Skip to content

Fix compaction splitting tool-use/result pairs#326

Merged
Gui-Yue merged 2 commits into
shareAI-lab:mainfrom
wangle201210:fix/snip_compact
Jun 7, 2026
Merged

Fix compaction splitting tool-use/result pairs#326
Gui-Yue merged 2 commits into
shareAI-lab:mainfrom
wangle201210:fix/snip_compact

Conversation

@wangle201210
Copy link
Copy Markdown
Contributor

@wangle201210 wangle201210 commented May 29, 2026

Summary

Fixes #325.

This updates context compaction so it preserves complete tool rounds instead of trimming raw message slices.

The bug was that snip_compact and reactive_compact could separate an assistant message containing tool_use blocks from the following user message containing tool_result blocks. That leaves orphaned tool results in history and can break the next model request.

Changes

  • add message-span helpers in s08_context_compact, s09_memory, and s20_comprehensive
  • make snip_compact trim by complete tool rounds instead of raw message count
  • make reactive_compact retain a safe tail instead of messages[-5:]
  • update s08_context_compact docs and generated site docs to describe the constraint correctly

Validation

  • python3 -m py_compile s08_context_compact/code.py s09_memory/code.py s20_comprehensive/code.py

@vercel
Copy link
Copy Markdown

vercel Bot commented May 29, 2026

@wangle201210 is attempting to deploy a commit to the crazyboym's projects Team on Vercel.

A member of the Team first needs to authorize it.

@Gui-Yue
Copy link
Copy Markdown
Collaborator

Gui-Yue commented May 30, 2026

@wangle201210
Hi! The issue itself is valid, but the current PR implementation is too heavy. The current changes noticeably increase the complexity of the teaching code and make the README examples harder to read.This problem is better suited for a small boundary fix, rather than introducing a relatively heavy span/window helper system separately into s08/s09/s20.

@wangle201210
Copy link
Copy Markdown
Contributor Author

wangle201210 commented Jun 1, 2026 •
edited
Loading

@wangle201210 Hi! The issue itself is valid, but the current PR implementation is too heavy. The current changes noticeably increase the complexity of the teaching code and make the README examples harder to read.This problem is better suited for a small boundary fix, rather than introducing a relatively heavy span/window helper system separately into s08/s09/s20.

@Gui-Yue 现在简化了很多写法，用来做学习应该问题不大了
主要是这个问题要是不修，之前的代码很可能跑不通

@Gui-Yue Gui-Yue self-assigned this Jun 3, 2026
@Gui-Yue Gui-Yue force-pushed the fix/snip_compact branch from b1751ee to fc08df5 Compare June 7, 2026 04:16
wangle201210 and others added 2 commits June 7, 2026 12:35
@wangle201210 @Gui-Yue
fix: keep tool-use/result pairs intact during compaction
7891c25 
Preserve assistant tool_use / user tool_result adjacency when compaction trims message history.

Fixes shareAI-lab#325.

Squashed original PR commits:

- Fix compaction breaking tool-use/result pairs

- Simplify compaction boundary fix
@Gui-Yue
fix: harden compaction pair handling
99a1e52
@Gui-Yue Gui-Yue force-pushed the fix/snip_compact branch from fc08df5 to 99a1e52 Compare June 7, 2026 04:36
@Gui-Yue Gui-Yue merged commit 20e7cbb into shareAI-lab:main Jun 7, 2026
3 of 5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

@Gui-Yue Gui-Yue

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

Compaction can split tool_use/tool_result pairs and break follow-up requests

2 participants

@wangle201210 @Gui-Yue