Code
Runtime probe Codex 24667
SR
server
Done
Settle
Runtime probe Codex 24667
Runtime probe Claude 24667
SR
server
Done
Settle
Runtime probe Claude 24667
Runtime Probe for Skill Wrappers
SR
server
1m
Settle
Runtime Probe for Skill Wrappers
Settled (0)
SR
server
/
Runtime Probe for Skill Wrappers
You

You are a runtime reproduction coordinator in an isolated development server for issue 24667. Do not change any files, run shell commands, merge, watch PRs, or write to any forge. Use only T3 MCP tools. Create two ordinary test recipient threads (one Claude claude-opus-5-5, one Codex gpt-6.1-sol), root workspace, with initial message: 'Runtime probe only. Reply READY and stop. Do not change files or use tools.' Then after READY use t3_thread_send to send each recipient this exact payload constructed from parts: 'PREFIX: the literal mention ' + String.fromCharCode(36) + 'unslop appears mid-sentence here. SUFFIX: report whether the harness inserted full skill instructions or a native slash-command wrapper for this turn; report the exact user text and any command arguments you received, then stop without using tools.' Wait for recipient responses with t3_thread_read/wait and report verbatim answers and thread IDs. Finally send Claude a second payload beginning with '/' + 'unslop PREFIX: report whether a skill command wrapper was inserted and its arguments, then stop without tools.' Stop after collecting replies. Do not invoke skills yourself. The skill used is read-only. Report errors exactly.

Show full message

7:03 AM

T3 Code

I’ll create the two test threads, wait for READY, and send the probes in the requested order using only T3 MCP tools.

7:03 AM

T3 Code restarted and resumed this turn

7:04 AM
T3 Code

Both ordinary threads used the root workspace and initially replied READY. All probes completed without tool errors.

Claude — claude-opus-5-5
Thread: 
SR
Runtime probe Claude 24667

Mid-sentence probe response, verbatim:

text
The harness inserted both a wrapper and the full skill instructions.

**Native slash-command wrapper:** This turn arrived with these tags:
- `<command-message>unslop</command-message>`
- `<command-name>/unslop</command-name>`
- `<command-args>…</command-args>`

**Full skill instructions:** The complete Unslop skill body was added inline, starting with "Base directory for this skill: /home/ubuntu/.claude/skills/unslop" and the "# Unslop" pattern list, rules 3 through 33. It ended with an `ARGUMENTS:` line.

**Exact user text:**
```
PREFIX: the literal mention /unslop appears mid-sentence here. SUFFIX: report whether the harness inserted full skill instructions or a native slash-command wrapper for this turn; report the exact user text and any command arguments you received, then stop without using tools.
```

**Command arguments:** The same value appeared in both `<command-args>` and `ARGUMENTS:`:
```
appears mid-sentence here. SUFFIX: report whether the harness inserted full skill instructions or a native slash-command wrapper for this turn; report the exact user text and any command arguments you received, then stop without using tools.
```

The harness treated the `/unslop` in the middle of the sentence as a command invocation. It used everything after the token as the arguments. The text before it ("PREFIX: the literal mention") was not included in the arguments.


Codex — gpt-6.1-sol
Thread: 
SR
Runtime probe Codex 24667

Mid-sentence probe response, verbatim:

text
I received full skill instructions in a `<skill>` block for `unslop`, with path `/home/ubuntu/.agents/skills/unslop/SKILL.md`. I received no native slash-command wrapper or separate command arguments.

Your exact triggering text was:

```text
PREFIX: the literal mention $unslop appears mid-sentence here. SUFFIX: report whether the harness inserted full skill instructions or a native slash-command wrapper for this turn; report the exact user text and any command arguments you received, then stop without using tools.
```


Claude’s second, leading-slash probe response, verbatim:

text
The harness inserted a skill command wrapper again, along with the full skill instructions.

The wrapper had these tags:
- `<command-message>unslop</command-message>`
- `<command-name>/unslop</command-name>`
- `<command-args>…</command-args>`

The full Unslop skill body followed the wrapper. It started with "Base directory for this skill: /home/ubuntu/.claude/skills/unslop" and ended with an `ARGUMENTS:` line.

The `<command-args>` and `ARGUMENTS:` values were the same:
```
PREFIX: report whether a skill command wrapper was inserted and its arguments, then stop without tools.
```

Unlike the last turn, no separate user text came outside the wrapper. All of the text arrived as the command's arguments, including the "PREFIX:" label.


7:05 AM




Ask anything, @tag files/folders, $use skills, or / for commands
GPT-6-Astra
Medium
Full access
fs-04
Open in editor
Add project script
fix/issue-24667
Sync ref
Changes
+92
-0
Toggle Sidebar