[elijahbenizzy]

Summary

• add /agentforce-doctor to reconstruct intended use cases, report only actionable findings, apply minimal repairs, and compare the unchanged baseline with the candidate
• add diagnostic coverage for instruction overrides, prompt text that impersonates control flow, routing, action contracts, outputs, state, lifecycle, transitions, side effects, architecture, capability honesty, and evaluation integrity
• require large agents to use bounded repair batches with parser checks and affected/unaffected regression cases between batches
• integrate the skill with plugin discovery, file-copy installation, routing, agent manifests, and user documentation
• add public AgentScript guidance for misleading pipe-text indentation, fake step sequencing, ungated actions, turn-ending setters, stale state, and competing side-effect owners

Evaluation integrity

• derive expected behavior from user requirements, approved specs, existing tests, and reachable behavior—in that order
• label inferred use cases instead of silently turning them into requirements
• run the same use cases and evaluator against exact baseline and candidate revisions
• inspect action availability, invocation, outputs, state, transitions, and external effects—not only final response text
• never treat simulated actions as evidence of a live side effect

Validation

• PYENV_VERSION=agentscript python -m pytest -q — 116 passed
• Agentforce Doctor skill validator — passed
• git diff --check
• independent forward tests against an intentionally flawed generic agent and a large isolated AgentScript agent; no source or organization-specific artifacts are included in this PR
• pinned open-source AgentScript asset validation (same CI command as the merge gate)

[almandsky]

Please create the PR to the internal repo.