Bias triage and batching toward issue types with higher agent-PR merge rates #20

Problem

Empirical evidence: autonomous-agent PRs merge at substantially higher rates for some issue types (docs, CI, build, small refactors) than others (bug fixes, performance work). The current template's triage and batching (Phases 1–2) treat all issue types equally and require the agent to make value judgments about issues without surfacing the reasoning.

Motivation

Per RESEARCH.md §2 and §7:

• Where Do AI Coding Agents Fail? 2026 (arXiv:2601.15195): merge rate strongly conditioned on issue type and PR size.
• LLMs for Issue Classification (arXiv:2506.00128): LLM triage accuracy 70–85%. Useful as filter, not decision-maker.

Confidence: medium-high for issue-type effects; medium for LLM-triage accuracy.

Proposal

Two changes to Phase 1 (Triage) and Phase 2 (Work selection):

1. Emit triage reasoning

When the agent decides to skip an issue, it must say so explicitly with the reason — either in the PR body of whatever it does pick, or as a comment on the skipped issue. Don't silently filter. This makes the triage auditable without requiring perfect accuracy.

2. Bias batching toward higher-merge-rate types

In Phase 2, when choosing between issues of comparable scope, bias toward (in order): documentation fixes, CI/build fixes, small refactors, then bug fixes, then performance work. Encode as a tiebreaker rule, not a hard exclusion — a clearly-scoped bug fix is still better than no progress.

This is especially important for the first 3–5 cycles after a new dev-loop is generated for a repo, when the agent has minimal prior context.

Acceptance criteria

• Phase 1 says explicitly: any issue not picked must have its skip reason recorded
• Phase 2 has a tiebreaker rule listing issue-type preference order with one-sentence rationale
• Phase 2 marks the "early cycle bias" as a soft preference, not a hard rule
• Test plan: run against a repo with a mix of doc/bug/perf issues; observe the picks match the bias

Retrofit

Per CLAUDE.md's promotion rule, retrofit PRs needed for all four child skills.

Related

• RESEARCH.md §2, §7
• Top-5 ranking in RESEARCH.md identifies this as the Add step to record new skills in example-skills-catalog #3 highest-impact change

drafted by Claude on behalf of Daniel Stephenson