codex-pr-thresholdlisted

## When to Use This skill is invoked after each batch's state update to determine whether the accumulated raw learnings are sufficient for distillation. ## Non-Negotiable Rules | # | Rule | Violation Example | Correct Pattern | |---|------|-------------------|-----------------| | 1 | Threshold is **exactly 200 lines** in learnings-raw.md | Using subjective criteria like "enough learnings" | `wc -l < ~/.claude/codex-pr-lessons/learnings-raw.md` > 200 | | 2 | Count ALL lines including headers and separators | Only counting observation lines | `wc -l` on the entire file | | 3 | Report the exact count and decision | Silently triggering or skipping | Always output: "{N} lines — {DISTILL/SKIP}" | | 4 | Never trigger distillation on first batch | Distilling from a single batch of 5 PRs | Minimum 2 batches must exist (check state.total_processed >= 10) | ## Threshold Definition ``` DISTILL if: (line_count > 200) AND (state.total_processed >= 10) SKIP otherwise ``` **Why 200 lines?** At ~25-30 lines per PR observation block plus batch summaries, 200 lines represents roughly 6-8 analyzed PRs with cross-batch observations — enough data to generalize but not so much that raw learnings become unwieldy. **Why minimum 10 PRs?** Patterns from fewer than 10 PRs risk being anecdotal. Two full batches provide enough contrast between merged/rejected, large/small, and well-described/poorly-described PRs. ## Verification Command ```bash LINE_COUNT=$(wc -l < ~/.claude/codex-pr-lessons/l