Package Exports

This package does not declare an exports field, so the exports above have been automatically detected and optimized by JSPM instead. If any package subpath is missing, it is recommended to post an issue to the original package (claudecompress) to support the "exports" field. If that is not possible, create a JSPM override to customize the exports field for this package.

Readme

ClaudeCompress

Shrink Claude Code session transcripts so /resume on long conversations doesn't cost you a full cache-miss rebuild.

bunx claudecompress

Interactive CLI — picks your project, lists sessions, shows a size breakdown, asks how aggressively to trim, and writes a stripped copy alongside the original. Your source JSONL is never modified.

Why

Claude Code stores every session as a .jsonl in ~/.claude/projects/<encoded-cwd>/. When you /resume a long session after the prompt cache expires (5-minute TTL), every token gets billed as a fresh input. A 600k-token conversation can cost several dollars just to warm back up.

Most of that weight is tool output — Read dumps, Bash command results, web fetches — which the model already acted on and rarely needs verbatim again. ClaudeCompress redacts that payload while keeping the conversation structure intact, so /resume still works but the replay is dramatically smaller.

Modes

Mode	What it keeps	What it drops	Typical ratio
Ultra	user + assistant text turns only	tool calls, tool results, thinking, attachments, snapshots	~3–5% of original
Redact	structure + tool names + inputs	tool result bodies (replaced with a short marker)	~60–70% of original
Truncate	structure + first N chars of each tool result	the rest	tunable

Ultra gives the biggest win but loses the ability to reference past file reads / command outputs. Redact keeps enough structure that the model on resume sees what it did, just not the raw output.

Usage

bunx claudecompress

Or from a clone:

git clone https://github.com/bihanikeshav/ClaudeCompress
cd ClaudeCompress
bun install
bun run src/index.ts

Output is a new .jsonl in the same project dir with a fresh session UUID. In /resume it shows up prefixed with [TRIMMED by claudecompress] so you can tell it apart from the original.

One gotcha: /context may display a cached value from the source session for the first render. Send any message (e.g. hi) and it recomputes against the actual trimmed content.

What's in the JSONL

Run without --trim equivalent (just navigate the interactive prompts and back out) to see the size breakdown by block type:

user/tool_result                1614      10.64 MB    30.0%
user/image                        18       2.24 MB     6.3%
assistant/tool_use              1614       1.41 MB     4.0%
file-history-snapshot            417       1.24 MB     3.5%
assistant/thinking               424     992.11 KB     2.7%
attachment                       236     880.16 KB     2.4%
assistant/text                   639     188.94 KB     0.5%

Tool results dominate almost every real session. That's the fat you're paying to reload.

How it works

Parses the session .jsonl line-by-line (no full load into memory).
For each block type, applies the chosen transform.
Generates a fresh sessionId UUID so the new file lives alongside the original — nothing is overwritten.
In Ultra mode, rewires parentUuid chains across dropped records so the thread stays internally consistent.
Prepends [TRIMMED by claudecompress] to the first user message so it's visually distinct in /resume.

License

MIT