stack-detect (Phase 1 — detect)

Classifies the project into one or more { paradigm, engine, orm, platform, source_of_truth, confidence, files } stacks. This is the deterministic front door of every audit: it picks the paradigm profile the scorer re-normalizes over (scripts/score.mjs → PROFILES) and the module set each auditor runs.

What to do

1. Resolve the target directory ($ARGUMENTS path, else the project root / cwd).

2. Run the detector and read its JSON stdout:

   node scripts/detect-stack.mjs --dir "<dir>"

   It returns { stacks: [...], files_scanned }. Never edit or re-implement it — it is tested foundation.

3. Empty stacks → there is nothing to audit offline. Do not guess an engine. Route the user to /claude-db:start (the guided wizard) or invite a plain-language description of the intended database. Surface the detector's hint verbatim.

Source-of-truth precedence (see references/detection-signals.md)

When several sources describe one database, authority is, in order:

1. Live Tier-1 introspection (via the introspect skill) — beats any file.
2. Declarative / generated artifact — schema.prisma, Drizzle *_snapshot.json, structure.sql, schema.rb, generated migration SQL → confidence: established.
3. Migration SQL over ORM program source.
4. ORM program source (schema.ts, models.py, Mongoose/CDK) → confidence: directional.

directional stacks parsed from program source never raise a severity-5 cap — they nudge toward a generated artifact or Tier-1. When a declarative artifact and ORM source disagree, that is schema drift, owned by db-migration-safety (M22).

Routing emitted per stack

For each detected stack, emit the routing the orchestrator needs:

• paradigm → selects the score.mjs profile and --paradigm flag (relational | document | key-value | wide-column | vector | time-series | graph).
• module set → M0 engine-selection (recommendation, unscored) + M1..M22 filtered to the paradigm's profile modules. Categories whose modules emit no scored finding go inactive — e.g. a document store is never penalised for missing foreign keys.
• source_of_truth + confidence → the parse reliability ceiling carried into every finding (Tier 0(a) reliable vs 0(b) best-effort).

Multi-store

Return all stacks (e.g. Postgres primary + Redis cache). The orchestrator audits each, scores each with its own --paradigm, and rolls up worst-of-across-stores per axis (see references/scoring-model.md → Multi-store rollup), naming the flooring store.

Output

Emit the raw stacks array, the per-stack routing (paradigm profile + module set), and the source_of_truth precedence note. Keep narration tight. Respond in the user's language (EN/ES).