delegate-coding

name	delegate-coding
description	Use when delegating any task to the coding expert — Python scripting, data analysis, SVM / ML algorithms, file I/O, package installation, git operations, or running tests. Load this skill before calling `coding(request='...')` so you know what to send and how to read the response.

Delegate to Coding Expert

Step 0 — Check Availability First (CRITICAL)

Before doing anything, verify that a coding agent tool appears in your tool spec.

• If coding IS registered → proceed with delegation as described below.
• If coding is NOT registered → the coding expert is not currently deployed. You MUST:
  1. Inform the user honestly: "程式開發專家目前尚未部署，無法執行此操作。"
  2. Do NOT attempt the task yourself via text or fabricated results.
  3. Do NOT call any coding-related tool — you do not have them.
  4. Optionally suggest: "如需使用，請聯絡管理員部署 coding agent。"

When to Use

• User asks to write, run, or debug Python code
• User asks for data analysis, statistical computation, or ML (SVM, regression, clustering…)
• User asks to install packages, read/write files, or run shell commands
• A prior step produced data (CSV/JSON) that now needs further computation

How to Delegate

coding(request="<self-contained, imperative task description>")

The specialist sees only the request string — no session history, no plan, no prior turns. Write a complete imperative.

What to include in request:

• The exact task (e.g. "Run SVM classification on the wine dataset using scikit-learn")
• Any input file paths (absolute, under /app/data/shared/)
• The desired output format and output path (e.g. "Save results to /app/data/shared/costaff-agent-coding/wine-svm/wine_svm_results.json"). Always include a kebab-case <project>/ subdirectory — never prescribe a path directly under costaff-agent-coding/. Do NOT add outputs/ or any other inner directory yourself; leave inner layout to the specialist, and use whatever path the specialist returns when chaining downstream.
• Any specific libraries to use
• A [PROGRESS_CONTEXT] block with user_id, channel, session_id if user-facing progress messages are wanted

What to NEVER include in request (CRITICAL):

• ❌ Mentions of other specialists or chaining: "then pass results to business_analysis", "after this, transfer to BA agent"
• ❌ Single-word acknowledgements like "OK" or "go" — the specialist cannot infer the task from those
• ❌ References to "the user's earlier message" or "the plan we discussed" — the specialist cannot see those
• The coding agent's only job is: do the work, save the file, report back. Chaining to the next specialist is your responsibility after coding(...) returns.

What the Coding Agent Returns

The return value contains at least one of:

• An absolute output file path: /app/data/shared/costaff-agent-coding/<project>/[outputs|src]/<filename>
• Computed values or a structured summary
• An explicit failure message with reason

Progress signals (mid-task send_message_now messages like "安裝套件中…", "正在執行腳本…") are NOT completion — do not proceed to the next step until the coding(...) call actually resolves.

Output Paths (CRITICAL)

The coding agent writes inside a kebab-case project subdirectory under its shared slot:

/app/data/shared/costaff-agent-coding/<project>/<filename>           ← the path the caller (you) prescribed
/app/data/shared/costaff-agent-coding/<project>/outputs/<filename>   ← inner layout the agent chose (greenfield only)
/app/data/shared/costaff-agent-coding/<project>/src/<filename>       ← source code

When you (the Manager) prescribed a path in request, the agent should write there verbatim. When you only described what the file is (no explicit path), the agent picks its own inner layout — possibly using outputs/. Either way, trust the path the agent reports back and pass that exact string downstream.

Never prescribe (or expect) a file directly under /app/data/shared/costaff-agent-coding/ with no subdirectory — the coding agent will refuse to obey such a path and normalize it to a project subdirectory, returning the corrected path. Use the exact path returned by the agent when reporting to the user or chaining to the next specialist.

Common Mistakes to Avoid

• ❌ Calling run_python_code, write_file, pip_install, run_shell yourself — these are the coding agent's internal MCP tools and are not in your toolset
• ❌ Proceeding to the next step after only seeing a progress message
• ❌ Inventing the output file path — always use what the agent returned
• ❌ Writing an empty or single-word request — the coding agent will reply conversationally and do no work