dsh-instruction-audit

In DeepSeek Harness environments with nested workspaces or frequently shifting boundaries, it is difficult to trace whether an AGENTS.md or CLAUDE.md instruction was actually injected into a session and at what logical scope. dsh-instruction-audit fills that audit gap by emitting a privacy-minimal, structured receipt for every instruction event the Harness genuinely places into a session, giving you a verifiable trail of instruction influence without combing through raw logs.

Each receipt records only the session ID, event sequence and timestamp, baseline flag, logical scope, display path, action type, and a producer-owned content digest. The plugin never reads instruction files, never stores instruction text, never modifies a workspace, and never alters prompt assembly; it passively observes the durable agent-instructions message source already emitted by the active Harness. The shipped bundle carries no runtime npm dependencies – it imports only the Cordis context type at build time – and is delivered as a standalone tarball.

It is most useful for teams that need to confirm which instruction files actually participated in a session after a nested-workspace boundary incident or an instruction-budget anomaly, and for compliance reviews that require a JSONL artifact proving no instruction body was persisted.