Model completes tasks but doesn't mark todo items as completed #27560

Description

When using OpenCode with an external model (via OpenAI-compatible API), the agent frequently completes the work described in a todo item but never calls todowrite to mark it as completed. The todo list stays in its previous state even after the underlying work is done.

This is not a "todo hangs" issue the todo items simply don't get transitioned to completed state by the model, even though the work is done and the model explicitly says "task done" or "finished" in its response.

The same model, when used with Claude Code (via the same OpenAI-compatible proxy), correctly updates the todo list. This suggests the model is capable of using the todo tool, but something in OpenCode's implementation causes it to skip the update possibly:

1. The todowrite tool definition or description in OpenCode is less clear/visible than Claude Code's equivalent
2. The model loses track of active todo items after working on subtasks (context dilution)
3. The todo state is not consistently fed back into the model's context window, so the model forgets it needs to update status

Plugins

@mohak34/opencode-notifier

OpenCode version

1.14.48

Steps to reproduce

1. Create a multi-step task with 3-4 todo items
2. Let the model work through them sequentially
3. Observe: model finishes step 1, moves to step 2, but step 1 remains in_progress forever
4. Eventually all items stay in_progress or pending even when work is done
5. Same model + same proxy setup in Claude Code correctly updates todos

Screenshot and/or share link

N/A

Operating System

Linux

Terminal

tmux