Features · Senior-dev agents
Scoped subagents that Claude delegates to. Each has its own tool allowlist, system prompt, and fresh context window per invocation. Stateless — continuity lives in artifacts, not agent memory.
Installed agents
code-reviewerReviews changes for the Next.js + Supabase failure modes. Knows to check RLS on new tables, credential shapes in diffs, Server Action return types.
plannerDrafts implementation plans before code. Sole writer to __plan/ so plan state has one canonical home.
test-runnerRuns the test suite, interprets failures, suggests minimal fixes. Doesn't write code — returns a verdict plus the relevant failure block.
security-reviewerRead-only auditor for license, URL, hash, and consent surfaces. No memory across invocations — fresh eyes every time.
The 7-agent default set
The kit installs four agents for the Next.js + Supabase stack. The underlying framework ships seven standard roles. Any project can instantiate the full set — or add its own specialists.
plannerSole writer to __plan/. Enforces Protocol 0 audit before any epic. Tools: Read, Edit/Write (in __plan/ only), Grep, Glob.
docs-writerDrafts task files and source-of-truth from planner outlines. Never decides scope. Tools: Read, Edit, Write, Grep, Glob. Model: Sonnet.
explore-codebaseFast read-only search. Tools: Read, Grep, Glob, Bash. Model: Haiku — cost optimisation.
domain-specialistImplementation in a specific domain. One instance per domain area. e.g. mcp-specialist (server), web-specialist (landing + app), claude-code-specialist (kit content). Prevents two agents touching the same file.
quality-auditorRead-only review of license, URL, hash, and consent surfaces. e.g. security-reviewer in claudekit. No Write tool.
claude-setup-curatorOptional. Only when the project consumes __claude_setup/ directly. Not relevant for installed kits.
+ your ownTester, planner alternates, data-pipeline specialists. Add by dropping an agent markdown file in .claude/agents/.
Agent frontmatter fields
nameAgent identifier. Used when dispatching from chat.
descriptionWhen to delegate to this agent. Read by Claude when choosing which agent to invoke.
modelhaiku / sonnet / opus. Default: sonnet. Haiku for search-heavy tasks; Opus for multi-step planning.
toolsExplicit tool allowlist. Defaults to * if omitted. Restrict here — not in CLAUDE.md.
memory: projectOptional. Gives the agent an append-only JSONL memory file scoped to the project.
isolation: worktreeOptional. Agent works in a temporary git worktree. Edits do not collide with the main checkout.
Model-cost principle
Read-only search (grep, glob, read)
Haiku
Fast and cheap. No generation needed.
Drafting prose or code
Sonnet
Good quality/cost ratio for most implementation.
Architecture, multi-step planning
Opus
Reserved for high-stakes decisions that benefit from deeper reasoning.
Domain-vs-stack scoping
A web-specialist owns the landing page and dashboard — all web surfaces, regardless of stack. A mcp-specialist owns the server and tool schemas. Scoping by domain, not by framework version or file extension, prevents the "two agents touching the same file" race condition.
The standard forwarding convention is "pass to the planner: <finding>". Any dialog can hand the planner a discovery from its domain. Only the planner writes to __plan/. Cross-agent communication routes through a written artifact — never through agent-to-agent calls.
Agent memory
Agents with memory: project write to an append-only JSONL file, one per agent, scoped to the project. An AST-level test enforces that no delete, update, clear, or truncate operations appear in the memory code path. Memory entries can accumulate; they cannot be removed programmatically.
On read, entries are recency-weighted. A baseline_anchor field lets the agent mark a recurring decision so it is loaded on every session without counting against the recency budget.
Pattern-reinforcement defense: if memory keeps returning the same answer, the kit alerts the human. A memory that always says the same thing is either a baseline anchor working correctly — or a poisoning signal. The alert surfaces the distinction.
Worktree isolation
isolation: worktree creates a temporary git worktree per agent invocation. Two agents editing the same file in parallel work in separate directory trees. Their edits do not collide.
The worktree is cleaned up automatically if the agent makes no changes. If it does make changes, the worktree is kept for explicit merge or discard — not silently merged.
Why scoped agents
A reviewer that also writes code drifts into justifying its own changes. A planner that also runs tests loses focus on the plan. Tight scope means tight accountability — each agent can be evaluated on exactly one outcome.
The tool allowlist enforces the scope at the runtime level. security-reviewer has no Write tool. It cannot make changes even if asked. That's not a limitation — it's the spec.