# Tools — Complete End-to-End Lifecycle

This document describes the complete tool lifecycle in the YiemAgent framework, from definition through execution, for framework authors and maintainers who need a thorough understanding of the architecture.

---

## Table of Contents

1. [Quick Start: Tool Lifecycle](#1-quick-start-tool-lifecycle)
2. [Overview](#2-overview)
3. [Tool Definition — The `agentTool` Struct](#3-tool-definition--the-agenttool-struct)
4. [Tool Registration — Per-Agent Tool Stores](#4-tool-registration--per-agent-tool-stores)
5. [The Agent Loop — High-Level Flow](#5-the-agent-loop--high-level-flow)
6. [Message Processing Pipeline](#6-message-processing-pipeline)
7. [Tool Call Extraction from LLM Response](#7-tool-call-extraction-from-llm-response)
8. [The Per-Call Pipeline — Prepare, Execute, Finalize](#8-the-per-call-pipeline--prepare-execute-finalize)
9. [Execution Modes — Sequential vs Parallel](#9-execution-modes--sequential-vs-parallel)
10. [Tool Call Batches & Termination Logic](#10-tool-call-batches--termination-logic)
11. [Tool Result Message Creation](#11-tool-result-message-creation)
12. [Error Handling & Recovery Pattern](#12-error-handling--recovery-pattern)
13. [Event System — Tool Lifecycle Events](#13-event-system--tool-lifecycle-events)
14. [Agent Lifecycle Hooks](#14-agent-lifecycle-hooks)
15. [Self-Modifying Tools](#15-self-modifying-tools)
16. [Complete End-to-End Example](#16-complete-end-to-end-example)
17. [Tool File Contract](#17-tool-file-contract)
18. [Appendix: Type Reference](#18-appendix-type-reference)

---

## 1. Quick Start: Tool Lifecycle

This section shows the complete lifecycle from tool registration through execution and result extraction. Each step maps to the detailed sections below.

### Step 1: Discover — `listTools`

The agent calls the `listTools` tool to see available tools and detect name collisions before creating new ones.

```julia
# The listTools tool is auto-injected via listTool(store) — no manual registration needed
tool = listTool(store)  # Returns an agentTool that, when executed, lists all tools in the store
```

**Result extraction:**
```julia
result = tool.execute("call-1", Dict{String,Any}(), nothing, x->x)
# result.content[1].text => "Available tools:\n- getTime: Time Lookup — Get current local time...\n- getWeather: Weather Lookup — Fetch current weather..."
```

**Source:** `toolRegistry.jl:54-82`

---

### Step 2: Load — `loadTools()`

Load all tool modules from a directory into a `toolStore`. Each `.jl` file must define `getTool()::agentTool`. `listTool` is auto-registered so the LLM can discover available tools.

```julia
using YiemAgent, YiemAgent.toolRegistry

store = toolStore(name="myAgent")
tools = loadTools(store, "src/tools")
# Scans src/tools/ for .jl files, wraps each in a submodule, calls getTool(), registers in store.tools
# Also auto-registers listTools for runtime discovery
```

**Result extraction:**
```julia
all_tools = getTools(store)  # OrderedDict{String, agentTool}
# Keys: "getTime", "getWeather", "writeTool", "listTools"
getTime_tool = all_tools["getTime"]

# Manual registration (alternative to loadTools)
registerTool(store, my_tool)
clearTools(store)  # Clear all tools from store
```

**Source:** `toolRegistry.jl:126-178`

---

### Step 2.5: Create Agent with Tools

Wire the loaded tools into a new `yiemAgent` instance. The `tools` parameter is deep-copied into `agent._state.tools`; `_tool_store` is kept for runtime registration.

```julia
using YiemAgent, YiemAgent.type, YiemAgent.toolRegistry

# 1. Set up toolStore and load tools (auto-registers listTools)
store = toolStore(name="myAgent")
loadTools(store, "src/tools")

# 2. Create agent — pass tools + _tool_store
agent = yiemAgent(
    systemPrompt = "You are a helpful assistant.",
    model = my_model,
    tools = getTools(store),          # OrderedDict{String, agentTool}
    llmCall = my_llm_call,            # Function that calls the LLM API
    agentEventSink = my_event_sink,   # Function for TUI/logging
    _tool_store = store,              # For runtime registerTool() calls
)
```

**Manual registration** (without `loadTools`):

```julia
store = toolStore(name="myAgent")
registerTool(store, getTime_tool)
registerTool(store, getWeather_tool)
registerTool(store, listTool(store))  # needed for manual registration

agent = yiemAgent(
    tools = getTools(store),
    llmCall = my_llm_call,
    agentEventSink = my_event_sink,
    _tool_store = store,
)
```

**Key constructor parameters:**

| Parameter | Type | Required | Purpose |
|-----------|------|----------|---------|
| `systemPrompt` | `String` | No (default: "You are helpful assistant.") | System prompt text |
| `model` | `llmModel` | No | LLM model config |
| `tools` | `OrderedDict{String, agentTool}` | No | Available tools (deep-copied) |
| `messages` | `Vector{agentMessage}` | No (default: empty) | Initial conversation history |
| `llmCall` | `Function` | **Yes** | `(messages::Dict) -> assistantMessage` — invokes the LLM |
| `agentEventSink` | `Function` | **Yes** | `(event) -> nothing` — receives tool lifecycle events |
| `_tool_store` | `toolStore` | No | Runtime tool registry for `registerTool()` |

Optional hooks: `prepareContext`, `formatMsgForLLM`, `beforeToolCall`, `afterToolCall`, `sessionId`, `maxRetryDelayMs`, `parallelToolExecute`.

**Source:** `type.jl:609-657`, `toolRegistry.jl:38-40, 191-195`

---

### Step 3: Use — Tool Execution

Tools can be used in two ways:

**Direct execution (testing / standalone):**
```julia
using YiemAgent.type

sig = nothing
op = x -> x  # no-op partial result callback

# Execute a loaded tool directly
result = getTime_tool.execute("call-1", Dict("city" => "Tokyo"), sig, op)
```

**Via agent loop (production):**
```
user message → runAgent(agent, Dict("role"=>"user", "content"=>...))
  → _agent_loop detects message → @spawn _process_message(agent)
    → prepareContext → formatMsgForLLM → llmCall
      → LLM returns tool_calls
        → executeToolCalls(context, response, tool_call_list, config, signal, emit)
          → prepareToolCall → executePreparedToolCall → finalizeExecutedToolCall
            → createToolResultMessage → batch.messages (toolResultMessage[])
```

**Source:** Direct: `test/toolTest.jl:81-99` | Agent: `agentCore.jl:35-311`

---

### Step 4: Extract Result

**`agentToolResult`** (raw tool output, `type.jl:429-434`):
```julia
result = getTime_tool.execute("call-1", Dict("city" => "Tokyo"), nothing, x->x)

result.content[1]      # textContent("Current time in Tokyo: ...")
result.content[1].text # "Current time in Tokyo: 2026-08-10T..."
result.details         # Dict{Any,Any}() — tool-specific metadata
result.usage           # nothing — llmUsage tracking (optional)
result.terminate       # false — signals loop termination
```

**`toolResultMessage`** (wrapped for conversation history, `type.jl:152-191`):
```julia
msg = batch.messages[1]  # toolResultMessage

msg.toolCallId   # "call-1"
msg.toolName     # "getTime"
msg.content      # Vector{messageContent}
msg.isError      # false
msg.details      # tool-specific metadata
msg.timestamp    # DateTime
```

---

## 2. Overview

The tool system follows a **three-phase pipeline** per tool call:

```
PREPARE → EXECUTE → FINALIZE
```

Each phase has a single responsibility and produces an intermediate result:

| Phase | Function | Input | Output | Purpose |
|-------|----------|-------|--------|---------|
| Prepare | `prepareToolCall()` | `agentContext`, `assistantMessage`, `agentToolCall`, `agentLoopConfig`, `abortSignal` | `preparedToolCall` or `immediateOutcome` | Resolve tool, validate args, run pre-hook |
| Execute | `executePreparedToolCall()` | `preparedToolCall`, `abortSignal`, `emit` | `executedOutcome` | Call `tool.execute()`, stream partial results |
| Finalize | `finalizeExecutedToolCall()` | `agentContext`, `assistantMessage`, `preparedToolCall`, `executedOutcome`, `agentLoopConfig`, `abortSignal` | `finalizedOutcome` | Run post-hook, emit end event |

The pipeline ensures that **every tool call produces a result**, even on failure. Errors are captured as `immediateOutcome`, `executedOutcome`, or `finalizedOutcome` with `isError=true`, then converted to `toolResultMessage` objects that are fed back to the LLM conversation history.

---

## 3. Tool Definition — The `agentTool` Struct

**Source:** `type.jl:261-281`

Every tool is an `agentTool` struct with the following fields:

```julia
struct agentTool
    name::String                              # Unique identifier (e.g. "getWeather")
    label::String                             # Human-readable name (e.g. "Weather Lookup")
    description::String                       # What the tool does (shown to the LLM for selection)
    inputSchema::Any                          # JSON Schema (MCP format) describing parameters
    execute::Function                         # Core execution: (toolCallId, args, signal, onPartialResult) -> agentToolResult
    prepareArguments::Union{Function, Nothing}  # Optional: (args) -> modified_args (before validation)
    validateRequiredArgs::Union{Function, Nothing}  # Optional: (args) -> Union{Nothing, String} error
    parallelToolExecute::Bool                 # Override: run in parallel with other tools
end
```

### Field Details

| Field | Required | Signature | Purpose |
|-------|----------|-----------|---------|
| `name` | Yes | `String` | Unique key for tool lookup in `context.tools[name]` |
| `label` | Yes | `String` | Human-readable name for display |
| `description` | Yes | `String` | Shown to LLM for tool selection decisions |
| `inputSchema` | Yes | `Dict{String,Any}` | JSON Schema (MCP format) with `type`, `properties`, `required` |
| `execute` | Yes | `Function` | The actual tool logic (see signature below) |
| `prepareArguments` | No | `Function` | Transforms args before validation; `(args::Dict) -> Dict` |
| `validateRequiredArgs` | No | `Function` | Custom validation; `(args::Dict) -> Union{Nothing, String}` |
| `parallelToolExecute` | No | `Bool` | Default `false`. When `true`, allows parallel execution |

### Execute Function Signature

```julia
execute(toolCallId::String,
        args::Dict{String,Any},
        signal::Union{Nothing,abortSignal},
        onPartialResult::Function)::agentToolResult
```

| Parameter | Description |
|-----------|-------------|
| `toolCallId` | Unique ID from the LLM's tool call (e.g., `"call_abc123"`) |
| `args` | Validated arguments from the LLM, already passed through `prepareArguments` and `validateRequiredArgs` |
| `signal` | Optional `abortSignal` for cancellable operations. Check `signal.aborted` to abort early. |
| `onPartialResult` | Callback for streaming progress: `onPartialResult(partial_data)` emits `toolExecUpdateEvent` |

### Returns — `agentToolResult`

**Source:** `type.jl:429-434`

```julia
struct agentToolResult
    content::Vector{messageContent}   # Output content (textContent, etc.)
    details::Dict{Any,Any}             # Tool-specific metadata (e.g., counts, IDs)
    usage::Union{llmUsage, Nothing}    # Token usage tracking (optional)
    terminate::Bool                    # If true, signals the tool requested loop termination
end
```

The `terminate` flag is checked at the batch level. See [Section 10](#10-tool-call-batches--termination-logic) for details.

---

## 4. Tool Registration — Per-Agent Tool Stores

**Source:** `toolRegistry.jl`

### How `toolStore` Works

The registry uses **per-agent isolated storage** via the `toolStore` struct. Each agent gets its own store, so tool registration is independent — `registerTool(store, tool)` only affects that agent's tool set.

```julia
struct toolStore
    tools::OrderedDict{String, agentTool}  # keyed by name for O(1) lookup + ordered iteration
    name::String                           # identifier for debugging/logs
end
```

`store.tools` is an `OrderedDict` — it provides O(1) lookup by tool name and preserves insertion order for iteration. `getTools(store)` returns this `OrderedDict` directly (not a copy), so mutations on the returned value affect the store.

### How `loadTools(store, dir)` Works

```julia
function loadTools(store::toolStore, dir::String)::OrderedDict{String, agentTool}
```

**Source:** `toolRegistry.jl:126-178`

1. **Scans** `dir` for `.jl` files (excluding files matching `registry` in name)
2. **Sorts** filenames alphabetically for deterministic registration order
3. **Wraps** each file in a dynamically created submodule:
    ```julia
    # For "getWeather.jl" → module _tool_getWeather
    module _tool_getWeather
        using ..type
        using Dates, UUIDs, DataStructures, JSON
        # (file contents here)
    end
    ```
4. **Evaluates** `getTool()` within the submodule scope using `Core.eval(mod, :(getTool()))` — this avoids world-age issues
5. **Validates** the return value is an `agentTool` instance
6. **Registers** the tool in `store.tools`
7. **Auto-registers** `listTool(store)` so the LLM can discover available tools at runtime

### Why Submodules?

Each tool file is loaded into its own **namespaced submodule**. This means:
- `validateRequiredArgs`, `prepareArguments`, `executeTool`, and helper functions defined in `getTime.jl` are scoped under `_tool_getTime`
- No name collisions between tools — `getTime.validateRequiredArgs` is distinct from `getWeather.validateRequiredArgs`
- The module reference is kept alive by the functions stored in `agentTool` (closures in `execute`, `validateRequiredArgs`, `prepareArguments`) so they don't get garbage collected

### Registration API

```julia
# Create per-agent stores
store1 = toolStore(name="agent1")
store2 = toolStore(name="agent2")

# Load tools into specific stores (auto-registers listTools)
tools1 = loadTools(store1, "src/tools/weather_tools")   # agent1 only
tools2 = loadTools(store2, "src/tools/wine_tools")       # agent2 only

# Manual registration (per-store)
registerTool(store1, my_tool)

# Query (returns OrderedDict keyed by tool name, in registration order)
all_tools = getTools(store1)  # OrderedDict{String, agentTool} — O(1) lookup + deterministic order

# Clear (per-store)
clearTools(store1)  # only clears store1
```

`getTools(store)` returns the internal `OrderedDict` directly, giving callers:
- O(1) lookup by tool name
- Deterministic iteration order (registration order)
- Consistency with `agentState.tools` (also `OrderedDict{String, agentTool}`)
- No copy overhead — mutations on the returned value affect the store

### Per-Agent Isolation

Each `toolStore` is completely independent — tools registered in one store do not appear in another:

```julia
storeA = toolStore(name="A")
storeB = toolStore(name="B")

registerTool(storeA, getTime_tool)
registerTool(storeB, getWeather_tool)

getTools(storeA)  # only contains getTime
getTools(storeB)  # only contains getWeather

clearTools(storeA)  # storeB is unaffected
```

This ensures that `yiemAgent` instances with different `tool_store` references operate with completely isolated tool sets.

---

## 5. The Agent Loop — High-Level Flow

**Source:** `agentCore.jl:35-145`

The `_agent_loop()` function runs as a background `@spawn` task, created when `yiemAgent` is constructed.

### Channel Architecture

```
yiemAgent struct contains:
  - inputChannel (Channel, capacity 16)   ← user sends messages here via runAgent()
  - followUpChannel (Channel, capacity 32) ← user sends follow-ups here via followUp()
  - outputChannel (Channel, capacity 16)   ← agent sends responses here via takeResponse()
  - _tool_store (toolStore)                ← per-agent isolated tool registry
```

### Loop States

The loop tracks 6 states (documented at `agentCore.jl:39-75`):

| State | `processingTask` | `activeRun` | `inputChannel` | `followUpChannel` | Behavior |
|-------|-----------------|-------------|----------------|-------------------|----------|
| 1 | `nothing` | `false` | empty | empty | Idle, waiting |
| 2 | `nothing` | `false` | has msg | empty | New message → spawn `_process_message` |
| 3 | running | `true` | empty | empty | Processing, no new input |
| 4 | running | `true` | has msg | empty | New message while busy → queued |
| 5 | running | `true` | empty | has msg | Follow-up while busy → queued |
| 6 | done | `false` | empty | empty | Task completed → send result, reset |

### Loop Logic (simplified)

```julia
function _agent_loop(agent::yiemAgent)
    while true
        # 1. Wait for message from inputChannel (blocking poll)
        msg = fetch!(agent.inputChannel)  # agentCore.jl:84

        # 2. Handle shutdown signal
        if msg === :shutdown
            drain both channels, break loop
        end

        # 3. If agent is idle, spawn _process_message
        if agent._state.activeRun == false
            processingTask = Threads.@spawn _process_message(agent)
            agent._state.activeRun = true
        end

        # 4. While processing: check for followUp messages
        if istaskdone(processingTask) == false && isready(agent.followUpChannel)
            drain followUpChannel → push to inputChannel
            continue  # wait for current processing to finish
        end

        # 5. When processing completes
        if istaskdone(processingTask) == true
            result = fetch(processingTask)
            put!(agent.outputChannel, result)
            agent._state.activeRun = false
            processingTask = nothing
        end
    end
end
```

**Key design:** The loop always checks `inputChannel` before `followUpChannel`. Follow-up messages are merged into `inputChannel` only when the current processing task is active, ensuring they are processed after the primary message completes but before new input arrives.

---

## 6. Message Processing Pipeline

**Source:** `agentCore.jl:175-311`

`_process_message(agent)` is the core function that processes a batch of user messages through the LLM pipeline.

### Pipeline Steps

```julia
function _process_message(agent::yiemAgent)::assistantMessage
    final_response = nothing

    while true  # Loop until LLM returns response without tool calls
        # ── Step 1: Drain inputChannel ──────────────────────────────
        while isready(agent.inputChannel)
            raw_msg = take!(agent.inputChannel)
            if raw_msg === :shutdown
                put!(agent.inputChannel, :shutdown)
                break
            end
            user_msg = OpenAiToUserMessage(raw_msg)  # Convert Dict → userMessage
            push!(agent._state.messages, user_msg)
        end

        # ── Step 2: Prepare context ─────────────────────────────────
        preparedContext = agent.prepareContext(agent._state)
        # Default: deep copies systemPrompt, messages, tools from agentState → agentContext
        # Override point: filter tools, inject context, modify system prompt

        # ── Step 3: Format for LLM ──────────────────────────────────
        formatted_messages = agent.formatMsgForLLM(preparedContext)
        # Converts agentContext → Dict("messages" => [...]) in OpenAI format
        # Wraps systemPrompt as system role, converts each messageContent block

        # ── Step 4: Call LLM ────────────────────────────────────────
        response = agent.llmCall(formatted_messages)
        # Returns assistantMessage with content::Vector{messageContent}
        # Each content block has a type: "text", "thinking", or "tool_call"

        # ── Step 5: Extract tool calls ──────────────────────────────
        has_tool_calls, tool_call_list = extract_tool_calls(response.content)
        # Inspects content blocks for "tool_calls" or "tool_call" Dict entries

        # ── Step 6: Execute tool calls or return ────────────────────
        if has_tool_calls && !isempty(tool_call_list)
            # Build context and config
            context = agentContext(agent._state.systemPrompt, agent._state.messages, agent._state.tools)
            config = agentLoopConfig(agent._state.tools, agent.beforeToolCall, agent.afterToolCall, ...)
            signal = nothing
            emit = agent.agentEventSink

            # Execute tool calls (sequential or parallel)
            batch = executeToolCalls(context, response, tool_call_list, config, signal, emit)

            # Save results to conversation history
            for tool_result in batch.messages
                push!(agent._state.messages, tool_result)
            end

            # Check termination
            if batch.terminate
                final_response = build_final_response(batch)
                break
            end
            # Otherwise, loop back to Step 1 (drain any new input) and call LLM again
        else
            # No tool calls — this is the final response
            final_response = response
            break
        end
    end

    return final_response
end
```

### Debug Note

There is a deliberate `error(5555555)` at `agentCore.jl:214` that halts execution after the LLM call. This appears to be a debugging/staging marker. Remove or replace it before production use.

---

## 7. Tool Call Extraction from LLM Response

**Source:** `agentCore.jl:217-245`

After the LLM call, the agent inspects `response.content` (a `Vector{messageContent}`) for tool call blocks. Two formats are supported:

### Format 1: OpenAI `tool_calls` array

```julia
# Response content block:
Dict(
    :type => "tool_calls",
    :tool_calls => [
        Dict(:id => "call_1", :name => "getWeather", :arguments => Dict("city" => "Tokyo")),
        Dict(:id => "call_2", :name => "getTime", :arguments => Dict("timezone" => "Asia/Tokyo")),
    ]
)
```

### Format 2: Single `tool_call` block

```julia
Dict(
    :type => "tool_call",
    :id => "call_1",
    :name => "getWeather",
    :arguments => Dict("city" => "Tokyo")
)
```

### Extraction Logic

```julia
tool_call_list = agentToolCall[]

for content_block in response.content
    if content_block isa Dict
        # OpenAI format: array of tool calls
        if get(content_block, :type, "") == "tool_calls"
            for tc_data in get(content_block, :tool_calls, [])
                tc = agentToolCall(
                    type = "function",
                    id = get(tc_data, :id, string(uuid4())),  # fallback UUID
                    name = get(tc_data, :function, Dict())[:name],
                    arguments = get(tc_data, :function, Dict())[:arguments],
                )
                push!(tool_call_list, tc)
            end
        # Alternative format: single tool call
        elseif get(content_block, :type, "") == "tool_call"
            tc_data = content_block
            tc = agentToolCall(
                type = "function",
                id = get(tc_data, :id, string(uuid4())),
                name = get(tc_data, :name, ""),
                arguments = get(tc_data, :arguments, Dict{String,Any}()),
            )
            push!(tool_call_list, tc)
        end
    end
end
```

The `agentToolCall` struct is defined at `type.jl:362-367`:

```julia
struct agentToolCall
    type::String                              # Always "function"
    id::String                                # Unique tool call identifier
    name::String                              # Tool name (matches context.tools keys)
    arguments::Dict{String, Any}              # Parsed tool arguments
end
```

---

## 8. The Per-Call Pipeline — Prepare, Execute, Finalize

This is the core of the tool execution system. Each tool call (whether part of a batch or standalone) goes through exactly three phases.

### 8.1 Phase 1: Prepare — `prepareToolCall()`

**Source:** `agentCore.jl:511-547`

```julia
function prepareToolCall(
    context::agentContext,
    assistantMsg::assistantMessage,
    toolCall::agentToolCall,
    config::agentLoopConfig,
    signal::Union{Nothing, abortSignal},
)::Union{preparedToolCall, immediateOutcome}
```

**Steps:**

1. **Resolve tool by name** — `get(context.tools, toolCall.name, nothing)`
   - If `nothing` → `immediateOutcome(createErrorToolResult("Tool X not found"), true)`

2. **Prepare arguments** — `prepareToolCallArguments(tool, toolCall)`
   - Calls `tool.prepareArguments(toolCall.arguments)` if defined
   - Returns the toolCall with transformed arguments
   - If no hook or no change, returns original `toolCall`

3. **Validate arguments** — `validateToolArguments(tool, prepared)`
   - Calls `tool.validateRequiredArgs(prepared.arguments)` if defined, otherwise uses default `validateRequiredArgs(prepared.arguments, tool.inputSchema)`
   - Default validator checks `inputSchema["required"]` array
   - On failure: throws `ArgumentError(error_string)`, caught by the try-catch below

4. **Run `beforeToolCall` hook** — if `config.beforeToolCall !== nothing`
   - Passes `beforeToolCallContext(assistantMsg, toolCall, validatedArgs, context)` and `signal`
   - Hook can return `nothing` (proceed), or `Dict(:block => true, :reason => "...")` (reject)
   - If `signal.aborted == true` → `immediateOutcome(createErrorToolResult("Operation aborted"), true)`
   - If `before.block == true` → `immediateOutcome(createErrorToolResult(get(before, :reason, "blocked")), true)`

5. **Return success** → `preparedToolCall(tool, toolCall, validatedArgs)`

**Source:** `type.jl:681-685` — `preparedToolCall` holds the resolved tool, original call metadata, and validated arguments together.

**Key design principle:** Preparation **never throws**. Every failure path returns an `immediateOutcome` with `isError=true`, ensuring the agent loop always has a valid result to feed back to the LLM.

### 8.2 Phase 2: Execute — `executePreparedToolCall()`

**Source:** `agentCore.jl:589-617`

```julia
function executePreparedToolCall(
    prep::preparedToolCall,
    signal::Union{Nothing, abortSignal},
    emit::Function,
)::executedOutcome
```

**Steps:**

1. **Initialize streaming state:**
    ```julia
    updateEvents = promise[]  # vector to collect update event handles
    accepting = true          # guard to prevent duplicate emissions
    ```

2. **Call `tool.execute()`:**
    ```julia
    result = prep.tool.execute(
        prep.toolCall.id,
        prep.args,
        signal,
        partialResult -> begin
            if accepting
                push!(updateEvents,
                      emit(toolExecUpdateEvent(prep.toolCall.id, prep.toolCall.name,
                                               prep.toolCall.arguments, partialResult)))
            end
        end
    )
    ```

3. **Wait for streaming to settle:**
    ```julia
    accepting = false
    wait.(updateEvents)  # wait for all pending update event handlers
    return executedOutcome(result, false)
    ```

4. **On error:**
    ```julia
    catch err
        accepting = false
        wait.(updateEvents)
        return executedOutcome(createErrorToolResult(sprint(showerror, err)), true)
    end
    ```

**Streaming design:** The `accepting` guard prevents emitting updates after the call completes. If the tool's `execute` function yields after emitting updates but before returning, no duplicate or stale updates are emitted.

### 8.3 Phase 3: Finalize — `finalizeExecutedToolCall()`

**Source:** `agentCore.jl:675-706`

```julia
function finalizeExecutedToolCall(
    context::agentContext,
    assistantMsg::assistantMessage,
    prep::preparedToolCall,
    executed::executedOutcome,
    config::agentLoopConfig,
    signal::Union{Nothing,abortSignal},
)::finalizedOutcome
```

**Steps:**

1. **Extract execution result:**
    ```julia
    result = executed.result
    isError = executed.isError
    ```

2. **Run `afterToolCall` hook** — if `config.afterToolCall !== nothing`:
    - Passes `afterToolCallContext(assistantMsg, prep.toolCall, prep.args, result, isError, context)` and `signal`
    - Hook can mutate the result:
      ```julia
      after = config.afterToolCall(afterToolCallContext(...))
      if after !== nothing
          result = merge(result, dict(
              :content => get(after, :content, result.content),
              :details => get(after, :details, result.details),
              :usage => get(after, :usage, result.usage),
              :terminate => get(after, :terminate, result.terminate)
          ))
          isError = get(after, :isError, isError)
      end
      ```
    - Common use cases: mask sensitive data, normalize usage, flip `terminate` based on business logic
    - On error: `result = createErrorToolResult(sprint(showerror, err)); isError = true`

3. **Return:**
    ```julia
    return finalizedOutcome(prep.toolCall, result, isError)
    ```

**Source:** `type.jl:787-791` — `finalizedOutcome` holds the original tool call reference, final result (post-hook), and error status.

### 8.4 Emission — `emitToolExecutionEnd()`

**Source:** `agentCore.jl:736-739`

```julia
function emitToolExecutionEnd(finalized::finalizedOutcome, emit::Function)
    emit(toolExecEndEvent(finalized.toolCall.id, finalized.toolCall.name,
                          finalized.result, finalized.isError))
end
```

This is called immediately after finalization, before building the `toolResultMessage`.

---

## 9. Execution Modes — Sequential vs Parallel

**Source:** `agentCore.jl:795-936, 988-1011`

### Dispatcher — `executeToolCalls()`

```julia
function executeToolCalls(
    context::agentContext,
    assistantMsg::assistantMessage,
    toolCalls::Vector{agentToolCall},
    config::agentLoopConfig,
    signal::Union{Nothing, abortSignal},
    emit::Function,
)::agentToolCallBatch
```

**Decision logic** (`agentCore.jl:997-1010`):

```julia
hasSequential = false
for tc in toolCalls
    t = get(context.tools, tc.name, nothing)
    if t !== nothing && !t.parallelToolExecute
        hasSequential = true
        break
    end
end

if config.toolExecution == "sequential" || hasSequential
    return executeToolCallsSequential(...)
else
    return executeToolCallsParallel(...)
end
```

**Rule:** If the global config is `"sequential"` **OR** any tool in the batch has `parallelToolExecute = false`, the entire batch runs sequentially. Sequential is the safe default.

### Sequential Execution — `executeToolCallsSequential()`

**Source:** `agentCore.jl:795-829`

```julia
function executeToolCallsSequential(...)::agentToolCallBatch
    finalizedCalls = finalizedOutcome[]
    messages = toolResultMessage[]

    for tc in toolCalls
        emit(toolExecStartEvent(tc.id, tc.name, tc.arguments))

        prep = prepareToolCall(context, assistantMsg, tc, config, signal)

        if prep isa immediateOutcome
            finalized = finalizedOutcome(tc, prep.result, prep.isError)
        else
            executed = executePreparedToolCall(prep, signal, emit)
            finalized = finalizeExecutedToolCall(context, assistantMsg, prep, executed, config, signal)
        end

        emitToolExecutionEnd(finalized, emit)
        push!(messages, createToolResultMessage(finalized))
        push!(finalizedCalls, finalized)

        if signal !== nothing && signal.aborted
            break  # abort: skip remaining calls
        end
    end

    return agentToolCallBatch(messages, shouldTerminate(finalizedCalls))
end
```

### Parallel Execution — `executeToolCallsParallel()`

**Source:** `agentCore.jl:888-936`

```julia
function executeToolCallsParallel(...)::agentToolCallBatch
    entries = union{finalizedOutcome, task{finalizedOutcome}}[]

    for tc in toolCalls
        emit(toolExecStartEvent(tc.id, tc.name, tc.arguments))

        prep = prepareToolCall(context, assistantMsg, tc, config, signal)

        if prep isa immediateOutcome
            finalized = finalizedOutcome(tc, prep.result, prep.isError)
            emitToolExecutionEnd(finalized, emit)
            push!(entries, finalized)  # immediate outcome — no task
        else
            task = task() do
                executed = executePreparedToolCall(prep, signal, emit)
                finalized = finalizeExecutedToolCall(context, assistantMsg, prep, executed, config, signal)
                emitToolExecutionEnd(finalized, emit)
                return finalized
            end
            schedule(task)
            push!(entries, task)  # pending task
        end

        if signal !== nothing && signal.aborted
            break  # abort during preparation — skip remaining
        end
    end

    # Collect results in original order
    finalizedCalls = finalizedOutcome[]
    for entry in entries
        outcome = entry isa task ? fetch(entry) : entry
        push!(finalizedCalls, outcome)
    end

    messages = toolResultMessage[]
    for f in finalizedCalls
        push!(messages, createToolResultMessage(f))
    end

    return agentToolCallBatch(messages, shouldTerminate(finalizedCalls))
end
```

**Key design:** All tool calls are **prepared** sequentially (validation, hooks), then prepared calls are **executed** concurrently as separate tasks. Results are collected in the original call order via `fetch()`.

**Trade-off:** Parallel execution reduces wall-clock time for independent tools but can overwhelm external resources (rate limits, connection pools, disk I/O).

---

## 10. Tool Call Batches & Termination Logic

**Source:** `type.jl:793-838`

### `agentToolCallBatch`

```julia
struct agentToolCallBatch
    messages::Vector{toolResultMessage}  # Tool result messages for this batch
    terminate::Bool                      # Whether the batch should terminate the loop
end
```

### Termination Logic — `shouldTerminate()`

**Source:** `agentCore.jl:409-411`

```julia
function shouldTerminate(batches::Vector{finalizedOutcome})::Bool
    return !isempty(batches) && all(b -> b.result.terminate, batches)
end
```

**Rule:** `terminate` is `true` **only when every tool in the batch** has `result.terminate == true`. This prevents a single tool that sets `terminate: true` (e.g., for metadata purposes) from accidentally stopping the agent when other tools did not intend to terminate.

### When to Set `terminate = true`

From the type documentation (`type.jl:803-815`):

| Use Case | Description |
|----------|-------------|
| Task completion | A tool like `deploy` or `submit` finishes its work and signals the agent to stop |
| Unrecoverable error | A tool hits a fatal condition (auth token expired, database connection lost) |
| Async handoff | A tool triggers a long-running external operation; the external system will later resume via `continue()` |

### Batch Processing in `_process_message()`

**Source:** `agentCore.jl:266-307`

```julia
batch = executeToolCalls(context, response, tool_call_list, config, signal, emit)

# Save results to conversation history
for tool_result in batch.messages
    push!(agent._state.messages, tool_result)
end

if batch.terminate
    # Build final response from tool results
    final_content = [textContent("Tool execution completed.")]
    for tool_result in batch.messages
        for content_block in tool_result.content
            if content_block isa textContent
                append!(final_content, [content_block])
            elseif content_block isa Dict
                if haskey(content_block, :text)
                    push!(final_content, textContent(content_block[:text]))
                end
            end
        end
    end
    final_response = assistantMessage(
        role = "assistant",
        content = final_content,
        api = response.api,
        model = response.model,
        usage = response.usage,
        stopReason = "tool_use_terminated",
        errorMessage = if any(x -> x.isError, batch.messages)
            "One or more tool calls failed"
        else
            nothing
        end,
        timestamp = now(),
    )
    break  # exit the while loop
end
# If batch.terminate == false, loop back to call LLM again with tool results
```

---

## 11. Tool Result Message Creation

**Source:** `agentCore.jl:373-379`

### `createToolResultMessage()`

```julia
function createToolResultMessage(f::finalizedOutcome)::toolResultMessage
    return toolResultMessage(
        "toolResult",                     # role
        f.toolCall.id,                    # toolCallId
        f.toolCall.name,                  # toolName
        f.result.content,                 # content (Vector{messageContent})
        f.result.details,                 # details
        f.result.usage,                   # usage
        get(f.result, :addedToolNames, string[]),  # addedToolNames (for dynamic tools)
        f.isError,                        # isError
        nowMillis(),                      # timestamp
    )
end
```

### `toolResultMessage` Struct

**Source:** `type.jl:152-191`

```julia
struct toolResultMessage <: agentMessage
    role::String                            # Always "tool"
    toolCallId::String                      # ID matching the original tool call
    toolName::String                        # Name of the executed tool
    content::Vector{messageContent}         # Tool output content
    details::Any                            # Additional tool-specific details
    usage::Union{llmUsage, Nothing}            # Token usage if applicable
    addedToolNames::Union{Vector{String}, Nothing}  # Tools added during execution
    isError::Bool                           # Whether the tool call resulted in an error
    timestamp::Timestamp                    # When the result was recorded
end
```

### Conversation History After Tool Execution

```
[system] "You are a helpful assistant."
[user]     "What's the weather in Tokyo?"
[assistant] {tool_calls: [getWeather(city="Tokyo")]}
[tool]      tool_call_id="call_1", tool_name="getWeather", content="Weather in Tokyo: Sunny, 22°C"
```

On the next loop iteration, `formatMsgForLLM()` converts `toolResultMessage` to OpenAI format:

```julia
Dict(
    "role" => "tool",
    "tool_call_id" => "call_1",
    "content" => [Dict("type" => "text", "text" => "Weather in Tokyo: Sunny, 22°C")]
)
```

---

## 12. Error Handling & Recovery Pattern

The framework uses a **result-based error handling** pattern instead of exceptions for tool call failures. This ensures the LLM always receives a tool result message, giving it the information to recover.

### Error Flow

```
Tool call fails at any phase
         │
         ▼
┌─────────────────────────┐
│ Phase: Prepare          │ → immediateOutcome(error_result, true)
│ Phase: Execute          │ → executedOutcome(error_result, true)
│ Phase: Finalize (hook)  │ → finalizedOutcome(tc, error_result, true)
└─────────────────────────┘
         │
         ▼
emitToolExecutionEnd(finalized, emit)
         │
         ▼
createToolResultMessage(finalized)
         │
         ▼
push!(agent._state.messages, toolResultMessage)
         │
         ▼
formatMsgForLLM() → LLM receives error as tool result
         │
         ▼
LLM can: retry with corrected args, report failure, or ask user for clarification
```

### `createErrorToolResult()`

**Source:** `agentCore.jl:339-341`

```julia
function createErrorToolResult(msg::String)::agentToolResult
    return agentToolResult([textContent("text", msg)], Dict{Any,Any}())
end
```

Returns an `agentToolResult` with a single `textContent` block containing the error message. This is wrapped in an `immediateOutcome`, `executedOutcome`, or `finalizedOutcome` depending on where the error occurred, then converted to `toolResultMessage` for the LLM.

**Key design:** Returning a result instead of throwing allows the LLM to see the error and decide whether to retry, re-issue the call with different arguments, or report failure to the user.

---

## 13. Event System — Tool Lifecycle Events

**Source:** `type.jl:480-516`

Each tool call emits a three-event lifecycle:

```
toolExecStartEvent → [zero or more toolExecUpdateEvent] → toolExecEndEvent
```

### Event Types

```julia
struct toolExecStartEvent
    toolCallId::String                    # ID of the tool call
    toolName::String                      # Name of the tool
    arguments::Dict{String, Any}          # Tool arguments
end

struct toolExecUpdateEvent
    toolCallId::String                    # ID of the tool call
    toolName::String                      # Name of the tool
    arguments::Dict{String, Any}          # Tool arguments
    partialResult::Any                    # The partial result data
end

struct toolExecEndEvent
    toolCallId::String                    # ID of the tool call
    toolName::String                      # Name of the tool
    result::agentToolResult               # The final tool result
    isError::Bool                         # Whether execution resulted in an error
end
```

### Event Emission Points

| Event | Emitted From | When |
|-------|-------------|------|
| `toolExecStartEvent` | `executeToolCalls*()` loop | Before `prepareToolCall()` for each tool call |
| `toolExecUpdateEvent` | `executePreparedToolCall()` | Inside `onPartialResult` callback during `tool.execute()` |
| `toolExecEndEvent` | `emitToolExecutionEnd()` | After `finalizeExecutedToolCall()` for each tool call |

### Event Sink

The `emit` function is passed through the entire call chain:

```julia
emit = agent.agentEventSink  # set during yiemAgent construction
```

The `agentEventSink` function is a user-provided callback that receives all events. This is typically used by:
- **TUI (Terminal UI):** Display real-time progress, tool names, results
- **Logging systems:** Record tool execution history
- **Monitoring:** Track tool usage, execution times, error rates
- **Audit trails:** Log all tool calls with arguments and results

---

## 14. Agent Lifecycle Hooks

### Hook Types

| Hook | Signature | Called | Purpose |
|------|-----------|--------|---------|
| `prepareContext` | `(state::agentState) -> agentContext` | Before each LLM call | Filter tools, inject context, modify system prompt |
| `formatMsgForLLM` | `(ctx::agentContext) -> Dict` | After `prepareContext` | Convert to LLM-specific format |
| `llmCall` | `(messages::Dict) -> assistantMessage` | After formatting | Actually invoke the LLM API |
| `beforeToolCall` | `(msgCtx::beforeToolCallContext, signal) -> Union{Nothing, Dict}` | In `prepareToolCall` | Ask for user permission, block execution, abort |
| `afterToolCall` | `(afterToolCallContext::afterToolCallContext, signal) -> Union{Nothing, Dict}` | In `finalizeExecutedToolCall` | Mutate result, mask data, flip `terminate` |
| `agentEventSink` | `(event) -> nothing` | Throughout lifecycle | Emit events for TUI, logging, monitoring |

### `beforeToolCall` Hook

**Source:** `agentCore.jl:530-541`

```julia
if config.beforeToolCall !== nothing
    before = config.beforeToolCall(
        beforeToolCallContext(assistantMsg, toolCall, validatedArgs, context), signal
    )
    if signal !== nothing && signal.aborted
        return immediateOutcome(createErrorToolResult("Operation aborted"), true)
    end
    if before !== nothing && before.block
        return immediateOutcome(
            createErrorToolResult(get(before, :reason, "Tool execution was blocked")), true)
    end
end
```

**Return values:**
- `nothing` — proceed with execution
- `Dict(:block => true, :reason => "...")` — block execution, error fed back to LLM
- `Dict(:block => false)` — proceed (explicit allow)

### `afterToolCall` Hook

**Source:** `agentCore.jl:687-703`

```julia
if config.afterToolCall !== nothing
    try
        after = config.afterToolCall(
            afterToolCallContext(assistantMsg, prep.toolCall, prep.args, result, isError, context), signal
        )
        if after !== nothing
            result = merge(result, dict(
                :content => get(after, :content, result.content),
                :details => get(after, :details, result.details),
                :usage => get(after, :usage, result.usage),
                :terminate => get(after, :terminate, result.terminate)
            ))
            isError = get(after, :isError, isError)
        end
    catch err
        result = createErrorToolResult(sprint(showerror, err))
        isError = true
    end
end
```

**Common use cases:**
- Mask sensitive data from result content before the LLM sees it (e.g., removing API keys from error messages)
- Normalize usage tracking data into a consistent format
- Inspect the result and decide to flip `terminate: true` based on business logic
- Wrap an error result in a friendlier message for the LLM to understand

### `prepareContext` Hook

**Source:** `utils.jl:111-125`

```julia
function prepareContext(state::agentState)::agentContext
    # TODO: filter tools from state.tools based on user intent
    filteredTools = state.tools

    # TODO: add filtered tools to the current system prompt / modify systemPrompt
    preparedSystemPrompt = state.systemPrompt

    # TODO: add system prompt, adjust/modify and inject additional context into messages
    preparedMessages = deepcopy(state.messages)

    return agentContext(preparedSystemPrompt, preparedMessages, filteredTools)
end
```

**Override points for customization:**
- Filter tools based on user intent (e.g., only show "wine" tools when user asks about wine)
- Modify the system prompt dynamically (e.g., inject current time, user preferences)
- Inject additional context (e.g., retrieved documents, current user state)
- Prune or reorder messages before formatting for the LLM

### `formatMsgForLLM` Hook

**Source:** `utils.jl:158-219`

Default implementation converts `agentContext` to OpenAI-compatible format:

```julia
function formatMsgForLLM(ctx::agentContext)::Dict{String, Any}
    messages = Vector{Dict{String, Any}}()

    # System prompt as system message
    if !isempty(ctx.systemPrompt)
        push!(messages, Dict(
            "role" => "system",
            "content" => [Dict("type" => "text", "text" => ctx.systemPrompt)]
        ))
    end

    # Conversation messages
    for msg in ctx.messages
        if msg isa userMessage
            push!(messages, _userMessageToOpenAI(msg))
        elseif msg isa assistantMessage
            push!(messages, _assistantMessageToOpenAI(msg))
        elseif msg isa toolResultMessage
            push!(messages, _toolResultMessageToOpenAI(msg))
        end
    end

    return Dict("messages" => messages)
end
```

Override this to produce custom LLM message formats for different APIs/providers (e.g., Anthropic, Google, Ollama).

---

## 15. Self-Modifying Tools

The framework supports tools that modify the tool system itself at runtime.

### `writeTool` — Create New Tool Files

**Source:** `tools/writeTool.jl`

`writeTool` is a **file writer**, not a code generator. The LLM provides the tool logic as `executeCode` (Julia code body), and `writeTool` wraps it in Julia boilerplate:
1. Converts `inputSchema` Dict into `Dict{String,Any}(...)` string literal
2. Indents `executeCode` with 4 spaces
3. Wraps it inside `function executeTool(...)::agentToolResult ... end`
4. Appends `getTool()` returning an `agentTool` struct
5. Writes the combined string to `src/tools/<name>.jl`

### `listTool` — Discover Available Tools

**Source:** `toolRegistry.jl:55-82`

Each `toolStore` gets its own `listTool` instance bound to that store via `listTool(store)`, so each agent sees only its own tools. `loadTools` auto-registers one, so the LLM can discover available tools at runtime. Also useful for **collision detection** before creating a new tool via `writeTool`.

### Self-Tooling Workflow

```
1. Agent detects no existing tool handles the user's request
2. Agent calls writeTool with:
   - name: "searchWine"
   - label: "Wine Search"
   - description: "Search a wine database..."
   - inputSchema: { ... }
   - executeCode: "query = args[\"query\"]\nresult = search(query)\n..."
   - (optional) validateCode, prepareCode
3. writeTool generates src/tools/searchWine.jl
4. Agent restarts (or hot-reloads) → loadTools(agent._tool_store, "src/tools") picks up the new file
5. Agent calls searchWine(query="cabernet")
6. Result: "Found 5 cabernet wines..."
```

### `writeTool` Input Schema

| Field | Type | Required | Description |
|-------|------|----------|-------------|
| `name` | `String` | Yes | Valid Julia identifier (letters, digits, underscores) |
| `label` | `String` | Yes | Human-readable tool name |
| `description` | `String` | Yes | What the tool does |
| `inputSchema` | `Dict` | Yes | JSON Schema in MCP format |
| `executeCode` | `String` | Yes | Julia code for `executeTool` body (NOT wrapped in function) |
| `validateCode` | `String` | No | Custom validation Julia code |
| `prepareCode` | `String` | No | Argument preparation code |
| `parallel` | `Bool` | No | Whether the tool can run in parallel (default: `false`) |

---

## 16. Complete End-to-End Example

### Full Lifecycle: User Message to Tool Result

```
USER SENDS MESSAGE
  └─> runAgent(agent, "What's the weather in Tokyo?")
       └─> put!(agent.inputChannel, Dict("role" => "user", "content" => [...]))


LOOP ITERATION 1 — LLM DECIDES TO USE A TOOL
  └─> _agent_loop: detects msg in inputChannel
  └─> Threads.@spawn _process_message(agent)

  ── _process_message ──────────────────────────────────────────────
  │
  │ Step 1: Drain inputChannel
  │   raw_msg = Dict("role" => "user", "content" => [...])
  │   user_msg = OpenAiToUserMessage(raw_msg)
  │   push!(agent._state.messages, user_msg)
  │
  │ Step 2: prepareContext
  │   ctx = agent.prepareContext(agent._state)
  │   → agentContext(systemPrompt, messages, tools)
  │
  │ Step 3: formatMsgForLLM
  │   formatted = agent.formatMsgForLLM(ctx)
  │   → Dict("messages" => [
  │       Dict("role" => "system", "content" => [...]),
  │       Dict("role" => "user", "content" => [...]),
  │   ])
  │
  │ Step 4: llmCall
  │   response = agent.llmCall(formatted)
  │   → assistantMessage(content = [
  │       Dict(:type => "tool_calls", :tool_calls => [
  │           Dict(:id => "call_1", :name => "getWeather",
  │                 :arguments => Dict("city" => "Tokyo"))
  │       ])
  │   ])
  │
  │ Step 5: Extract tool calls
  │   tool_call_list = [agentToolCall("function", "call_1", "getWeather", ...)]
  │
  │ Step 6: Execute tool calls
  │   context = agentContext(systemPrompt, messages, tools)
  │   config = agentLoopConfig(tools, beforeToolCall, afterToolCall, "sequential")
  │   batch = executeToolCalls(context, response, tool_call_list, config, nothing, emit)
  │
  │   ── executeToolCallsSequential ──────────────────────────────
  │   │
  │   │ For tc = agentToolCall("call_1", "getWeather", ...):
  │   │
  │   │ emit(toolExecStartEvent("call_1", "getWeather", {"city": "Tokyo"}))
  │   │
  │   │ PREPARE:
  │   │   tool = context.tools["getWeather"]  → found!
  │   │   validatedArgs = validateToolArguments(tool, tc)
  │   │   → validateRequiredArgs(Dict("city" => "Tokyo"), inputSchema)  → passes
  │   │   beforeToolCall_hook(...) → nothing (skipped)
  │   │   → preparedToolCall(tool, tc, {"city" => "Tokyo"})
  │   │
  │   │ EXECUTE:
  │   │   result = tool.execute("call_1", {"city" => "Tokyo"}, nothing, onPartialResult)
  │   │   → agentToolResult([textContent("Weather in Tokyo: Sunny, 22°C")], {}, nothing, false)
  │   │   → executedOutcome(result, false)
  │   │
  │   │ FINALIZE:
  │   │   afterToolCall_hook(...) → nothing (skipped)
  │   │   → finalizedOutcome(tc, result, false)
  │   │
  │   │ emit(toolExecEndEvent("call_1", "getWeather", result, false))
  │   │ msg = createToolResultMessage(finalized)
  │   │ → toolResultMessage("tool", "call_1", "getWeather", [...], {}, nothing, [], false, ts)
  │   │
  │   └─> agentToolCallBatch([msg], false)
  │
  │ Save results:
  │   for tool_result in batch.messages
  │       push!(agent._state.messages, tool_result)
  │   end
  │
  │ batch.terminate == false → loop back to Step 1
  │


LOOP ITERATION 2 — LLM RETURNS FINAL TEXT RESPONSE
  ── _process_message (second iteration) ───────────────────────────
  │
  │ Step 1: Drain inputChannel → empty
  │
  │ Step 2-3: prepareContext → formatMsgForLLM
  │   → messages now include:
  │     [system] "You are a helpful assistant."
  │     [user]   "What's the weather in Tokyo?"
  │     [tool]   tool_call_id="call_1", content="Weather in Tokyo: Sunny, 22°C"
  │
  │ Step 4: llmCall
  │   → assistantMessage(content = [Dict(:type => "text", :text => "The weather in Tokyo is sunny, 22°C.")])
  │
  │ Step 5: Extract tool calls → none
  │
  │ Step 6: has_tool_calls == false → break, return final_response
  │
  └─> return final_response


AGENT LOOP: SEND RESPONSE TO USER
  └─> put!(agent.outputChannel, final_response)
  └─> takeResponse(agent) → assistantMessage("The weather in Tokyo is sunny, 22°C.")
```

---

## 17. Tool File Contract

Each `.jl` file in `src/tools/` must conform to the following contract:

### Required Function

```julia
function getTool()::agentTool
    # Must return an agentTool instance
end
```

### Optional Functions

```julia
# Argument preparation (before validation)
function prepareArguments(args::Dict{String,Any})::Dict{String,Any}
    # Return modified args, or args unchanged
    return args
end

# Custom validation (before execution)
function validateRequiredArgs(args::Dict{String,Any})::Union{Nothing,String}
    # Return nothing to pass, or error string to fail
    return nothing
end

# Core execution
function executeTool(toolCallId::String,
                     args::Dict{String,Any},
                     signal::Union{Nothing,abortSignal},
                     onPartialResult::Function)::agentToolResult
    # Return agentToolResult with content, details, usage, terminate
    return agentToolResult([textContent("result")], Dict{Any,Any}(), nothing, false)
end
```

### File Structure

```julia
# src/tools/myTool.jl

using Dates  # ← tool declares its own dependencies (registry injects only `using ..type`)

# Optional: helper functions
function helper_function(...)
    ...
end

# Optional: prepareArguments
function prepareArguments(args::Dict{String,Any})::Dict{String,Any}
    return args
end

# Optional: validateRequiredArgs
function validateRequiredArgs(args::Dict{String,Any})::Union{Nothing,String}
    return nothing
end

# Required: executeTool
function executeTool(toolCallId::String, args::Dict{String,Any},
                     signal::Union{Nothing,abortSignal},
                     onPartialResult::Function)::agentToolResult
    ...
end

# Required: getTool
function getTool()::agentTool
    return agentTool(
        name = "myTool",
        label = "My Tool",
        description = "What this tool does",
        inputSchema = Dict{String,Any}(
            "type" => "object",
            "properties" => Dict(...),
            "required" => [...]
        ),
        execute = executeTool,
        prepareArguments = prepareArguments,
        validateRequiredArgs = validateRequiredArgs,
        parallelToolExecute = false
    )
end
```

### Dependencies

Each tool file **declares its own dependencies** via `using` statements at the top of the file. The registry does **not** inject any standard library packages — if a tool needs `Dates`, `JSON`, `HTTP`, `CSV`, or any other package, it must include its own `using` statements.

```julia
# src/tools/getTime.jl
using Dates

function executeTool(...)
    now()  # Dates.now requires `using Dates`
end
```

```julia
# src/tools/myApiTool.jl
using HTTP, JSON

function executeTool(...)
    response = HTTP.get("https://api.example.com")
    data = JSON.parse(String(response.body))
    ...
end
```

### Module Isolation

When `loadTools()` loads a file, it wraps it in a dynamically created submodule. The registry injects **only** `using ..type` to make core types (`agentTool`, `textContent`, `agentToolResult`, `abortSignal`, etc.) available:

```julia
# User writes in src/tools/myTool.jl:
using Dates, HTTP, JSON  # ← tool's own dependencies

function getTool()::agentTool ... end

# loadTools() creates:
module _tool_myTool
    using ..type            # ← injected by registry (core types only)
    using Dates, HTTP, JSON # ← from tool file
    # (user's code here)
end
```

All functions in the file are scoped under `_tool_myTool`, preventing name collisions with other tools. The module reference is kept alive by the function objects stored in `agentTool`, preventing garbage collection of closures.

---

## 18. Appendix: Type Reference

### Message Types

| Type | Source | Description |
|------|--------|-------------|
| `messageContent` | `type.jl:67` | Abstract base for message content |
| `textContent` | `type.jl:69` | Plain text content (`text::String`) |
| `imageContent` | `type.jl:73` | Image content (`data::String`, `mimeType::String`) |
| `agentMessage` | `type.jl:82` | Abstract base for all messages |
| `userMessage` | `type.jl:84` | User message (`role`, `content`, `timestamp`) |
| `assistantMessage` | `type.jl:111` | LLM response (`role`, `content`, `api`, `provider`, `model`, `usage`, `stopReason`, `errorMessage`, `timestamp`) |
| `toolResultMessage` | `type.jl:152` | Tool result (`role`, `toolCallId`, `toolName`, `content`, `details`, `usage`, `addedToolNames`, `isError`, `timestamp`) |

### Tool Types

| Type | Source | Description |
|------|--------|-------------|
| `agentTool` | `type.jl:261` | Tool definition (name, label, description, schema, execute, hooks) |
| `agentToolCall` | `type.jl:362` | Tool call from LLM (type, id, name, arguments) |
| `agentToolResult` | `type.jl:429` | Tool execution result (content, details, usage, terminate) |
| `agentToolCallBatch` | `type.jl:835` | Batch of tool results (messages, terminate) |

### Lifecycle Outcome Types

| Type | Source | Description |
|------|--------|-------------|
| `preparedToolCall` | `type.jl:681` | After prepare: (tool, toolCall, args) |
| `immediateOutcome` | `type.jl:719` | Failed before execution: (result, isError) |
| `executedOutcome` | `type.jl:753` | After execute, before finalize: (result, isError) |
| `finalizedOutcome` | `type.jl:787` | After all phases: (toolCall, result, isError) |

### Context & Config Types

| Type | Source | Description |
|------|--------|-------------|
| `agentContext` | `type.jl:299` | Conversation snapshot (systemPrompt, messages, tools) |
| `agentState` | `type.jl:310` | Mutable runtime state (systemPrompt, model, tools, messages, pendingToolCalls, activeRun, errorMessage) |
| `agentLoopConfig` | `type.jl:403` | Loop config (tools, beforeToolCall, afterToolCall, toolExecution) |
| `abortSignal` | `type.jl:416` | Abort flag (`aborted::Bool`) |
| `beforeToolCallContext` | `type.jl:445` | Context for beforeToolCall (message, toolCall, args, context) |
| `afterToolCallContext` | `type.jl:463` | Context for afterToolCall (message, toolCall, args, result, isError, context) |

### Event Types

| Type | Source | Description |
|------|--------|-------------|
| `toolExecStartEvent` | `type.jl:480` | (toolCallId, toolName, arguments) |
| `toolExecUpdateEvent` | `type.jl:495` | (toolCallId, toolName, arguments, partialResult) |
| `toolExecEndEvent` | `type.jl:511` | (toolCallId, toolName, result, isError) |

### Agent Types

| Type | Source | Description |
|------|--------|-------------|
| `agent` | `type.jl:522` | Abstract base type |
| `yiemAgent` | `type.jl:527` | High-level agent wrapper (state, channels, callbacks, task) |