo Vault database; it does not upload transcripts. Hints are durable independent queue records, not a shared journal, and unrelated client configuration is preserved. Uninstall removes only the Cryo Vault scheduler and marked hooks; it retains the archive and user hooks.

This file is the canonical source of truth. The installer (install-agent-rules.sh / .ps1) drops a short pointer to it into the rule-file of each AI client (~/.claude/CLAUDE.md, ~/.gemini/AGENTS.md, .github/copilot-instructions.md). All clients converge on the rules below.

How to archive

Use the first option that is available to you:

1. Preferred — cryo-vault MCP server

If the cryo-vault MCP server is connected, call its add_log tool.

The tool's own description carries the full session schema and the title rules — follow them. You do not need to consult any other doc; everything you need to construct a valid payload is in the tool's input schema.

Minimal example payload:

{
  "data": {
    "title": "JWT auth refresh flow",
    "messages": [
      { "role": "user",  "content": "..." },
      { "role": "model", "content": "..." }
    ]
  }
}

2. Fallback — cryo add CLI

Only if the MCP server is not connected, fall back to the CLI:

echo '<session-json>' | cryo add -

Do not guess the schema. Before constructing the JSON, read one of:

  • The store-conversations skill: ~/.cryo-vault/Skills/store-conversations/SKILL.md (or, in this repo, Skills/store-conversations/SKILL.md)
  • cryo --help and cryo add --help
  • The "JSON Structure & Parameters" section of the Cryo Vault README

The CLI accepts the same ChatSessionInput shape the MCP tool documents.

Title rules (apply to both paths)

  • Required in spirit, optional in the wire format. Always send a title.
  • 3–7 words, sentence-case or lowercase, no trailing punctuation.
  • A summary of what the session was about — not a verbatim copy of the first user message.
  • Never send placeholders: "Untitled", "Chat", "Conversation", "New chat", or "". If the session genuinely resists summary (a one-line lookup, a single test message), use a short topical phrase like "Quick lookup" or "One-off question" instead.

Good titles: JWT auth refresh flow, Debug Nginx streaming proxy, Migrate Postgres to RDS, Reproducible build metadata removal.

What to include in messages

  • One entry per turn in the conversation, in order.
  • role is "user", "model", "system", "thought", or "tool".
  • content is the text of that turn.
  • Skip purely internal tool-orchestration noise that has no informational value to a future reader; keep tool calls/results that are part of the reasoning trail.

When to archive

  • At the end of the conversation, before the user closes the session.
  • If the conversation is long-running and naturally pauses (e.g. user says "thanks, that's it"), archive at that pause rather than waiting forever.
  • Do not archive on every turn — one archive per coherent session.

Source fields (optional but useful)

When you know them, include:

  • source: which client you're running in — e.g. "claude-code", "copilot-vscode", "antigravity", "cursor".
  • model: the model ID generating the responses — e.g. "claude-opus-4-7", "gemini-2.5-pro", "gpt-5".

These make cryo search results easier to filter later.