robot_lab 0.2.8 → 0.3.1

data/tool_manifest_plan.md

@@ -1,155 +0,0 @@
-# ToolManifest Evolution Plan
-
-## Concept: Dynamic Tool Selection in Pipelines
-
-There is a pattern where the total number of tools available within RobotLab numbers in the hundreds -- too many to send to any single LLM call. A pipeline uses a "selector" robot that reviews the next robot's prompt and determines which tools should be made available. The selected subset is then applied to the target robot via `with_tools(...)`.
-
-### Example Flow
-
-```
-Selector Robot (step 1)
-  - Receives the user prompt destined for step 2
-  - Has access to a summary of all available tools (name + description)
-  - Outputs a list of tool names relevant to the task
-
-Target Robot (step 2)
-  - Receives the selected tools via with_tools(...)
-  - Runs with only the relevant subset (e.g., 5-10 tools instead of hundreds)
-```
-
-## Current State
-
-`ToolManifest` (lib/robot_lab/tool_manifest.rb, 218 lines) is a name-keyed Hash wrapper with:
-- Lookup by name (`[]`, `fetch`)
-- `Enumerable` support
-- Merge/replace operations
-- Serialization (`to_h`, `to_json`, `from_hash`)
-
-It is not currently used by any production code path. `Robot` stores tools as plain arrays (`@local_tools`, `@mcp_tools`) and `all_tools` concatenates them.
-
-## What It Needs to Become
-
-### 1. Global Registry Entry Point
-
-A singleton or module-level registry where all available tools accumulate:
-
-```ruby
-RobotLab.tool_registry  # => ToolManifest (global instance)
-RobotLab.tool_registry.add(MyTool)
-RobotLab.tool_registry.names  # => ["my_tool", "weather", "calculator", ...]
-```
-
-### 2. Metadata for Selection
-
-Categories, tags, or other metadata so the selector robot can reason about which tools are relevant without seeing full schemas:
-
-```ruby
-# Registration with metadata
-RobotLab.tool_registry.add(MyTool, tags: [:database, :read_only])
-
-# Filtering
-RobotLab.tool_registry.tagged(:database)  # => subset manifest
-```
-
-### 3. Search-Friendly Summary Output
-
-The selector robot needs a compact summary it can reason over -- name + description pairs, not full JSON schemas:
-
-```ruby
-RobotLab.tool_registry.summary
-# => [{ name: "get_weather", description: "Get current weather for a location" },
-#     { name: "calculate", description: "Performs basic arithmetic" }, ...]
-```
-
-### 4. Subset Extraction by Name
-
-Resolve a list of tool names (the selector robot's output) into actual tool objects:
-
-```ruby
-names = ["get_weather", "calculate"]  # from selector robot
-tools = RobotLab.tool_registry.select_by_names(names)
-target_robot.with_tools(tools)
-```
-
-### 5. Integration with Robot
-
-`Robot#all_tools` and `with_tools(...)` should accept a ToolManifest or be able to pull from the global registry.
-
-## Design Decisions Still Needed
-
-- Should MCP-discovered tools auto-register in the global manifest?
-- How does the selector robot receive the tool summary -- as part of its system prompt, injected into the user message, or via a tool call?
-- Should categories/tags be defined on the Tool class itself (e.g., `Tool.create(name: ..., tags: [...])`), or only at registration time?
-- Does the manifest need versioning or hot-reload support for long-running processes?
-
----
-
-## Review Notes
-
-### The Problem Statement is Strong
-
-The "selector robot" pattern — an LLM choosing tools from hundreds based on a compact summary — is a real need. Sending 200+ tool schemas in every API call is both expensive and degrades LLM accuracy. The plan correctly identifies that `ToolManifest` as it exists today is a passive container that nothing actually uses.
-
-### What's Solid
-
-- **Global registry** (section 1) — The right call. Tools currently scatter across `@local_tools` and `@mcp_tools` per-robot with no central view. `RobotLab.tool_registry` is the obvious home.
-- **Summary output** (section 3) — Essential for the selector pattern. Name + description pairs are the right granularity for LLM reasoning.
-- **Subset extraction** (section 4) — `select_by_names` maps cleanly to what `ToolConfig.filter_tools` already does, but at the registry level.
-
-### Gaps and Concerns
-
-**1. Tags/metadata may be premature.**
-
-Section 2 proposes tags (`tags: [:database, :read_only]`) and `.tagged(:database)` filtering. This adds a classification system that someone has to maintain. In the selector-robot pattern, the LLM is doing the classification — that's the whole point. If you also hand-tag tools, you're duplicating the selector's job.
-
-A simpler alternative: let `summary` be the only interface the selector sees. If you later want pre-filtering (e.g., "only show the selector database-related tools"), add it then. Right now it's YAGNI.
-
-**2. MCP tool auto-registration is the key decision, not a footnote.**
-
-The plan lists "Should MCP-discovered tools auto-register in the global manifest?" as an open question. This is actually the central architectural decision. Looking at `MCPManagement#discover_mcp_tools`, MCP tools are created dynamically per-robot, per-connection. They hold closures over their specific `mcp_client` instance. If you auto-register them globally, you need to handle:
-
-- **Lifecycle** — MCP tools are only valid while their client is connected. The global registry would hold dead references after `disconnect`.
-- **Duplicates** — Two robots connecting to the same MCP server would register the same tool names with different client closures.
-- **Scope** — An MCP tool created for robot A shouldn't be callable by robot B unless they share the client.
-
-Suggestion: MCP tools should **not** auto-register. The global registry should hold "tool definitions" (name, description, schema) — not live callable instances. The selector robot picks names; the target robot resolves those names against its own live tools.
-
-**3. The plan doesn't address the split between "tool definition" and "tool instance."**
-
-Currently `Tool` is both the definition (name, description, parameters) and the callable (the `execute` method with its closure). For a global registry, you want the definition without the closure. The selector robot doesn't need to call tools — it just needs to read summaries and output names.
-
-This suggests `ToolManifest` should store lightweight descriptors, and `Robot#filtered_tools` continues resolving names to live instances from `all_tools`.
-
-**4. Integration with `Robot#with_tools` needs more thought.**
-
-The plan says "`with_tools(...)` should accept a ToolManifest." But `with_tools` currently delegates to `@chat.with_tools` from RubyLLM, which expects tool class/instance objects. Accepting a manifest means either:
-
-- Converting manifest entries back to tool instances (requires the live tool lookup), or
-- Passing tool names and letting the robot resolve internally (which is what `filtered_tools` already does)
-
-The second path is simpler and already mostly works. The plan could just say: "The selector robot outputs tool names. Pass them to `robot.run(..., tools: names)`." That's the existing `tools:` parameter.
-
-**5. Missing: how the selector robot itself gets built.**
-
-The plan describes the pattern but not the API. Is this a special network task type? A built-in robot template? A method on `ToolManifest`? Something like:
-
-```ruby
-network = RobotLab.create_network(name: "smart_pipeline") do
-  task :selector, selector_robot, depends_on: :none
-  task :worker, worker_robot, tools: :from_selector, depends_on: [:selector]
-end
-```
-
-The `tools: :from_selector` (or similar) is the glue that needs designing.
-
-### Suggested Simplification
-
-Instead of evolving ToolManifest into a full registry with tags, a narrower v1:
-
-1. **`RobotLab.tool_registry`** — Global `ToolManifest` that accumulates tool descriptors (name + description only, no live instances)
-2. **`ToolManifest#summary`** — Returns the compact format for selector prompts
-3. **`ToolManifest#select_by_names(names)`** — Returns a sub-manifest (already close to what exists)
-4. **Auto-registration of local tools at `Robot.new`** — Push descriptors into the registry when robots are created
-5. **No auto-registration of MCP tools** — Too coupled to connection lifecycle
-
-Skip tags, skip `with_tools(manifest)` integration, skip hot-reload. Those can come later if needed.