[BUG] Advisor tool returns redacted result despite Opus 4.7 model selected (docs say plaintext)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The advisor tool always returns advisor_redacted_result (encrypted) in Claude Code, but the API docs explicitly state that Claude Opus 4.7 as advisor returns plaintext advisor_result.

What Should Happen?

The expected behavior (per API docs) is

Which variant you receive depends on the advisor model: - advisor_result with text — "The advisor model returns plaintext (for example, Claude Opus 4.7)."

With claude-opus-4-7 as the advisor model, the response should contain advisor_result with a readable text field.

Actual behavior

The response always contains advisor_redacted_result with encrypted_content. This is true for Opus 4.6 as executor with Opus 4.7 as advisor, on a Max 20x subscription via Claude Code CLI (v2.1.119).

The CLI renders this as: "Advisor has reviewed the conversation and will apply the feedback" — the actual advice text is inaccessible even in verbose mode.

Error Messages/Logs

Steps to Reproduce

1. Set "advisorModel": "claude-opus-4-7" in ~/.claude/settings.json (or run /advisor claude-opus-4-7)
2. Start a new Claude Code session with Opus 4.6 as executor
3. Give it a multi-step task that triggers an advisor call, or just instruct it to "Call the advisor tool"
4. Observe the advisor result displays as: "Advisor has reviewed the conversation and will apply the feedback"
5. Press Ctrl+O to expand transcript mode: nothing happens, the full text is not available
6. The same happens if starting claude in --verbose mode.

Expected: In verbose mode or via Ctrl+O, the advisor's plaintext response should be visible.

Actual: The response is always redacted.

Claude Model

claude-opus-4-6

Is this a regression?

No

Last Working Version

_No response_

Claude Code Version

2.1.119 (Claude Code)

Platform

Anthropic API

Operating System

macOS and Linux

Terminal/Shell

iTerm2

Additional Information

Why this matters

• Transparency: In any other service, paying for compute entitles you to the output. Here you're paying for Opus 4.7 inference at Opus rates, and the product of that computation is deliberately hidden from you while being consumed on your behalf. It's a novel and somewhat uncomfortable arrangement, essentially asking customers to trust a black box.
• Debuggability: When the executor makes a questionable decision, knowing whether it followed or ignored the advisor's guidance is essential for understanding what went wrong.
• Trust: Encrypting the response (not just collapsing it in the UI) actively prevents inspection even from users who explicitly want it.

Request

If redaction is intentional for Claude Code / Max users despite the docs, please document that distinction and provide an opt-in to receive plaintext responses.

If this is a bug, the fix should be straightforward since the codebase already handles advisor_result rendering in verbose mode.

curl -s https://api.anthropic.com/v1/messages \
  --header "x-api-key: $ANTHROPIC_API_KEY" \
  --header "anthropic-version: 2023-06-01" \
  --header "anthropic-beta: advisor-tool-2026-03-01" \
  --header "content-type: application/json" \
  --data '"'"'{"model":"claude-sonnet-4-6","max_tokens":256,"tools":[{"type":"advisor_20260301","name":"advisor","model":"claude-opus-4-7"}],"messages":[{"role":"user","content":"Call advisor."}]}'"'"'