research: CAP model handling, model-as-tool and skill delegation (cap-protocol, +59 sources, 2 conclusions)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
2026-10-25 18:24:40 +01:00
parent 0ac0584a96
commit 636886201b
64 changed files with 2026 additions and 5 deletions
@@ -0,0 +1,51 @@
+# Claude Code: Custom subagents
+
+## Metadata
+
+| Field | Value |
+|---|---|
+| Institute | Molodetz Institute — Research & Retrieval Division |
+| Subject | cap-protocol |
+| Document type | Source Dossier |
+| Tier | 1 |
+| Source title | Claude Code: Custom subagents |
+| Source URL | https://code.claude.com/docs/en/sub-agents |
+| Source publisher / author | Anthropic |
+| Source type | vendor docs |
+| Source date (published) | Unknown — not stated |
+| Date researched | 2026-10-25 |
+| Researched by | tai (agent), model: Claude Sonnet 5, on behalf of retoor |
+| Research method | direct page fetch |
+| Confidence | Medium |
+| Status | Active |
+
+## Summary
+
+Claude Code's documentation of subagents: file format, the `model` field, model resolution order, delegation modes and context isolation. Only the first 100000 of 116349 characters were read.
+
+## Key findings
+
+- Frontmatter: `name`, `description` (when to delegate), `tools` (inherits all if omitted), `model` (`sonnet`, `opus`, `haiku`, `fable`, a full id, or `inherit`).
+- Model resolution order: per-invocation `model` parameter, then the definition's `model`, then the `CLAUDE_CODE_SUBAGENT_MODEL` environment variable, then the main conversation's model; before v2.1.251 the environment variable came first.
+- Delegation is automatic from descriptions (include 'use proactively' to encourage it), by natural language, by @-mention (guarantees the run), or session-wide.
+- Each subagent has its own context window, system prompt, tool access and permissions; a non-fork subagent does not see the conversation history and works from a delegation message; only a summary returns.
+- Subagent requests count toward the same usage limits as the main conversation; costs are controlled by routing tasks to faster, lower-cost models like Haiku.
+- Combined subagent descriptions above 15,000 tokens trigger a startup warning; use the main conversation for frequent back-and-forth or quick targeted changes.
+
+## Pros / strengths
+
+A shipped product's explicit, ordered precedence for the model of a delegated worker, and a cost-routing rationale.
+
+## Cons / limitations / caveats
+
+Unread tail of the page; the precedence change between versions shows the rule is unstable across releases.
+
+## Relevance to tai
+
+Precedent for a tai rule: per-call model beats the agent's configured model beats a global default beats the parent's model. Also shows 'description decides delegation' costs context, which argues for few, well-described delegation tools.
+
+## Changelog
+
+| Date | Change |
+|---|---|
+| 2026-10-25 | Initial dossier created. |