WebSearch tool returns "Did 0 searches in 3s" — no searches performed

[dtereshenko]

Description

The WebSearch tool in Claude Code completes without error but performs zero
actual searches. The UI shows "Did 0 searches in 3s" every time the tool is
invoked, regardless of query. Web search results are never returned.

Steps to Reproduce

1. Open Claude Code (CLI) connected to the CodeMie ai-run-sso provider
2. Ask a question that triggers the WebSearch tool (e.g. "search the web for X")
3. Observe the tool output: "Did 0 searches in 3s"

Expected Behavior

WebSearch should execute the query and return relevant search results.

Actual Behavior

Tool reports "Did 0 searches in 3s" — silently completes with no results.

Environment

Field	Value
OS	Darwin 25.4.0 arm64
Node.js	v24.13.0
npm	11.11.0
CodeMie CLI	0.1.0
Shell	/bin/zsh
Terminal	Apple_Terminal

Active Profile

• Provider: ai-run-sso
• CodeMie URL: https://codemie.lab.epam.com
• Model: claude-sonnet-4-6

Installed Agents

• CodeMie Code (0.0.47)
• Claude Code (2.1.114)

CodeMie Doctor Output

Full doctor output

Node.js: ✓ v24.13.0
npm: ✓ 11.11.0
Python: ✓ 3.9.6
uv: ○ not found (optional)
AWS CLI: ⚠ not found (optional)
Active Profile: default | Provider: ai-run-sso | Model: claude-sonnet-4-6
CodeMie SSO: ✓ operational, session expires in 8h
Installed Agents: CodeMie Code (0.0.47), Claude Code (2.1.114)

Recent Errors

WARN/ERROR entries from recent logs

[2026-05-15] [WARN] [hook:SessionEnd] No transcript_path in event, skipping incremental sync
[2026-05-15] [ERROR] [ConversationApiClient] Non-retryable error (403): [object Object]
[2026-05-15] [ERROR] [conversation-sync] Failed to sync conversation: API error: 403

Full Debug Logs

https://gist.github.com/dtereshenko/163827f4db3aacda93de0afa88846bf4

[aolshevskiy]

same here: Debian Linux Testing, Claude Code 2.1.116.

[tomke92z]

Same here. Claude's analysis of the problem:

  what happened:                                                                
                                                                                
  The Problem                                                                   
                                                                                
  When I called WebSearch with these queries:                                   
  (...)
                                                                                
  Instead of returning web search results, the tool returned responses that look
   like Claude (me) responding to those queries as if they were jailbreak       
  attempts:                                                 
                                                                                
  "I appreciate your interest, but I need to clarify something important:
  I'm Claude, an AI assistant made by Anthropic. I'm not actually "Claude Code" 
  or a CLI tool..."                                                             
                                                                                
  Why This is Wrong                                                             
                                                            
  The WebSearch tool should return:                                             
  - Actual web page results                                 
  - Links and snippets from real websites                                       
  - Search engine results                
                                                                                
  But instead it's returning:                                                   
  - Meta-commentary about Claude's identity                                     
  - Responses treating the queries as attempts to confuse the AI                
  - No actual search results at all                                             
                                                                                
  Root Cause Analysis                                       
                                                                                
  It appears the WebSearch tool is somehow:                                     
  1. Passing the query to another Claude instance (or reflecting back to me)    
  2. That instance interprets technical queries about "Claude Code" as identity 
  confusion                                                                    
  3. It responds defensively rather than performing an actual web search        
  4. The "search results" are actually Claude's own responses           
                                                                                
  This is clearly a bug in the WebSearch tool implementation - it's not actually
   searching the web, it's feeding queries to an AI that responds as if being   
  jailbroken.```