Deep Learn Mode (WP2 — full-codebase priming scan)

Deep Learn Mode is an on-demand, read-only scan of the current project (.claude/forge-bin/forge-deeplearn.cjs) that produces an honest project-summary and risk-list BEFORE the Lead commits to a PRD or a big architectural change. It never writes into the scanned tree — its only output is its own artifact under .claude/ (with --store) and optional dashboard events.

When the Lead runs it

• A new or unfamiliar project (first /forge run, or "gebruik forge voor dit project" on a folder with no memory yet).
• Before generating a PRD or acceptance criteria for a non-trivial feature.
• Before a large refactor, migration, or architecture change.
• Not needed for a 1-2 line fix, or a project already scanned this session — check .claude/forge-artifacts/index.jsonl first.

How

node .claude/forge-bin/forge-deeplearn.cjs --path . --run <run_id> --store

• --path <dir> — defaults to the current directory.
• --run <run_id> — logs deep_learn_started / deep_learn_completed to that run's dashboard events (omit to run standalone).
• --store — persists the full result via forge-store.cjs (.claude/forge-artifacts/deeplearn-<epoch>.json).

What it outputs

Stack (node/python/go/rust/dotnet/php/static + framework hints from package.json deps), file counts by category, likely entry points, tests presence, the 5 largest code files, and a risk-list (high/med/low) covering missing tests, oversized files (>800 lines), a missing README, TODO/FIXME density, secret-looking strings, and an unignored .env. Feed the summary + risks straight into the PRD's acceptance criteria, and hand any high risk to the Security Boss / security-reviewer; a secret-looking hit under a test/fixture/docs location (test/, tests/, __tests__/, test-evidence/, fixtures/, docs/, *.test.*, *.spec.*, *.md) is instead reported as a med secret-pattern-fixture-looking risk with a note pointing at forge-doctor.cjs's leak scan (forge-secret-scrub.cjs) as the real authority.

Honesty rule (non-negotiable)

• Read-only. It only ever reads files under --path; it never edits, deletes, or writes into the scanned project. Its only write is the explicit --store artifact under .claude/forge-artifacts/.
• Real findings only — every risk is backed by an actual file/pattern match; nothing is invented.
• Never prints or stores a raw secret. A secret-looking match is reported as {file, pattern_name} only, never the matched text — forge-store.cjs's own redaction is a second safety net when --store is used.

Deeper priming (optional, not required)

For genuinely deep semantic understanding beyond this structural scan, the Lead MAY additionally invoke the graphify skill (knowledge graph of the codebase) or the ECC learn-codebase skill (full-file read-through). A full-file read-through is costly: use it only for a large, unfamiliar codebase before a big change; otherwise the structural scan is enough. When reading big files, use chunks rather than loading entire files at once. Deep Learn Mode is fast and structural; those are slower and semantic — use them together on large/unfamiliar codebases, not as a replacement for each other.