💡 Multi-Agent Contribution Attribution & PR Quality Governance

[github-actions[bot]]

Summary

Extend the existing PR limits infrastructure to track per-agent contribution metrics (PR count, CI pass rate, merge rate, revert rate, fix-review loop stalls) across all agent authors (dev-lead, Claude Code, Copilot, Dependabot). Enable data-driven quality governance that throttles unreliable agents and promotes reliable ones, going beyond the current blunt PR count cap.

Market Signal

GitHub reports that Copilot coding agent and third-party agents now rank among the top contributors by PR volume in many enterprises. Repository-level Copilot usage metrics are now GA (July 2026), shifting reporting from "who used Copilot" to "where Copilot is influencing delivery." The governance gap — agents acting faster than humans can review — is recognized as the central challenge of agentic CI in 2026. Enterprises deploying multiple agent types need per-agent quality signals, not just aggregate throughput caps.

User Signal

Issues #892 ("dev-lead fix-review won't act on advisory-bot nitpick threads") and #888 ("dev-lead fix-review ignores CI-check regressions it introduced") show recurring quality gaps in agent fix-review loops. The PR limits standard (pr-limits.json) caps concurrent automation PRs but doesn't distinguish between agent types or measure quality. The org can't currently answer: "which agent produces the most stalled PRs?" or "which agent's fix-review loop has the highest failure rate?" These are questions the team is already investigating manually via individual issues.

Technical Opportunity

The existing PR limit gate (scripts/lib/pr-limit-gate.sh) already counts PRs by actor and maintains an exempt-actor list. Extending it to emit per-agent attribution data (agent name → PR count → CI pass rate → merge rate → time-to-merge → revert rate → fix-review iterations) is an incremental extension. The daily org status report (scripts/org_status.sh, issue #891) already aggregates PR data — adding agent attribution columns is straightforward. The gh pr list data already collected contains author information; grouping by author type is a filter, not a new data source.

Assessment

Dimension	Score	Rationale
Feasibility	high	Extends existing PR limit gate and org status report. Uses data already collected by gh pr list.
Impact	med	Enables quality-based agent governance instead of quantity-only caps. Provides signal to prioritize agent improvements.
Urgency	med	Agent PR volume is growing. Active issues (#892, #888) show quality gaps that attribution data would illuminate.

Adversarial Review

Strongest objection: We already have PR limits and the canary system. Adding per-agent metrics is over-engineering for a small org with 3-4 repos.

Rebuttal: The PR limits system is a blunt cap that treats all agents equally. Per-agent attribution answers the questions the org is already asking via issues #892 and #888: "why does dev-lead stall on advisory findings?" and "why does it ignore its own CI regressions?" Without attribution data, these questions require manual investigation of each PR. A simple dashboard (agent → PR count → CI pass rate → merge rate → stall rate) costs minimal engineering and provides the signal needed to prioritize agent improvements. Start with the daily org report; graduate to a standard when patterns emerge.

Suggested Next Step

Extend the daily org status report (scripts/org_status.sh) to include a per-agent-author breakdown: count of open PRs, merged PRs (7d), CI pass rate, and median time-to-merge. Use the existing gh pr list data already collected — group by author login and filter for known bot/agent authors. Once the data reveals patterns over 2-3 weeks, codify quality thresholds into a governance standard.