Decisions Router

Decisions: Router and Agents

This page holds the decisions D12, D13, and D25. Read it to learn how dotclaude sets the model and the effort of each subagent. The index is Decisions. The reversed parts of D25 are in Decisions-Reversed.

D12 Agent Frontmatter Search

Problem: Each subagent needs one model and one effort. The values must be known before a run starts.

Choice: The front matter of each file in src/plugins/dotclaude/slices/agents/agents/ sets model and effort. The 1.0.0 values are in .dotclaude/spec/changes/1.0.0/workflow.md, in the section Agents. The values are fixed, so a run uses the same pair each time.

Alternatives: A data file or a code list that maps a task class to a model and an effort. The choice rejects this. The front matter already holds the choice, and a second copy would drift (Agents).

How the values are set: The evals run the cases of one agent role. They try pairs from the cheapest. They stop at the first pair with no loss in the resolve rate, the over-claim rate, and the policy-violation rate. The values of 0.28.0 came from one eval report and the effort guide of Anthropic. The user decided this on 2026-10-09 (Evals History). The 1.0.0 values come from the 1.0.0 design. The research gives no effort cost index for the 5.5 models. Each row above a measured pair is inference, not a measurement.

Evidence for the pass condition: The effort guide says that Haiku 5.5 at low can skip a search, stop early, or skip a check. The research of 2026-10 recorded this statement. The source is external/dossier/agents-roster.md, section 4, and it is reported, not measured. So the over-claim rate is a pass condition for each pair.

Evidence for the price of long prompts: The price table in external/dossier/agents-roster.md, section 4, gives prices for Haiku 5.5 above 100K prompt tokens. These are an input price and an output price. Each of these prices is 5 times the price up to 100K. This is derived from the table. dotclaude has no rule that moves a long task to another model.

Model ID: claude-haiku-4-5 is not in the availableModels list of the settings snippet (Settings Snippet).

D13 No Router, the Main Loop Keeps the Session Choice

Problem: Something must set the model and the effort of a turn or of a subagent. A router that does this adds a rule that a test cannot check on its own.

Choice: The main loop keeps the model and the effort that the user sets. The user sets them with /model, /effort, or effortLevel. dotclaude does not write model or effort in a hook or in a command.

The code paths are these:

• The only reads of CLAUDE_CODE_EFFORT_LEVEL and CLAUDE_CODE_SUBAGENT_MODEL_FORCE are in the drift check. They are in src/plugins/dotclaude/slices/setup/adapters/mod.ts and src/plugins/dotclaude/slices/setup/domain/facts.ts.
• The decision model of src/plugins/dotclaude/slices/decide/adapters/mod.ts sets its own effort. It sets the effort for its own call, not the effort of the main loop.

Alternatives: A deterministic router on prompt.submit and agent.spawn, with a table of model and effort pairs. The first plan of 0.28.0 had this router. The choice does not keep it, for these reasons:

• The router needs the table of D12.
• A hook that rewrites model or effort adds a rule that only a probe can prove.
• A router on tool calls was wrong on about 1 route in 6 in a cited study. The study is arXiv 2605.07112. It is cited in .dotclaude/spec/changes/1.0.0/workflow.md, and this run did not check it.

The file .dotclaude/spec/changes/1.0.0/change.md lists dynamic model routing as out of scope. The reason is that the evidence does not support it.

Subagent fallback: The snippet sets CLAUDE_CODE_SUBAGENT_MODEL to claude-sonnet-5-5. This value applies only to a subagent that has no model of its own (Settings Snippet Keys). Each agent file of the roster sets model. So the fallback applies to none of them.

Forced variables: The snippet never sets CLAUDE_CODE_EFFORT_LEVEL or CLAUDE_CODE_SUBAGENT_MODEL_FORCE. Each one overrides the front matter. The drift check reports when either variable is set (Architecture). Binary fact, Claude Code 2.1.296: the bundle reads CLAUDE_CODE_SUBAGENT_MODEL_FORCE. When it is set, the coordinator text of the bundle says that the model parameter is ignored. This was found by a text search of external/cc-2.1.296/strings.txt. The 2.1.294 finding that the variable also drops model from the agent files is not checked again here.

Full model ID: The agent files use the full ID claude-haiku-5-5. The Claude Code docs say that the haiku alias resolves to Haiku 5.5 on the Anthropic API. They say that it resolves to Haiku 4.5 on Bedrock, Vertex, and Foundry. The docs are recorded in Open Items. This page does not probe them.

D25 Agent Roster

Problem: Which subagents exist, and which work each one takes.

Choice: A fixed roster of six agents. Each agent has a set model, effort, turn bound, and tool list. The roster is the set of files in src/plugins/dotclaude/slices/agents/agents/. The test-runner agent of workflow.md, in the section Agents, is replaced by the dotclaude-run-checks CLI of the checks slice. The reason is item 8 of .dotclaude/spec/changes/1.0.0/evidence/zero-token.md.

• The roster, the turn limits, and the pins are in Agents.
• The turn bounds are in src/plugins/dotclaude/slices/agents/limits.ts as MAX_TURNS. Each bound leaves room for the report. The design of Agents describes this.
• No agent has the Agent tool. Each file of src/plugins/dotclaude/slices/agents/agents/ lists Agent in disallowedTools.
• Planning is a subagent in 1.0.0. It is the planner agent, which uses Opus at high and writes only under spec/. Orchestration stays in the main loop. The 0.x rule that kept planning in the main loop is reversed (Decisions-Reversed).
• A coordinator that needs a stronger model for a failed fix or an audit may pass model in the Agent call. The 2.1.296 bundle has a model parameter on that call (see D13). Whether the call takes effort is not settled in this run. The 1.0.0 workflow also says that the main session does not set model on a spawn. So the two rules must be read together before the release.
• Fable 5.1, or Mythos 5.1 for an account in CVP, has no agent. It runs only when the user selects it (Settings Snippet).
• The settings snippet turns off the five built-in agents. Explore and Plan are turned off with CLAUDE_CODE_DISABLE_EXPLORE_PLAN_AGENTS. general-purpose, claude-code-guide, and statusline-setup are turned off with one permissions.deny entry each. Each entry is Agent(<name>) (Settings Snippet). The 2.1.296 bundle has all five names. This is a text search of external/cc-2.1.296/strings.txt. The effect of each deny entry is not probed in this run.
• Each listed built-in agent adds a line to each request (Off Switches). The cost is per request, not per start.

Related Pages

• Agents
• Usage Evidence