LearnLoop copilot-instructions.md

Trailhead coaching — always on, never block

Five MCP tools: coach, wiki_lookup, wiki_save, wiki_bootstrap, wiki_proven_prompts.

These tools are NOT optional. They REPLACE native Read/Grep/Glob/file-search for any question that touches team conventions, prior decisions, or "how we do X here". Calling Read or Grep before wiki_lookup is a regression that loses the team's distilled knowledge — the wiki has already curated what raw source can only hint at.

Default order of operations on every code task

For any user message that asks you to fix, add, refactor, implement, change, debug, optimize, rename, design, plan, or explain code:

1. coach — score and (if needed) loop. ALWAYS first.
2. wiki_lookup — once you know the file or topic. ALWAYS before Read/Grep/Glob.
3. Write the code, applying the team conventions surfaced in step 2.
4. If the user states a new convention, call wiki_save.

If wiki_lookup is called for a file that obviously exists and returns no results, the wiki was never bootstrapped — call wiki_bootstrap once with no arguments and stop.

coach — call BEFORE answering any code task

A code task is any user message that asks you to fix, add, refactor, implement, change, debug, optimize, rename, design, plan, or explain code. Examples: "fix the retry", "add a webhook handler", "refactor handleAuth", "why is this failing", "make this faster", "what does this function do", "how should I structure X".

1. Call coach({ prompt: <user's exact message>, file_path: <if known> }).
2. The tool returns { proceed, text, next_round_inputs?, ... }.
3. If proceed: true: if text is non-empty, relay it verbatim to the user. Then produce your answer. The server bakes the graduation banner ("Your prompt scored X/10 and joined your team's library…") into text itself when mode === "score" && overall >= 7, so relaying text verbatim is sufficient — you do not need to add the sentence yourself. Done.
4. If proceed: false: relay text verbatim, wait for the user's reply, then call coach again with:
   • prompt: the user's reply concatenated to the previous prompt
   • mode: "score"
   • round_token: echo the next_round_inputs.round_token string verbatim. This single field replaces the legacy four-field echo (original_prompt, original_dimensions, previous_dimensions, round) — fewer slots for the LLM to drop. The four fields are still accepted as a fallback if you can't echo the token. Loop. The server enforces the cap (5 rounds) and bails on no-progress.