Spaces:
Sleeping
Sleeping
Commit History
-
fix(ui+search): approachable starter prompts, animation timing, search robustness
Starter queries: casual everyday topics instead of technical jargon. Quick actions: short action-oriented labels. Chat animation: wrap resetStreaming in useCallback, defer isStreaming=false until word animation completes so input stays locked during playback. Search: catch network errors in webSearch, log empty DDG results for debugging, use realistic User-Agent to avoid bot detection. Made-with: Cursor24cf70c
-
fix(agent): remove content moderation from auditor
Keep only structural checks (step count, HTTP budget, empty inputs) and LLM policy audit. No query-level filtering. Made-with: Cursor26ad573 -
fix(agent): deterministic audit, synthesizer grounding, structured output reliability
Auditor: two-layer audit with deterministic rule checks (content safety, step limits, request budget) before LLM audit. Harmful queries hard-rejected by regex. Repetition-loop detection flags degenerate output. Transparent summary for inconclusive LLM audits. Synthesizer: skip LLM when all steps returned no results, emit structured insufficient-data report. Grounding instructions added for partial results. Thinker: lower temperature to 0.3, simplified prompt to field-spec format for small model reliability. LLM pipeline: add ChatOptions (temperature, max_new_tokens) to chatComplete. Made-with: Cursor0829ca0 -
fix(agent): thinker prompt injects real query, auditor nonsense-check tightened
The SmolLM2-360M model was copying the example JSON verbatim instead of using the actual user query. Pre-fill the example with the real query so the model copy-tendency works in our favor. Auditor: auto-approve now checks only violations (suggestions are advisory), and recognizes DeepTrust Default Policy as a parroted heading. Made-with: Cursorc8019ed -
fix(agent): HITL resume via Command, cap reasoning, auditor nonsense filter, real web search
- Resume interrupted graph with Command(resume + humanApproved) for approve flow - appendReasoning returns single entry; reducer merges and caps at MAX_REASONING_ENTRIES - Broaden auditor placeholder/nonsense detection for small models - web_search: DuckDuckGo HTML + optional Google CSE fallback Made-with: Cursor974f8b2 -
fix: simplify tools to web_search only and break auditor rejection loop
Narrow tool enum from 4 to web_search only. Detect auditor placeholder parroting and auto-approve. Cap reasoning at 20 entries. Sanitize fallback rationale. Simplify thinker prompt. Made-with: Cursorcd7228e
-
fix: enhance auditor's JSON parsing and error handling
Added robust error handling for auditor responses, including a retry mechanism for JSON parsing and a fallback to extract verdicts from partial outputs. Improved the extractJSON utility to handle cases where JSON objects may be emitted without braces.f49206c
-
fix: normalize auditor policy feedback
Made-with: Cursora19123a -
fix: improve auditor JSON handling and worker config
Made-with: Cursor5e84fc9 -
fix: use null resume for HITL stream
Made-with: Cursor1223bfe -
chore: lower max_new_tokens for SmolLM
Made-with: Cursor64e51f1 -
Deploy: DeepTrust workspace (v2 clean history)
Made-with: Cursor8eb816f