4.3 KiB
SKILL: Sub-Agent Manager (Concurrency & Parallelism)
Overview
The Sub-Agent Manager enables the Neurosymbolic Lisp Machine to handle multiple concurrent thoughts. It allows the primary kernel to "spawn" lightweight, isolated Lisp threads (sub-agents) to perform long-running or background tasks (research, massive refactors, etc.) without blocking the main event bus.
Phase A: Demand (PRD)
1. Purpose
Define the interfaces for parallel cognitive execution and thread lifecycle management.
2. User Needs
- Non-Blocking Execution: Spawn background threads for long-running tasks.
- Context Isolation: Sub-agents must have their own execution context to prevent parent context poisoning.
- Communication Loop: Sub-agents must inject a "Return Stimulus" upon completion.
- Observability: Ability to list and terminate active sub-agents.
3. Success Criteria
TODO Successful spawning of a non-blocking background thread
TODO Verification of context isolation (distinct local variables)
TODO Autonomous injection of :sub-agent-complete stimulus
TODO Thread safety verification using bordeaux-threads locks
Phase B: Blueprint (PROTOCOL)
Phase B: Blueprint (PROTOCOL)
1. Architectural Intent
The Sub-Agent Manager is designed as a facade over a thread management library (initially `bordeaux-threads`). It provides a high-level API for spawning, managing, and monitoring sub-agents. The core principle is to create isolated Lisp environments for each sub-agent, encapsulating all state and preventing interference with the main system or other sub-agents. Communication back to the main kernel occurs through a standardized `:sub-agent-complete` stimulus injected into the event bus. Thread safety, enforced with locks where necessary, is paramount.
2. Semantic Interfaces (Lisp Signatures)
`spawn-sub-agent (task-fn &key name)`
- Purpose: Creates and starts a new sub-agent thread.
-
Parameters:
- `task-fn`: A function of no arguments that contains the code to be executed in the sub-agent.
- `name`: (optional) A symbol representing the name of the sub-agent for identification and debugging.
- Returns: A sub-agent object (e.g., a struct) representing the spawned thread, containing its ID, status, and other metadata.
- Side Effects: Creates a new thread and starts the execution of `task-fn` within it.
`kill-sub-agent (sub-agent)`
- Purpose: Terminates a running sub-agent.
-
Parameters:
- `sub-agent`: The sub-agent object (returned by `spawn-sub-agent`) representing the thread to terminate.
- Returns: `T` if the sub-agent was successfully terminated, `NIL` otherwise.
- Side Effects: Attempts to terminate the specified thread, potentially releasing any resources held by the sub-agent.
`list-sub-agents ()`
- Purpose: Returns a list of all active sub-agents.
- Parameters: None
- Returns: A list of sub-agent objects, each representing a running sub-agent.
`sub-agent-status (sub-agent)`
- Purpose: Returns the current status of a sub-agent.
-
Parameters:
- `sub-agent`: The sub-agent object to query.
- Returns: A symbol representing the status of the sub-agent (e.g., `:running`, `:completed`, `:terminated`, `:error`).
`inject-sub-agent-completion-stimulus (result &key sub-agent)`
- Purpose: This PRIVATE function (not exposed directly) is called by the sub-agent, to inject knowledge of the result of its process into the stimulus stream.
-
Parameters:
- `result`: The result of the sub-agent's computation.
- `sub-agent`: The current sub-agent (optional).
- Returns: `T` if stimulus was injected successfully
- Side Effects: Injects a `:sub-agent-complete` stimulus into the event bus. The stimulus will contain the `result` and any metadata associated with the `sub-agent` (including its name/id). The stimulus will be of the form `(:type :sub-agent-complete :result <result> :sub-agent <sub-agent>)`