Files
passepartout/harness/reason.org

502 lines
20 KiB
Org Mode

#+PROPERTY: header-args:lisp :tangle (expand-file-name "harness/reason.lisp" (expand-file-name "harness/"))
#+TITLE: Stage 2: Reason (reason.lisp)
#+AUTHOR: Amr
#+FILETAGS: :harness:reason:
#+STARTUP: content
* Stage 2: Reason (reason.lisp)
** Architectural Intent: The Dual-Engine Cognitive Architecture
The Reason stage implements the core innovation of OpenCortex: the separation of probabilistic reasoning (neural/LLM) from deterministic verification (logic/safety).
This dual-engine design solves a fundamental problem in AI safety:
1. *Probabilistic Engine* - Uses LLMs for semantic understanding, natural language generation, and complex reasoning. It is powerful but can hallucinate, make syntax errors, or propose unsafe actions.
2. *Deterministic Engine* - Uses formal verification (skills) to check LLM proposals before execution. It is slower but provably correct.
The LLM proposes; the skills verify. This is the "Bouncer Pattern" - the deterministic engine is literally a bouncer that checks the LLM's proposals at the door before letting them through to execution.
** Why Plists for Communication?
The Reason stage communicates exclusively through property lists (plists). This design choice reflects the homoiconic nature of Lisp - plists are native data structures that can be read, written, and manipulated by the same code that processes them.
A plist message like:
: (TYPE :REQUEST TARGET :CLI PAYLOAD (ACTION :MESSAGE TEXT "Hello"))
Is simultaneously:
- Human-readable text
- Machine-parseable data structure
- Executable Lisp code
This means the reasoning pipeline can generate, modify, and execute its own communication protocol without external parsing.
* Package Context
#+begin_src lisp
(in-package :opencortex)
#+end_src
* Probabilistic Engine (Neural/LLM Integration)
The probabilistic engine is responsible for all neural/LLM operations. It maintains a registry of provider backends and implements a cascading failover mechanism.
** Backend Registry Variables
#+begin_src lisp
(defvar *probabilistic-backends* (make-hash-table :test 'equal)
"Registry mapping provider keywords (:openrouter, :ollama) to their calling functions.")
(defvar *provider-cascade* nil
"Ordered list of provider keywords to try. First available provider wins.")
(defvar *model-selector-fn* nil
"Optional function that selects a specific model for each provider.
Signature: (funcall fn provider context) => model-name-string")
(defvar *consensus-enabled-p* nil
"When T, run multiple providers and compare results for critical decisions.")
#+end_src
** register-probabilistic-backend: Backend Registration
#+begin_src lisp
(defun register-probabilistic-backend (name fn)
"Register a neural provider backend.
NAME is a keyword like :openrouter or :ollama.
FN is a function with signature: (funcall fn prompt system-prompt &key model)
returning either:
- (list :status :success :content \"response text\")
- (list :status :error :message \"error description\")
- a simple string on success
Example registration:
(register-probabilistic-backend :openrouter #'openrouter-call)"
(setf (gethash name *probabilistic-backends*) fn))
#+end_src
** probabilistic-call: Cascade Dispatch
#+begin_src lisp
(defun probabilistic-call (prompt &key
(system-prompt "You are the Probabilistic engine.")
(cascade nil)
(context nil))
"Dispatch a neural request through the provider cascade.
PROMPT - The user's query or task description.
SYSTEM-PROMPT - Instructions for how the LLM should behave.
CASCADE - Override the default provider cascade.
CONTEXT - Current signal context (for model selection).
Returns the LLM response as a string, or a failure plist if all providers fail.
The cascade mechanism ensures reliability: if OpenRouter is rate-limited,
it automatically falls back to OpenAI, then Anthropic, etc."
(let ((backends (or cascade *provider-cascade*)))
(or (dolist (backend backends)
(let ((backend-fn (gethash backend *probabilistic-backends*)))
(when backend-fn
(harness-log "PROBABILISTIC: Attempting backend ~a..." backend)
;; Optional model selection based on context
(let* ((model (when *model-selector-fn*
(funcall *model-selector-fn* backend context)))
(result (if model
(funcall backend-fn prompt system-prompt :model model)
(funcall backend-fn prompt system-prompt))))
;; Normalize result format
(cond ((and (listp result) (eq (getf result :status) :success))
(return (getf result :content)))
((stringp result)
(return result))
(t
(harness-log "PROBABILISTIC: Backend ~a failed: ~a"
backend (getf result :message))))))))
;; All providers failed
(list :type :LOG
:payload (list :text "Neural Cascade Failure: All providers exhausted.")))))
#+end_src
* Cognitive Proposal Generation (Think)
The `think` function is the heart of the probabilistic engine. It constructs a prompt from context, sends it to the LLM, and parses the response into a structured action.
** strip-markdown: Clean LLM Output
#+begin_src lisp
(defun strip-markdown (text)
"Strip markdown formatting from LLM output.
LLMs often wrap their responses in code fences (```lisp ...```).
This function removes those markers to extract the raw plist.
Handles:
- Leading code fences with language tags: ```lisp
- Trailing code fences: ```
- Orphan closing fences: ```"
(if (and text (stringp text))
(let ((cleaned text))
(setf cleaned (cl-ppcre:regex-replace-all "^```[a-z]*\\n" cleaned ""))
(setf cleaned (cl-ppcre:regex-replace-all "\\n```$" cleaned ""))
(setf cleaned (cl-ppcre:regex-replace-all "```" cleaned ""))
(string-trim '(#\Space #\Newline #\Tab) cleaned))
text))
#+end_src
** normalize-plist-keywords: Fix LLM Keyword Output
#+begin_src lisp
(defun normalize-plist-keywords (plist)
"Normalize all keys in a plist to keywords.
LLMs often return plists with unquoted keys: (TYPE REQUEST ...)
instead of keyword syntax: (:TYPE :REQUEST ...)
This function converts all symbol keys to their keyword equivalents,
making the plist compatible with standard Lisp property accessors.
Example transformation:
(TYPE REQUEST PAYLOAD (ACTION MESSAGE TEXT \"Hi\"))
=> (:TYPE :REQUEST :PAYLOAD (:ACTION :MESSAGE :TEXT \"Hi\"))"
(when (listp plist)
(loop for (k . rest) on plist by #'cddr
collect (if (and (symbolp k) (not (keywordp k)))
(intern (string k) :keyword)
k)
collect (car rest))))
#+end_src
** think: Generate Action Proposal
#+begin_src lisp
(defun think (context)
"Generate a Lisp action proposal based on current context.
This is the core cognitive function. It:
1. Finds the most relevant skill based on context
2. Assembles global awareness (memory context, system logs)
3. Constructs a detailed prompt with available tools
4. Calls the LLM via probabilistic-call
5. Parses the LLM response into a structured action plist
The LLM is instructed to respond with exactly ONE plist, never prose.
This constraint makes parsing deterministic and prevents rambling.
Returns a plist with structure:
(:TYPE :REQUEST :TARGET :CLI :PAYLOAD (:ACTION :MESSAGE :TEXT \"...\"))"
;; Gather context components
(let* ((active-skill (find-triggered-skill context))
(tool-belt (generate-tool-belt-prompt))
(global-context (context-assemble-global-awareness))
(system-logs (context-get-system-logs))
(assistant-name (or (getenv "MEMEX_ASSISTANT") "Agent"))
(rejection-trace (proto-get (proto-get context :payload) :rejection-trace)))
;; Generate prompt from skill or raw text
(let* ((prompt-generator (when active-skill
(skill-probabilistic-prompt active-skill)))
(raw-prompt (if prompt-generator
(funcall prompt-generator context)
;; Fallback: use raw user input
(let ((p (proto-get (proto-get context :payload) :text)))
(if (and p (stringp p))
p
"Maintain metabolic stasis."))))
;; Inject Reflection Loop feedback if a previous proposal was rejected
(reflection-feedback (if rejection-trace
(format nil "~%~%PREVIOUS PROPOSAL REJECTED:~%Your previous proposal was rejected by the deterministic safety gates.~%Rejection Trace: ~a~%You MUST fix the syntax or logic error described above and try again." rejection-trace)
""))
(system-prompt (format nil
"IDENTITY: ~a~a
You are a component of the OpenCortex neurosymbolic AI agent.
Your task is to generate exactly ONE valid Lisp plist response.
MANDATE: Respond with ONE Lisp plist. Never output prose.
IMPORTANT: To reply to the user, you MUST use:
(:TYPE :REQUEST :PAYLOAD (:ACTION :MESSAGE :TEXT \"<Response Text>\"))
To call a tool, you MUST use:
(:TYPE :REQUEST :TARGET :TOOL :ACTION :CALL :TOOL \"<name>\" :ARGS (:arg1 \"val\"))
MANDATORY VALIDATION RULE: Before declaring any Lisp code edit complete,
you MUST call the `:validate-lisp` tool with the proposed code. If the tool
returns `:status :error`, read the `:reason` and `:failed` fields, fix the
defect, and re-validate. You are strictly forbidden from relying on your
own paren-balancing or syntax intuition.
PROVIDER RULE: Always use the default cascade provider unless a specific
model or capability is required for the task.
AVAILABLE TOOLS:
~a
GLOBAL CONTEXT:
~a
RECENT LOGS:
~a"
assistant-name
reflection-feedback
tool-belt
global-context
system-logs)))
;; Call LLM and process response
(let* ((thought (probabilistic-call raw-prompt
:system-prompt system-prompt
:context context))
(cleaned (strip-markdown thought))
(meta (proto-get context :meta))
(source (proto-get meta :source)))
(when cleaned
(harness-log "THINK: LLM raw output = ~a"
(subseq cleaned 0 (min 200 (length cleaned)))))
;; Parse LLM response
(if (and cleaned (stringp cleaned) (> (length cleaned) 0))
(let ((*read-eval* nil))
(if (char= (char cleaned 0) #\()
;; Response starts with paren - try to parse as plist
(handler-case
(let ((parsed (read-from-string cleaned)))
(when parsed
(harness-log "THINK: parsed = ~a" parsed)
;; Normalize keyword keys (LLM often returns TYPE instead of :TYPE)
(let ((parsed-normalized (normalize-plist-keywords parsed))
(type (proto-get parsed :TYPE))
(target (or (proto-get parsed :TARGET)
(proto-get parsed :target))))
(cond
;; Recognized message type - use directly
((member type '(:REQUEST :EVENT :STATUS :RESPONSE))
(unless (proto-get parsed :target)
(setf (getf parsed :target) (or source :CLI)))
parsed-normalized)
;; Tool call detected - wrap in standard envelope
((or (eq target :TOOL)
(eq target :tool)
(getf parsed :TOOL)
(getf parsed :tool)
(and (listp parsed)
(listp (car parsed))
(keywordp (caar parsed))))
(list :TYPE :REQUEST
:TARGET :TOOL
:PAYLOAD (normalize-plist-keywords parsed)))
;; Unknown format - treat as user message
(t
(list :TYPE :REQUEST
:TARGET (or source :CLI)
:PAYLOAD (list :ACTION :MESSAGE :TEXT cleaned)))))))
(error (c)
(harness-log "THINK ERROR: ~a" c)
(list :TYPE :REQUEST
:TARGET (or source :CLI)
:PAYLOAD (list :ACTION :MESSAGE :TEXT cleaned))))
;; No leading paren - treat as plain text message
(list :TYPE :REQUEST
:TARGET (or source :CLI)
:PAYLOAD (list :ACTION :MESSAGE :TEXT cleaned))))
;; No response from LLM
thought)))))
#+end_src
* Deterministic Engine (Formal Verification)
The deterministic engine runs all registered skills' verification functions. This is where safety checks, policy enforcement, and skill-specific processing happen.
** deterministic-verify: Skill Chain Verification
#+begin_src lisp
(defun deterministic-verify (proposed-action context)
"Run all skill deterministic gates on a proposed action.
Each skill can define a deterministic function that either:
- Passes the action through unchanged
- Modifies the action (adds explanation, changes target, etc.)
- Blocks the action (returns a :LOG message instead)
Skills are sorted by priority (highest first). A skill with higher
priority can intercept and modify actions before lower-priority
skills see them.
The Bouncer Pattern: If any skill returns a :LOG or :EVENT type,
processing stops and that message is returned immediately. This
allows skills to veto actions.
Example skill chain:
1. Policy skill (priority 500) - checks for missing explanations
2. Protocol validator (priority 95) - validates message schema
3. Shell actuator guard (priority 50) - checks command whitelist"
(let ((current-action proposed-action)
(skills nil))
;; Collect all skills with deterministic functions
(maphash (lambda (name skill)
(declare (ignore name))
(when (skill-deterministic-fn skill)
(push skill skills)))
*skills-registry*)
;; Sort by priority (highest first)
(setf skills (sort skills #'> :key #'skill-priority))
;; Run each skill's gate
(dolist (skill skills)
(let ((trigger (skill-trigger-fn skill))
(gate (skill-deterministic-fn skill)))
;; Skill activates if no trigger or trigger returns true
(when (or (null trigger)
(ignore-errors (funcall trigger context)))
;; Run the gate
(let ((next-action (funcall gate current-action context)))
(let ((original-type (proto-get current-action :type)))
;; Check if skill intercepted (returned LOG/EVENT instead of REQUEST)
(when (and (listp next-action)
(member (proto-get next-action :type)
'(:LOG :EVENT :log :event))
(or (not (member original-type '(:LOG :EVENT :log :event)))
(not (eq next-action current-action))))
;; Skill blocked or modified - stop processing
(harness-log "DETERMINISTIC: Intercepted by skill '~a'"
(skill-name skill))
(return-from deterministic-verify next-action)))
;; Action passed through - continue to next skill
(setf current-action next-action)))))
;; Return final action (may be modified by skills, or original if all passed)
current-action))
#+end_src
* Reason Gate (Pipeline Stage)
** reason-gate: The Stage Function
#+begin_src lisp
(defun reason-gate (signal)
"Stage 2 of the metabolic pipeline: Reason.
Transforms perceived signals into approved actions by combining:
1. Probabilistic reasoning (LLM generates proposal)
2. Deterministic verification (skills validate proposal)
Only processes :EVENT signals with :user-input or :chat-message sensors.
Other signals pass through unchanged (heartbeats, tool outputs, etc.).
Modifies the signal in place by setting:
- :approved-action - The final verified action, or NIL
- :status - :reasoned
Returns the modified signal."
(let* ((type (proto-get signal :type))
(payload (proto-get signal :payload))
(sensor (proto-get payload :sensor)))
;; Only reason about user input, not internal signals
(unless (and (eq type :EVENT)
(member sensor '(:user-input :chat-message)))
(return-from reason-gate signal))
;; Reflection Loop: Retry up to 3 times if deterministic gates reject
(let ((retries 3)
(current-signal (copy-tree signal))
(last-rejection nil))
(loop
(when (<= retries 0)
(harness-log "REASON: Reflection loop exhausted. Final rejection.")
(setf (getf signal :approved-action) last-rejection)
(setf (getf signal :status) :reasoned)
(return signal))
(when last-rejection
(setf (getf (getf current-signal :payload) :rejection-trace) last-rejection))
(let ((candidate (think current-signal)))
(harness-log "REASON: candidate type = ~a" (type-of candidate))
(if (and candidate
(listp candidate)
(or (keywordp (car candidate))
(eq (car candidate) 'TYPE)
(eq (car candidate) 'type)))
(let ((verified (deterministic-verify candidate current-signal)))
(if (member (getf verified :type) '(:LOG :EVENT :log :event))
(progn
(harness-log "REASON: Proposal rejected by gate. Retrying (~a left)." (1- retries))
(decf retries)
(setf last-rejection verified))
(progn
(setf (getf signal :approved-action) verified)
(setf (getf signal :status) :reasoned)
(return signal))))
(progn
(harness-log "REASON: Invalid candidate type ~a, dropping" (type-of candidate))
(setf (getf signal :approved-action) nil)
(setf (getf signal :status) :reasoned)
(return signal))))))))
#+end_src
* Test Suite
These tests verify the Reason (cognitive) pipeline. Run with:
~(fiveam:run! 'pipeline-reason-suite)~
#+begin_src lisp :tangle (expand-file-name "harness/pipeline-reason-tests.lisp" (concat (concat (or (getenv "INSTALL_DIR") ".") "/harness") "/tests"))
(defpackage :opencortex-pipeline-reason-tests
(:use :cl :fiveam :opencortex)
(:export #:pipeline-reason-suite))
(in-package :opencortex-pipeline-reason-tests)
(def-suite pipeline-reason-suite
:description "Test suite for Reason pipeline")
(in-suite pipeline-reason-suite)
(test test-decide-gate-safety
"Decide gate should block unsafe LLM proposals."
;; Setup: clear skills and register mock
(clrhash opencortex::*skills-registry*)
(opencortex::defskill :mock-safety
:priority 50
:trigger (lambda (ctx) t)
:probabilistic (lambda (ctx) "Mock probabilistic")
:deterministic (lambda (action ctx)
(list :type :LOG :payload (list :text "Action rejected by skill heuristics"))))
(let* ((candidate (list :type :REQUEST :payload (list :action :eval :code "(shell-command \"rm -rf /\")")))
(signal (list :type :EVENT :candidate candidate))
(result (deterministic-verify candidate signal)))
(is (eq :LOG (getf result :type)))
(is (search "Action rejected by skill heuristics" (getf (getf result :payload) :text)))))
#+end_src