Fix validation errors and runtime exceptions in podcast generation workflow

[YosiElias]

Fix validation errors and runtime exceptions in podcast generation workflow

This PR addresses six bugs that cause validation errors and runtime exceptions during podcast generation. Each fix resolves a specific failure point that prevents successful podcast creation.

Note: Each fix is implemented as a separate commit for easier review.

---

Bug Fixes

1. Fix Pydantic validation error from missing scratchpad field

Files: frontend/shared/shared/podcast_types.py, shared/shared/podcast_types.py

Bug: The Conversation.scratchpad field was required but not instructed to be provided in the prompt, causing Pydantic validation errors.

Impact: Podcast generation fails with validation error when scratchpad is omitted.

Solution: Added empty string default (scratchpad: str = "") to make the field optional.

---

2. Fix duration unit mismatch between prompt and Pydantic model

Files: services/AgentService/podcast_prompts.py

Bug: The prompt instructed the LLM to output duration in minutes, but the Pydantic model expected integer seconds. This type mismatch caused validation failures.

Impact: Every outline generation fails with validation error due to incompatible duration units.

Solution: Updated prompt to clearly specify duration must be integer seconds, matching the model's expectation.

---

3. Fix schema regeneration that discards filename enum

Files: services/AgentService/podcast_flow.py

Bug: Code dynamically added filename enum to schema, then immediately regenerated the schema, discarding the enum modifications.

Impact: Schema regeneration overrides the dynamic modifications, causing the filename enum constraints to not be applied.

Solution: Removed the redundant schema = PodcastOutline.model_json_schema() call that was overriding the enum modifications.

---

4. Align models.json with llmmanager.py defaults

Files: models.json

Bug: Configuration mismatch between models.json and hardcoded defaults in llmmanager.py:

• JSON role: models.json used llama-3.1-8b-instruct, code defaults to llama-3.1-70b-instruct
• Iteration role: models.json used llama-3.1-70b-instruct, code defaults to llama-3.1-405b-instruct

Impact: Using the 8B model for JSON generation causes frequent Pydantic validation errors due to lower model capacity (NVIDIA hosted models).

Solution: Updated models.json to match the code defaults, ensuring consistent model usage.

---

5. Fix TypeError when recipient field is None

Files: frontend/__main__.py

Bug: The recipient field is None by default unless the user clicks the accordion. Code attempts len(recipient) without null check, causing TypeError: object of type 'NoneType' has no len().

Impact: Podcast generation fails immediately unless user explicitly opens the recipient accordion (even if not providing a value).

Solution: Added null coalescing: recipient = recipient or "" to handle the default None case.

---

6. Fix Gradio InvalidPathError for audio and transcript files

Files: frontend/__main__.py

Bug: Gradio raises InvalidPathError when attempting to serve generated audio and transcript files.

Impact: Users cannot download generated podcast audio or transcripts through the web UI.

Solution: Resolved path handling to ensure Gradio can correctly serve generated files.

---

Testing

All fixes have been tested and verified to resolve their respective issues.

Impact Summary

These fixes enable successful end-to-end podcast generation by resolving:

• 4 Pydantic validation errors
• 2 runtime exceptions

Without these fixes, the podcast generation workflow fails at multiple points.