maintain

Maintain Skill — Apply Detected Tooling Drift and Approved Lessons

You are running onboard's maintain apply step. The caller has already run maintain-detect.sh (a script, no model) and, optionally, collected lessons its owner approved. Your job is narrow: write the report's apply items and the approved lessons into tooling — verbatim, one line per fact, in each file's own style — and record what you did. The run is non-interactive: nobody will answer a question, so never ask one. Unsure → defer, never guess.

You may be running in the caller's main session or inside a subagent; the steps are the same (you need only the paths and the repository).

Input

File paths only, given as detect=<path> out=<path> [lessons=<path>] or as a JSON object with those keys:

Who writes what

.claude/ is a protected path in Claude Code: your Write and Edit tools may not write there in an unattended run. So:

• You edit CLAUDE.md files (Step 3) with the Edit tool.
• maintain-write.sh writes lesson entries (Step 4) and records every result entry; maintain-guard.sh after --result writes the result file (Step 5). Never write under .claude/ or the out file with your own tools.

Read files — the inputs, tooling files, and this skill's references/ — with the Read tool. Run every helper from the repository root as its own bash "<script>" … call — no cd, no shell variables, no ; or && chains, no pipes. Pass every free-text value — lesson text, evidence summary and ref, the summaries you compose — in single quotes, exactly as given, and write an apostrophe inside it as '\'' (don't → 'don'\''t'). Never double quotes: inside them bash runs `backticks` and expands $names before the helper sees the text, and lesson text is markdown, full of backticks: