- _parse_messages: include last 6 history entries as context for coherent narrative
- generate_llm_choices: verb-based action choices with history context, bare-line output format
- _llm_health_delta: LLM rates risk of player action (-15/0/+10), falls back to random
- _run_turn: player choice feeds into story instruction, history passed to choices
- _parse_choices: strip numbering, handle bare-line format
- api_start_game/api_make_choice: pass history to choices generator
fix(tinybard): use correct logger name (log not logger)
Hermes Bot
Unverified signature
e78038c
fix(tinybard): simplify choice prompt for 1.5B model
Hermes Bot
Unverified signature
3169187
- System prompt now asks for bare lines (no numbering, no pipes)
- 1.5B model generates better with simple newline format
- Temperature lowered 0.8→0.6 for structured output
- Added debug logging for raw LLM output and parsed choices
- Parser unchanged — already handles bare newline format
- Force-clear cooldown in generate_llm_story() and generate_llm_choices() so
story+choices always use LLM, never fall back to procedural due to cooldown
- Merge duplicate click handlers (api_start_game + handle_start_game) into
single handler per button with api_name for MCP
- Remove cooldown short-circuit from _run_turn() — always attempt LLM first
- Update system prompt: Nanaboozhoo trickster narrator with 'Be unpredictable'
- Default INFERENCE_PROVIDER to featherless-ai (supports small instruct models)
fix: pin huggingface_hub>=0.36 to avoid old router endpoint
Hermes Bot
Unverified signature
1420fd3
fix: pass model at chat_completion call time, not constructor (HF router fix)
Hermes Bot
Unverified signature
bc6b501
fix: use Llama-3.2-1B-Instruct (free HF Inference), auto-detect provider
Hermes Bot
Unverified signature
2922094
Merge branch 'develop'
Hermes Bot
Unverified signature
b1b7ff5
# Conflicts:
# projects/tinybard/app.py
feat: Anishinaabe Solarpunk Gradio UI with choice input, save/load, and CRT aesthetic
Hermes Bot
Unverified signature
62071c9
fix: skip uvicorn on HF Spaces, use demo.launch() instead
Hermes Bot
Unverified signature
49bd205
fix: import BaseModel from pydantic for UserConfig class
Hermes Bot
Unverified signature
bca6cbf
fix: restore HF Space YAML front matter in root README
- Remove step >= 4 win condition; adventure cycles infinitely via step % 3
- Add genre field to all API response dicts
- Add 4 save/load endpoints (POST save, GET saves, POST load, DELETE save)
- Saves stored as JSON in projects/tinybard/saves/
- Add X exit button with save dialog (Save & Exit / Exit Without Saving / Cancel)
- Add SAVES dropdown listing server + localStorage saves
- Add RESUME buttons on genre selector for recent saves
- Add localStorage backup for persistence across server restarts
- All new UI follows cedar-copper solarpunk CRT aesthetic
- Remove step >= 4 win condition; adventure cycles infinitely via step % 3
- Add genre field to all API response dicts
- Add 4 save/load endpoints (POST save, GET saves, POST load, DELETE save)
- Saves stored as JSON in projects/tinybard/saves/
- Add X exit button with save dialog (Save & Exit / Exit Without Saving / Cancel)
- Add SAVES dropdown listing server + localStorage saves
- Add RESUME buttons on genre selector for recent saves
- Add localStorage backup for persistence across server restarts
- All new UI follows cedar-copper solarpunk CRT aesthetic
feat: add projects/tinybard/modal_inference.py and update llms.txt
Hermes Bot
Unverified signature
6f876a9
feat: add support for custom OpenAI-compatible endpoints in TinyBard
Hermes Bot
Unverified signature
4d181c9
- Update shared/inference_client.py to support custom endpoints via httpx
- Support token, model, and custom_endpoint overrides in generate()
- Add custom endpoint configuration field in CONFIG modal UI
- Save and reload config in localStorage
- Fallback to Qwen 1.5B default, or procedural if inference fails
fix: target huggingface.co domain for serverless API to bypass DNS blocks
Hermes Bot
Unverified signature
4e120ef
fix: resolve base_url and model conflict in InferenceClient
Hermes Bot
Unverified signature
f4f2e74
Omit the model parameter from the constructor and pass the direct
model-specific serverless API endpoint as base_url. This forces the
use of the free serverless Inference API and resolves the conflict.
fix: route chat completions to free serverless Inference API
Hermes Bot
Unverified signature
93e094a
By default, recent huggingface_hub versions route InferenceClient.chat_completion()
to router.huggingface.co, which requires paid providers. Explicitly passing
base_url='https://api-inference.huggingface.co' forces it to use the free serverless
inference endpoint.
fix: lock default model to Qwen/Qwen2.5-1.5B-Instruct
Hermes Bot
Unverified signature
dd78128
Newer candidates (Phi-4-mini, SmolLM3-3B, Gemma-2-9B-it) return
'model_not_supported' from the HF Inference router on this account.
Keep the free-tier working default; users can still override via config.
fix: revert default model to Qwen/Qwen2.5-1.5B-Instruct
Hermes Bot
Unverified signature
04b08cb
Qwen 2.5 7B requires paid Inference Providers on this account.
Revert to 1.5B free-tier model; users can still override via config modal.
fix: use Qwen/Qwen2.5-7B-Instruct as default model
Hermes Bot
Unverified signature
e382b5e
Gemma 2 9B is not supported by the HF Inference router on this account.
Qwen 2.5 7B is supported, still well under 32B, and strong for storytelling.
feat: upgrade default model to google/gemma-2-9b-it
Hermes Bot
Unverified signature
6f6cdba
9B parameters, newer architecture, better storytelling quality
Still under 32B constraint for Tiny Titan + Thousand Token Wood tracks
feat: LLM-driven infinite adventure loop
Hermes Bot
Unverified signature
ceada34
- Replace story+choices two-call pattern with single generate_llm_turn()
- Each turn now produces story, 3 choices, and health delta from the model
- Remove fixed 4-step win/lose cap; game continues until health <= 0
- Procedural fallback remains when inference is in cooldown or fails
