Discover Patterns

Discover Patterns

Analyze a codebase to discover and document its implicit architectural and implementation patterns.

Workflow

1. Scope and explore

• Confirm scope with the user: full codebase (default), specific packages/directories, or specific concern.
• Explore the scoped code using parallel explore subagents. Read source files, test files, config files, and build files.
• Classify files by type: has_go, has_proto, has_ts, has_infra, etc.

2. Identify candidate patterns (Opus)

Launch up to 4 concurrent subagents (subagent_type="generalPurpose", model: opus per subagent-model-routing — architecture-level pattern recognition requires deep reasoning across the codebase), each covering a group of pattern categories from the taxonomy. Each subagent reads all relevant files and identifies candidate patterns with a signature the counting stage can apply mechanically.

Each subagent receives:

• The scoped file list for its relevant file types
• Reference: reference-pattern-taxonomy.md for what to look for

For each candidate pattern, record:

• Name: short descriptive name
• Description: what the pattern is and how it works
• Search signature: a concrete way to find further occurrences. A grep regex, file glob, AST predicate (e.g. "calls to context.Background() outside main.go"), or symbol pattern. Must be specific enough that the counting stage applies it mechanically.
• Seed exemplars: 1-2 file:line references the discovery subagent already saw

This stage does NOT compute confidence or gather exhaustive exemplars. Step 2.5 handles that.

2.5. Count occurrences per pattern (Haiku fan-out)

For each candidate pattern from step 2, spawn an Explore subagent (model: haiku per subagent-model-routing — applying a known signature is mechanical). Run up to 4 concurrently; if more than 4 patterns, launch the first 4 and the rest after one completes.

Each subagent receives (per subagent-prompt-contract):

• Goal: apply the search signature to the scoped file list. Return match count, total relevant files, additional exemplars, counter-examples.
• Inline context: pattern name + description + search signature + the scoped file list (paste; do not re-derive scope).
• Output shape: structured per-pattern record (counts + 2-3 exemplars + up to 3 counter-examples), ≤80 words.
• Constraints: read-only; no further subagents.
• Return: prefixed with Status: DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT.

Parent computes confidence from the returned counts:

• ESTABLISHED: matches in >80% of relevant files
• EMERGING: 50-80%
• INCONSISTENT: <50% but ≥3 matches

Drop patterns with fewer than 3 total matches (anecdotal, not a pattern).

3. Consolidate

Merge per-pattern records from step 2.5 (confidence, exemplars, counter-examples) with the descriptions from step 2. Deduplicate overlapping patterns (e.g. a DI pattern that also appears as a testing pattern). Resolve confidence levels across the merged dataset.

4. Write output

If REVIEW_DIR was provided by the review-all orchestrator, use it. Otherwise, resolve it:

REVIEW_DATE=$(date +%Y-%m-%d)
REVIEW_DIR=".reviews/${REVIEW_DATE}"
if [ -d "$REVIEW_DIR" ]; then REVIEW_DIR=".reviews/${REVIEW_DATE}-$(date +%H%M)"; fi
~/.claude/scripts/ensure-gitignore.sh '.reviews/'
mkdir -p "$REVIEW_DIR"

Write to ${REVIEW_DIR}/PATTERNS.md. Present the report to the user.

Output Template

# Codebase Patterns

> Discovered on YYYY-MM-DD. Scope: [full codebase | specific paths].

## STRUCT — Structure & Layering

### [Pattern Name]
**Confidence:** ESTABLISHED
**Description:** Brief description of the pattern.
**Exemplars:**
- `path/to/file.go:42` — explanation
- `path/to/other.go:15` — explanation

### [Pattern Name]
**Confidence:** INCONSISTENT
**Description:** Brief description.
**Exemplars:**
- `path/to/file.go:10` — majority approach
**Counter-examples:**
- `path/to/outlier.go:25` — deviates because...

## ERR — Error Handling
...

## OBS — Observability
...

## CFG — Configuration
...

## DI — Dependency Injection
...

## TEST — Testing
...

## XPORT — Transport
...

Guidelines

• Focus on patterns that are actionable; skip trivially obvious things (e.g., "files end with .go").
• A pattern requires at least 3 exemplars to be reported. Fewer is anecdotal, not a pattern.
• When confidence is INCONSISTENT, note which approach is more common and which is the outlier.
• Keep the output concise: 2-3 exemplars per pattern, not exhaustive file lists.
• For detailed discovery heuristics per category, see reference-pattern-taxonomy.md.
• Per ~/.claude/rules/probe-not-assume.md: confirm via tool/command before recommending; do not infer.