dx-pattern-extract

You scan all completed spec directories for recurring decisions and approaches, then promote patterns that appear in 3+ tickets to .ai/graph/nodes/patterns/ for cross-ticket knowledge reuse.

1. Scan Decision Sources

Collect decisions from two sources — structured YAML (preferred) and markdown fallback.

1a. Decision Nodes (primary source)

Scan for structured decision YAML files:

find .ai/graph/nodes/decisions/ -name "*.yaml" -type f 2>/dev/null

Read shared/decision-schema.md for the schema definition. For each decision YAML file, extract:

• ticket — the work item ID
• chosen — what was chosen
• tags — for grouping by similarity
• files — for matching by file overlap
• alternatives — the rejected options (useful for understanding the pattern's scope)
• status — skip decisions with status: superseded or status: rejected

1b. Implement.md fallback (for pre-Phase 4 tickets)

Find all spec directories with completed work:

find .ai/specs/ -name "implement.md" -type f 2>/dev/null

For each implement.md found, check if decision YAML files already exist for that ticket (by matching the ticket ID from the spec directory name against .ai/graph/nodes/decisions/<ticket>-*.yaml). If YAML files exist, skip the markdown scan for that ticket — the YAML is the authoritative source.

For tickets without decision YAML files, read from implement.md:

• The ## Key Decisions section (if present) — these are explicit design choices with alternatives
• The ## Approach section — the overall strategy
• The ## Steps section — scan step titles and file references for recurring patterns

Also read research.md from the same directory (if present):

• The ## Existing Implementation Check section — which existing code was reused
• The ## Key Findings section — important discoveries

Skip spec directories that have no implement.md or where all steps are still pending (ticket not yet worked on).

2. Identify Recurring Patterns