Skip to content

Specsmith

Lean Applied Epistemic Engineering governance for AI-assisted development.

Intelligence proposes. Constraints decide. The ledger remembers.

Specsmith sits around the coding agent and tools you already use. It keeps requirements durable, links them to enforceable tests, bounds epistemic context, and records evidence so a plausible model answer cannot silently become a fact.

The essential loop

requirement -> linked test -> preflight -> native implementation
            -> observed validation -> verification -> checkpoint

The host agent owns code editing, Git, browsers, deployment, and framework skills. Specsmith owns scope, test traceability, uncertainty, confidence gates, and durable evidence.

Quick start

pipx install specsmith
cd your-project
specsmith import --project-dir .
specsmith audit --project-dir .
specsmith preflight "Fix the pagination boundary. Scope: REQ-123" --json
# Edit and test with your normal tools.
specsmith verify --project-dir .
specsmith checkpoint --project-dir .

Generate a focused host integration with specsmith integrate, or start Grace with specsmith run when a terminal/local-model fallback is useful.

Core capabilities

  • durable requirements with explicit epistemic boundaries;
  • tests linked to accepted requirements;
  • deterministic mutation, destructive-operation, and release gates;
  • compact context that preserves knowns, unknowns, and provenance;
  • bounded verification based on observed diffs and test results;
  • SQLite evidence storage with an optional ChronoMemory backend;
  • MCP and focused integrations for common coding agents;
  • safe repair of older Specsmith-managed Zoo/Roo Code configuration;
  • equivalent behavior on Windows, Linux, and macOS.

Grace

Grace is the friendly optional REPL, not a separate orchestration layer. It uses the same preflight, context, verification, and evidence paths as every external integration. /help, /status, /why, provider switching, and actionable error guidance make the first local session straightforward.

Governance efficiency evidence

The published benchmark reports task-dependent results rather than a universal claim. Current evidence identifies where lightweight governance helps, where it adds token cost, and which paths need improvement. See the governance efficiency report and model comparison.

Documentation