This commit is contained in:
2026-08-10 14:55:09 +07:00
parent 1b69f69c7d
commit c78f4b023d
5 changed files with 242 additions and 119 deletions
+67 -34
View File
@@ -8,7 +8,7 @@ This document describes the complete tool lifecycle in the YiemAgent framework,
1. [Overview](#1-overview)
2. [Tool Definition — The `agentTool` Struct](#2-tool-definition--the-agenttool-struct)
3. [Tool Registration — The Global Registry](#3-tool-registration--the-global-registry)
3. [Tool Registration — Per-Agent Tool Stores](#3-tool-registration--per-agent-tool-stores)
4. [The Agent Loop — High-Level Flow](#4-the-agent-loop--high-level-flow)
5. [Message Processing Pipeline](#5-message-processing-pipeline)
6. [Tool Call Extraction from LLM Response](#6-tool-call-extraction-from-llm-response)
@@ -110,66 +110,98 @@ The `terminate` flag is checked at the batch level. See [Section 9](#9-tool-call
---
## 3. Tool Registration — The Global Registry
## 3. Tool Registration — Per-Agent Tool Stores
**Source:** `tools/registry.jl`
### How `loadTools()` Works
### 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
function loadTools(dir::String)::OrderedDict{String, agentTool}
struct ToolStore
tools::Vector{agentTool} # ordered tool list (for listTool iteration)
modules::Vector{Module} # keeps tool submodules alive to prevent GC
name::String # identifier for debugging/logs
end
```
**Source:** `tools/registry.jl:95-160`
### How `loadTools(store, dir)` Works
```julia
function loadTools(store::ToolStore, dir::String)::OrderedDict{String, agentTool}
```
**Source:** `tools/registry.jl:115-180`
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
```
```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. **Stores** the module reference in `_tool_modules` to prevent GC of closures
7. **Registers** the tool in `_registry` and returns an `OrderedDict{String, agentTool}`
6. **Stores** the module reference in `store.modules` to prevent GC of closures
7. **Registers** the tool in `store.tools` and returns an `OrderedDict{String, agentTool}`
### 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 in `_tool_modules` so closures (in `execute`, `validateRequiredArgs`, `prepareArguments`) don't get garbage collected
### Auto-Registration
The `_listTool()` is auto-registered in `__init__()` (line 16-18), so `listTools` is always available without explicit loading.
- The module reference is kept alive in `store.modules` so closures (in `execute`, `validateRequiredArgs`, `prepareArguments`) don't get garbage collected
### Registration API
```julia
# Auto-load from directory (returns OrderedDict keyed by tool name)
tools = loadTools("src/tools") # OrderedDict{String, agentTool}
# Create per-agent stores
store1 = ToolStore(name="agent1")
store2 = ToolStore(name="agent2")
# Manual registration (adds to global _registry)
registerTool(my_tool)
# Load tools into specific stores
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() # OrderedDict{String, agentTool} — O(1) lookup + deterministic order
all_tools = getTools(store1) # OrderedDict{String, agentTool} — O(1) lookup + deterministic order
# Clear
clearTools() # Empties _registry
# Clear (per-store)
clearTools(store1) # only clears store1
```
**Why `OrderedDict` for `getTools()`?** The internal `_registry` is a `Vector{agentTool}` for ordered iteration (used by `listTools`). `getTools()` builds an `OrderedDict` from `_registry` so callers get:
**Why `OrderedDict` for `getTools()`?** The internal `store.tools` is a `Vector{agentTool}` for ordered iteration (used by `listTool`). `getTools()` builds an `OrderedDict` from `store.tools` so callers get:
- O(1) lookup by tool name
- Deterministic iteration order (registration order: `listTools` auto-registered first, then tools loaded alphabetically by filename)
- Deterministic iteration order (registration order)
- Consistency with `agentState.tools` (also `OrderedDict{String, agentTool}`)
### 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.
---
## 4. The Agent Loop — High-Level Flow
@@ -185,6 +217,7 @@ yiemAgent struct contains:
- inputChannel (Channel, capacity 16) ← user sends messages here via run_agent()
- followUpChannel (Channel, capacity 32) ← user sends follow-ups here via follow_up()
- outputChannel (Channel, capacity 16) ← agent sends responses here via take_response()
- _tool_store (ToolStore) ← per-agent isolated tool registry
```
### Loop States
@@ -1083,11 +1116,11 @@ The framework supports tools that modify the tool system itself at runtime.
4. Appends `getTool()` returning an `agentTool` struct
5. Writes the combined string to `src/tools/<name>.jl`
### `listTools` — Discover Available Tools
### `listTool` — Discover Available Tools
**Source:** `tools/registry.jl:24-52`
**Source:** `tools/registry.jl:54-82`
Returns all registered tools. Primarily useful for **collision detection** before creating a new tool via `writeTool`.
Each `ToolStore` gets its own `listTool` instance bound to that store via `listTool(store)`, so each agent sees only its own tools. Primarily useful for **collision detection** before creating a new tool via `writeTool`.
### Self-Tooling Workflow
@@ -1101,7 +1134,7 @@ Returns all registered tools. Primarily useful for **collision detection** befor
- 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("src/tools") picks up the new file
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..."
```
@@ -1362,7 +1395,7 @@ module _tool_myTool
end
```
All functions in the file are scoped under `_tool_myTool`, preventing name collisions with other tools. The module reference is kept alive in `_tool_modules` to prevent garbage collection of closures.
All functions in the file are scoped under `_tool_myTool`, preventing name collisions with other tools. The module reference is kept alive in `store.modules` to prevent garbage collection of closures.
---
+108 -55
View File
@@ -1,27 +1,57 @@
module toolRegistry
export loadTools, registerTool, getTools, clearTools
export ToolStore, loadTools, registerTool, getTools, clearTools, listTool
using Dates
using JSON, DataStructures
using ..type
# Global registry — populated at runtime by loadTools() or registerTool()
const _registry = Vector{agentTool}()
"""
Per-agent isolated tool storage.
# Module references — kept alive to prevent GC of tool code that closures depend on
const _tool_modules = Vector{Module}()
Each agent gets its own `ToolStore` so tool registration is independent —
`registerTool(store, tool)` only affects that agent's tool set.
# Auto-register the built-in listTools tool
function __init__()
registerTool(_listTool())
# Fields
- `tools::Vector{agentTool}` — ordered tool list (for `listTool` iteration)
- `modules::Vector{Module}` — keeps tool submodules alive to prevent GC of closures
- `name::String` — identifier for debugging/logs
"""
struct ToolStore
tools::Vector{agentTool}
modules::Vector{Module}
name::String
end
"""
Create a new isolated tool store.
# Keyword Arguments
- `name::String`: Identifier for this store (default: "default")
# Examples
```julia
store = ToolStore(name="agent1")
tools = loadTools(store, "src/tools")
registerTool(store, my_tool)
agent = yiemAgent(tools=getTools(store), llmCall=..., _tool_store=store)
```
"""
function ToolStore(; name::String="default")::ToolStore
ToolStore(agentTool[], Module[], name)
end
"""
List tool definition — lets the agent query available tools for collision detection
when creating new tools via writeTool.
# Arguments
- `store::ToolStore`: The tool store to list from
Each `ToolStore` gets its own `listTool` instance bound to that store,
so each agent sees only its own tools.
"""
function _listTool()::agentTool
function listTool(store::ToolStore)::agentTool
return agentTool(
name = "listTools",
label = "List Tools",
@@ -32,7 +62,7 @@ function _listTool()::agentTool
"required" => Any[]
),
execute = (toolCallId, args, signal, onPartialResult) -> begin
tools = getTools()
tools = getTools(store)
if isempty(tools)
result_text = "No tools registered."
else
@@ -52,47 +82,37 @@ function _listTool()::agentTool
end
"""
Load all tool modules from a directory.
Load all tool modules from a directory into a specific ToolStore.
Scans `dir` for `.jl` files. Each file must define a function named
`getTool()::agentTool`. Files are sorted alphabetically so tool
`getTool()::agentTool`. Files are sorted alphabetically so tool
registration order is deterministic.
Each `.jl` file is loaded into its own **submodule** so that all functions
defined in the file (`validateRequiredArgs`, `prepareArguments`, `executeTool`,
and any helper functions) are namespaced and never collide with other tools.
# Tool file format
Each `.jl` file defines one function `getTool()` that returns an `agentTool`.
Inside the file you can freely define as many helper functions as you need —
they will all be scoped under the tool's submodule.
```julia
# src/tools/getWeather.jl
# These are namespaced — no collision with getTime.validateRequiredArgs, etc.
function validateRequiredArgs(args::Dict{String,Any})::Union{Nothing,String}
...
end
function getTool()::agentTool
return agentTool(
name = "getWeather",
...
)
end
```
# Arguments
- `store::ToolStore`: The tool store to register tools into
- `dir::String`: Directory path to scan for `.jl` tool files
# Returns
- `Vector{agentTool}`: All loaded tools
- `OrderedDict{String, agentTool}`: All loaded tools keyed by name
# Errors
- Throws `ArgumentError` if a tool file does not define a `getTool` function
# Examples
```julia
julia> store = ToolStore(name="agent1")
julia> tools = loadTools(store, "src/tools")
OrderedDict{String, agentTool} with 3 entries:
"getWeather" => agentTool(...)
"getTime" => agentTool(...)
"listTools" => agentTool(...)
```
"""
function loadTools(dir::String)::OrderedDict{String, agentTool}
function loadTools(store::ToolStore, dir::String)::OrderedDict{String, agentTool}
if !isdir(dir)
throw(ArgumentError("Tool directory does not exist: $dir"))
end
@@ -141,10 +161,10 @@ function loadTools(dir::String)::OrderedDict{String, agentTool}
# Keep module reference alive — closures in the agentTool (execute,
# validateRequiredArgs, prepareArguments) may reference module-scoped
# functions. Without this, GC could collect the module.
push!(_tool_modules, mod)
push!(_registry, tool)
push!(store.modules, mod)
push!(store.tools, tool)
tools[tool.name] = tool
println("[toolRegistry] Loaded tool: $(tool.name) $(tool.label)")
println("[$(store.name)] Loaded tool: $(tool.name) ($(tool.label))")
catch e
if e isa UndefVarError || occursin("getTool", sprint(showerror, e))
throw(ArgumentError(
@@ -160,42 +180,75 @@ function loadTools(dir::String)::OrderedDict{String, agentTool}
end
"""
Register a single agentTool into the global registry.
Register a single agentTool into a specific ToolStore.
# Arguments
- `store::ToolStore`: The tool store to register into
- `tool::agentTool`: The tool to register
# Returns
- `Vector{agentTool}`: Updated registry
- `Vector{agentTool}`: Updated tool list for this store
# Examples
```julia
julia> store = ToolStore(name="agent1")
julia> registerTool(store, my_tool)
[toolRegistry:agent1] Registered tool: my_tool
```
"""
function registerTool(tool::agentTool)::Vector{agentTool}
push!(_registry, tool)
println("[toolRegistry] Registered tool: $(tool.name)")
return _registry
function registerTool(store::ToolStore, tool::agentTool)::Vector{agentTool}
push!(store.tools, tool)
println("[$(store.name)] Registered tool: $(tool.name)")
return store.tools
end
"""
Get all registered tools as an `OrderedDict{String, agentTool}` keyed by tool name.
Get all registered tools from a specific ToolStore as an
`OrderedDict{String, agentTool}` keyed by tool name.
The internal `_registry` is a `Vector` for ordered iteration (used by `listTools`).
This function builds an `OrderedDict` from `_registry` so callers get:
The internal vector is for ordered iteration (used by `listTool`).
This function builds an `OrderedDict` so callers get:
- O(1) lookup by name
- Deterministic iteration order (registration order: alphabetical by filename)
- Deterministic iteration order (registration order)
- Consistency with `agentState.tools` (also `OrderedDict{String, agentTool}`)
# Arguments
- `store::ToolStore`: The tool store to query
# Returns
- `OrderedDict{String, agentTool}`: Copy of the registry keyed by tool name, in registration order
- `OrderedDict{String, agentTool}`: Tools keyed by name, in registration order
# Examples
```julia
julia> getTools(store)
OrderedDict{String, agentTool} with 3 entries:
"listTools" => agentTool(...)
"getWeather" => agentTool(...)
"getTime" => agentTool(...)
```
"""
function getTools()::OrderedDict{String, agentTool}
return OrderedDict{String, agentTool}(t.name => t for t in _registry)
function getTools(store::ToolStore)::OrderedDict{String, agentTool}
return OrderedDict{String, agentTool}(t.name => t for t in store.tools)
end
"""
Clear all registered tools from the global registry.
Clear all registered tools from a specific ToolStore.
# Arguments
- `store::ToolStore`: The tool store to clear
# Returns
- `nothing`
# Examples
```julia
julia> clearTools(store)
[toolRegistry:agent1] Registry cleared
```
"""
function clearTools()::Nothing
empty!(_registry)
println("[toolRegistry] Registry cleared")
function clearTools(store::ToolStore)::Nothing
empty!(store.tools)
println("[$(store.name)] Registry cleared")
return nothing
end
+3 -3
View File
@@ -7,7 +7,7 @@ The agent can use this tool when it encounters a task that no existing tool
can handle. Provide the tool's name, label, description, inputSchema, and
execute logic as Julia code. The tool is written to `src/tools/<name>.jl`.
After calling this tool, restart the agent so `loadTools("src/tools")` picks
After calling this tool, restart the agent so `loadTools(agent._tool_store, "src/tools")` picks
up the new file. The new tool is immediately available.
# Example
@@ -250,13 +250,13 @@ function getTool()::agentTool
tool_code = join(parts)
# Write the file — tool is loaded on next agent restart via loadTools()
# Write the file — tool is loaded on next agent restart via loadTools(store, "src/tools")
write(filepath, tool_code)
onPartialResult(Dict("status" => "Done"))
return agentToolResult(
[textContent("Tool '$(tool_name)' written to $filepath. Restart the agent so loadTools() picks it up, then call listTools to verify.")],
[textContent("Tool '$(tool_name)' written to $filepath. Restart the agent so loadTools(agent._tool_store, \"src/tools\") picks it up, then call listTools to verify.")],
Dict{Any,Any}(
"file" => filepath,
"name" => tool_name,
+11 -6
View File
@@ -567,9 +567,10 @@ mutable struct yiemAgent <: agent # High-level agent wrapper
# prepareNextTurnWithContext::Union{Function, Nothing} # Same but receives context
sessionId::Union{String, Nothing} # Optional session identifier
maxRetryDelayMs::Union{Int64, Nothing} # Maximum delay between retries (ms)
parallelToolExecute::Bool # Default: false
agentEventSink::Function # agent emits its status via this function
end
parallelToolExecute::Bool # Default: false
agentEventSink::Function # agent emits its status via this function
_tool_store::Any # Reference to the ToolStore for runtime registration
end
"""
Create a new yiemAgent instance with a background loop task.
@@ -593,15 +594,17 @@ on `inputChannel` and `followUpChannel` channels concurrently.
- `maxRetryDelayMs::Union{Int64, Nothing}`: Maximum delay between retries in milliseconds (default: `nothing`)
- `parallelToolExecute::Bool`: Run tool calls in parallel (default: `false`)
- `agentEventSink::Function`: Callback to receive agent events
- `tool_store::Union{Any, Nothing}`: ToolStore for runtime tool registration (default: `nothing`)
# Returns
- A new `yiemAgent` instance with an active background task
# Examples
```julia
julia> tools = loadTools("src/tools")
julia> agent = yiemAgent(systemPrompt="You are a helpful assistant", model=my_model, tools=tools, llmCall=...)
yiemAgent(agentState(...), Channel(...), Channel(...), Channel(...), ..., ...)
julia> store = ToolStore(name="agent1")
julia> tools = loadTools(store, "src/tools")
julia> agent = yiemAgent(systemPrompt="You are a helpful assistant", model=my_model, tools=tools, llmCall=..., tool_store=store)
yiemAgent(agentState(...), Channel(...), Channel(...), Channel(...), ..., store)
"""
function yiemAgent(
; systemPrompt::String="You are helpful assistant.",
@@ -619,6 +622,7 @@ function yiemAgent(
maxRetryDelayMs::Union{Int64, Nothing}=nothing,
parallelToolExecute::Bool=false,
agentEventSink::Function,
tool_store::Union{Any, Nothing}=nothing,
)
# Create channels: input (user -> agent), followUp (async queue), output (agent -> user)
inputChannel = Channel(16)
@@ -643,6 +647,7 @@ function yiemAgent(
maxRetryDelayMs,
parallelToolExecute,
agentEventSink,
tool_store,
)
# Spawn the background loop and attach it
+53 -21
View File
@@ -6,12 +6,13 @@ using YiemAgent.type
# Path to the real tools directory
TOOLS_DIR = joinpath(@__DIR__, "..", "src", "tools")
@testset "loadTools" begin
@testset "loadTools with ToolStore" begin
# ------------------------------------------------------------------ #
# 1. loadTools throws on non-existent directory #
# ------------------------------------------------------------------ #
@test_throws ArgumentError loadTools("/nonexistent/dir/that/does/not/exist")
store = ToolStore(name="test1")
@test_throws ArgumentError loadTools(store, "/nonexistent/dir/that/does/not/exist")
# ------------------------------------------------------------------ #
# 2. loadTools throws if a .jl file does not define getTool() #
@@ -20,12 +21,13 @@ TOOLS_DIR = joinpath(@__DIR__, "..", "src", "tools")
# ------------------------------------------------------------------ #
bad_dir = mktempdir()
write(joinpath(bad_dir, "noTool.jl"), "x = 42\n")
@test_throws ArgumentError loadTools(bad_dir)
@test_throws ArgumentError loadTools(store, bad_dir)
# ------------------------------------------------------------------ #
# 3. loadTools loads actual tool files from src/tools/ #
# ------------------------------------------------------------------ #
loaded = loadTools(TOOLS_DIR)
store2 = ToolStore(name="test2")
loaded = loadTools(store2, TOOLS_DIR)
@test !isempty(loaded)
@test length(loaded) == 3
@@ -98,20 +100,29 @@ TOOLS_DIR = joinpath(@__DIR__, "..", "src", "tools")
@test occursin("72°F", result_w2.content[1].text)
# ------------------------------------------------------------------ #
# 7. getTools / registerTool / clearTools #
# 7. getTools / registerTool / clearTools (per-store isolation) #
# ------------------------------------------------------------------ #
registry_tools = getTools()
@test !isempty(registry_tools)
@test "getTime" in keys(registry_tools)
@test "getWeather" in keys(registry_tools)
# listTools is auto-registered via __init__() → first key, then tools loaded alphabetically
@test collect(keys(registry_tools))[1] == "listTools"
@test collect(keys(registry_tools))[2] == "getTime"
@test collect(keys(registry_tools))[3] == "getWeather"
@test collect(keys(registry_tools))[4] == "writeTool"
store3 = ToolStore(name="test3")
registry_tools = getTools(store3)
@test isempty(registry_tools)
clearTools()
@test isempty(getTools())
# listTool is not auto-registered anymore — each store starts empty
# Register tools manually
registerTool(store3, loaded["getTime"])
registerTool(store3, loaded["getWeather"])
registerTool(store3, loaded["writeTool"])
reg = getTools(store3)
@test !isempty(reg)
@test "getTime" in keys(reg)
@test "getWeather" in keys(reg)
@test "writeTool" in keys(reg)
@test collect(keys(reg))[1] == "getTime"
@test collect(keys(reg))[2] == "getWeather"
@test collect(keys(reg))[3] == "writeTool"
clearTools(store3)
@test isempty(getTools(store3))
test_tool = agentTool(
name = "manualTool",
@@ -124,8 +135,8 @@ TOOLS_DIR = joinpath(@__DIR__, "..", "src", "tools")
validateRequiredArgs = nothing,
parallelToolExecute = true
)
registerTool(test_tool)
reg = getTools()
registerTool(store3, test_tool)
reg = getTools(store3)
@test haskey(reg, "manualTool")
@test length(reg) == 1
@test reg["manualTool"].parallelToolExecute == true
@@ -133,9 +144,30 @@ TOOLS_DIR = joinpath(@__DIR__, "..", "src", "tools")
# ------------------------------------------------------------------ #
# 8. getTools returns deep copy (mutations don't affect registry) #
# ------------------------------------------------------------------ #
copy1 = getTools()
copy2 = getTools()
copy1 = getTools(store3)
copy2 = getTools(store3)
@test copy1 !== copy2
empty!(copy1)
@test !isempty(getTools())
@test !isempty(getTools(store3))
# ------------------------------------------------------------------ #
# 9. Per-store isolation — two stores don't share tools #
# ------------------------------------------------------------------ #
storeA = ToolStore(name="isolationA")
storeB = ToolStore(name="isolationB")
registerTool(storeA, loaded["getTime"])
registerTool(storeB, loaded["getWeather"])
regA = getTools(storeA)
regB = getTools(storeB)
@test "getTime" in keys(regA)
@test "getWeather" keys(regA)
@test "getWeather" in keys(regB)
@test "getTime" keys(regB)
clearTools(storeA)
@test isempty(getTools(storeA))
@test !isempty(getTools(storeB)) # storeB unaffected
end