harness-synthesizer
Installation
SKILL.md
Harness Synthesizer Skill
Purpose
Synthesize executable validation harnesses for agent tool calls, inspired by AutoHarness (Google DeepMind, arxiv 2603.03329). Generates code-level verifiers that check action validity before or after execution, reducing agent errors through structured constraint enforcement.
Default mode is advisory (verifier). Hard enforcement requires explicit --hard-enforce opt-in per R021.
Three Modes
| Mode | Flag | Behavior | Enforcement |
|---|---|---|---|
verifier |
default | Post-hoc check: validates tool call results after execution | Advisory only |
filter |
--mode filter |
Pre-execution check: blocks invalid tool calls | Opt-in, requires --hard-enforce |
policy |
--mode policy |
Suggests the best valid action from available options | Advisory only |
Verifier Mode (Default)
Generates a YAML harness that describes post-execution checks for each tool the agent uses. Checks are emitted as advisory warnings — they do not block execution.