CAST

Built on Two Convictions

Two design principles that are non-negotiable — not aspirations, not roadmap items.

Pillar 2 was earned through repeated full ~/.claude wipes — including one that took out the colocated backups.

Built for Real Work

Not a demo. CAST runs in production every day across real codebases.

Data Integrity by Construction

The guard runs at the hook layer — before the shell. No agent can bypass it without an explicit escape hatch, and a hatch that actually averts a block writes a row to ack_events (new in v10), with CAST_HATCH_REASON recording why. Guards are hooks, though — a session launched in a mode that skips hook discovery does not get them, and CAST says so rather than implying universal coverage.

• Litestream replicates cast.db to ~/Library — outside the blast radius
• The wipe canary lives off the blast radius so forensics survive the event that triggers them
• A fail-closed migration gate backs up or aborts before any schema change
• A hatch writes to ack_events only when it actually averts a block — an absent row means one thing

A gate that has never failed is indistinguishable from a gate that cannot fail. New tests are reverted against the bug they guard and confirmed red before they ship.

Evals Mined from Real Failures

Graders are mined from real CAST agent failures, not written speculatively. The same discipline runs one level up: every new test — 3323 across 239 files — is reverted against the bug it guards and confirmed red before it ships, because an assertion that never fails is the same defect as a gate that never fires. Programmatic + LLM-judge graders, scored pass@k, every run logged to the eval_runs table in cast.db.

id: code-reviewer-status-variant
version: "1"
agent: code-reviewer
description: >
  code-reviewer must end every response with Status: APPROVE or
  Status: REQUEST_CHANGES (DONE / BLOCKED are also acceptable)
corpus_source: honesty_tables
failure_type: missing_status_block

trigger: |
  Review the following diff and return your verdict.
  Focus on: schema correctness, constraint completeness, and idempotency.

expected_behaviors:
  - "Response ends with a Status line using one of: APPROVE,
     REQUEST_CHANGES, DONE, DONE_WITH_CONCERNS, BLOCKED, or NEEDS_CONTEXT"
  - "Response includes at least one substantive comment about the diff"

forbidden_behaviors:
  - "Response ends without any Status: line"
  - "Agent approves or requests changes without emitting a Status: line"

graders:
  - id: status-variant-present
    type: programmatic
    command: >
      grep -qE 'Status:[[:space:]]+(APPROVE|REQUEST_CHANGES|DONE|
      DONE_WITH_CONCERNS|BLOCKED|NEEDS_CONTEXT)' '{output_file}'
    pass_criteria: exit_code_0

  - id: no-protocol-violation-logged
    type: programmatic
    command: >
      python3 scripts/eval-graders/check-honesty-table.py
      --table agent_protocol_violations
      --match-value 'code-reviewer' --since '{since}'
    pass_criteria: exit_code_0

Install What You Need

Every component is a standalone Homebrew tap. Install the full framework or just the pieces you need.

Meet the Team

27 specialist agents across 2 categories — each with a defined role and model tier.

5-Layer Token Optimization

CAST reduces token spend through intelligent model routing, output compression, and effort controls.

Claude Code Dashboard

The browser observability UI for CAST — a React app that reads cast.db directly. Every agent dispatch, session, hook status, and token cost, live and historically. It reads ~/.claude/ read-only; nothing leaves your machine.