AGENTS.md

Hermes SkillEval Contributor Instructions

• Read CONTEXT.md, README.md, the relevant docs/v0.3/ protocol, and the nearest existing implementation or docs before editing.
• Keep one PR scope at a time. Do not bundle unrelated refactors or release work into protocol, benchmark, runtime, or documentation changes.
• Preserve existing CLI behavior, release gates, tests, and historical Phase 10-18 artifacts unless a later approved change explicitly updates them.
• Do not modify src/hermes_skilleval/**, tests/**, historical phase docs, or release logic for v0.3 PR-0.
• Treat the SkillRouter scored set as evaluation-only: no training, tuning, threshold selection, variant selection, or hard-negative mining on final scored labels.
• Preserve Phase 10 as deterministic offline replay evidence. Do not describe it as live-agent evidence.
• Use deterministic verifiers as the primary live-agent success judge. LLM or human review may diagnose failures but must not replace verifier outcomes.
• Keep VALID_EVIDENCE, INVALID_EVIDENCE, and REVIEW_REQUIRED separate from router promotion decisions such as KEEP_BASELINE and PROMOTE_CANDIDATE.
• Never commit external full datasets, model checkpoints, embedding caches, credentials, raw auth files, private host details, or unsanitized agent traces.
• Mark unavailable evidence as UNAVAILABLE; do not fabricate source, temporal, cost, token, activation, or verifier fields.
• Do not claim a benchmark result, live-agent result, release decision, or public-readiness posture that was not produced by a committed command and recorded artifact.