Roo Code Test of constitution Command Fails

[hiarcs]

Roo Code Version: 3.28.15 (7734b29c)
VSC Version: 1.104.3 Mac

Test Method
Create a new project using a command similar to specify init sk4 --ai roo --script sh, then select different models to execute /speckit.constitution SOME DESCRIPTION with Roo Code. The situation is confusing:

1. GPT-5-Codex
   constitution.md is modified successfully, but at the same time, three templates in .specify/templates are also modified. I'm not sure if the latter should be modified. If they should, it means there's redundant information in the specification file. If not, it indicates that this tool does not provide read-only protection for specific files, and its prompt has factual defects.
   Additionally, in my test, I used a Chinese description, so constitution.md was modified to a state mixed with Chinese and English. I'm not sure if the modified language in the section titles will affect subsequent operations.
   The process consumed 400k tokens, which was somewhat unexpected.

2. Qwen3-Next-80B-A3B-Thinking
   constitution.md was modified, with some content deleted, but none of the expected content was written. A total of four files in .specify/templates were modified in a similar manner.
   The issue here is the same as with Claude and Qwen3 235B later: the project name and constitution requirements were not passed correctly. Since I had auto-approval enabled, the AI automatically chose to use blank values for replacement.

3. Qwen3-Coder-30B-A3B-Instruct
   Unable to complete the command

4. Grok-Coder-Fast-1
   Unable to complete the command

5. Claude 4.5
   After consuming 120k tokens, the model asked me to provide the project name, core principles, and whether I want to create a new file or update it.

6. GPT-OSS-120B
   Unable to complete the command

7. Qwen3-235B-A22B-Instruct-2507
   After consuming 180k tokens, similar to Claude, it asked for the project name and core principles.

Is this tool really adapted for Roo Code?

Furthermore, writing a project's constitution content is not a high-difficulty task. Logically speaking, even relatively weak models should be capable of handling it. I'm not sure if the failure above is due to the model, Roo Code, or Spec Kit itself, but I suggest adding recommended models in the README.

[hiarcs]

GLM4.6, performing similarly to Codex, first updated all 5 template files and then constitution.md, consuming a total of 770k tokens.

During execution, despite explicitly including error handling strategies in constitution.md, when executing the /speckit.specify command, it still asked me what to do if an error was encountered. This is quite absurd. Furthermore, in the output document of the planning task, its designed behavior further deviated from the requirements set forth during the constitution and specification process.

4.5M tokens were consumed before the first source code file was even generated. By the time 7.5M tokens were consumed and the implementation was declared complete, the overall scope and implementation approach had significantly deviated from my original requirements.

[clsferguson]

Real quick, the constitution command includes modifications to those 3 templates. So codex did it right. Have a look at the prompt to see.