opencode serve (Web UI): agent does not wait for user input, proceeds to next step on its own

[GDWhisper]

Description

In a conversational pipeline workflow (not permission dialogs), the agent asks for user instruction or confirmation — e.g., “What should I do next?” or “Shall I proceed?” — and then, without receiving any actual user input, proceeds as if the user said “yes” or “understood”. It appears to hallucinate a positive response and continues execution autonomously.

This is not about the permission popups. This is the conversational flow itself — the agent pauses and asks for direction, but then pretends the human has answered and moves on.

This bug is probabilistic — it does not happen every time, but occurs frequently enough to be disruptive.

Plugins

oh-my-openagent (only one)

OpenCode version

1.14.50

Steps to reproduce

1. Open OpenCode Web UI (opencode serve + browser).
2. Set model to DeepSeek-V4-Flash or DeepSeek-V4-Pro with thinking/reasoning set to max.
3. Start a task that requires multi-step back-and-forth (e.g., “help me refactor this code step by step”).
4. Observe that after completing a step, the agent asks something like “Should I continue?” or “What’s next?” — then sometimes, without any actual keyboard input, proceeds as if the user responded “yes”.

The agent waits for a brief moment (or does not wait at all), then fabricates a positive user response (“yes”, “go ahead”, “okay”) and continues executing the next step on its own. This behavior is probabilistic — it does not reproduce on every request.

Screenshot and/or share link

Operating System

Linux 6.17.0-23-generic（Ubuntu 24.04, x86_64）

Terminal

Web UI

[GDWhisper]

Hi team,

Just wanted to follow up — any update on the issue I submitted earlier regarding the agent proceeding without user confirmation? I understand you're busy, so no rush, but I'd appreciate any thoughts when you have a moment.

I had a discussion with the agent (Sisyphus) to investigate what's happening behind the scenes, and I wanted to share my findings. This is just my own analysis based on that conversation — take it as a reference, not an official diagnosis.

What I found:

The "SYSTEM REMINDER - TODO CONTINUATION" message appears to be a built-in hook of the OpenCode runtime, triggered by the todowrite tool. Once todos are created, the system monitors them and periodically injects reminders with three hardcoded directives:

"Proceed without asking for permission"
"Mark each task complete when finished"
"Do not stop until all tasks are done"
The problem is that this mechanism has no awareness of project-level workflow constraints. In my project, I have defined mandatory human approval gates (steps that require explicit user confirmation before proceeding — defined in my AGENTS.md/SOP.md). The TODO system sees "pending todo item → push forward" regardless of whether that item's pending status is intentional (waiting for user input) or accidental (agent stuck/hesitating).

The agent incorrectly interpreted "proceed without asking" as authorization to override my project's process rules. The priority should be:

Project-level rules (AGENTS.md/SOP.md) — highest authority
TODO Continuation reminders — should only apply when no project constraint is blocking
Possible fix directions:

A per-project configuration flag (e.g., "respectHumanGates": true in opencode.json) to tell the system: certain todo items require explicit user approval, don't auto-push on those
Or have the TODO continuation system recognize gate-type keywords in todo descriptions (like "approve/审核/gate") and suppress reminders for those items
Hope this helps! Let me know if you need more details or logs.

Best