---

Selective offline compaction for Claude Code sessions. You choose a split point - by topic or by count - and decant summarizes everything before it while keeping the rest intact. The output matches Claude Code's native format, so claude --resume just works.

Unlike the built-in compaction, which summarizes the entire session, decant lets you preserve recent context and only compress the old parts. Pick a topic ("keep everything about the API refactor") or a count ("keep the last 5 exchanges") and it handles the rest.

Install

git clone https://github.com/TKasperczyk/decant.git
cd decant
python -m venv .venv
source .venv/bin/activate
pip install -e .

Usage

List sessions

decant list
decant list --project claude-memory

Preview a session's exchanges

decant show <session-id>
decant show <session-id> --full

Compact by topic

Keep everything from a topic onward, summarize everything before it:

decant compact <session-id> --topic "API refactor"
decant compact <session-id> --topic "the bug fix" --model sonnet

Compact by count

Keep the last N user turns, summarize the rest:

decant compact <session-id> --last 5
decant compact <session-id> --last 10 --model opus

Options

• --model haiku|sonnet|opus - which model to use for boundary finding and summarization (default: haiku)
• --strip - remove noise (progress ticks, thinking blocks, metadata, oversized tool output) before compaction
• --dry-run - preview what would happen without touching anything
• --no-backup - skip creating a .bak file (not recommended)

Session IDs can be the full UUID, a prefix (minimum 6 chars), or a direct path to a .jsonl file.

How it works

1. Parses the session JSONL and reconstructs the conversation tree via parentUuid chains
2. Finds the boundary message, either by asking an LLM to locate a topic, or by counting user turns from the end
3. Optionally strips noise (progress ticks, thinking blocks, metadata, oversized tool output)
4. Sends the head section to the LLM for summarization
5. Writes a summary record matching Claude Code's native format ({type, summary, leafUuid})
6. Sets the boundary message as the new tree root (parentUuid: null)
7. Drops all head messages and their sidechains, keeps everything from the boundary forward

A timestamped backup is created before any modifications.

Authentication

Decant needs Anthropic API access for the LLM calls. It checks these in order:

1. ANTHROPIC_API_KEY or OPENCODE_API_KEY environment variable
2. ANTHROPIC_AUTH_TOKEN environment variable
3. Claude Code OAuth credentials (~/.claude/.credentials.json)

If you're already authenticated with Claude Code, it just works. Decant reuses those credentials with automatic token refresh.

Why "decant"?

You're pouring the conversation through a filter. The clear, recent context comes through. The sediment stays behind as a summary.