/learn

name: learn
description: Explore a codebase with parallel Haiku agents — clone, read, and document. Modes — --fast (1 agent), default (3), --deep (5). Use when user says "learn [repo]", "explore codebase", "study this repo", or shares a GitHub URL to study. Do NOT trigger for finding projects (use /trace), session mining (use /dig), or cloning for active development (use /incubate).
argument-hint: "<repo-url> [--fast | --deep]"

/learn - Deep Dive Learning Pattern

Explore a codebase with 3 parallel Haiku agents → create organized documentation.

Usage

/learn [url]             # Auto: clone via ghq, symlink origin/, then explore
/learn [slug]            # Use slug from ψ/memory/slugs.yaml
/learn [repo-path]       # Path to repo
/learn [repo-name]       # Finds in ψ/learn/owner/repo
/learn --init            # Restore all origins after git clone (like submodule init)

Depth Modes

| Flag | Agents | Files | Use Case | |------|--------|-------|----------| | `--fast` | 1 | 1 overview | Quick scan, "what is this?" | | (default) | 3 | 3 docs | Normal exploration | | `--deep` | 5 | 5 docs | Master complex codebases |

/learn --fast [target]   # Quick overview (1 agent, ~2 min)
/learn [target]          # Standard (3 agents, ~5 min)
/learn --deep [target]   # Deep dive (5 agents, ~10 min)

Directory Structure

ψ/learn/
├── .origins             # Manifest of learned repos (committed)
└── owner/
    └── repo/
        ├── origin       # Symlink to ghq source (gitignored)
        ├── repo.md      # Hub file - links to all sessions (committed)
        └── YYYY-MM-DD/  # Date folder
            ├── 1349_ARCHITECTURE.md      # Time-prefixed files
            ├── 1349_CODE-SNIPPETS.md
            ├── 1349_QUICK-REFERENCE.md
            ├── 1520_ARCHITECTURE.md      # Second run same day
            └── ...

**Multiple learnings**: Each run gets time-prefixed files (HHMM_), nested in date folder.

**Offload source, keep docs:**

unlink ψ/learn/owner/repo/origin   # Remove symlink
ghq rm owner/repo                  # Remove source
# Docs remain in ψ/learn/owner/repo/

/learn --init

Restore all origins after cloning (like `git submodule init`):

ROOT="$(pwd)"
# Read .origins manifest and restore symlinks
while read repo; do
  ghq get -u "https://github.com/$repo"
  OWNER=$(dirname "$repo")
  REPO=$(basename "$repo")
  mkdir -p "$ROOT/ψ/learn/$OWNER/$REPO"
  ln -sf "$(ghq root)/github.com/$repo" "$ROOT/ψ/learn/$OWNER/$REPO/origin"
  echo "✓ Restored: $repo"
done < "$ROOT/ψ/learn/.origins"

Step 0: Detect Input Type + Resolve Path

**CRITICAL: Capture ABSOLUTE paths first (before spawning any agents):**

date "+🕐 %H:%M %Z (%A %d %B %Y)" && ROOT="$(pwd)"
echo "Learning from: $ROOT"

**IMPORTANT FOR SUBAGENTS:** When spawning Haiku agents, you MUST give them TWO literal paths: 1. **SOURCE_DIR** (where to READ code) — the **`readlink`-RESOLVED ghq path**. **Never the `origin/` symlink path.** See ⚠️ below — this is not cosmetic. 2. **DOCS_DIR** (where to WRITE docs) - the parent directory, NOT inside origin/

⚠️ **BUG 1 (writes)**: If you only give agents `origin/` path, they cd into it and write there → files end up in WRONG repo!

⚠️ **BUG 2 (reads — worse, and silent)**: The `origin/` path is *nested inside this oracle repo* (`ROOT/ψ/learn/...`). An agent handed that path reads the target as a subdirectory of us and describes it as a variant of our repo. Measured on `mattpocock/skills` (2026-08-16): the arm given the nested path produced 3 contaminated docs / 26 false claims; the arm given the resolved ghq path produced **zero**. Same target, same minute, same model. Resolve it:

SOURCE_DIR="$(readlink "$ROOT/ψ/learn/$OWNER/$REPO/origin")"   # → ghq path

**FIX**: Always give BOTH paths as LITERAL absolute values (no variables!):

Example: ROOT=/home/user/ghq/.../my-oracle, learning acme-corp/cool-library, TODAY=2026-02-04, TIME=1349:

READ from:  /home/user/ghq/github.com/acme-corp/cool-library/     ← resolved, NOT under ψ/
WRITE to:   .../ψ/learn/acme-corp/cool-library/2026-02-04/1349_[FILENAME].md

Tell each agent: "Read from [SOURCE_DIR]. Write to [DOCS_DIR]/[TIME]_[FILENAME].md"

⚠️ MANDATORY: Isolation preamble (prepend to EVERY agent prompt)

Agents spawned here inherit **this oracle repo's** `CLAUDE.md` as project instructions. They will silently describe the *target* repo using *our* vocabulary. Measured on `mattpocock/skills` (2026-08-16): 3 of 5 docs contaminated, 26 false structural claims — including a doc that wrote *"(not in mattpocock/skills; reference from oracle-skills-cli CLAUDE.md)"* and then asserted the claim anyway. **Detection is not containment.** The agent noticing the mismatch does not stop it shipping.

Prepend this verbatim to every agent prompt:

ISOLATION RULE — read before anything else.
You are running inside an unrelated repo whose CLAUDE.md is in your context.
Its conventions describe YOUR HOST, not the target you are analyzing.
Document ONLY what you can cite from a file under SOURCE_DIR.
Before writing any structural claim (build tooling, versioning scheme,
directory layout, curation/lifecycle model, CI gates), verify it with a
concrete check — `ls`, reading package.json, reading the config file.
If you cannot cite it, write "not present" — never substitute a mechanism
you know from elsewhere, and never describe the target as a variant of
another repo.

**Do not soften this to "be careful."** The failure mode is confident and fluent; only a citation requirement catches it.

⚠️ Absent-referent rule (the biggest single cause)

A mandated section with no material in the target is what actually produces contamination. `TESTING.md` was demanded for a repo with **zero tests**; the agent correctly wrote "no test infrastructure" — then filled the rest of the page with our CalVer, our `bun run compile`, our "public shelf". The two docs whose topics were fully s