Small elephant in the room - SlashCommand tool

[rokicool]

The original GSD uses SlashCommand tool from Claude Code. The tool allows executing a command from another command or prompt.

invoke SlashCommand("/gsd:plan-phase [X+1]")

The conversion procedure I worked on replaced this call with

invoke Command("/gsd:plan-phase [X+1]")

And the 'original' GSD literally replaces it with

 invoke skill("/gsd:plan-phase [X+1]")

And the elephant that I was running around yesterday evening is none of the solutions worked. OpenCode does not support any of these.

Fortunately, TACHES moved most of the logic to shared workflows and don't use invoke SlashCommand often. That is why the elephant is tiny. But it is here, with us, in this very room.

And I don't know how to solve it yet. Try to create a custom tool that would execute the command? Translate a command to a skill and then execute a skill?

Ideas are appreciated!

[rokicool]

@dpearson2699 Hey, Derek. I would really appreciate your input, if you are still here. :)

[dpearson2699]

Hey @rokicool — did some digging on this with context7 and Opus fact checking me. Here's what I found in the OpenCode docs, what I'm still not sure about, and what probably needs actual testing.

---

What the docs say

Found this in the OpenCode docs (from packages/opencode/src/tool/task.txt):

Use the Task tool to execute custom slash commands. The entire prompt should be the slash command invocation, which can include arguments.

Task(description="Check the file", prompt="/check-file path/to/file.py")

So there IS a mechanism to invoke commands via Task(). That part is documented.

Also found:

• @agent syntax for manually invoking agents (@general help me search)
• subtask: true config option to force subagent invocation
• REST API /session/{id}/command for external triggers

---

What I'm NOT sure about

Here's where it gets fuzzy:

1. Does Task() exit the current command or nest under it? The docs don't make this clear. If it spawns a subagent that returns to the parent, that's different from what the original GSD does (exit and start fresh).

2. Context pollution — If Task() keeps the parent alive, we might be building up nested context instead of getting a clean slate. Haven't tested this.

3. Does the prompt actually get interpreted as a command? The example shows /check-file but I haven't verified this works with our /gsd-* commands specifically.

---

What needs testing

Before we commit to a solution, probably need to actually try:

• Task(prompt="/gsd-plan-phase 1") from inside another command — does it work at all?
• If it works, does it return to the parent command after or does it exit cleanly?
• What happens to context window — does it accumulate or reset?
• Does subtask: true in opencode.json change this behavior?

---

The 2 spots affected

Just so we know the scope — found these in transition.md:

• Line 436: Exit skill and invoke Command("/gsd-plan-phase [X+1]")
• Line 481: Exit skill and invoke Command("/gsd-complete-milestone {version}")

Both are in YOLO mode auto-continuation paths.

---

[rokicool]

Yep.

• Line 436: Exit skill and invoke Command("/gsd-plan-phase [X+1]")
• Line 481: Exit skill and invoke Command("/gsd-complete-milestone {version}")

These two lines is exactly the reason I started researching about that. They don't work. :)

[rokicool]

So, for now I found the only workaround.

Imagine you have custom command '/cmd':

---
name: cmd
description: Test command 
---

Use bash tool to execute the next command: `date +'%Y-%m-%d %H:%M test command here'`

And you like to execute /cmd command from /cmd-outer command.

There is no way you can do it in OpenCode.

However, you can create a skill with, say, 'cmd-skill' name:

---
name: cmd-skill
description: Test command to show idea
---

Use bash tool to execute the next command `date +'%Y-%m-%d %H:%M test command here'`

and then you can rewrite the '/cmd' file like this:

---
name: cmd
description: Test command 
---

Load and execute skill: skill(name="cmd-skill")

So, when it is necessary to execute '/cmd' command from another command or prompt - you should

Load and execute skill: skill(name="cmd-skill")

That is not pretty. But it works.

[dpearson2699]

There are some open PRs on the OpenCode GitHub that could fix this problem, but not sure if they'll be merged.

[PwccaCode]

There are some open PRs on the OpenCode GitHub that could fix this problem, but not sure if they'll be merged.

unfortunately opencode does not seen very keen to actually take in pull requests, i feel like they only ever listen to themselves

[rokicool]

I ended up creating a 'cloned' skill for every command that is called via the skill('gsd-command') method.
It is not a solution, but it is a functional workaround.