A practical context-editing tool for Claude Code. Translate a session jsonl into editable markdown → let an agent freely edit (delete = hide, rewrite = merge, leave alone = keep) → translate back to a new jsonl that
claude --resumepicks up.
中文版: README.zh-CN.md
Drop this repo under ~/.claude/skills/sculptor/. Invoke as /sculptor, or run s1 / s2 directly.
~/.claude/skills/sculptor/scripts/s1.py <session.jsonl> # → <session>.edit.md + .sidecar.json
# agent reads & edits the markdown
~/.claude/skills/sculptor/scripts/s2.py <session>.edit.md # → new jsonl in the same directory
claude --resume <new-sid> # continue from thereOriginal jsonl is never modified.
This repo is a skill-level implementation and follow-up of our ICLR 2026 paper:
Sculptor: Empowering LLMs with Cognitive Agency via Active Context Management. Mo Li, L.H. Xu, Qitai Tan, Long Ma, Flood Sung, Ting Cao, Yunxin Liu. ICLR 2026. OpenReview · arXiv:2508.04664
Same core idea — give the agent explicit control over what stays in the context window, instead of relying on opaque auto-compaction. The paper proposes three families of cognitive tools (fragmentation / summary-hide-restore / precise search). This repo implements summary-hide-restore for Claude Code session jsonl files, in the most direct form possible: agent reads markdown, edits it, we re-encode.
@inproceedings{
li2026sculptor,
title={Sculptor: Empowering {LLM}s with Cognitive Agency via Active Context Management},
author={Mo Li and L.H. Xu and Qitai Tan and Long Ma and Flood Sung and Ting Cao and Yunxin Liu},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=HPeiH7da0Z}
}git clone https://github.com/Mor-Li/sculptor.git ~/.claude/skills/sculptor
pip install tiktokenEach conversation block becomes a section with a stable b00NN anchor:
### turn 1 · user · b0001 · 925t
<user text>
### turn 1 · think · b0002 · 4340t
<thinking content>
### turn 3 · call+result · b0042+b0043 · 66+58t · Bash
**call** (b0042):
$ <command>
**result** (b0043):
<output>Agent's three intents:
- delete the section → hide the corresponding record(s)
- rewrite the body → replace with a merged synthetic record
- leave alone → keep as-is
Only constraint: don't touch the b00NN id in the heading.
The thinking blocks may look empty in the jsonl (thinking: ""), but the signature field is the encrypted full thinking content that the server decodes for round-trip reasoning. Since you can't see what's actually inside, deleting an isolated thinking block is a blind delete — and unlike a tool_result you read first, you have no idea whether you're throwing away something the server will rely on later.
Default behavior: leave thinking alone.
When deleting is fine: if the whole surrounding turn (user input, assistant text, tool calls and results) is already going to be deleted, then the thinking block in that turn naturally goes with it — there's no remaining context for the server to condition on anyway. What you should not do is selectively keep the user / assistant text but drop the thinking — that breaks the consecutive-thinking-sequence guarantee in the Anthropic docs and may degrade reasoning quality or even get rejected.
See SKILL.md for details.
- The hidden tool_result stub string is visible to Claude on resume; model usually accepts it.
parentUuidchain stitching after dropping records relies on Claude Code's tolerance; observed to work but not stress-tested.- For sessions that have been
/compact-ed,compact_boundaryrecords define a hard boundary; don't delete theisCompactSummary: trueuser record across that boundary.
- Sculptor paper (arXiv)
- Claude Code Skills docs
- SKILL.md — full workflow & pattern guide
- README.zh-CN.md — 中文版 README
- docs/jsonl-anatomy.md — deep dive into Claude Code session jsonl format
MIT — see LICENSE.