From 59ef28ffab4d0c923fcd473d7e32bd4f5c1b2c9e Mon Sep 17 00:00:00 2001
From: Tara Menon <tara.menon@gyannetra.com>
Date: Thu, 11 Jun 2026 22:55:19 +0530
Subject: [PATCH] =?UTF-8?q?=F0=9F=94=A7=20Fix=20tool=20calling=20+=20capab?=
 =?UTF-8?q?ility=20awareness?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Three fixes:
1. OllamaProvider now sends tools[] in native /api/chat requests
2. OllamaProvider parses tool_calls from response (OpenAI-compatible
   format with function.name, function.arguments)
3. System prompt now lists available tools so LLM knows it can use
   bash, curl, read/write files — stops saying "I cannot do that"
---
 crates/gyandoot-agent/src/agent_loop.rs |  2 +-
 crates/gyandoot-providers/src/ollama.rs | 21 +++++++++++++++++++--
 2 files changed, 20 insertions(+), 3 deletions(-)

diff --git a/crates/gyandoot-agent/src/agent_loop.rs b/crates/gyandoot-agent/src/agent_loop.rs
index 94b6e0e..64c65fd 100644
--- a/crates/gyandoot-agent/src/agent_loop.rs
+++ b/crates/gyandoot-agent/src/agent_loop.rs
@@ -196,7 +196,7 @@ impl AgentLoop {
         let mut system_prompt = self.config.system_prompt.clone();
         if !agent_name.is_empty() {
             system_prompt = format!(
-                "{}\n\n## Your Identity\n- Name: {}\n- Role: {}\n- You are speaking to: {} ({})\n\n## Memory\nYou have a verbatim dialogue log with your user. Use everything from past conversations as context when answering.",
+                "{}\n\n## Your Identity\n- Name: {}\n- Role: {}\n- You are speaking to: {} ({})\n\n## Your Capabilities\nYou have access to tools that let you:\n- Read, write, and edit files\n- Execute shell commands (bash) — use this for curl, git, package managers\n- Search files and directories\n- Use `bash` to run curl for web requests when needed\n\nWhen a user asks you to do something that requires external data, use your tools. Do NOT say you can't do something if you have a tool that can do it.\n\n## Memory\nYou have a verbatim dialogue log with your user. Use everything from past conversations as context when answering.",
                 system_prompt,
                 agent_name,
                 if agent_role.is_empty() { "AI Assistant" } else { &agent_role },
diff --git a/crates/gyandoot-providers/src/ollama.rs b/crates/gyandoot-providers/src/ollama.rs
index d9d2a9b..dc9597a 100644
--- a/crates/gyandoot-providers/src/ollama.rs
+++ b/crates/gyandoot-providers/src/ollama.rs
@@ -154,7 +154,7 @@ impl Provider for OllamaProvider {
                 })
                 .collect();
 
-            let body = serde_json::json!({
+            let mut body = serde_json::json!({
                 "model": self.model,
                 "messages": messages,
                 "stream": false,
@@ -163,6 +163,11 @@ impl Provider for OllamaProvider {
                 }
             });
 
+            // Add tools if present (for function calling)
+            if !context.tools.is_empty() {
+                body["tools"] = serde_json::to_value(&context.tools).unwrap_or_default();
+            }
+
             let url = format!("{}/api/chat", self.normalized_base_url());
             let mut request = self
                 .client
@@ -192,9 +197,21 @@ impl Provider for OllamaProvider {
                 .to_string();
             let finish_reason = data["done_reason"].as_str().unwrap_or("stop").to_string();
 
+            // Parse tool calls from response (qwen3.5 returns OpenAI-compatible format)
+            let tool_calls = data["message"]["tool_calls"].as_array().map(|calls| {
+                calls.iter().map(|tc| {
+                    let func = &tc["function"];
+                    gyandoot_core::ToolCall {
+                        id: tc["id"].as_str().unwrap_or_default().to_string(),
+                        name: func["name"].as_str().unwrap_or_default().to_string(),
+                        arguments: func["arguments"].clone(),
+                    }
+                }).collect()
+            });
+
             Ok(ProviderResponse {
                 content,
-                tool_calls: None,
+                tool_calls,
                 finish_reason,
             })
         }