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 +1,52 @@
+# Claude Code: Escalate hard decisions with the advisor tool
+
+## Metadata
+
+| Field | Value |
+|---|---|
+| Institute | Molodetz Institute — Research & Retrieval Division |
+| Subject | cap-protocol |
+| Document type | Source Dossier |
+| Tier | 1 |
+| Source title | Claude Code: Escalate hard decisions with the advisor tool |
+| Source URL | https://code.claude.com/docs/en/advisor |
+| 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 | High |
+| Status | Active |
+
+## Summary
+
+Documentation of the advisor tool: the main model consults a second, typically stronger model at decision points and applies the guidance.
+
+## Key findings
+
+- Claude decides when to call the advisor (before committing to an approach, when an error recurs, before declaring done); the advisor receives the full conversation including tool calls and results.
+- It is a server tool, experimental, Anthropic API only; the user picks the advisor model (`/advisor`, `advisorModel`, `--advisor`); an advisor must rank at or above the main model.
+- Cost: advisor tokens bill at the advisor model's rates on top of the main model's; there is no setting to cap or force advisor calls.
+- The advisor adds less value on short tasks or when every turn needs the strongest model; for those, switch the main model instead.
+- Claude does not follow advice blindly: when its own evidence contradicts a claim it surfaces the conflict.
+- Comparison table: advisor (decision points, model-called), `opusplan` (plan mode), subagents with `model` (whole subtask), `/model` (from the next request).
+- Subagents inherit the configured advisor and the same pairing check applies to their own model.
+
+## Pros / strengths
+
+Shows model-as-tool shipped in a major product, with the failure modes (unavailable, declined) surfaced to the user.
+
+## Cons / limitations / caveats
+
+Experimental, vendor-specific, and the pairing table is tied to Claude models. No cap on calls.
+
+## Relevance to tai
+
+Concrete model-as-tool precedent and the comparison of four ways to combine models. The lack of a call cap is a gap tai should close with a static ceiling.
+
+## Changelog
+
+| Date | Change |
+|---|---|
+| 2026-10-25 | Initial dossier created. |