Files
passepartout/org/core-pipeline.org
Amr Gharbeia c227877302 v0.8.3: TUI stabilization — box calls, package fixes, sandbox, configure
Bug fixes:
- Fix box() calls: set color-pair before box, pass ACS default chtype integers
- Fix markdown functions: move to passepartout.channel-tui package where
  Croatoan is imported; use add-attributes/remove-attributes instead of
  :bold/:underline kwargs to add-string; call theme-color in gate-trace-lines
  to convert theme keys to Croatoan colors
- Fix sandbox: remove dex:get/dex:post from restricted symbols
  (blocked neuro-provider from loading)
- Export *log-lock* from passepartout (was unbound in jailed skill packages)
- Fix configure: always deploy to XDG, skip cp when source==dest
- Fix bash crash handler format string (~~ escaping)
- Revert test reorder in 28 files (caused package leakage in skill loader)

Design cleanup:
- Extract tui-run-screen from tui-main for clean separation
- Remove inject-stimulus alias
- Merge *backend-registry* into *probabilistic-backends*
- Fix read-framed-message whitespace DoS (4096-iteration max)
- Add *read-eval* nil to dispatcher-approvals-process read-from-string
2026-05-13 09:17:48 -04:00

433 lines
19 KiB
Org Mode

#+TITLE: The Metabolic Loop (loop.lisp)
#+AUTHOR: Agent
#+FILETAGS: :harness:loop:
#+STARTUP: content
#+PROPERTY: header-args:lisp :tangle ../lisp/core-pipeline.lisp
* Overview: Architectural Intent
The Metabolic Loop is the cranial nerve reflex of Passepartout. While skills provide specialized intelligence, the loop provides the fundamental rhythm of existence: the continuous processing of signals from perception through cognition to action.
Every signal flows through three stages:
1. **Perceive** — normalize raw input into a standard Signal format
2. **Reason** — think (LLM) then verify (deterministic gates)
3. **Act** — dispatch the approved action to the appropriate actuator
If a stage produces a new signal (e.g., the Act stage produces a tool-output event), that signal feeds back into Perceive and the loop continues. This is how the agent has multi-step conversations: each LLM response produces an action, which produces a tool output, which feeds back as a new perception, which triggers the next reasoning cycle.
** Why Separate Stages?
A single function that called the LLM, checked safety, and executed the result would be simpler to write. But it would be impossible to:
- Test each stage independently (a bug in the LLM call would block safety testing)
- Insert new stages between P and R or R and A (adding consensus means adding a gate in the middle)
- Recover from failures mid-pipeline (an LLM timeout shouldn't prevent safety checks on the next cycle)
The stage separation is the functional equivalent of the "thin harness" principle: each stage is a pure function that transforms a signal. The loop is the composition of these functions.
** Why the Depth Limit?
A signal that generates another signal that generates another signal can infinite-loop. The depth limit (max 10) prevents this. If depth exceeds 10, the signal is silently dropped. This is the metabolic loop's circuit breaker.
The three-tier error recovery model, now backed by a condition hierarchy
that skills can hook into via ~handler-bind~:
1. **Transient errors** (tool failures, network timeouts) — recoverable, generate a :loop-error signal at higher depth for retry. Use the ~skip-signal~ or ~use-fallback~ restart.
2. **Critical errors** (undefined functions, malformed data) — require memory rollback to the last snapshot.
3. **Recursive loops** (signals generating more signals indefinitely) — depth limit enforcement.
Condition types available for structured error handling:
- ~pipeline-error~ — any Perceive→Reason→Act failure
- ~llm-error~ — provider timeout, cascade exhaustion, API error (slots: provider, cascade, attempt-count)
- ~gate-error~ — dispatcher blocked a proposed action (slots: gate-name, rejected-action)
- ~budget-error~ — session cap exceeded (slots: remaining, requested)
- ~protocol-error~ — malformed message or framing failure
** Contract
1. (loop-process signal): the full pipeline loop — Perceive → Reason
→ Act. Enforces depth limit (10). Catches errors with rollback and
~:loop-error~ re-injection on non-terminal errors below depth 2.
Establishes restart options: ~skip-signal~ (drop the event),
~use-fallback text~ (inject canned response), ~abort-pipeline~
(clean exit). Skills can invoke these restarts from ~handler-bind~
clauses on the condition hierarchy.
2. (process-signal signal): thin alias for ~loop-process~.
3. (diagnostics-startup-run): runs health check on startup, sets
~*system-health*~ to ~:healthy~, ~:degraded~, or ~:unhealthy~.
4. *passepartout-error* condition hierarchy: ~pipeline-error~,
~llm-error~ (provider, cascade, attempt-count slots), ~gate-error~
(gate-name, rejected-action slots), ~budget-error~ (remaining,
requested slots), ~protocol-error~ (raw-message slot). All carry a
~:message~ string via the root ~passepartout-error~.
* Implementation
** Package Context
#+begin_src lisp
(in-package :passepartout)
#+end_src
** Error Condition Hierarchy
The pipeline defines a condition hierarchy so callers can distinguish
failure modes without inspecting raw error strings. Every pipeline
condition carries structured slots for telemetry and restart selection.
Skills install ~handler-bind~ for specific conditions (e.g., a provider
health monitor that records ~llm-error~ failures per backend). The
restarts registered in ~loop-process~ enable structured recovery:
skip the signal, retry with a modified prompt, inject a fallback
response, or abort the cycle.
#+begin_src lisp
(define-condition passepartout-error (error)
((message :initarg :message :reader error-message))
(:report (lambda (c s) (format s "Passepartout error: ~a" (error-message c))))
(:documentation "Root of the pipeline error hierarchy."))
(define-condition pipeline-error (passepartout-error)
((signal :initarg :signal :reader pipeline-error-signal :initform nil))
(:report (lambda (c s) (format s "Pipeline error: ~a" (error-message c))))
(:documentation "Any error during the Perceive→Reason→Act cycle."))
(define-condition llm-error (pipeline-error)
((provider :initarg :provider :reader llm-error-provider)
(cascade :initarg :cascade :reader llm-error-cascade :initform nil)
(attempt-count :initarg :attempt-count :reader llm-error-attempt-count :initform 0))
(:report (lambda (c s) (format s "LLM error (~a): ~a" (llm-error-provider c) (error-message c))))
(:documentation "LLM provider failure: timeout, cascade exhaustion, or API error."))
(define-condition gate-error (pipeline-error)
((gate-name :initarg :gate-name :reader gate-error-gate-name)
(rejected-action :initarg :rejected-action :reader gate-error-rejected-action))
(:report (lambda (c s) (format s "Gate ~a blocked action: ~a" (gate-error-gate-name c) (error-message c))))
(:documentation "Deterministic gate blocked a proposed action."))
(define-condition budget-error (pipeline-error)
((remaining :initarg :remaining :reader budget-error-remaining :initform 0.0)
(requested :initarg :requested :reader budget-error-requested :initform 0.0))
(:report (lambda (c s) (format s "Budget exhausted: $~,4f remaining, $~,4f requested" (budget-error-remaining c) (budget-error-requested c))))
(:documentation "Session budget cap has been reached."))
(define-condition protocol-error (passepartout-error)
((raw-message :initarg :raw-message :reader protocol-error-raw-message :initform nil))
(:report (lambda (c s) (format s "Protocol error: ~a" (error-message c))))
(:documentation "Malformed message, framing failure, or schema violation."))
#+end_src
** Global Interrupt State
Thread-safe interrupt flag. The ~*loop-interrupt-lock*~ mutex protects access so that the signal handler and the main loop don't race on shutdown.
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defvar *interrupt-flag* nil
"Atomic flag set by signal handlers to trigger graceful shutdown.")
#+end_src
** *loop-interrupt-lock*
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defvar *loop-interrupt-lock* (bt:make-lock "harness-interrupt-lock")
"Mutex protecting *interrupt-flag* access.")
#+end_src
** *heartbeat-thread*
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defvar *heartbeat-thread* nil
"Handle to the heartbeat thread.")
#+end_src
#+end_src
** Core Engine (loop-process)
The entry point to the metabolic pipeline. Each cycle runs Perceive → Reason → Act. If Act produces feedback (a new signal), the loop continues with that signal at the same depth.
The function handles four failure modes:
- **Depth exceeded**: signal dropped, nil returned
- **Interrupt flag**: graceful shutdown, nil returned
- **Handler error**: caught by handler-case, logged, and depending on the sensor type and depth:
- Normal errors at low depth → memory rollback + retry as :loop-error
- :loop-error and :tool-error at any depth → dropped (avoids infinite retry loops)
- High-depth errors (depth > 2) → dropped (avoids cascading failures)
- **Unhandled error**: the handler-case catches everything, preventing any single bad signal from crashing the agent
*** loop-process
The main pipeline entry point.
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defun loop-process (signal)
"The entry point to the Metabolic Pipeline: Perceive -> Reason -> Act."
(let ((current-signal signal))
(loop while current-signal do
(let ((depth (getf current-signal :depth 0))
(meta (getf current-signal :meta)))
(when (> depth 10)
(log-message "METABOLISM ERROR: Max recursion depth reached.")
(return nil))
(when (bt:with-lock-held (*loop-interrupt-lock*) *interrupt-flag*)
(log-message "METABOLISM: Interrupted by shutdown signal.")
(return nil))
(restart-case
(handler-bind
((pipeline-error (lambda (c)
(log-message "PIPELINE ERROR: ~a" (error-message c)))))
(handler-case
(progn
(setf current-signal (perceive-gate current-signal))
(setf current-signal (reason-gate current-signal))
(let ((feedback (act-gate current-signal)))
(if feedback
(progn
(unless (getf feedback :meta) (setf (getf feedback :meta) meta))
(setf current-signal feedback))
(setf current-signal nil))))
(error (c)
(let ((sensor (ignore-errors (getf (getf current-signal :payload) :sensor))))
(log-message "METABOLISM CRASH [~a]: ~a" (or sensor :unknown) c)
(unless (member sensor '(:loop-error :tool-error :syntax-error))
(log-message "CRITICAL ERROR: Initiating Micro-Rollback.")
(rollback-memory 0))
(if (or (> depth 2) (member sensor '(:loop-error :tool-error)))
(setf current-signal nil)
(setf current-signal
(list :type :EVENT :depth (1+ depth) :meta meta
:payload (list :sensor :loop-error :message (format nil "~a" c) :depth depth))))))))
(skip-signal ()
:report "Drop the current signal and continue the loop."
(setf current-signal nil))
(use-fallback (text)
:report "Inject a canned response instead of the LLM result."
(setf current-signal
(list :type :EVENT :depth (1+ depth) :meta meta
:payload (list :sensor :loop-error :message text :depth depth))))
(abort-pipeline ()
:report "Terminate the cognitive cycle cleanly."
(return nil)))))))
#+end_src
*** process-signal (backward-compatibility alias)
The pipeline entry point was originally named ~process-signal~. Code
that still uses the old name can call this alias. New code should call
~loop-process~.
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defun process-signal (signal)
(loop-process signal))
#+end_src
** Heartbeat Mechanism
The heartbeat is a background thread that fires every N seconds (configurable via ~HEARTBEAT_INTERVAL~ env var, default 60). On each tick, it:
1. Increments the save counter and saves memory to disk when the counter exceeds the auto-save interval (default 300s)
2. Injects a ~:heartbeat~ signal into the pipeline
The heartbeat signal is how background skills (Gardener, Scribe) get triggered without user input. These skills have triggers that match ~:sensor :heartbeat~ and run maintenance tasks during idle cycles.
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defvar *memory-auto-save-interval* 300)
#+end_src
** *heartbeat-save-counter*
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defvar *heartbeat-save-counter* 0)
#+end_src
** heartbeat-start
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defun heartbeat-start ()
"Starts the background heartbeat thread."
(let ((interval (or (ignore-errors (parse-integer (uiop:getenv "HEARTBEAT_INTERVAL"))) 60))
(auto-save (or (ignore-errors (parse-integer (uiop:getenv "MEMORY_AUTO_SAVE_INTERVAL"))) *memory-auto-save-interval*)))
(setf *memory-auto-save-interval* auto-save)
(setf *heartbeat-save-counter* 0)
(setf *heartbeat-thread*
(bt:make-thread
(lambda ()
(loop
(sleep interval)
(incf *heartbeat-save-counter*)
(when (>= *heartbeat-save-counter* (/ *memory-auto-save-interval* interval))
(setf *heartbeat-save-counter* 0)
(save-memory-to-disk))
(stimulus-inject
(list :type :EVENT :payload (list :sensor :heartbeat :unix-time (get-universal-time))))))
:name "passepartout-heartbeat"))))
#+end_src
#+end_src
** Shutdown Save Flag
Controls whether memory is saved on shutdown. Useful for testing when you want a clean state on next boot.
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defvar *shutdown-save-enabled* t)
#+end_src
** System Health Status
Used by the health check protocol and the daemon's status endpoint. Set by ~diagnostics-startup-run~ during boot.
- ~:healthy~ — all checks passed
- ~:degraded~ — checks found issues but the daemon can still run
- ~:unhealthy~ — checks failed, the daemon may not function correctly
- ~:unknown~ — health check hasn't run yet
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defvar *system-health* :unknown
"Current system health status: :healthy, :degraded, :unhealthy, or :unknown.")
#+end_src
** *health-check-ran*
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defvar *health-check-ran* nil
"Flag indicating if initial health check has completed.")
#+end_src
#+end_src
** Proactive Doctor
Runs the doctor diagnostics automatically at startup. If the doctor finds issues (missing dependencies, misconfigured providers), it prints a diagnostic message but does NOT block the daemon from starting. The user can see the issues and run ~passepartout doctor --fix~ to repair.
This is the "fail open" principle applied to boot: the system should start even with problems, not refuse to start until everything is perfect.
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defun diagnostics-startup-run ()
"Runs the doctor diagnostics on startup. Returns health status."
(format t "~%")
(format t "==================================================~%")
(format t " DOCTOR: Running Startup Health Check~%")
(format t "==================================================~%")
(handler-case
(progn
(when (fboundp 'diagnostics-run-all)
(let ((result (diagnostics-run-all :auto-install nil)))
(setf *health-check-ran* t)
(if result
(progn
(setf *system-health* :healthy)
(format t "DAEMON: Health check passed. Starting services.~%"))
(progn
(setf *system-health* :degraded)
(format t "DAEMON: Health check found issues.~%")
(format t " Run 'passepartout diagnostics' to repair.~%")))))
(setf *health-check-ran* t))
(error (c)
(format t "DIAGNOSTICS ERROR: ~a~%" c)
(setf *system-health* :unhealthy)
(setf *health-check-ran* t)))
(format t "==================================================~%~%"))
#+end_src
** Main Entry Point (main)
The top-level entry point. Called by ~passepartout daemon~ and ~passepartout tui~.
Boot sequence:
1. Load environment variables from ~.config/passepartout/.env~
2. Load persisted memory state from disk
3. Register core actuators (:system, :tool, :tui)
4. Initialize all skills (tangging .lisp or loading from XDG)
5. Run the proactive health check
6. Start the heartbeat thread (background maintenance)
7. Start the TCP daemon (listens for CLI/TUI connections)
8. Install the SIGINT handler (graceful shutdown on Ctrl+C)
9. Enter the idle sleep loop (wakes on interrupt)
;; REPL-VERIFIED: 2026-05-03T13:00:00
#+begin_src lisp
(defun main ()
"Entry point for Passepartout. Initializes the system and enters idle loop."
(let* ((home (uiop:getenv "HOME"))
(env-file (uiop:merge-pathnames* ".config/passepartout/.env" (uiop:ensure-directory-pathname home))))
(when (uiop:file-exists-p env-file)
(cl-dotenv:load-env env-file)))
(load-memory-from-disk)
(actuator-initialize)
(skill-initialize-all)
;; Run proactive diagnostics before starting services
(diagnostics-startup-run)
(when (fboundp 'events-start-heartbeat)
(events-start-heartbeat))
(start-daemon)
#+sbcl
(sb-sys:enable-interrupt sb-unix:sigint
(lambda (sig code scp)
(declare (ignore sig code scp))
(log-message "SHUTDOWN: SIGINT received. Saving memory...")
(when *shutdown-save-enabled* (save-memory-to-disk))
(uiop:quit 0)))
(let ((sleep-interval (or (ignore-errors (parse-integer (uiop:getenv "DAEMON_SLEEP_INTERVAL"))) 3600)))
(loop
(when (bt:with-lock-held (*loop-interrupt-lock*) *interrupt-flag*)
(log-message "SHUTDOWN: Interrupt flag set. Saving memory...")
(when *shutdown-save-enabled* (save-memory-to-disk))
(return))
(sleep sleep-interval))))
#+end_src
* Test Suite
Verifies that the immune system (error handling) correctly catches and reports errors from the cognitive pipeline.
#+begin_src lisp
(eval-when (:compile-toplevel :load-toplevel :execute)
(ql:quickload :fiveam :silent t))
(defpackage :passepartout-immune-system-tests
(:use :cl :fiveam :passepartout)
(:export #:immune-suite))
(in-package :passepartout-immune-system-tests)
(def-suite immune-suite :description "Verification of the Immune System (Core Error Hooks)")
(in-suite immune-suite)
(test loop-error-injection
"Contract 1: a crash in think/decide triggers :loop-error stimulus."
(clrhash passepartout::*skill-registry*)
(passepartout:defskill :evil-skill
:priority 100
:trigger (lambda (ctx) (eq (getf (getf ctx :payload) :sensor) :user-input))
:probabilistic (lambda (ctx) (declare (ignore ctx)) (error "CRITICAL BRAIN FAILURE"))
:deterministic nil)
(passepartout:loop-process '(:type :EVENT :payload (:sensor :user-input)))
(let ((logs (if (fboundp 'passepartout::context-get-system-logs)
(passepartout:context-get-system-logs 20)
nil)))
(is (or (null logs) ; no log service available — degraded but not broken
(not (null (find-if (lambda (line) (search "CRITICAL BRAIN FAILURE" line)) logs)))))))
(test test-process-signal-normal-path
"Contract 1: a valid signal passes through the pipeline without crash."
(clrhash passepartout::*skill-registry*)
(handler-case
(let ((signal (list :type :EVENT :depth 0 :payload (list :sensor :heartbeat))))
(process-signal signal)
(pass))
(error (c)
(fail "Pipeline crashed on normal signal: ~a" c))))
(test test-loop-process-returns-nil-on-deep
"Contract 1: depth > 10 returns nil from loop-process."
(let ((result (loop-process '(:type :EVENT :depth 11 :payload (:sensor :heartbeat)))))
(is (null result))))
#+end_src