Model Parameter Validation & Temperature Error Flooding

[gltanaka]

Problem

When running pdd generate with a single API key, users see a flood of temperature-related error messages. This is the most common user experience (single key) and creates a poor first impression.

From Dec 13 Benchmarking Meeting:

"When I ran PDD generate after doing setup with no changes at all manually, it gave me a bunch of error messages, a flood. So I got a bunch of temperature messages... 4 out of 5 models giving me some sort of temperature error"

Root Cause

• System iterates through models based on strength parameter
• Models have different parameter constraints (temperature ranges, required values)
• Anthropic thinking models require temperature=1.0
• Some models do not accept temperature parameter at all
• No validation before attempting API call

Proposed Solution

Add model parameter constraint metadata:

# Option A: Add columns to llm_model.csv
temperature_min,temperature_max,temperature_required_value

# Option B: Separate model_constraints.yaml
claude-sonnet-4-5:
  temperature:
    min: 0.0
    max: 1.0
    required_with_thinking: 1.0
  thinking:
    type: budget
    max_tokens: 64000

Files to Modify

• pdd/llm_invoke.py (lines ~1645-1705) - add constraint validation
• .pdd/llm_model.csv - add constraint columns

Acceptance Criteria

• No temperature errors printed when scanning models with valid single key
• Model constraints documented in CSV or separate config
• Auto-correct parameters when possible (e.g., force temp=1.0 for thinking models)
• Unit tests for parameter validation

Related

• PDD fails with 'All candidate models failed' error - configured with non-existent OpenAI models #71 (model names) - that issue covers invalid model names, this covers parameter constraints
• cache all LLMLite request-response pairs and support replay mode #47 (replayability) - identified as related critical infrastructure in meeting