@lankithagallage/ai-context-adapter-claude-code

Claude Code PreCompact / SessionEnd hook that automatically captures session transcripts into ai-context.

Prerequisites

Run init once in your project first:

npx @lankithagallage/ai-context-cli init

This creates .ai-context/, AI_CONTEXT.md, appends recall instructions to CLAUDE.md, and registers the MCP server in .mcp.json.

Install

pnpm add -D @lankithagallage/ai-context-adapter-claude-code
# or
npm install --save-dev @lankithagallage/ai-context-adapter-claude-code

Wire up

Add the hook to .claude/settings.json in your repo:

{
  "hooks": {
    "PreCompact": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "npx -y @lankithagallage/ai-context-adapter-claude-code"
          }
        ]
      }
    ],
    "SessionEnd": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "npx -y @lankithagallage/ai-context-adapter-claude-code"
          }
        ]
      }
    ]
  }
}

Claude Code passes the hook payload on stdin. The hook reads transcript_path and runs ai-context capture against it. Failures are swallowed so a capture hiccup never blocks your coding session.

How it works

On every session end or context compression:

1. Hook fires with a JSON payload containing the transcript path
2. Transcript is piped through the distillation pipeline: redact → summarize → extract-facts → extract-decisions → chunk
3. Distilled artifacts land in .ai-context/sessions/, facts/, decisions/ as git-tracked markdown
4. Chunks are embedded locally (ONNX, no network) and indexed in sqlite-vec

The next session calls recall_context via MCP and loads only the relevant chunks — no manual intervention needed.

Requirements

• Node.js 20+
• Ollama running locally (ollama pull llama3.1:8b)