[BUG] [v0.1.0] handle_reject() in actions.rs never sends rejection result back to LLM, stalling the agentic loop

[climax-dev-1]

Bug Description

When a user rejects a tool execution via the Reject keybinding (handle_reject() in actions.rs), the rejection result is never fed back to the LLM as a pending tool result. This means the agentic loop stalls after a rejection — the LLM never learns the tool was rejected and cannot adapt or try an alternative approach.

Location

src/cortex-tui/src/runner/event_loop/actions.rs, lines 218–245

Root Cause

handle_reject() calls self.app_state.update_tool_result() to update the UI status, but it never calls self.app_state.add_pending_tool_result() to queue the rejection for the next LLM turn. Compare with every other tool completion path in the codebase:

• handle_tool_completed in tools.rs:407-408 → calls add_pending_tool_result
• handle_tool_failed in tools.rs:528-529 → calls add_pending_tool_result
• cancel_question_prompt in mouse.rs:542-547 → calls add_pending_tool_result
• submit_question_answers in mouse.rs:517-518 → calls add_pending_tool_result
• All subagent completion/failure paths → call add_pending_tool_result

But handle_reject() only does:

async fn handle_reject(&mut self) -> Result<()> {
    if let Some(approval) = self.app_state.reject() {
        self.app_state.update_tool_result(
            &approval.tool_call_id,
            "Tool execution rejected by user".to_string(),
            false,
            "Rejected".to_string(),
        );
        // Legacy bridge handling...
        // MISSING: self.app_state.add_pending_tool_result(...)
        // MISSING: continuation check (continue_with_tool_results)
    }
    Ok(())
}

Reproduction Steps

1. Start a conversation that triggers a tool requiring approval (e.g., Edit, Write, Execute)
2. When the approval prompt appears, press the Reject key
3. Observe that the conversation stalls — the LLM never receives the rejection and cannot continue

Expected Behavior

After rejecting a tool, the rejection message should be sent back to the LLM as a tool result (with success: false), allowing the agentic loop to continue. The LLM can then acknowledge the rejection and try an alternative approach.

Fix

Add add_pending_tool_result and a continuation check after the rejection:

async fn handle_reject(&mut self) -> Result<()> {
    if let Some(approval) = self.app_state.reject() {
        self.app_state.update_tool_result(
            &approval.tool_call_id,
            "Tool execution rejected by user".to_string(),
            false,
            "Rejected".to_string(),
        );

        // Feed rejection back to LLM for agentic loop continuation
        self.app_state.add_pending_tool_result(
            approval.tool_call_id.clone(),
            approval.tool_name.clone(),
            "Tool execution rejected by user".to_string(),
            false,
        );

        // Continue agentic loop if no other tools are running
        if self.running_tool_tasks.is_empty() && self.running_subagents.is_empty() {
            if self.app_state.has_pending_tool_results() {
                let _ = self.continue_with_tool_results().await;
            }
        }

        // Legacy bridge handling...
    }
    Ok(())
}

Impact

Any time a user rejects a tool during an agentic conversation, the entire conversation becomes stuck. The user must manually send a new message to resume interaction. This defeats the purpose of the approval/rejection flow in the agentic loop.

Version

v0.1.0