[BUG] Orchestrator AI prompts overwritten by static YAML config

[dhirmadi]

Bug Report

Severity: High (affects all orchestrator-mode tenants with custom AI prompts)
Component: publisher_v2/src/publisher_v2/services/ai.py
Discovered: 2025-12-30 during integration testing for lbd.shibari.photo

---

Problem

Custom AI prompts (system_prompt, role_prompt) sent by the orchestrator are ignored. The Publisher always uses the static YAML defaults instead of tenant-specific prompts.

Root Cause

In CaptionGeneratorOpenAI.__init__() (ai.py lines 209-231), the orchestrator prompts are loaded first, then overwritten by static config:

# Step 1: Load from orchestrator config ✅
self.system_prompt = config.system_prompt   # Tenant's custom prompt
self.role_prompt = config.role_prompt

# Step 2: OVERWRITE with static YAML ❌ (wrong order!)
static_prompts = get_static_config().ai_prompts
if static_prompts.caption.system:
    self.system_prompt = static_prompts.caption.system   # ← Overwrites tenant prompt!
if static_prompts.caption.role:
    self.role_prompt = static_prompts.caption.role       # ← Overwrites tenant prompt!

Evidence

Orchestrator sends custom prompts for lbd.shibari.photo:

"ai": {
    "system_prompt": "Using the image analysis, write a tiny caption between 100-150 characters. 1. Wrap the message in ✨ sparkles ✨. 2. Start with a theme: Memory Monday, Titty Tuesday...",
    "role_prompt": "You are a classy, fun, and playful 30-year-old woman posting on FetLife. Your style is 'short and sweet'—sophisticated yet flirty..."
}

But Publisher uses static defaults from config/static/ai_prompts.yaml:

caption:
  system: >
    You are a senior social media copywriter. Write authentic, concise, platform-aware captions.
  role: "Write a caption for:"

Expected Behavior

Orchestrator-provided prompts (tenant-specific) should take precedence over static YAML defaults. Static config should only be used as a fallback when the orchestrator doesn't provide prompts.

Proposed Fix

Change the conditional logic to use static config as fallback, not override:

# Load from orchestrator config first
self.system_prompt = config.system_prompt
self.role_prompt = config.role_prompt

# Only use static config as FALLBACK if orchestrator didn't provide prompts
static_prompts = get_static_config().ai_prompts
if not self.system_prompt and static_prompts.caption.system:
    self.system_prompt = static_prompts.caption.system
if not self.role_prompt and static_prompts.caption.role:
    self.role_prompt = static_prompts.caption.role

Same pattern needed for:

• sd_caption_system_prompt (lines 228-229)
• sd_caption_role_prompt (lines 230-231)

Affected Code Locations

File	Lines	Issue
publisher_v2/src/publisher_v2/services/ai.py	224-227	Caption prompts overwritten
publisher_v2/src/publisher_v2/services/ai.py	228-231	SD caption prompts overwritten

Acceptance Criteria

• Orchestrator-provided system_prompt takes precedence over static YAML
• Orchestrator-provided role_prompt takes precedence over static YAML
• Same for sd_caption_system_prompt and sd_caption_role_prompt
• Static YAML prompts are only used when orchestrator provides null/empty
• Add unit test to verify prompt precedence

Testing Notes

To reproduce:

1. Configure a tenant in orchestrator with custom ai.system_prompt and ai.role_prompt
2. Run Publisher in orchestrator mode
3. Observe generated captions use generic "social media copywriter" style instead of custom prompts

---

Related: Feature 022 (Orchestrator Schema V2 Integration)

[dhirmadi]

Confirmed root cause in publisher_v2/src/publisher_v2/services/ai.py: static YAML prompt overrides were applied unconditionally, overwriting tenant (orchestrator) prompts.

Fix implemented:

• Static YAML prompts are now fallback only.
• Tenant-provided prompts (from orchestrator config) take precedence for:
  • system_prompt / role_prompt
  • sd_caption_system_prompt / sd_caption_role_prompt
• Added regression tests to lock precedence behavior.

Commit: e44d3950c0d58210c0703fbdbdf2c28755af07f2
Branch: bugfix/issue48-ai-prompt-precedence
Tests:

• uv run pytest -q publisher_v2/tests/test_ai_prompt_precedence.py ✅
• uv run pytest -q publisher_v2/tests/test_ai_sd_generate.py publisher_v2/tests/test_ai_error_paths.py ✅

You can open a PR from: https://github.com/dhirmadi/SocialMediaPythonPublisher/pull/new/bugfix/issue48-ai-prompt-precedence

[dhirmadi]

✅ QA Verification Complete

Branch: bugfix/issue48-ai-prompt-precedence
Tested: 2026-01-03

---

Test Results

Test Suite	Result
Full test suite	✅ 439 passed in 38.99s
New regression tests	✅ 2/2 passed

New Tests Added

publisher_v2/tests/test_ai_prompt_precedence.py

✅ test_orchestrator_prompts_take_precedence_over_static_yaml
   → Verifies tenant prompts (TENANT_SYSTEM, TENANT_ROLE) are NOT overwritten by static YAML

✅ test_static_yaml_is_used_as_fallback_when_config_is_defaults
   → Verifies static YAML prompts are applied when config uses schema defaults

Code Fix Verified

The fix correctly uses schema defaults comparison to detect tenant-provided vs default prompts:

# Before (BUG): unconditional override
if static_prompts.caption.system:
    self.system_prompt = static_prompts.caption.system

# After (FIX): only use static as fallback when config is defaults
default_cfg = OpenAIConfig()
if self.system_prompt == default_cfg.system_prompt and static_prompts.caption.system:
    self.system_prompt = static_prompts.caption.system

Acceptance Criteria Status

• Orchestrator-provided system_prompt takes precedence over static YAML
• Orchestrator-provided role_prompt takes precedence over static YAML
• Same for sd_caption_system_prompt and sd_caption_role_prompt
• Static YAML prompts are only used when orchestrator provides defaults
• Unit tests added to verify prompt precedence

---

Ready to merge. PR can be opened from branch bugfix/issue48-ai-prompt-precedence.