Spaces:
Sleeping
Sleeping
Commit History
-
feat: JSON-mode structured outputs for critic scoring
- llm_client.query gains json_mode: response_format json_object on Groq/OpenRouter, responseMimeType application/json on Gemini; validated against all three live providers - Critic evaluation now requests JSON mode; existing brace-slice parsing retained as fallback - Fix stale LANGCHAIN_PROJECT name in .env.example04accf1 -
fix: Handle gpt-oss reasoning-token budget in Groq calls and deep health
gpt-oss models spend max_tokens on hidden reasoning before emitting content; tiny budgets return empty content that reads as provider failure and triggers needless fallback. Set reasoning_effort=low for gpt-oss on Groq and give /health/deep a 512-token budget.96a9373
-
fix: Migrate to living LLM models and defuse stale production config
- Default models updated: Groq llama-3.1-8b-instant (deprecated 2026-06-17) -> openai/gpt-oss-20b; Gemini gemini-2.0-flash (shut down 2026-06-01) -> gemini-flash-lite-latest; OpenRouter delisted gemma-3n free model -> openai/gpt-oss-20b:free - Remove frontend/.env.production pointing at the torn-down AWS API Gateway (would have been baked into the next HF Space build) - Fix stale A2A researcher URL in .env.example - Cap langchain/langgraph/langchain-core below 1.0 to match the known-good deployed build until proper lockfile pinning lands - Add docs/audit_checklist.md from the 2026-07 technical auditf26a13c
-
Change OpenRouter model to Gemma 3n E4B IT
google/gemma-3n-e4b-it:free - optimized for efficient execution, supports 32K context, multimodal inputs. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>6a70bab -
Fix OpenRouter model: use deepseek-chat-v3 (was invalid model)
The previous model openai/gpt-oss-120b:free doesn't exist (404). Changed to deepseek/deepseek-chat-v3-0324:free which is a valid free model on OpenRouter. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>784089a -
Rotate between LLM providers instead of consecutive retries
Before: Groq → Groq → Groq → Gemini → Gemini → Gemini After: Groq → Gemini → OpenRouter → Groq → Gemini → OpenRouter - 3 rounds of rotation through all providers - 10s delay between each attempt - Gives each provider time to recover from rate limits Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>494ea7a -
Increase LLM retry delays and add fallback delay for all errors
- Exponential backoff: 10s, 20s, 40s (was 5s, 10s, 20s) - Provider fallback: 10s delay on ANY error (was only on 429) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>8d5bad3 -
Update default models: Gemini stable + OpenRouter GPT fallback
- Gemini: gemini-2.0-flash (stable instead of experimental) - OpenRouter: openai/gpt-oss-120b:free (different model for true fallback diversity) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>87e16be -
Add 10s delay between provider fallbacks after rate limit (429)
When a provider returns 429 rate limit error, wait 10 seconds before trying the next provider in the cascade. This prevents rapid-fire requests that trigger rate limits on all providers. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>59c9b47
-
Increase exponential backoff from 3s to 5s base
Backoffs now: 5s, 10s, 20s (was 3s, 6s, 12s) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>6810cb2 -
Fix: Increase rate limit delays and fix revision detection
1. Increase pre-call delays: 2s → 5s 2. Increase backoff times: 2s,4s,8s → 3s,6s,12s 3. Fix revision mode detection: check critique_details.status instead of revision_count (which was still 0 on first loop) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>efcfc92 -
Fix: Add rate limit handling with retry and delays
1. LLM client: Add exponential backoff retry on 429 errors - MAX_RETRIES=3, delays of 2s, 4s, 8s - New _request_with_retry() helper method 2. Critic: Add 2s delay before LLM call - Avoids hitting rate limit after Analyzer's call 3. Analyzer: Add 2s delay before revision LLM call - Avoids hitting rate limit after Critic's call This prevents 429 errors from cascading through the workflow. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>de66f3b
-
Rename financials to fundamentals for schema consistency
- Update analyzer.py to use fundamentals_all instead of financials_all - Align with Researcher-Agent schema normalization changes Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>3e0da39
-
Initial commit: Instant SWOT Agent
Multi-agent strategic analysis system with: - Research Gateway (A2A client to Researcher-Agent) - Analyzer, Critic, Editor workflow nodes - Data grounding enforcement in all LLM calls - React frontend with real-time progress 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>0c591a7