[BUG] Cowork — account enrolled in two A/B variants (`testfoo` + `0526`) causing skill misfires + connector failure; `ENABLE_TOOL_SEARCH` mismatch

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?

Summary

My Cowork account (Max plan, macOS) has been unusable since May 26, 2026. Through direct inspection of local Cowork session files I've traced this to my account being enrolled in two concurrent A/B prompt variants with a configuration mismatch in the ENABLE_TOOL_SEARCH / deferred-tool-loading flags. Verified across two separate accounts on the same machine, confirming the issue is environment-bound rather than account-cohort-specific. Anthropic's Fin AI Agent has separately confirmed this requires human-engineer intervention to fix.

Posting publicly per the suggestion of other folks in similar threads to (a) surface other affected users so we can compare fingerprints and (b) provide internal engineering visibility separate from the support queue.

Environment

• Product: Cowork (Claude Desktop)
• Claude Desktop version: 1.9255.2
• OS: macOS 25.5.0
• Plan: Max (also reproduced on a fresh Pro account on a different email, same machine)
• First noticed broken: May 26, 2026
• Installed config: 33 user skills, 11 personal plugins from "My Uploads" marketplace, 11 claude.ai cloud connectors

Symptoms

Symptom 1 — All claude.ai cloud connectors fail at runtime

Connectors show "Connected" in Customize → Connectors but tool dispatch fails. Verified failing: Granola, Notion, Slack, Gmail, Google Calendar, Google Drive, Canva, Excalidraw, Figma, Higgsfield, Firecrawl. Same connectors work normally in Chat mode (different code path).

Exact error from Cowork's runtime:

Error: No such tool available: ToolSearch.
ToolSearch exists but is not enabled in this context.
Use one of the available tools instead.

And on direct tool calls:

Error: No such tool available: mcp__6ee4aa57-023b-4349-ae1f-d0da6b23ce8f__list_meetings

(UUID is the Granola connector's internal ID — the registration exists, the runtime cannot dispatch.)

Symptom 2 — Skill auto-router fires wrong skills on weak keyword matches

User skills (and Anthropic system skills like init, consolidate-memory) fire on prompts that don't match their intent. Examples from my session log:

| Prompt | Skill auto-fired | Should have fired |
|---|---|---|
| "Open Figma on my machine and teach me how to make a button" | init | computer-use / teach mode |
| "Use computer use via toolsearch" | youtube-transcript | (nothing — meta question) |
| "Can you see my meetings in granola?" | search, then youtube-transcript | Granola MCP tools |
| "See what's in my scripts database in Notion" | blackman-method:notion-review | Notion MCP tools |

Root cause analysis

Finding 1 — Account enrolled in two A/B prompt variants

From local_*.json spVariantPrompts field on the most recent task:

{
  "0526":    { "mode": "replace", "length": 47774 },
  "testfoo": { "mode": "append",  "length": 46918 }
}

The variant name testfoo does not match Anthropic's documented variant naming conventions (no tengu_* prefix, no date prefix). Whatever it is, the name is anomalous.

Finding 2 — Both variants contain a "promiscuously" directive

Both variant prompts contain the identical instruction:

"...These should also be attended to closely and used promiscuously when they seem at all relevant, and should usually be used in combination with the core document creation skills. This is extremely important, so thanks for paying attention to it."

Position in 0526: char 46391.
Position in testfoo: char 45827.

Applied twice (replace + append), this directive reinforces aggressive skill firing — which is the proximate cause of the misfire pattern documented above.

Finding 3 — Feature flag mismatch

From local cowork-gb-cache.json:

{
  "tengu_claudeai_mcp_connectors": true,
  "tengu_mcp_retry_failed_remote": false,
  "tengu_tool_search_unsupported_models": ["haiku"],
  "tengu_auto_mode_config": {
    "enabled": "enabled",
    "twoStageClassifier": true
  }
}

This combination means: cloud connectors are enabled at the feature level, but failed-MCP-loads do NOT retry, and the ToolSearch mechanism that the system prompt instructs the model to use ("load via ToolSearch first") is not actually exposed in the session. So deferred tools never become available.

Finding 4 — Bug follows the machine, not the account

I created a fresh Anthropic account on a different email (Pro plan, same Mac). Verified from its session JSON:

{
  "0526":    { "mode": "replace", "length": 47774 },
  "testfoo": { "mode": "append",  "length": 46918 }
}

Identical variant assignment. Same modes, same character lengths, same flags. This rules out account-cohort assignment as the sole cause and indicates either machine-fingerprint-based cohort assignment, or a wider rollout than the variant name testfoo would suggest.

What I've ruled out

• Fresh Claude Desktop reinstall (May 27)
• Sign out and back in
• Disconnect/reconnect every cloud connector individually
• Local feature-flag cache override (server resync overwrites local edits on app start)
• Local file edits to project memory CLAUDE.md (reverted by server resync within hours)
• Local skill deletion (re-installed by server resync)
• Switch to a different account on a different email on the same machine — same bug appears

Requested actions

1. Remove affected accounts from the testfoo variant cohort. If testfoo is not intended for production assignment, investigate whether the cohort targeting rule needs cleanup.

1. Either set ENABLE_TOOL_SEARCH to false for affected accounts (load MCP tools eagerly), or enable ToolSearch in Cowork sessions so deferred loading works as the system prompt assumes.

1. Review the 0526 prompt variant. If it's an intended production rollout, the "promiscuously" directive for user skills appears to cause runaway misfires on accounts with many installed skills. Consider gating the directive on installed-skill count or removing it entirely.

Are you also hitting this?

If you're seeing similar Cowork behavior and have access to your local session JSON (~/Library/Application Support/Claude/local-agent-mode-sessions/.../local_*.json on macOS), please comment with:

• Your spVariantPrompts keys (just the names, not the content)
• Number of installed user skills
• Whether connectors are failing for you in Cowork
• Whether Chat mode works normally

That'll help establish whether this is account-isolated or a broader cohort issue.

What Should Happen?

Cowork sessions should initialize with all claude.ai cloud connectors (Notion, Granola, Slack, Gmail, Google Calendar, Google Drive, Canva, Excalidraw, Figma, Higgsfield, Firecrawl) callable at runtime once shown as "Connected" in Customize → Connectors.

The skill auto-router should fire skills only on strong semantic matches to user intent — not fire arbitrary skills on weak keyword overlaps.

Both behaviors worked normally on this account prior to May 26, 2026.

Error Messages/Logs

From Cowork's runtime, attempting any cloud connector tool call (Granola example):

Error: No such tool available: mcp__6ee4aa57-023b-4349-ae1f-d0da6b23ce8f__list_meetings

When the model attempts to load deferred tools via the documented mechanism:

Error: No such tool available: ToolSearch. ToolSearch exists but is not enabled in this context. Use one of the available tools instead.

These errors persist across: fresh Claude Desktop reinstall, sign-out/in, disconnect+reconnect each connector, local feature-flag cache override, and switching to a different Anthropic account on the same machine.

Steps to Reproduce

This is an account/environment-state bug, not a code bug, so it isn't reproducible via a script. Reproduction on the affected account:

1. Open Claude Desktop (version 1.9255.2 on macOS)
2. Switch to Cowork mode (top of left sidebar)
3. Start a new task
4. Confirm Customize → Connectors shows Granola (or Notion, Slack, etc.) as "Connected"
5. In the task, ask: "What's my latest Granola meeting?" (or any cloud-connector tool)
6. Observe: model attempts the tool call, runtime returns "No such tool available" error
7. Also: ask any prompt with a weak keyword match to an installed skill (e.g., "what's in my scripts database in Notion") — observe the wrong skill fires (in my case blackman-method:notion-review fired on a Notion-data query)

For root cause diagnostic, see the body of the issue — local session JSON shows account is enrolled in two A/B prompt variants (testfoo + 0526) and the tengu_mcp_retry_failed_remote flag is false while tengu_claudeai_mcp_connectors is true.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Pre-May-26-2026. Exact prior Claude Desktop version unknown due to auto-update; the regression coincided with what appears to be a server-side A/B prompt variant rollout (variant 0526) rather than a client build change. Current Claude Desktop: 1.9255.2.

Claude Code Version

N/A — this issue is in Claude Desktop (Cowork mode), not Claude Code CLI. Claude Desktop version: 1.9255.2 (macOS).

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Cross-references:

• Anthropic Fin AI Agent confirmed (separate support conversation, May 27 2026) that this is a server-side feature flag mismatch requiring human engineer intervention: "a human agent from our engineering team will need to adjust the feature flags for your Cowork sessions to either enable ToolSearch or disable deferred loading." Human escalation requested via support; ticket open.
• Related GitHub issues with overlapping patterns (Cowork connector failures, OAuth/MCP loading, skill misfires): #20377, #23736, #27492, #45306, #58172, #61344.

Note: this bug-report template is optimized for Claude Code CLI bugs. Several fields (Claude Code Version, Terminal/Shell) don't map cleanly to Claude Desktop Cowork mode. Filling them with closest-match values rather than leaving blank.

Happy to share full session JSON, gb-cache snapshot, complete system prompt extracts, or join a screenshare with an engineer. The issue body contains the verified flag values, variant assignments, and runtime errors.