Skip to content
Cameron Mills

All projects

A small checked task handover that Codex and Claude Code can both read

Threadhaven

Part of Life Tracker

Open source, checked task handovers, so Codex and Claude Code can each resume the other's work. It saves a task's goal, decisions, next action and evidence as one checked snapshot.

Code on GitHub

Problem

A long task often has to move between sessions, or between Codex and Claude Code. The new session starts without the goal, the decisions already made or the next step, and a pasted transcript is too long to trust.

Approach

Keep the handover small and check it. A snapshot is the complete state of one task, kept to roughly 200 to 500 words, with links to the detailed sources rather than copies of them.

Architecture

Each task is a project and a task name. A save writes a new immutable record that names its parent, under a local lock, in a private store on one computer. A read returns the latest record and checks the files it refers to.

What I built

A snapshot carries the goal, decisions, constraints, progress, unfinished work, next action and evidence. Both tools use the same local storage engine and file format. A second command reads context-window readings and drops any taken before a compaction. It runs locally on Python's standard library, with no model calls or API keys.

What it guarantees

Saved records do not change under normal use, and a checksum shows accidental edits. A save has to name the snapshot it builds on, so a stale writer fails instead of replacing the other session's work. A read compares the files a snapshot refers to with their current bytes.

Results

This is the 0.1.0 release candidate. All 51 tests and an installed-wheel smoke check pass in CI on Linux, macOS and Windows, with Python 3.9 and 3.13.

What didn't work

It does not move a native conversation, compact a session or reset a usage quota. The summary is still written by a model or a person, so Threadhaven cannot tell whether it kept every important fact. Token savings and equal reasoning in the receiving session have not been shown.

Stack

  • Python
  • Claude Code
  • Codex

How these figures are sourced

The outcome figure is agent-reported: it is taken from a test log or report written by the agents that built and ran the system. The status was read from the code repository at commit 6bbdb9d. Every figure and claim on this page is tied to a source file or a CV line in the site's content.