Claude Agent Specialist Team

CAST

27 specialist agents that plan, implement, review, and commit — on a local-first SQLite record that isn't just observability: it acts.

$ /plugin install cast

Copies marketplace + install + enable.

Star on GitHub Read the Docs

0Agents

0Tests

0Repos

Scroll

cast-session

$ cast status

CAST v9.5.3 — native Claude Code plugin

Agents: 27 (16 haiku · 10 sonnet · 1 opus)

Commands: 21 · Skills: 17 · cast.db: 39 tables

Last dispatch: code-writer → code-reviewer ✓ DONE

Quality gates: raw git commit/push hard-blocked

Backups: Litestream active

Plugin: cast@cast enabled

Ledger: signed ✓ · chain verified

Built on Two Convictions

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

Local-first by construction

The core development loop never has to leave your machine. CAST runs inside ~/.claude/ with the cloud strictly opt-in; it ships as a native Claude Code plugin and works fully offline.

Data integrity by construction

An agent platform should be structurally unable to destroy its own evidence. Backups live outside the failure domain and the runtime cannot delete itself.

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.

Agent Teams

27 specialists across 2 tiers — dispatched natively through Claude Code's own Agent tool with isolated execution contexts and quality gates. No custom orchestration layer.

Quality Gates

Raw git commit and push are hard-blocked by hooks. Structured Output schemas validate every agent response. Code ships through the commit agent or not at all.

Full Observability

cast.db tracks 39 tables across sessions, agent runs, telemetry, and quality gates. A browser dashboard and a native desktop app read from it — real-time SSE streaming, dashboard views, and an embedded PTY terminal.

Token Optimization

5-layer cost reduction: model tiering (16 Haiku / 10 Sonnet / 1 Opus), Laconic Mode, RTK output compression, Effort Routing, and Compact Discipline.

The Record That Acts

cast.db is not just observability — it is the substrate decisions run on. cast ask searches the full execution record. cast predict reads your telemetry to surface routing suggestions. cast ledger signs every session with SHA-256. A weekly record-review job mines that same telemetry to propose its own fixes, then re-audits its own work.

Extensible

Agents are markdown files. Hooks are bash scripts. Plans are JSON. 17 skills, 21 slash commands, and cast mcp — a read-only stdio MCP server over cast.db that any MCP-aware tool can query.

Provenance & Audit

cast ledger stamps every session with a SHA-256-signed receipt. cast verify-chain builds a hash-chain across sessions — a tamper-evident provenance trail for every agent dispatch and commit.

Local-First

Zero cloud lock-in. Everything runs on your machine — agents, hooks, memory, database. Ollama fallback via LiteLLM proxy. All data stays in SQLite.

Data Integrity by Construction

The guard runs at the hook layer — before the shell. No agent can bypass it without an explicit escape hatch.

cast-command-guard

$ rm -rf ~/.claude

**[CAST]** Catastrophic `rm -rf` of a protected path blocked. Escape hatch: prefix the command with CAST_RM_OK=1.

exit 2

$ pkill -9 node

**[CAST]** Dangerous process-kill blocked (pkill/killall, or kill of a process group / all processes). Escape hatch: prefix the command with CAST_KILL_OK=1.

exit 2

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

The destructive paths are tested by proving the system refuses them — not by proving the happy path works.

How It Works

Hooks intercept tool calls. Directives route to specialists. Every result lands in cast.db — and the record isn't passive: it drives decisions through dispatch prediction, cost attribution, and signed audit receipts.

User Prompt

CLAUDE.md Dispatch27-agent routing table

27-agent routing table

PreToolUse HooksBlocks raw git commit/push

pre-tool-guard.sh, cast-audit-hook.sh, cast-headless-guard.sh

Blocks raw git commit/push

The Record Actsask · predict · ledger · verify-chain · record-review

cast.db written by hooks, decisions run on the record

ask · predict · ledger · verify-chain · record-review

Agent DispatchSpecialist agent runs in isolated execution contexts (Managed Agents or worktrees)

SubagentStart hook fires, task_claimed → cast.db

Specialist agent runs in isolated execution contexts (Managed Agents or worktrees)

PostToolUse HooksInjects [CAST-REVIEW] after writes

Injects [CAST-REVIEW] after writes

Post-Chain Protocolcode-reviewer → commit → push

Quality gate validation, Structured Output schema check

code-reviewer → commit → push

{
  "files_changed": ["src/components/Hero.tsx"],
  "status": "DONE",
  "blockers": "none",
  "key_decisions": "Reused existing CopyButton"
}

Every agent emits this typed contract — the orchestrator injects it verbatim into the next agent's prompt, so failures can't cascade silently.

cast.db39 tables, WAL-mode, queryable

sessionsagent_runsquality_gatesrouting_eventsagent_memoriesdispatch_decisionsotel_eventsincidents

39 tables, WAL-mode, queryable

Cast DesktopNative macOS app — terminal + dashboard views

DashboardSessionsAnalyticsCostWork LogAgentsSystemDocsMemoryRecordProvenance

Native macOS app — terminal + dashboard views

cast.db

Written by hooks. Read by everything.

cast.db is a local SQLite database populated entirely by CAST hook scripts — no app writes it, no cloud touches it. Every agent run, quality gate result, cost, memory, and governance event lands here. Cast Desktop and the browser dashboard are read-only consumers. Any SQLite tool can query it directly.

39Tables

WALMode

SQLiteEngine

LocalOnly

Core

sessionsagent_runsrouting_eventsagent_memories

Quality

quality_gateshook_failuresagent_hallucinationsagent_protocol_violations

Governance

injection_logdispatch_decisionsunstaged_warningsschema_migrations

Telemetry

compaction_eventsstop_failure_eventsrate_limit_snapshotsfile_writeseval_runs

The Record Acts

cast ask runs FTS5 full-text search over every session, run, and memory. cast predict reads dispatch history to surface routing suggestions before you run a task.

A weekly cast record-review job goes further: it mines the week's telemetry for its own maintenance — proposing fixes from real data, shipping the ones a human approves, and catching blind spots in its own reasoning before they ship.

sqlite3 ~/.claude/cast.db .tables — see everything CAST has ever recorded about your agents.

Evals Mined from Real Failures

Graders are mined from real CAST agent failures, not written speculatively. Programmatic + LLM-judge graders, scored pass@k, every run logged to the eval_runs table in cast.db.

evals/cases/code-reviewer/code-reviewer-status-variant.yamlF19 · corpus: honesty_tables

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

pass@k

Scored across k agent runs, not a single best-of

2 grader types

Programmatic exit-code + LLM-judge graders per case

eval_runs

Every run logged to cast.db for longitudinal tracking

Install What You Need

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

CAST Core

v9.5.3

The complete multi-agent framework

brew tap ek33450505/cast && brew install cast

GitHub

Claude Code Dashboard

v2.5.0

React observability UI — sessions, agent analytics, hook health, memory browser, SQLite explorer.

brew tap ek33450505/cast && brew install claude-code-dashboard

GitHub

Cast Desktop

v1.2.12

Tauri 2 native app — embedded PTY terminal, command palette, and dashboard views.

brew install --cask ek33450505/cast/cast-desktop

GitHub

CAST Memory

v0.4.1

Persistent agent memory for Claude Code — FTS5 full-text search, weighted relevance, temporal validity, Ollama embeddings, and weekly consolidation over cast.db.

brew tap ek33450505/cast-memory && brew install cast-memory

GitHub

Claude's Journal

v0.3.1

Three-hook journaling for Claude Code (Stop/SessionStart/UserPromptSubmit) — maintains Claude's perspective and working memory across sessions as Obsidian-compatible markdown.

brew tap ek33450505/claudes-journal && brew install claudes-journal

GitHub

CAST Time

v0.1.2

Gives Claude Code a clock — injects local time, timezone, and a semantic time-of-day bucket at every SessionStart.

brew tap ek33450505/cast-time && brew install cast-time

GitHub

CAST Doctor

v0.1.3

Standalone read-only health check for any Claude Code install — validates hooks, MCP config, agent frontmatter, cast.db core schema, and stale memories without the full CAST framework.

brew tap ek33450505/cast-doctor && brew install cast-doctor

GitHub

CAST Ledger

v0.1.0

Signed, hash-chained, tamper-evident session receipts for Claude Code — SHA-256-stamped audit receipts from cast.db with --verify, plus an optional provenance hash-chain across sessions.

brew tap ek33450505/cast-ledger && brew install cast-ledger

GitHub

CAST MCP

v0.1.0

Read-only MCP server over the Claude Code execution record (cast.db) — dispatch decisions, incidents, cost, sessions, and full-text search as 5 MCP tools + 5 resources.

brew tap ek33450505/cast-mcp && brew install cast-mcp

GitHub

CAST Predict

v0.1.0

Telemetry-driven dispatch prediction for Claude Code — reads cast.db to predict a task's likely cost, suggest agents, and surface related past incidents before you run it.

brew tap ek33450505/cast-predict && brew install cast-predict

GitHub

Meet the Team

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

Agent	Category	Model	Effort	Purpose
Core (16)
planner	Core	sonnet	high	Sequenced task plans with Agent Dispatch Manifests
backend-writer	Core	sonnet	high	Backend feature implementation across files
frontend-writer	Core	sonnet	high	Frontend feature implementation across files
bash-specialist	Core	sonnet	high	Shell scripts, BATS tests, and hook scripts
debugger	Core	sonnet	high	Root-cause diagnosis and bug fixes
code-reviewer	Core	haiku	low	Per-unit diff review for correctness and conventions
frontend-qa	Core	haiku	low	React, TypeScript, and accessibility review
security	Core	sonnet	high	Auth, input validation, secrets, and vulnerability audit
test-writer	Core	haiku	low	Unit and integration test creation
test-runner	Core	haiku	low	Test suite execution and result gating
commit	Core	haiku	low	Semantic git commits with Co-Authored-By
push	Core	haiku	low	Remote push with branch safety checks
merge	Core	haiku	low	PR lifecycle — CI watch and squash-merge on approval
docs	Core	haiku	low	Documentation, READMEs, and changelogs
researcher	Core	sonnet	high	Multi-source analysis, gap reports, and citations
morning-briefing	Core	haiku	low	Daily git activity and system summary
Dev Workflow (11)
api-contract	Dev Workflow	haiku	low	REST API breaking-change detection
db-reader	Dev Workflow	sonnet	high	Read-only SQL exploration and reporting
dep-auditor	Dev Workflow	haiku	low	Dependency audit for CVEs, licenses, and compatibility
devops	Dev Workflow	haiku	low	CI/CD and GitHub Actions workflow authoring
email-drafter	Dev Workflow	haiku	low	Drafts Gmail messages and syncs showcase READMEs
eval-writer	Dev Workflow	sonnet	high	Eval and benchmark fixture author for agent prompts
infra-writer	Dev Workflow	haiku	low	Docker, IaC, and deployment configuration
migration-reviewer	Dev Workflow	opus	high	Database schema change safety and rollback plans
pr-reviewer	Dev Workflow	sonnet	high	Holistic PR-level review at PR-open time
release-notes	Dev Workflow	haiku	low	Structured changelog generation from git history
report-writer	Dev Workflow	haiku	low	Status updates, health checks, and chain summaries

Core (16)

planner

sonnet

Sequenced task plans with Agent Dispatch Manifests

backend-writer

sonnet

Backend feature implementation across files

frontend-writer

sonnet

Frontend feature implementation across files

bash-specialist

sonnet

Shell scripts, BATS tests, and hook scripts

debugger

sonnet

Root-cause diagnosis and bug fixes

code-reviewer

haiku

Per-unit diff review for correctness and conventions

frontend-qa

haiku

React, TypeScript, and accessibility review

security

sonnet

Auth, input validation, secrets, and vulnerability audit

test-writer

haiku

Unit and integration test creation

test-runner

haiku

Test suite execution and result gating

commit

haiku

Semantic git commits with Co-Authored-By

push

haiku

Remote push with branch safety checks

merge

haiku

PR lifecycle — CI watch and squash-merge on approval

docs

haiku

Documentation, READMEs, and changelogs

researcher

sonnet

Multi-source analysis, gap reports, and citations

morning-briefing

haiku

Daily git activity and system summary

Dev Workflow (11)

api-contract

haiku

REST API breaking-change detection

db-reader

sonnet

Read-only SQL exploration and reporting

dep-auditor

haiku

Dependency audit for CVEs, licenses, and compatibility

devops

haiku

CI/CD and GitHub Actions workflow authoring

email-drafter

haiku

Drafts Gmail messages and syncs showcase READMEs

eval-writer

sonnet

Eval and benchmark fixture author for agent prompts

infra-writer

haiku

Docker, IaC, and deployment configuration

migration-reviewer

opus

Database schema change safety and rollback plans

pr-reviewer

sonnet

Holistic PR-level review at PR-open time

release-notes

haiku

Structured changelog generation from git history

report-writer

haiku

Status updates, health checks, and chain summaries

5-Layer Token Optimization

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

Model Tiering

16 Haiku / 10 Sonnet / 1 Opus — right model for each task

Laconic Mode

Terse output (lite/full/ultra) cutting prose tokens

RTK Hook

Compression on large tool outputs

Effort Routing

Per-agent effort levels routed by task complexity — native Claude Code mechanism

Compact Discipline

Auto-compaction before the context window fills

optimization layers

16/10/1

Haiku/Sonnet/Opus split

2389

tests passing

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.

SessionsAnalyticsHooksMemorySQLite Explorer

localhost:5173

v2.5.0

Views

Home

Sessions

Analytics

Agents

Evals

Hooks

Memory

SQLite

Home

cast.db tables

agents

2389

tests passing

SSE stream active

Activity

code-writer

running

code-reviewer

queued

commit

idle

brew tap ek33450505/cast && brew install claude-code-dashboard

View on GitHub

Want a native macOS app instead? Cast Desktop is the desktop companion — the same cast.db, an embedded PTY terminal, and a command palette.