This page describes the six role agents of the core plugin claudetto. It says what each agent file sets, which hooks bound the agents, and which tests and evals pin them. It is for a contributor who edits an agent file or a test of the agents.

Each agent is one Markdown file in src/plugins/claudetto/slices/agents/agents/. The build copies each file to plugins/claudetto/agents/. The copy is generated, so edit only the source.

Agent	Model	Effort	Tools	Delegate it for
investigator	claude-haiku-5-5	xhigh	Bash, Read, Grep, Glob	Read-only answers from many files, logs, git history, or dependency data.
implementer	claude-haiku-5-5	xhigh	Read, Grep, Glob, Bash, Write, Edit	One scoped slice with a known check. It runs in a worktree.
debugger	claude-opus-5-5	medium	Read, Grep, Glob, Bash, Write, Edit, EnterWorktree, ExitWorktree	A failure with an unknown cause, or a fix that failed.
planner	claude-haiku-5-5	xhigh	Read, Grep, Glob, Bash, Write, Edit	A plan across several files, before any edit.
researcher	claude-haiku-5-5	xhigh	Bash, Read, Grep, Glob, WebFetch, WebSearch	A claim that must match current web, vendor, or official docs.
reviewer	claude-opus-5-5	medium	Read, Grep, Glob, Bash	A read-only review of code, security, an API design, or a plan.

The model and the effort come from the role table of the big sweep (D7). A test compares the front matter of each file with that table. The test fails when an agent on Opus has the effort xhigh or max. The main session keeps the model and the effort that the user set.

Each file has a tools allowlist, and no disallowedTools line. No agent has the Agent tool or AskUserQuestion. A test reads each tool that a prompt or a fragment names in inline code, and checks that the tool is in the tools line.

The planner writes a plan only in minispec/changes/<name>/. A tool.call hook in the agents slice denies a Write or Edit call of a planner agent for any other path. The hook also denies a path with a .. part. The hook cannot see a write that the planner does through Bash, so the prompt of the planner also tells it to use Write and Edit.

A plugin agent cannot carry a hook, and no setting bounds a subagent. For this reason the slice uses the Mod API for this limit and for the report limit below.

Each agent file includes the fragment report-size. It says that the final report has 400 words or fewer. The turn.step hook of the agents slice records the context size of each subagent. When a subagent has more than 100,000 tokens of context, or 85,000 tokens on Haiku, the next tool.call hook adds a report-now note. The hook adds the note one time for each agent. The bounds are in src/plugins/claudetto/slices/agents/limits.ts. Bounds lists them.

Each agent file includes the fragment settle-it. An agent settles what the code, the docs, or a check can settle, and it states its assumption. A subagent cannot ask the user. When a choice belongs to the user, the agent gives the options in its report, with one fact for each option. The caller then asks the user. The agent does not end its report with a question, and it does not give the user a command that it can run.

Three claude plugin eval cases are in src/plugins/claudetto/slices/agents/evals/:

• big-file-subagent sends a 1 MB log search to one investigator.
• delegation-related-grouped gives four related questions to one investigator.
• delegation-related-split gives each of four questions to its own investigator.

Each case has these graders:

• An answer grader reads ANSWERS.md line by line. It fails when a numbered line names another file, or when the file is a list of all files.
• An Agent grader checks the subagent_type of each Agent call. It fails when a call uses an agent type other than investigator.
• A final-message grader fails a run that ends with a question in plain text. It also fails a run that gives the user a decision, or a command, that the agent can settle.

The tests in tests/claudetto/agents/graders.test.ts run each grader on a transcript that passes and on transcripts that fail. The transcripts are in tests/claudetto/agents/transcripts/, and a person built them by hand.

• Overview
• Architecture
• Handoff
• Bounds
• Subagents