Files
passepartout/skills/org-skill-llm-gateway.org

190 lines
10 KiB
Org Mode

:PROPERTIES:
:ID: llm-gateway-skill
:CREATED: [2026-04-09 Thu]
:EDITED: [2026-04-11 Sat]
:END:
#+TITLE: SKILL: Unified LLM Gateway (Universal Literate Note)
#+STARTUP: content
#+FILETAGS: :llm:gateway:infrastructure:psf:
#+DEPENDS_ON: id:credentials-vault-skill
* Overview
The *Unified LLM Gateway* is the single sensory and reasoning interface for all neural backends. It consolidates the previously fragmented provider skills into a high-integrity dispatch layer, standardizing credential management, error handling, and payload formatting.
* Phase A: Demand (PRD)
:PROPERTIES:
:STATUS: SIGNED
:END:
** 1. Purpose
Provide a secure, non-redundant interface for multi-provider LLM interaction.
** 2. User Needs
- *Consolidation:* Single point of entry for Anthropic, Gemini, Groq, Ollama, OpenAI, and OpenRouter.
- *Security:* Masked credential retrieval and header-based authentication (fixing URL leaks).
- *Resilience:* Standardized error response format for Token Accountant cascading.
- *Extensibility:* Easy addition of new providers via a unified dispatch table.
* Phase B: Blueprint (PROTOCOL)
:PROPERTIES:
:STATUS: SIGNED
:END:
** 1. Architectural Intent
The gateway utilizes a functional dispatch pattern. A single entry point, `execute-llm-request`, resolves the provider-specific nuances (URLs, headers, JSON structures) while exposing a uniform interface to the harness.
** 2. Semantic Interfaces
#+begin_src lisp
(defun execute-llm-request (prompt system-prompt &key provider model)
"Executes a neural request. Returns (:status :success :content ...) or (:status :error :message ...).")
#+end_src
* Phase C: Success (QUALITY)
:PROPERTIES:
:STATUS: SIGNED
:END:
** 1. Success Criteria
- [ ] *Credential Safety:* API keys are never logged or hardcoded.
- [ ] *Header Integrity:* Correct headers (x-api-key, Bearer) for each provider.
- [ ] *Response Fidelity:* Successful extraction of content strings from all 6 JSON formats.
- [ ] *Resilience:* Standardized error return on timeout or 4xx/5xx responses.
** 2. TDD Plan
Verification will occur via `tests/llm-gateway-tests.lisp` using the FiveAM framework. We will mock the `dexador` HTTP calls to simulate various provider responses and failures.
* Phase D: Build (Implementation)
** Package Context
#+begin_src lisp
#+end_src
** Nested Extraction Helper (get-nested)
A robust utility to navigate deeply nested JSON alists produced by `cl-json`, handling both objects and arrays.
#+begin_src lisp
(defun get-nested (alist &rest keys)
"Recursively extracts nested values from an alist, handling both objects and arrays."
(let ((val alist))
(dolist (k keys)
;; If val is an array (a list where the first element is a list but NOT a pair),
;; descend into the first element.
(when (and (listp val) (listp (car val)) (not (keywordp (caar val))))
(setf val (car val)))
(let ((pair (assoc k val)))
(if pair
(setf val (cdr pair))
(return-from get-nested nil))))
val))
#+end_src
** Unified Request Executor (execute-llm-request)
This is the primary actuator for neural reasoning. It handles the specific JSON payload formats and HTTP headers required by each provider. It retrieves secrets from the [[file:org-skill-credentials-vault.org][Credentials Vault]], ensuring that API keys are masked in all diagnostic output.
#+begin_src lisp
(defun execute-llm-request (prompt system-prompt &key provider model)
"Unified entry point for all LLM providers."
(let ((api-key (vault-get-secret provider :type :api-key))
(full-prompt (format nil "~a~%~%Prompt: ~a" system-prompt prompt)))
(harness-log "PROBABILISTIC ENGINE: Requesting ~a (Model: ~a) [Key: ~a]"
provider (or model "default") (vault-mask-string api-key))
(case provider
(:gemini-web
(let ((res (uiop:symbol-call :org-agent.skills.org-skill-web-research :ask-gemini-web full-prompt)))
(if res (list :status :success :content res) (list :status :error :message "Web Research Failure"))))
(:ollama
(let* ((host (or (uiop:getenv "OLLAMA_HOST") "localhost:11434"))
(url (format nil "http://~a/api/generate" host))
(body (cl-json:encode-json-to-string `((model . ,(or model "llama3")) (prompt . ,full-prompt) (stream . :false)))))
(handler-case
(let* ((response (dex:post url :headers '(("Content-Type" . "application/json")) :content body :connect-timeout 5 :read-timeout 60))
(json (cl-json:decode-json-from-string response)))
(list :status :success :content (cdr (assoc :response json))))
(error (c) (list :status :error :message (format nil "Ollama Failure: ~a" c))))))
(t ;; Cloud Providers (Anthropic, Gemini API, Groq, OpenAI, OpenRouter)
(when (or (null api-key) (string= api-key ""))
(return-from execute-llm-request (list :status :error :message (format nil "API Key missing for ~a" provider))))
(let* ((endpoint (case provider
(:anthropic "https://api.anthropic.com/v1/messages")
(:gemini-api (format nil "https://generativelanguage.googleapis.com/v1/models/~a:generateContent" (or model "gemini-1.5-flash-latest")))
(:groq "https://api.groq.com/openai/v1/chat/completions")
(:openai "https://api.openai.com/v1/chat/completions")
(:openrouter "https://openrouter.ai/api/v1/chat/completions")))
(headers (case provider
(:anthropic `(("Content-Type" . "application/json") ("x-api-key" . ,api-key) ("anthropic-version" . "2023-06-01")))
(:gemini-api `(("Content-Type" . "application/json") ("x-goog-api-key" . ,api-key)))
(:openrouter `(("Content-Type" . "application/json") ("Authorization" . ,(format nil "Bearer ~a" api-key))
("HTTP-Referer" . "https://github.com/amr/org-agent") ("X-Title" . "org-agent Sovereign Kernel")))
(t `(("Content-Type" . "application/json") ("Authorization" . ,(format nil "Bearer ~a" api-key))))))
(body (case provider
(:anthropic (cl-json:encode-json-to-string `((model . ,(or model "claude-3-5-sonnet-20240620")) (max_tokens . 4096) (system . ,system-prompt) (messages . (( (role . "user") (content . ,prompt) ))))))
(:gemini-api (cl-json:encode-json-to-string `((contents . (((parts . (((text . ,full-prompt))))))))))
(t (cl-json:encode-json-to-string `((model . ,(or model (case provider (:groq "llama-3.3-70b-versatile") (:openai "gpt-4o") (t "openrouter/auto"))))
(messages . (( (role . "system") (content . ,system-prompt) ) ( (role . "user") (content . ,prompt) )))))))))
(handler-case
(let* ((response (dex:post endpoint :headers headers :content body :connect-timeout 10 :read-timeout 30))
(json (cl-json:decode-json-from-string response)))
(let ((content (case provider
(:anthropic (get-nested json :content :text))
(:gemini-api (get-nested json :candidates :parts :text))
(t (get-nested json :choices :message :content)))))
(if content
(list :status :success :content content)
(list :status :error :message (format nil "Failed to parse ~a response structure." provider)))))
(error (c) (list :status :error :message (format nil "LLM Gateway Failure (~a): ~a" provider c)))))))))
#+end_src
** Cognitive Tools
The `:ask-llm` tool exposes the gateway's power to Probabilistic Engine, allowing it to explicitly request reasoning from a specific provider when the default cascade is insufficient.
** Registration: Tool
Register the unified gateway as a cognitive tool.
#+begin_src lisp
(def-cognitive-tool :ask-llm
"Queries an LLM provider via the unified gateway."
((:prompt :type :string :description "The user prompt.")
(:system-prompt :type :string :description "The system instructions.")
(:provider :type :keyword :description "The provider (e.g., :gemini-api, :anthropic, :groq, :openai, :openrouter, :ollama, :gemini-web).")
(:model :type :string :description "Optional specific model ID."))
:body (lambda (args)
(execute-llm-request (getf args :prompt)
(or (getf args :system-prompt) "You are a helpful assistant.")
:provider (getf args :provider)
:model (getf args :model))))
#+end_src
Register each supported provider with the harness's neural registry.
#+begin_src lisp
(dolist (p '(:anthropic :gemini-api :gemini-web :groq :ollama :openai :openrouter))
(org-agent:register-probabilistic-backend p (lambda (prompt system-prompt &key model)
(execute-llm-request prompt system-prompt :provider p :model model))))
#+end_src
** Registration: Skill
Define the foundational skill entry for the gateway.
#+begin_src lisp
(defskill :skill-llm-gateway
:priority 150 ; Higher than individual old skills
:trigger (lambda (context) (declare (ignore context)) nil)
:probabilistic (lambda (context) (declare (ignore context)) nil)
:deterministic (lambda (action context) (declare (ignore context)) action))
#+end_src
* Phase E: Chaos (Verification)
** 1. Unit Tests (FiveAM)
Verification is performed in `tests/llm-gateway-tests.lisp` by mocking the `dex:post` client.
** 2. Chaos Scenarios
- *Scenario A (Key Exhaustion):* Use the `chaos` skill to temporarily clear an API key and verify the `token-accountant` successfully falls back to the next healthy provider.
- *Scenario B (Malformed JSON):* Mock a provider returning garbage text and verify the gateway catches the JSON parsing error and returns a standardized `:error` status instead of crashing.
* Phase F: Memory (RCA)
- *[2026-04-09 Thu]:* Refactored 6 providers into this unified gateway to solve the URL key-leakage security vulnerability and reduce boilerplate by 60%.
- *[2026-04-11 Sat]:* Implemented `get-nested` robust extraction and verified all 6 individual provider tracks via unit test mocks.