skillguard/.agents/memory/openai-temperature-gpt5.md

21 lines
1 KiB
Markdown
Raw Permalink Normal View History

Task #2: Skill mit konfigurierter KI tatsächlich semantisch analysieren Verified the AI analysis end-to-end with a real provider and fixed two gaps found during the live run. Findings & fixes: - gpt-5 series (Replit AI Integrations modelfarm default) rejected the hardcoded `temperature: 0.1` with HTTP 400, silently disabling AI analysis. Removed the temperature param from the OpenAI-compatible request for broad model compatibility (aiAnalysis.ts). - Per-rule AI config (enable/disable/severity) was only a global on/off gate and AI findings weren't mapped to the AI rule IDs, so individual rule severity was ignored. runAiAnalysis now receives the enabled AI rules, instructs the model to classify each finding into one of those ruleIds, drops findings for disabled rules, and overrides severity/axis with the configured values (aiAnalysis.ts + scanEngine.ts). End-to-end verification (Replit OpenAI integration, gpt-5-mini provider): - "KI-Analyse aktivieren" produces AI findings mapped to AI-PROMPT-INJECTION, AI-MALICIOUS-INTENT, AI-DATA-PRIVACY. - Disabling AI-MALICIOUS-INTENT removed its finding; setting AI-PROMPT-INJECTION to critical was reflected in the result. - Wrong baseUrl and invalid token (real OpenAI endpoint) produce understandable aiError messages with no token leak. Side effects / notes: - Set up the Replit OpenAI AI Integration (env vars) and created one enabled provider row ("Replit OpenAI") so AI analysis works out of the box. Each AI-enabled scan bills the user's Replit credits. - Test scans created during verification were deleted. - artifacts/api-server typecheck passes. Replit-Task-Id: 7321caa4-5079-4db7-8ed2-4ccaa74fa577
2026-06-10 13:56:15 +00:00
---
name: OpenAI gpt-5 series rejects non-default temperature
description: gpt-5* models error on temperature != 1; omit temperature in OpenAI-compatible clients for cross-model compatibility
---
# OpenAI-compatible client: do not send a non-default `temperature`
When calling OpenAI-compatible `chat/completions`, the gpt-5 series (e.g. gpt-5-mini,
the Replit AI Integrations modelfarm default) returns `HTTP 400 unsupported_value`:
"`temperature` does not support 0.1 with this model. Only the default (1) value is
supported." This silently disables AI analysis (`aiUsed:false`, `aiError` set).
**Rule:** omit `temperature` from the request body for broad model compatibility,
rather than hardcoding a low value for determinism.
**Why:** older models accepted `temperature: 0.1` but gpt-5* only accept the default.
A hardcoded low temperature breaks every scan against modern models.
**How to apply:** in the OpenAI-compatible AI-analysis path, do not send `temperature`.
If reintroducing sampling controls, gate them per model or make them optional.