Advisor always 'unavailable' in sessions with many deferred MCP tools — fails at ~56K tokens while slim-tool session succeeds at ~62K

Environment

• Claude Code versions: 2.1.200 – 2.1.203
• OS: Windows 11 Pro
• Main model: claude-fable-5
• Setting: "advisorModel": "fable" (valid pairing per the advisor docs)
• Entrypoints affected: cli, claude-vscode, and sidechain subagents

Bug description

Every advisor call in normal interactive sessions fails with:

● Advising using Fable 5
  ⎿  Advisor unavailable (unavailable)

The session transcript records the raw server result as:

{"type":"advisor_tool_result","tool_use_id":"srvtoolu_...","content":{"type":"advisor_tool_result_error","error_code":"unavailable"}}

Scanning all local transcripts over ~4 weeks: 352 advisor calls → 350 × unavailable, 2 × overloaded, 0 successes.

Key evidence: fails small with many deferred tools, works bigger with few

The account and model pairing are fine — the advisor works in minimal sessions. Same account, same fable + fable pairing, same day:

| Session type | Deferred tools present | Total context at advisor call | Result |
|---|---|---|---|
| Interactive (normal environment) | ~160 (built-ins + one stdio MCP server exposing ~130 tools + a browser MCP + 2 connectors) | ~56K tokens (input 8,851 + cache_read 18,017 + cache_creation 29,668) | unavailable |
| Interactive (normal environment) | ~160 | 56K – 460K | unavailable (every call) |
| Headless claude -p with --strict-mcp-config --mcp-config '{"mcpServers":{}}' | ~25 (built-ins only) | ~31K | success |
| Headless claude -p, default config (only an 8-tool MCP connector finished connecting; the 130-tool stdio server was still connecting) | ~33 | ~62K | success |

A 56K-token session with the full deferred-tool list fails while a 62K-token session with a slim list succeeds — so the trigger correlates with the deferred-tool load, not transcript size. (Long sessions here also exceed 100K tokens routinely, which may compound with #67609, but that can't explain the 56K failure.)

Repro note for anyone testing headless: slow stdio MCP servers (launched via npx) don't finish connecting in a one-shot claude -p run, so their deferred tools never register — headless tests understate the interactive deferred-tool load.

Expected behavior

Advisor calls succeed, or fail with a specific, actionable error code instead of the catch-all unavailable.

Related issues

Likely the same or overlapping root cause: #73923 (deferred tools present → advisor unavailable), #67609 (large transcript → unavailable), #73365 (deterministic Fable advisor failure), #66784 (all failures flattened to unavailable).

Happy to share request IDs from the failing transcripts privately if useful for server-side lookup.