Status: Proposed
Date: 2026-07-02

Context

The EvoSkill framework (Sentient AGI) automatically discovers and refines agent
skills through an evolutionary loop: propose changes → generate candidate files →
evaluate → retain improvements via Pareto frontier. Meanwhile, this project’s
remediation engine (reporter/remediation.go) produces static, template-based
advice — the same generic suggestions regardless of the skill’s specific gaps.

The gap is that we can tell a skill what is wrong and suggest generic fixes,
but we cannot discover new skill content, test whether a specific fix actually
helps, or iterate toward improvement.

A detailed integration analysis is in the findings document.

Decision

Port EvoSkill’s core evolutionary loop to Go as an --evolve flag on the
existing skill-auditor remediate command. The port scope is:

1. Proposer — LLM prompt templates that analyze (task, failure, current skill)
   and propose specific SKILL.md edits or new skill files.
2. Generator — apply proposals as structured skill folder edits (file I/O).
3. Loop controller — iterative: run eval → collect failures → propose →
   generate → re-evaluate, retaining top-K candidates per generation.

Do not port: harness runners (we run the CLI directly), dataset loaders (use
our existing eval scenario format), or Python-specific plumbing.

The native eval runner (cmd/eval.go) serves as the evaluation callback, and
the existing D1-D9 scorer provides a multi-objective fitness signal alongside
task accuracy.

Consequences

Positive:

• Remediation evolves from static advice to automated improvement
• Reuses existing infrastructure: eval runner, scorer, git operations
• Same command (remediate), different execution strategy (--evolve flag)

Negative:

• LLM costs per generation (proposer calls + eval runs)
• Proposer prompt templates will need iteration to produce quality proposals
• No-regression guarantee requires careful frontier management

Alternatives considered

• Keep static remediation only (no change) — loses the automation opportunity
• Bridge to Python EvoSkill via subprocess — rejected due to language boundary
  complexity and the project being Go-only