[BUG] Claude Code Issue: Subagent Responses Not Returned to User

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?

# Claude Code Issue: Subagent Responses Not Returned to User

## Summary
Claude Code 2.1.56 consistently fails to return responses from subagent tasks (Explore, Plan, general-purpose agents, etc.) when invoked via "Ask Claude Code" or slash commands. Tasks complete successfully but output is silently dropped, showing only "Claude Code finished without returning a reply."

## Steps to Reproduce
1. Open Claude Code Command Palette
2. Invoke "Ask Claude Code" or any slash command that spawns a subagent (e.g., `/plan`, an Agent tool call)
3. Allow the agent to complete its work
4. **Expected**: Response appears in the transcript
5. **Actual**: "Claude Code finished without returning a reply" — no output

## Environment
- **Claude Code Version**: 2.1.56 (Claude Code)
- **OS**: Windows 11 Pro (10.0.26100)
- **Model**: claude-haiku-4-5-20251001
- **Frequency**: 100% (occurs every time)

## Root Cause Analysis

Debug logs from `.claude/debug/` show:

2026-04-27T13:26:58.058Z [DEBUG] Getting matching hook commands for SubagentStart with query: Explore
2026-04-27T13:26:58.058Z [DEBUG] Matched 0 unique hooks for query "Explore" (0 before deduplication)
...
2026-04-27T13:27:04.849Z [DEBUG] Getting matching hook commands for SubagentStop with query: Explore
2026-04-27T13:27:04.849Z [DEBUG] Matched 0 unique hooks for query "Explore" (0 before deduplication)

**Key findings:**
1. `SubagentStop` lifecycle events fire successfully (agents complete their work)
2. Zero hooks are configured/matched for these events (expected — no config exists)
3. **The problem**: When `SubagentStop` fires with no handlers, the harness does not return the agent's output to the user
4. Feature flag `tengu_gleaming_fair: true` is enabled but has no corresponding response-handling logic
5. Disabling the flag does not resolve the issue

## Configuration State
- `.claude/settings.json`: No hooks defined
- `.claude.json`: `"tengu_gleaming_fair": true` (enabled)
- MCP servers: tolaria, orca-note, tana-local (configured but not the root cause)

## Expected Behavior
When a subagent completes, its output should be returned to the user in the transcript, regardless of hook configuration.

## Actual Behavior
SubagentStop events fire but the response is never delivered to the user.

## Attempted Mitigations
1. ✗ Disabled `tengu_gleaming_fair` feature flag — no effect
2. ✗ Attempted to add SubagentStop hook — hooks cannot inject output back into transcript (architectural limitation)

## Hypothesis
The SubagentStop event handler in the harness is missing or broken. Either:
- The response pipeline from agent → harness → user is disconnected for `SubagentStop` events
- The feature flag requires additional internal setup that is not present
- Regression introduced in 2.1.56 related to agent lifecycle management

## Additional Context
This completely blocks use of Claude Code's agent features ("Ask Claude Code", slash commands that spawn agents). The feature is non-functional in its current state.

What Should Happen?

Receive a reply in Tolaria.md on W11

Error Messages/Logs

Steps to Reproduce

Steps to Reproduce

1. Open Claude Code (version 2.1.56)
2. Open the Command Palette (usually Ctrl+Shift+P or Cmd+Shift+P)
3. Select "Ask Claude Code"
4. Type any prompt (example: "What files are in this directory?")
5. Press Enter/Send to submit the prompt
6. Wait for Claude Code to process (you'll see it's thinking/working)
7. Claude Code will finish execution

Expected Result:
The response/answer appears in the transcript or output area

Actual Result:
Message displays: "Claude Code finished without returning a reply"
No response is shown to the user

Claude Model

Not sure / Multiple models

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

Claude Code Version: 2.1.56

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

WSL (Windows Subsystem for Linux)

Additional Information

_No response_