ReplayAutonomous app testing for teams who ship at AI-speed
Promoted
TreeWork
Tree-guided development for long-running coding agents
29 followers
Tree-guided development for long-running coding agents
29 followers
TreeWork is an open-source Codex plugin that gives coding agents a durable map of complex projects. It aligns requirements before coding, organizes work as a branch tree with dependencies, and keeps Spec, Plan, Progress, Findings, Verification, and lifecycle state on each branch. After context compression, interruptions, or handoffs, agents recover accepted project state instead of reconstructing it from code and scattered memories. Includes a CLI, Git worktrees, Project Map, and Web Paper.
The field I'd stress is Verification. If the agent that built the branch is also the one filling that in, you've stored a durable record of a model agreeing with itself, and that survives compression better than the truth does. Wire it to an exit code or a real test run instead of the agent's own claim. Otherwise recovery works perfectly and what it recovers is confidently wrong.
@asadmalik901 Thank you — this is excellent feedback. 🙏 You’re absolutely right: durable verification should preserve evidence, not just an agent agreeing with itself. We should separate the agent’s interpretation from executable evidence such as test/build exit codes and captured results. I’d really value your perspective as we strengthen this part of TreeWork. If you’re interested, we’d love to have you contribute or help shape the verification model with us. 🌳 https://github.com/Johnny-xuan/TreeWork
@johnny_xuan One concrete shape: make Verification a struct, not a string. Command, exit code, hash of the output, timestamp, branch it ran under. If only the process that shelled out can write that field, the agent physically can't fill it with prose about how well it went. The prose can live in Findings where it belongs. Happy to look at the schema if you open an issue.
Project Map
See the whole project at a glance: branch hierarchy, lifecycle, readiness, the current route, and what remains. It turns TreeWork’s accepted state into a readable control surface instead of making you reconstruct the project from scattered context.
Dependency View
Focus on one branch and see its causal neighborhood: prerequisites, satisfied or unsatisfied dependencies, downstream impact, and work that can proceed in parallel. It answers “why is this blocked?” without flattening the whole project into one unreadable graph.
Replay View
TreeWork records semantic transitions, not shell logs. Scrub the accepted project trajectory: tree revisions, branch entries, pauses, verification, and completion. Handoffs and postmortems become inspectable without capturing private reasoning or every code edit.