Leeaoyin/dsh-agent-harness-audit

README excerpt

dsh-harness-audit English 中文 Audits an agent harness for the failure modes that break agent loops — and refuses any finding whose evidence isn't actually in the code. Fifteen checks, one throwaway subagent each, and a summary written by code rather than by a model. The plugin is self-contained: installing it installs the criteria too. The fifteen checks Each check is one failure mode of an agent loop, and states what to look at. State stays self-consistent Check What it examines C1 ★ Tool-call pairing completeness Whether every way tool execution can end — error, timeout, cancellation included — still records a result, and whether one failure in a parallel batch discards the rest. C2 History is append-only Whether anything edits a message after it was added, rather than only appending new ones. C3 Crash and checkpoint semantics What a reader sees if the process dies between the two halves of a multi-part write, and whether that half-written state is detectable on restart. Untrusted input is treated as untrusted Check What it examines C4 ★ Model output parsing How model output is parsed: whether malformed arguments, a truncated stream, or a repeated call id are handled rather than t…

Category