engraim-onboard

name	engraim-onboard
description	Set up a brand-new EngrAIm workspace for a folder/project. Use on first run (the SessionStart nudge), when the user runs /engraim:onboard, or when adopting EngrAIm into an existing project. Scans the project deterministically, seeds the memory frame (ENV/USER/SOUL), offers history backfill and optional semantic recall, and explains the workflow — without turning it into an interrogation.

Onboard a workspace

Goal: get a new workspace from empty stubs to genuinely useful in a couple of minutes, mostly from what can be detected rather than what the user has to type.

Steps

0. Opt in. Call the enable tool first — it creates .engraim/ if missing (idempotent). EngrAIm is inert in a project until this runs, and the SessionStart hook no longer adopts projects automatically, so this is what actually turns memory on here.
1. Scan. Call scan_project (read-only). Summarize for the user what this project is: languages, build tooling, package scripts, git remotes, notable top-level files.
2. Seed the frame by editing files in .engraim/frame/ directly:
   • ENV.md — the durable, bounded environment facts worth injecting every session (stack, tooling, key conventions detected). Keep it tight; it's a hot tier.
   • SOUL.md — operating rules that are obvious from the project (e.g. "this is a library, keep the public API stable") — only what you're confident about.
   • USER.md — ask one short question about working/communication preferences (or infer from how the user talks). Don't interrogate.
3. Offer backfill. Suggest /engraim:backfill to bootstrap memory from past Claude Code sessions in this project (note older transcripts may already be trimmed). 3b. Offer a .gitignore suggestion (don't write without asking): recommend the project ignore .engraim/store.db (rebuildable index) and .engraim/vendor/ (optional deps), and commit the rest of .engraim/ (wiki + frame are canonical, diffable knowledge).
4. Mention semantic recall as optional: /engraim:enable-semantic. Make clear keyword recall already works and semantic is a nice-to-have that adds a dependency.
5. Finish. Call mark_onboarded (stops the first-run nudge) and give a two-line tour: just work normally — sessions are captured and distilled (/engraim:curate), corrections become calibration (/engraim:calibrate), recurring procedures can become skills (/engraim:draft-skill), and /engraim:lint surfaces knowledge gaps.

Discipline

• Prefer detected facts over asked ones. Every question has a cost.
• Seed, don't overfit: a few solid frame lines beat a wall of speculation.
• Don't record secrets from config files you scan.