Skip to content

Triage agent should treat absent eval coverage as a prerequisite for issues that modify agent skill behavior #116

Description

@fullsend-ai-retro

What happened

The triage agent (run 29104087079) triaged issue #108 — a proposal to modify the review skill's behavior on rebase-only pushes — as sufficient with ready-to-code. This triggered the code agent, which produced PR #111. The human reviewer rejected the PR because the review skill has no eval coverage to validate behavioral changes. The triage agent did not check for eval coverage existence and treated the issue as ready for implementation.

What could go better

The triage agent already checks for blocking dependencies in step 2b (duplicate and blocking relationship search). It could extend this to check whether the affected skill has eval infrastructure. For the agents repo specifically, modifying skills/pr-review/SKILL.md without eval/review/eval.yaml existing is a prerequisite gap — the change cannot be safely deployed without validation. If the triage agent had emitted action: 'prerequisites' noting the need for review eval coverage, the code agent would not have been dispatched. Confidence: medium — this is specific to the agents repo's eval maturity. Once review evals exist, this gate would be satisfied and issues could proceed normally. The check might be too repo-specific for the general triage agent prompt, but it could be added to repo-level context (AGENTS.md or CLAUDE.md).

Proposed change

Add a section to AGENTS.md (the repo's agent-facing instructions) that documents the eval requirement for skill changes:

## 8. Eval coverage requirement for skill changes

Changes to agent skill behavior (`skills/*/SKILL.md`) require eval coverage for the affected skill. Before labeling a skill-modification issue as ready-to-code, verify that eval infrastructure exists (`eval/{skill-name}/eval.yaml`). If no evals exist for the target skill, the issue is blocked on eval creation.

Currently, only the triage skill has eval coverage (`eval/triage/`). Issues proposing behavioral changes to review, code, fix, or retro skills should be treated as having a prerequisite on eval infrastructure creation.

This approach leverages the triage agent's existing repo-context reading (step 2a: it reads CLAUDE.md and AGENTS.md) rather than modifying the general triage agent prompt. The triage agent would see this constraint during its context-gathering phase and factor it into its prerequisite analysis.

Validation criteria

On the next triage of an issue proposing behavioral changes to a skill that lacks eval coverage (e.g., review, code, fix), the triage agent should emit action: 'prerequisites' or note the eval gap in its reasoning rather than labeling the issue ready-to-code. Verify by checking the next 2 triage runs on skill-modification issues.


Generated by retro agent from #111

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions