axial-coding-new

Run this axial coding phase after open coding: it delegates bulk trace reading to a trace-reader subagent, groups failure notes by what actually went wrong, and drafts labeled clusters with definitions, counts, and severity for your approval before writing taxonomy.yaml.

Axial coding groups open codes (per-trace annotations) into higher-level failure categories derived from the data itself. This skill implements that as phase 3 of a discovery pipeline, producing a validated taxonomy.yaml and ranked backlog.

Yes, but the skill pushes back first because pre-supplied lists miss emergent categories. If you insist, it complies and notes in taxonomy.yaml that categories were user-supplied rather than data-derived.

It requires evals/open_codes.jsonl produced by the open-coding phase, plus evals/traces.jsonl for trace context. If open_codes.jsonl is missing or empty, the skill stops and reports that.

Each code is scored as frequency x severity weight (low=1 to critical=4) x fixability (1-5). Fixability is a judgment call the skill asks you to sanity-check when it materially changes the ranking.

Reading all traces into the main context is expensive and unnecessary. The trace-reader subagent batches and summarizes large trace collections down to what clustering needs, at bounded turn cost.