Generate random seeds by default to produce unique images on each run.
Users can still specify --seed option for reproducible results.

Updated documentation to explain seed behavior and reproducibility.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Use clearer section headers and separators in final answer:
- Story text displayed prominently first
- Clean separator line between sections
- Enhanced prompt and file paths grouped separately

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Provide full example story with proper spacing and formatting so LLM
knows exactly how to structure the final answer with story text.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Restore the technical introduction explaining multi-modal architecture,
tool composition through mixins, and Lemonade Server integration.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Removed redundant sections:
- Deleted "Quick Concepts" section (duplicated "What You'll Build")
- Deleted standalone "Choosing SD Models" section (moved to Step 2 accordion)
- Removed redundant "Run Your Agent" section (integrated into Step 3)

Added technical depth to "What You'll Build":
- MRO chain, HTTP endpoints, tool signatures
- Complete tool registry explanation
- Instance state details

Fixed technical inaccuracy:
- Corrected model formats: LLMs use GGUF, SD uses safetensors
- Added specific format details for each model

Added missing prerequisite:
- Virtual environment creation step

Result: 14% shorter, more accurate, better flow.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Clarify what runs where:
- GGUF models (Qwen3-8B, Qwen3-VL-4B) run on iGPU (Radeon) via Vulkan
- SDXL-Turbo (safetensors) currently runs on CPU

Added tabbed selection for venv activation (Windows/Linux).

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Add --python 3.12 flag to uv venv command for consistency with
quickstart guide and to ensure correct Python version.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Add proper 4-space indentation to method definitions in Step 2 so they
can be directly pasted under the class definition from Step 1.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Specify model_id="Qwen3-8B-GGUF" in super().__init__() to use the
model downloaded by gaia init --profile sd, not the default
Qwen3-Coder-30B which isn't included in the SD profile.

This fixes "model_load_error" when running the example.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Replace simple system prompt with the actual SDAgent prompts that include:
- Research-backed prompt enhancement strategies
- Model-specific guidelines (SDXL-Turbo, SD-Turbo, etc.)
- Workflow instructions for tool usage

Tool schemas (parameters, models) are auto-injected by Agent base class.

This fixes issues where LLM uses wrong models or parameters.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Add get_sd_system_prompt() method to SDToolsMixin so agents can
compose system prompts from inherited mixins instead of manually
importing prompt fragments.

Pattern:
- Mixins provide both tools AND domain-specific prompt fragments
- Agents compose them: return self.get_sd_system_prompt()
- Tool schemas auto-injected by Agent base class

Benefits: Single responsibility, reusability, composability.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Add template method pattern for automatic prompt composition:
- _get_mixin_prompts() - Auto-collects from inherited mixins
- _compose_system_prompt() - Composes mixin + agent prompts
- Mixins provide get_*_system_prompt() methods

Benefits:
- Mixins own their domain knowledge (tools + prompts)
- Agents automatically inherit behavior
- Can modify, extend, or override prompts
- Fully backwards compatible

SDToolsMixin now provides get_sd_system_prompt() with research-backed
prompt engineering guidelines.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Moved SD prompts from src/gaia/agents/sd/ to src/gaia/sd/ to keep
mixins self-contained (tools + prompts in same package).

Changes:
- Created src/gaia/sd/prompts.py (moved from agents/sd/)
- SDToolsMixin imports from new location
- Added get_vlm_system_prompt() to VLMToolsMixin for consistency
- SDAgent uses mixin's get_sd_system_prompt() (no manual import)
- Deprecated old prompts.py location

Benefits:
- Mixins are truly self-contained (src/gaia/sd/ has everything for SD)
- Cleaner agent implementations (just call mixin methods)
- Better separation of concerns

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Prompts now live with the mixin in src/gaia/sd/prompts.py.
Removed deprecated file from src/gaia/agents/sd/.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Added comprehensive documentation of composable system prompts:
- Part 1: Explains the simple usage (return self.get_sd_system_prompt())
- Part 2: Deep dive with 5 usage patterns and debugging guide

Patterns covered:
1. Use mixin prompts as-is (automatic)
2. Return mixin prompt explicitly
3. Extend with custom instructions
4. Modify mixin prompts
5. Custom composition order

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Updated Part 1 example to show:
- Using mixin tools directly (generate_image, analyze_image)
- Creating custom tools that wrap mixin functionality
- Adding custom prompt instructions

Example now creates create_story_about_image() that wraps VLM's
create_story_from_image with custom metadata, demonstrating the
wrapper pattern for building specialized tools.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Changed _get_system_prompt from abstract to concrete method with
empty string default. This allows agents using only mixin prompts
to avoid implementing it.

Fixed SDAgent duplicate prompt bug:
- Was returning self.get_sd_system_prompt() causing duplication
- Now returns "" to use automatic mixin composition

Backwards compatible: existing agents continue to work.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Enhanced comments and docstrings to better explain:
- Why create a custom tool (specialization vs. generic)
- What the wrapper pattern adds (fixed style, metadata, extensibility)
- How it calls mixin methods via inheritance

Fixed incorrect system prompt explanation to accurately describe
composition: mixin prompts + custom instructions.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Solved initialization order issues with hybrid approach:
- Mixins provide static base guidelines (no instance state needed)
- get_*_system_prompt() composes base + instance-specific gracefully
- Falls back to base if called before mixin initialization

Changes:
- SDToolsMixin: get_base_sd_guidelines() (static) + get_sd_system_prompt() (instance)
- VLMToolsMixin: get_base_vlm_guidelines() (static) + get_vlm_system_prompt() (instance)
- Agent: system_prompt now lazy property (composes on first access)
- Composition includes: mixin prompts + custom + tools + response format

Benefits:
- No initialization order issues
- Graceful degradation (works even if called early)
- Simple, robust, debuggable

Added debugging documentation with 5 methods to observe prompts.

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Remove redundant 're' import in _resolve_plan_parameters() - already imported at module level.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

The 8K context was insufficient for multi-step planning with dynamic parameters. Increased to 16K to handle:
- SD system prompts with placeholder examples
- VLM system prompts
- Agent custom prompts
- Multi-step plan with tool schemas
- Conversation history

Changes:
- SDAgentConfig.ctx_size: 8192 → 16384
- SD profile min_context_size: 8192 → 16384
- Fix Pylint W0201: Initialize _ctx_verified and _ctx_warning in __init__
- Add warning when context verification fails for LLM models
- Update all documentation references from 8K to 16K

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

When testing model inference during init, unload the model first if already loaded, then reload with save_options=True. This ensures recipe_options with ctx_size are properly persisted.

Also improved display to show warning emoji when context verification fails:
- ✓ Qwen3-8B-GGUF - OK (ctx: 16384)  ← Success
- ✓ Qwen3-8B-GGUF - OK ⚠️ Context unverified!  ← Warning

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Merged index.mdx and part-1-building-agent.mdx into a single streamlined playbook. Deleted parts 2 and 3 which were overly verbose and redundant.

Changes:
- Merged index intro, video, and quick test into main playbook
- Kept all hands-on tutorial content from part-1
- Removed architecture deep-dive (part-2) and variations (part-3)
- Updated docs.json navigation to show single page
- Result: One comprehensive, focused guide instead of fragmented multi-part tutorial

The consolidated guide maintains the strong structure of both original files while eliminating unnecessary complexity.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

After image generation and story creation complete, the agent was creating additional stories in a loop. Added explicit instruction to provide final answer immediately after both tools complete.

Fix prevents infinite loop detection warning when agent tries to call create_story_from_image multiple times.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

…opping

Changed from single-repeat detection to consecutive-count detection. Agent now allows up to 3 consecutive identical tool calls before triggering the loop detector.

Changes:
- Replaced last_tool_call with tool_call_history (tracks last 5 calls)
- Count consecutive identical calls
- Trigger after 3 consecutive repeats (was: 1 repeat)

This allows legitimate use cases like "create 3 robot designs" while still preventing infinite loops.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Added max_consecutive_repeats parameter to Agent.__init__() to allow customization of how many consecutive identical tool calls are allowed before triggering loop detection.

Default: 4 consecutive calls (increased from hardcoded 3)

This allows users to adjust sensitivity:
- Lower value (2-3): More aggressive loop detection
- Higher value (5-10): More tolerant of repetition
- For SD agent: Default 4 allows multiple variations while preventing infinite loops

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Use console.print_repeated_tool_warning() exclusively instead of also logging. Console provides better user visibility and Rich formatting.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Fixed outdated comment that still referenced 8K context.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>

Fixed all references to:
- 8K context → 16K context for multi-step planning
- Updated context size table for all models
- Updated troubleshooting with correct ctx-size command
- Updated playbook reference from "3-part" to single comprehensive guide
- Removed broken link to deleted part-2-architecture.mdx

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <[email protected]>