code

Code

Run a coding tool as a sub-agent, suitable for a variety of tasks such as code review, ideation, and implementation. This skill is powerful, and can handle code reviews, run commands (i.e. git diff, git rebase, etc.), and other tasks by itself internally.

Instructions

The user's request is in natural language. You must extract the following positional arguments and pass them to the script. Do NOT pass the raw user text as-is. If needed, write the prompt to a file and have the tool read that file instead.

SAMBA_PLUGIN_PYTHON is provided as an environment variable by the hook setup. DO NOT SET IT. THAT WOULD BE REALLY EMBARRASSING. Run: bash ${CLAUDE_SKILL_DIR}/scripts/code.sh <tool> <model> <cwd> <prompt> [--tool-arg <arg>...] DO NOT PERFORM THE TASK YOURSELF. THAT'S EMBARASSING, LIKE COVERING UP FOR A SUBORDINATE BY DOING THEIR TASK FOR THEM. DO NOT RUN THIS IN THE BACKGROUND. CLAUDE HAS EMBARASSING BUGS THAT CAUSE THE AGENT TO STOP AS SOON AS IT STARTS.

Arguments

Arg	Required	Description
tool	yes	Coding tool to use. See the section below under "Available Tools and Documentation" for more information.
model	yes	Model name as stored in the parameters database (use /list-models to check). Important: Use the bare model ID (e.g. MiniMax-M2.7), not a provider-prefixed name (e.g. sambanova/MiniMax-M2.7). The provider is configured automatically.
cwd	yes	Working directory for the tool. Default to the project root if not specified.
prompt	yes	The prompt to send to the tool. Quote it as a single shell argument. If the prompt contains shell-sensitive characters (e.g. $, <, >, (, ), `, !, {, }, `
--tool-arg	no	Extra arguments passed through verbatim to the underlying tool. Repeatable — each --tool-arg takes exactly one value. See the Passing --tool-arg section below for syntax details.
--max-tokens	no	Override the model's max_completion_tokens for this run. The database value is only a default; the actual limit is the model's context length. Use this to request more output tokens when needed.

Passing --tool-arg

--tool-arg uses argparse action="append", so each occurrence takes exactly one value. To pass a flag and its value to the underlying tool, use two separate --tool-arg entries.

Caveat: Values that start with - (i.e. flags) are ambiguous to argparse. Use the = syntax (--tool-arg="--flag") to avoid parsing errors.

# Basic usage with continue
code.sh continue MiniMax-M2.7 /project "prompt"

# Override max tokens to 64k
code.sh continue MiniMax-M2.7 /project "prompt" --max-tokens 65536

# Pass extra arguments through to continue
code.sh continue MiniMax-M2.7 /project "prompt" \
  --max-tokens 16000 \
  --tool-arg="--thinking"

# Attach a file to opencode (note: = syntax for the -f flag)
code.sh opencode MiniMax-M2.7 /project "prompt" \
  --tool-arg="-f" --tool-arg="/path/to/file.py"

# Multiple files with opencode
code.sh opencode MiniMax-M2.7 /project "prompt" \
  --tool-arg="-f" --tool-arg="src/main.py" \
  --tool-arg="-f" --tool-arg="src/utils.py"

Prompting Rules

Progress tracking

Include instructions in the prompt telling the sub-agent to write progress updates to a file, letting user check on progress. Do not poll for completion.

When constructing the prompt, always append something like:

As you work, write periodic progress updates to /tmp/samba-plugin/progress/progress_<unique_id>.md. Include what phase you are on, what you have completed, and what remains. Write to this file when starting the task to indicate that you have started, and update it periodically even if forward progress is not being made. Additionally, update this file whenever encountering an unexpected or noteworthy error or update.

Providing Context:

The coding tool is given the project directory and the prompt. It does not have access to the terminal history or the broader system context unless you explicitly provide it via the prompt or attached files. If the user provided instructions (including via AGENTS.md, CLAUDE.md, etc.) which are relevant to the task, include them as well. Additionally, if available, provide instructions for testing the code or verifying the task is complete.

Task Fusion

Since this skill is capable of using tools and calling commands, put the commands into the prompt instead of running them first and injecting their results.

Compound and fuse as many actions as reasonable into the instructions provided to the agent, including prelude and post-skill actions. TREAT THIS AGENT AS IF IT WERE A VERY INTELLIGENT INTERN. IT WOULD BE VERY EMBARRASSING IF YOU PRE-DIGEST THE ENTIRE TASK, OR PROVIDE EXCESSIVE HAND-HOLDING. IT WOULD ALSO BE EMBARRASSING IF YOU DIDN'T PROVIDE TESTABLE OUTCOMES OR ASKED THE INTERN TO MAKE ITS OWN. WHAT KIND OF MENTOR DOES THAT? For complex tasks with provided testable outcomes, prefer to include instructions such as:

After making changes, run the tests to verify they work. If they fail, iterate and refine until the tests pass. However, if the task seems intractable, summarize the issues and report them instead.

The agent is fully proficient in ALL common programming languages and file formats. Instructions that do not affect semantics such as "Add explicit template instantiations at the bottom of the .cpp:" or "Use this-> prefix for dependent names" are generally not needed.

As a general rule of thumb, prefer to think less about the minutae of the task, and instead let the agent deal with it.

Verifying Work

Since this skill is capable of using tools and calling commands, prefer to have the agent provide evidence of its results.

When writing prompts for this agent, use the WRITE skill. DO NOT use cat, echo, or other shell commands to write the prompt unless absolutely necessary.

File Paths

When providing paths to the agent, use absolute paths. This prevents confusion between /tmp (global temporary) and ./tmp (local temporary), and ensures that interactions with the agent are unambiguous.

Available Tools and Documentation:

• Continue
• Opencode