Sub-agent, Questions tool, and jumping to another agent role

[allanchan339]

Problem

Hi, I am using this plugin, this is powerful and work quite smooth like cc /goal. Thank for the tool and effort on it.

However, i have met couple friction while using it.

1. Sub-agent: Goal-Explorer is asking simple cli command e.g grep, sed, cat, etc. These basic command should be pre-set as ALLOW set ?
2. Questions tool: as Goal agent role, most of the time i expect they will undergo ralph loop and keep trying until success met. However, i have encountered that the goal agent ask me question and hold there. The question tool should be in FORBIDDEN?
3. Jumping to another agent role: after goal set (switching to goal agent role), working smoothly, however after a compaction, agent role switched to build which is no longer the pre-set goal agent (and model).

The main reason of 3 is because this prompt is not using goal agent

Proposed solution

Giving sub-agent same level of tool
Disable Questions tool for goal agent
Stricter control on obey the identity before and after compaction? (No idea, i havent try to develop any tool for opencode specifically)

Alternatives considered

No response

[devinoldenburg]

Thanks for the detailed report!

1. You're right. goal-explorer currently has bash: ask for everything, while similar agents like goal-mapper allow safe read-only commands (rg, git status, git diff). Basic read-only tools like grep, cat, sed, and head should not need approval every time.

2. Also fair. The goal agent currently has question: allow, and that means it's not programmatically enforced. I will set this to deny in the next release.

3. This one is partly addressed already. In v0.6.6, the version I just released, guard-driven continuations (Goal Guard messages after review cycles) now explicitly pass agent: "goal". Before that, continuations had no agent in the body, so after compaction OpenCode could default the session to build. This matches what you saw in the screenshot.

Could you upgrade and retest?

npm install -g opencode-goal-mode@0.6.6 && opencode-goal-mode --global

Then restart OpenCode and run through the same flow.
Thanks again, I'll track #1 and #2 for the next patch. Let me know if v0.6.6 resolves #3 on your side.

[devinoldenburg]

I resolved your stated issues in https://github.com/devinoldenburg/opencode-goal-mode/releases/tag/v0.6.7 and #3

Please update to the latest version and give me a short update.

[allanchan339]

Hi @devinoldenburg , thank for prompt reply, I am installing this new version.

One more issue it may need to address, the current goal contract seems not smooth enough to construct the goal loop. I see model is claiming they wait guard to do auto review. But it should be triggered by themselves instead of manual skill trigger?

No one do the review and then stop here.

[allanchan339]

@devinoldenburg sorry to bother. Per latest update 0.6.11, I think todo access should give back the goal agent? or it is design ?

2. git, ruff access for goal-quality-gate? (I am quite curious why it need to keep asking, maybe giving them bash access?)