Agent Harness Kit

The problem it solves: multi-agent workflows have a hidden failure mode where dependency state changes between sessions and nobody notices until the Builder is halfway through an implementation that targets the wrong version of a library. The Explorer mapped things correctly — at that time. By the next session, someone ran an install, something got bumped, and now the plan is subtly out of sync with reality.

The Consultant is a read-only advisory agent that runs conditionally between Explorer and Builder. It uses two new MCP tools exclusive to its role: deps.snapshot (saves a baseline of package.json to .harness/deps-lock.json) and deps.check (diffs the current state against that baseline, returning added/removed packages and major version bumps with a significant: true flag). When deps.check returns significant changes — or when the task explicitly touches package.json — the Lead invokes the Consultant before Builder starts. The Consultant reads the Explorer's analysis, checks deps, writes structured advisory covering patterns to follow, risks to avoid, and dependency-specific notes, then hands off. It does not write code. It cannot. No write permissions, by design.

The release also includes a mandatory docs/README analysis criterion on every task — a final acceptance item the Reviewer enforces before anything closes. Forces explicit documentation accountability without making it a policy nobody reads.