Agents

Agents

dotclaude defines five agents in plugins/dotclaude/agents/. This page lists the roster, the pins, and the rules for the files. The decisions are D12, D13, and D25 in Decisions.

Roster

Agent	Work	Model and effort in the file	maxTurns
investigator	Read-only answers from files, logs, git history, dependency data, or web pages.	Haiku 5.5 medium	31
implementer	One scoped slice: a feature, a fix with a known cause, tests, docs, or a specified bulk change.	Haiku 5.5 medium	61
debugger	A failure with an unknown cause, or a failed fix.	Haiku 5.5 high	61
reviewer	Read-only review of code, security, an API, a plan, or PR comments.	Sonnet 5.5 medium	41
test-runner	Runs checks and returns the exact failures. It has omitClaudeMd.	Haiku 5.5 low	16

• web-researcher does not exist. investigator takes its web lookups.
• No agent has the Agent tool, so no agent starts another agent.
• Each maxTurns holds one spare turn for the report. A subagent that reaches maxTurns returns no report.
• The file is the only source of the choice. No router rewrites it, and a coordinator can pass model and effort in an Agent call (Decisions: router and agents).
• Planning and orchestration stay in the main loop.
• Fable 5.1 and Mythos 5.1 are in no file. They run only when the user selects them.
• The Explore, Plan, general-purpose, claude-code-guide, and statusline-setup agents are off, so the listing holds only these five (Off switches).

Pins

The frontmatter of agents/*.md is the only source of model, effort, and maxTurns. No code list repeats them. The pin test in tests/dotclaude/agents.test.mjs does these checks:

• The model is one of claude-opus-5-5, claude-sonnet-5-5, and claude-haiku-5-5. claude-haiku-4-5 fails the test.
• The effort is in SUBAGENT_EFFORTS for that model, and it is not max. SUBAGENT_EFFORTS is in features/statusline/limits.mjs, and the status line marks an effort that is not in the list.
• The roster is these five names and no other.

The turn bound reads maxTurns from the file, and tests/dotclaude/context/turn-bound.test.mjs tests it.

Description rules

• A description has 300 characters or fewer. AGENT_DESCRIPTION_MAX_CHARS in tools/limits.mjs holds the bound, because only tests read it.
• A description says what the agent does and when to delegate to it.
• A description uses no forceful word.
• The agent listing goes in each request, so a long description costs tokens for each request.

Body rules

A subagent does not inherit the output style or the shared sections of the system prompt. Each body therefore repeats the rules that matter for its role:

• The scope rule and the verification rule for the role.
• The read rule: search first with Grep, then read the matched range with Read and its offset and limit (Context bound).
• The report bound REPORT_MAX_WORDS (400 words), with the value from features/context/limits.mjs.
• The keep-the-work rule and the block rule come in the agent.spawn block, and not in the body (System prompt).
• What to do when a check fails: debugger and implementer fix the code or report the failure with the evidence. They do not raise a bound, weaken or skip a test, or edit a check, a config, or a hook unless the brief says that the user asked.
• debugger does not change a test. When it thinks that a test is wrong, it reports that with the evidence, and the parent decides.
• reviewer looks for tests of prose, a raised bound, a skipped test, and an edited lint or CI config.

Related pages

• Decisions: router and agents
• Core plugin
• Context bound