qwen-sci 0.2.2__py3-none-any.whl

src/config/default.yaml

@@ -0,0 +1,1122 @@
+# ============================================
+# Qwen-Sci Unified Configuration
+# ============================================
+
+# Global settings
+project:
+  name: "qwen-sci"
+  root: "${repo_root}"
+
+# ============================================
+# Workspace Configuration
+# ============================================
+workspace:
+  root: "workspace"
+
+# ============================================
+# Shared API Configuration
+# ============================================
+llm:
+  default_provider: "${oc.env:QWENSCI_LLM_PROVIDER,'qwen'}"
+  providers:
+    openai:
+      api_key_env: "OPENAI_API_KEY"
+      api_key: "${oc.env:OPENAI_API_KEY,''}"
+      base_url: "${oc.env:OPENAI_BASE_URL,''}"
+      api_style: "chat_completions"
+      openhands_model_prefix: "openai"
+      tokenizer_fallback: "o200k_base"
+      token_limit_parameter: ""
+      default_models:
+        survey: "gpt-5.4-mini"
+        judge: "gpt-5.5"
+        blog: "gpt-5.5"
+        idea: "gpt-5.5"
+        idea_generation: "gpt-5-mini"
+        idea_evaluation: "gpt-5.5"
+        experiment: "gpt-5.4"
+        experiment_design: "gpt-5.4"
+        author: "gpt-5.4"
+        planner: "gpt-5.4"
+        worker: "gpt-5-mini"
+        reviewer: "gpt-5.4"
+        master: "gpt-5.4"
+        memory_filter: "gpt-5-mini"
+        memory_router: "gpt-5-mini"
+        memory_synthesis: "gpt-5.4"
+    qwen:
+      api_key_env: "DASHSCOPE_API_KEY"
+      api_key: "${oc.env:DASHSCOPE_API_KEY,''}"
+      base_url: "${oc.env:DASHSCOPE_BASE_URL,'https://dashscope.aliyuncs.com/compatible-mode/v1'}"
+      api_style: "chat_completions"
+      openhands_model_prefix: "openai"
+      tokenizer_fallback: "utf8_bytes"
+      token_limit_parameter: "max_tokens"
+      default_models:
+        survey: "qwen3.6-flash"
+        judge: "qwen3.8-max"
+        blog: "qwen3.8-max"
+        idea: "qwen3.8-flash"
+        idea_generation: "qwen3.8-flash"
+        idea_evaluation: "qwen3-max-2026-01-23"
+        experiment: "qwen3.7-plus"
+        experiment_design: "qwen3.8-flash"
+        author: "qwen3.8-flash"
+        planner: "qwen3.7-plus"
+        worker: "qwen3.7-plus"
+        reviewer: "qwen3.7-plus"
+        master: "qwen3.7-plus"
+        memory_filter: "qwen3.6-flash"
+        memory_router: "qwen3.6-flash"
+        memory_synthesis: "qwen3.8-max"
+  models:
+    qwen3-max-2026-01-23:
+      provider: "qwen"
+      api_style: "responses"
+      capabilities:
+        responses: true
+        reasoning: true
+        json_schema: true
+        json_object: true
+    qwen3.7-plus:
+      provider: "qwen"
+      api_style: "chat_completions"
+      capabilities:
+        chat_completions: true
+        streaming: true
+        tools: true
+        streaming_tools: true
+        vision: true
+        json_object: true
+    qwen3.8-max:
+      provider: "qwen"
+      api_style: "chat_completions"
+      capabilities:
+        chat_completions: true
+        streaming: true
+        vision: true
+        json_object: true
+    qwen3.8-flash:
+      provider: "qwen"
+      api_style: "chat_completions"
+      capabilities:
+        chat_completions: true
+        streaming: true
+        json_object: true
+      max_output_tokens: 128000
+    qwen3.6-flash:
+      provider: "qwen"
+      api_style: "chat_completions"
+      capabilities:
+        chat_completions: true
+        streaming: true
+        json_object: true
+    qwen3-vl-plus:
+      provider: "qwen"
+      api_style: "chat_completions"
+      capabilities:
+        chat_completions: true
+        streaming: true
+        vision: true
+        json_object: true
+    qwen3-vl-flash:
+      provider: "qwen"
+      api_style: "chat_completions"
+      capabilities:
+        chat_completions: true
+        streaming: true
+        vision: true
+        json_object: true
+    wan2.7-image-pro:
+      provider: "qwen"
+      api_style: "images"
+      capabilities:
+        image_generation: true
+    qwen-image-3.0-pro:
+      provider: "qwen"
+      api_style: "images"
+      capabilities:
+        image_generation: true
+    z-image-turbo:
+      provider: "qwen"
+      api_style: "images"
+      capabilities:
+        image_generation: true
+
+vision:
+  provider: "${oc.env:VISION_LLM_PROVIDER,'qwen'}"
+  quality_model: "${oc.env:VISION_QUALITY_MODEL,'qwen3-vl-plus'}"
+  batch_model: "${oc.env:VISION_BATCH_MODEL,'qwen3-vl-flash'}"
+  api_key: "${oc.env:DASHSCOPE_API_KEY,''}"
+  base_url: "${oc.env:DASHSCOPE_BASE_URL,'https://dashscope.aliyuncs.com/compatible-mode/v1'}"
+  timeout_seconds: 120
+  max_tokens: 2048
+
+image_generation:
+  provider: "${oc.env:IMAGE_GENERATION_PROVIDER,'qwen'}"
+  api_key: "${oc.env:DASHSCOPE_API_KEY,''}"
+  base_url: "${oc.env:DASHSCOPE_IMAGE_BASE_URL,'https://dashscope.aliyuncs.com/api/v1'}"
+  timeout_seconds: 180
+  poll_interval_seconds: 1
+  role_models:
+    academic_figure: "${oc.env:IMAGE_ACADEMIC_FIGURE_MODEL,'qwen-image-3.0-pro'}"
+    text_rich_figure: "${oc.env:IMAGE_TEXT_RICH_FIGURE_MODEL,'qwen-image-3.0-pro'}"
+    draft: "${oc.env:IMAGE_DRAFT_MODEL,'z-image-turbo'}"
+  models:
+    wan2.7-image-pro:
+      endpoint_style: "dashscope_multimodal"
+      endpoint_path: "services/aigc/multimodal-generation/generation"
+      supports_reference: false
+      supports_edit: false
+      supports_mask: false
+      supports_4k: true
+    qwen-image-3.0-pro:
+      endpoint_style: "dashscope_multimodal"
+      endpoint_path: "services/aigc/multimodal-generation/generation"
+      supports_reference: false
+      supports_edit: false
+      supports_mask: false
+      supports_4k: false
+    z-image-turbo:
+      endpoint_style: "dashscope_multimodal"
+      endpoint_path: "services/aigc/multimodal-generation/generation"
+      supports_reference: false
+      supports_edit: false
+      supports_mask: false
+      supports_4k: false
+
+api:
+  openai:
+    api_key: "${oc.env:OPENAI_API_KEY,''}"
+    api_base: "${oc.env:OPENAI_API_BASE,''}"
+    base_url: "${oc.env:OPENAI_BASE_URL,''}"
+  semantic_scholar:
+    api_key: "${oc.env:SEMANTIC_SCHOLAR_API_KEY,''}"
+    base_url: "https://api.semanticscholar.org"
+  github_ai:
+    token: "${oc.env:GITHUB_AI_TOKEN,''}"
+  serper:
+    api_key: "${oc.env:SERPER_API_KEY,''}"
+  jina:
+    api_key: "${oc.env:JINA_API_KEY,''}"
+  tavily:
+    api_key: "${oc.env:TAVILY_API_KEY,''}"
+    remote_url_template: "https://mcp.tavily.com/mcp/?tavilyApiKey={api_key}"
+  huggingface:
+    endpoint: "https://hf-mirror.com"
+
+# ============================================
+# Survey Agent Configuration
+# ============================================
+survey:
+  output:
+    root_dir: "${repo_root}/src/agents/survey_agent/outputs"
+  multimodal_evidence:
+    enabled: false
+    allow_remote_perception: false
+    quality_model: "qwen3-vl-plus"
+    max_data_anchored_sh: 3
+    max_vl_calls: 8
+    max_preview_pixels: 1600000
+    max_preview_bytes: 4194304
+    max_records_per_modality: 24
+    max_input_file_bytes: 52428800
+    max_total_input_bytes: 268435456
+    sampling_policy: "stratified_by_label_group_condition_timepoint_round_robin_v1"
+    input_spec: {}
+    local_input_context: {}
+    runtime_evidence: {}
+  BasicInfo:
+    topic: ""
+    research_title: ""
+    declared_domain: ""
+    research_objective: ""
+    research_brief: ""
+    research_context_path: ""
+    research_context_use_llm: true
+    research_context: {}
+    research_design_inventory: {}
+    project_context_artifact_path: ""
+    survey_run_id: ""
+    subhypotheses: []
+    subhypothesis_decomposition: {}
+    subhypothesis_decomposition_path: ""
+    subhypothesis_retrieval: {}
+    cache_path: "./database"
+    debug: false
+    error_conservatism_mode: false
+    skip_exist: false
+    adapter_mode: false
+  APIInfo:
+    llm_provider: "${oc.env:QWENSCI_LLM_PROVIDER,'qwen'}"
+    llm_api_key: ""
+    llm_api_base_url: ""
+    llm_model_name: "${oc.env:SURVEY_LLM_MODEL,''}"
+    llm_max_context_length: 880000
+    use_stream_mode: false
+    # Keep remote LLM connection pressure below the token-plan endpoint's
+    # stable operating range. Survey presets may not exceed this default.
+    batch_chat_agent_worker: 4
+    chat_timeout: 900
+    batch_chat_timeout: 1800
+    # OpenAlex is the primary literature-discovery provider. API keys are
+    # required by the current OpenAlex API; the adapter sends this value only
+    # as an Authorization header and never writes it to logs or cache artifacts.
+    openalex_base_url: "https://api.openalex.org"
+    openalex_api_key: "${oc.env:OPENALEX_API_KEY,''}"
+    # Optional contact address sent as an HTTP From header. It is no longer a
+    # substitute for API-key authentication or a rate-limit policy.
+    openalex_email: "${oc.env:OPENALEX_EMAIL,''}"
+    # Keep all local OpenAlex request starts below the configured service budget.
+    # This is a request-rate cap, not a worker-count setting.
+    openalex_requests_per_second: 8
+    # PyAlex itself does not pass a requests timeout. The OpenAlex adapter owns
+    # timeout and retry so every attempt is observable and rate-limited.
+    openalex_connect_timeout_seconds: 10
+    openalex_read_timeout_seconds: 30
+    openalex_api_max_retry: 2
+    openalex_retry_base_delay_seconds: 1
+    openalex_retry_max_delay_seconds: 60
+    openalex_search_per_page: 10
+    openalex_graph_candidate_per_page: 100
+    openalex_graph_recent_quota: 15
+    openalex_graph_high_impact_quota: 15
+    openalex_graph_cache_schema_version: 2
+    # Unpaywall resolves DOI-backed open-access PDFs before provider URLs. Its
+    # public API requires a contact email and is disabled when this is unset.
+    unpaywall_base_url: "https://api.unpaywall.org/v2"
+    unpaywall_email: "${oc.env:UNPAYWALL_EMAIL,''}"
+    unpaywall_timeout: 30
+    semantic_scholar_api_key: "${oc.env:SEMANTIC_SCHOLAR_API_KEY,''}"
+    low_flow_mode: false
+    low_flow_latency: 0
+    exponential_backoff: true
+    exponential_backoff_time: 1
+    exponential_backoff_max_time: 60
+    arxiv_api_max_retry: 3
+    semantic_scholar_api_max_retry: 3
+  ModuleInfo:
+    WorkCollector:
+      download_safe_mode: false
+      download_timeout: 60
+      download_in_parallel: true
+      download_parallel_workers: 4
+      pdf_parse_batch_size: 1
+      # Legal OA full-text acquisition is independent from paper/SH relevance.
+      # `access_context_generation` must change when a legitimate TDM or other
+      # authorized provider is introduced, so anonymous 403 caches are ignored.
+      fulltext_resolution_enabled: true
+      fulltext_unpaywall_all_locations_enabled: true
+      # Lower-priority public metadata routes; direct PDF URLs are byte-checked
+      # and non-PDF landing URLs require an explicit OA declaration.
+      fulltext_generic_metadata_urls_enabled: true
+      # Runs immediately after Unpaywall candidates, with a bounded DOI
+      # redirect chain; never use it to bypass subscription access controls.
+      fulltext_doi_landing_fallback_enabled: true
+      fulltext_max_candidates_per_paper: 12
+      fulltext_max_declared_pdf_links_per_landing_page: 8
+      fulltext_max_doi_landing_recovery_pdf_attempts: 4
+      fulltext_landing_page_timeout_seconds: 20
+      fulltext_landing_page_max_bytes: 1500000
+      fulltext_landing_page_max_redirects: 5
+      fulltext_pdf_max_bytes: 50000000
+      fulltext_per_host_concurrency: 2
+      # Bounded, polite retries are only for transient GET failures (429/5xx
+      # and transport errors), never for authentication or subscription walls.
+      fulltext_http_max_retries: 2
+      fulltext_http_retry_base_delay_seconds: 1
+      fulltext_http_retry_max_delay_seconds: 10
+      fulltext_connect_timeout_seconds: 10
+      fulltext_read_timeout_seconds: 60
+      fulltext_oa_resolution_cache_ttl_seconds: 604800
+      fulltext_pdf_success_cache_ttl_seconds: 2592000
+      fulltext_failure_404_ttl_seconds: 86400
+      fulltext_failure_access_denied_ttl_seconds: 1800
+      fulltext_failure_rate_limited_ttl_seconds: 900
+      fulltext_failure_non_pdf_ttl_seconds: 21600
+      fulltext_failure_signed_url_ttl_seconds: 300
+      fulltext_failure_transient_first_ttl_seconds: 60
+      fulltext_failure_transient_ttl_seconds: 180
+      fulltext_host_denial_threshold: 2
+      fulltext_host_denial_window_seconds: 600
+      fulltext_host_circuit_ttl_seconds: 3600
+      fulltext_access_context_generation: "anonymous-v1"
+      invalidate_fulltext_failure_cache: false
+      fulltext_provenance_enabled: true
+      cache_enabled: true
+      max_seed_paper_num: 18
+      use_seed_filter_LLM: true
+      LLM_seed_threshold: 4
+      auto_decompose_subhypotheses: true
+      subhypothesis_decomposition_provider: "qwen"
+      subhypothesis_decomposition_model: "qwen3.8-max"
+      subhypothesis_decomposition_min_count: 3
+      subhypothesis_decomposition_max_count: 6
+      subhypothesis_decomposition_temperature: 0.1
+      subhypothesis_decomposition_max_output_tokens: 16000
+      subhypothesis_decomposition_cache_enabled: true
+      enable_subhypothesis_retrieval: true
+      # Persist normalized provider results for SH slot/query-variant lanes.
+      # This cache avoids duplicate OpenAlex/Semantic Scholar requests only;
+      # every run still recomputes local merge, SH semantic assessment and seed
+      # selection. Empty completed requests use a short TTL, while failed
+      # provider calls are never written as empty scientific results.
+      sh_retrieval_cache_enabled: true
+      sh_retrieval_cache_success_ttl_seconds: 604800
+      sh_retrieval_cache_empty_ttl_seconds: 21600
+      # Set refresh=true for one run to bypass reads and update entries. Set
+      # invalidate=true only for an explicit one-time cache clear, then reset it.
+      refresh_sh_retrieval_cache: false
+      invalidate_sh_retrieval_cache: false
+      # LLM classifies a paper's semantic contribution to an SH before seed
+      # selection. It does not require a complete causal chain or all slots.
+      enable_sh_paper_semantic_assessment: true
+      sh_paper_semantic_assessment_cache_enabled: true
+      # Temporarily keep arXiv discovery lanes offline; OpenAlex remains active.
+      enable_arxiv_discovery: false
+      subhypothesis_relevance_threshold: 3
+      subhypothesis_max_unique_papers: 10
+      subhypothesis_max_slots_per_paper: 2
+      # A slot's query variants are alternative recall paths, never a joint
+      # eligibility condition for one paper. Values can only narrow the
+      # contract-level safety caps of 5 variants and 6 terms.
+      max_query_variants_per_slot: 5
+      max_terms_per_query_variant: 6
+      min_slot_candidates_before_relaxation: 5
+      enable_adjacent_discipline_precision_lane: true
+      max_adjacent_discipline_fields: 3
+      enable_slot_semantic_scholar_fallback: true
+      # Run independent slot tasks concurrently. Their OpenAlex requests still
+      # share the process-wide requests-per-second limiter under APIInfo.
+      slot_retrieval_parallel_workers: 4
+      reference_graph_depth: 1
+      # Exploration seeds are useful graph roots, but receive a bounded
+      # traversal budget distinct from direct evidence seeds.
+      exploration_seed_reference_graph_depth: 1
+      sentence_transformer_model: models/bge-m3
+      sentence_transformer_batch_size: 4
+      expand_in_local_paper_graph: false
+      advanced_filter_in_local_paper_graph_expansion: true
+      related_work_top_k: 30 
+      log_related_work_num: -1
+      related_work_threshold: 0.5
+      related_work_threshold_for_llm: 4
+      # `related_work_top_k` is a per-seed embedding pre-filter.  This is the
+      # run-level cap applied only after all graph candidates pass LLM
+      # relatedness filtering, so only the globally highest-related papers
+      # enter PDF download, parsing, and deep reading.
+      fulltext_download_max_papers: 70
+      # Citation expansion is only a discovery route.  A selected expanded
+      # paper may be promoted into its own SH evidence slot only after a
+      # complete-section keynote and a separate SH assessment.  Inspect up to
+      # twice an SH's remaining writing deficit, never more than this cap.
+      enable_fulltext_expanded_promotion: true
+      expanded_fulltext_promotion_candidate_multiplier: 2
+      expanded_fulltext_promotion_max_candidates_per_sh: 40
+      relatedness_temperature: 0.1
+      RAG_source_use_embedding_filter: true
+      RAG_source_use_LLM_filter: false
+      RAG_source_downloadable_only: false
+      use_ds_when_graph_fail: true
+    Database:
+      default_top_k: 20
+    PaperGraphRetriever:
+      db_path: data/processed/graph.db
+    WorkAnalyzer:
+      abstract_only_mode: false
+      abstract_when_full_text_fail: true
+      cache_enabled: true
+      use_local_paper_graph_keynotes: false
+      paper_reading_temperature: 0.3
+      paper_reading_max_retry: 5
+      llm_max_context_overhead_length_in_paper_reading: 20000
+      # MinerU uses one ## heading for every physical paper section.  Keep
+      # whole sections together and split very long papers across requests
+      # instead of token-prefix truncating a chapter in the middle.
+      fulltext_section_packing_enabled: true
+      fulltext_section_max_tokens: 512000
+      fulltext_section_prompt_reserve_tokens: 24000
+      fulltext_section_max_output_tokens: 16000
+      fulltext_section_batch_worker: 10
+      # Keep up to 10 workers for ordinary short packets, but cap the total
+      # estimated input token load admitted to each request wave.  This avoids
+      # ten near-512k contexts being sent concurrently to the provider.
+      fulltext_section_max_in_flight_tokens: 800000
+      oversized_unsplittable_section_policy: abstract
+      clustering_batch_size: 8
+      clustering_temperature: 1.0
+      clustering_in_steps: true
+      clustering_batch_size_in_creation: 6
+      clustering_batch_size_in_assignment: 10
+      cluster_assign_fast_mode: true
+      paper_clustering_max_retry: 10
+      paper_clustering_creation_max_retry: 3
+      paper_clustering_assignment_max_retry: 4
+      propose_question_temperature: 0.5
+      anwer_question_temperature: 0.2
+      intra_cluster_analysis_max_retry: 5
+      intra_cluster_analysis_temperature: 1.0
+      intra_clustering_probelm_proposing_max_retry: 10
+      graph_keynote_extraction_batch_retry: 7
+      graph_graph_keynote_extraction_temperature: 0.1
+      use_ds_keynotes_when_graph_fail: true
+    SurveyGenerator:
+      include_initial_analysis: false
+      include_relation_graph: true
+      include_relation_table: true
+      include_code_report: false
+      include_env_report: false
+      omit_error_preserve_retry_time: 2
+      outline_generation_in_steps: true
+      outline_generation_draft_max_retry: 3
+      outline_generation_draft_batch_size: 20
+      outline_generation_draft_max_iterations: 5
+      outline_generation_draft_empty_keynotes_iteration: 1
+      outline_draft_RAG_topk: 60
+      outline_generation_assign_max_retry: 10
+      outline_generation_assign_batch_size: 20
+      outline_assign_RAG_topk: 20
+      outline_assign_fast_mode: true
+      outline_generation_temperature: 0.5
+      # Preflight outline prompts against component and total input budgets.
+      # Full SH provenance remains in artifacts; LLM prompts receive a compact
+      # projection that preserves only writing-relevant evidence constraints.
+      outline_prompt_max_input_tokens: 240000
+      outline_max_output_tokens: 16000
+      outline_evidence_plan_max_input_tokens: 100000
+      outline_current_outline_max_input_tokens: 20000
+      outline_keynotes_max_input_tokens: 70000
+      outline_analysis_max_input_tokens: 12000
+      outline_rag_max_input_tokens: 12000
+      outline_repair_previous_response_max_input_tokens: 8000
+      outline_repair_current_outline_max_input_tokens: 8000
+      # Outline drafting sees SH-covering representatives; paper assignment
+      # still receives the full admitted set after the shape is accepted.
+      outline_representative_papers_per_sh: 4
+      outline_representative_max_papers: 30
+      outline_representative_include_context_papers: true
+      # Bound the complete survey at 30k words, then derive per-subsection
+      # budgets from the actual outline. The current 33.3k/40k total-survey
+      # contract permits up to 2k words per subsection for the target outline.
+      survey_target_words: 32000
+      survey_max_words: 40000
+      # Every SH gets one bounded writing whitelist. Direct/qualified/context
+      # non-expanded papers fill it first; complete-section-promoted citation
+      # expansions may fill only the remaining positions.
+      writing_max_papers_per_sh: 20
+      outline_min_sections: 5
+      outline_target_sections: 6
+      outline_max_sections: 7
+      outline_min_subsections_per_section: 1
+      outline_target_subsections_per_section: 3
+      outline_max_subsections_per_section: 3
+      subsection_target_min_words: 875
+      subsection_target_max_words: 2000
+      subsection_target_citations: 3
+      subsection_max_citations: 5
+      section_preamble_target_words: 250
+      section_preamble_max_words: 400
+      section_preamble_target_citations: 1
+      section_preamble_max_citations: 2
+      outline_generation_batch_size: 16
+      outline_generation_max_retry: 3
+      outline_generation_max_retry_in_generation_loop: 6
+      subsection_draft_temperature: 0.5
+      subsection_draft_max_retry: 5
+      # Temporary operational bypass: keep visible citation checks, but do not
+      # block drafting on LLM-generated claim-trace metadata.
+      claim_trace_validation_enabled: false
+      # A structurally invalid evidence trace repairs metadata only.  It never
+      # regenerates an otherwise accepted long subsection.
+      claim_trace_repair_max_attempts: 2
+      claim_trace_repair_max_output_tokens: 2048
+      section_draft_max_retry: 3
+      always_omit_error_in_draft_validation: true
+      section_draft_temperature: 0.1
+      draft_refinement_temperature: 1.0
+      section_draft_max_error_info_length: 500
+      subsection_draft_max_error_info_length: 500
+      outline_max_error_info_length: 500
+      llm_max_context_overhead_length_generation: 20000
+      llm_max_context_overhead_length_outline_generation: 10000
+      # Legacy keys retained for external overrides; the dynamic total budget
+      # above controls new drafts instead of per-unit hard minima.
+      subsection_least_words: 875
+      subsection_least_citations: 3
+      section_least_words: 4250
+      section_least_citations: 3
+      section_preamable_least_words: 250
+      section_preamble_least_words: 250
+      section_preamble_least_citations: 3
+      draft_length_relax_ratio: 0.1
+      use_full_text_in_survey_generation: false
+      include_other_relevant_papers_RAG_in_outline: true
+      outline_RAG_topk: 200
+      include_other_relevant_papers_RAG: true
+      section_RAG_topk: 60
+      subsection_RAG_topk: 60
+      use_title_in_draft: true
+      enable_review_and_revise: true
+      agentic_refine_section: true
+      section_review_temperature: 1.0
+      section_review_retry: 3
+      # In evidence-bounded writing, run a safe, section-local quality pass
+      # instead of the legacy free-form review/revision agent.  It never
+      # blocks completion: sections still below the score target are retained
+      # after this bounded number of accepted revisions.
+      evidence_bounded_section_quality_review_enabled: true
+      evidence_bounded_section_quality_score_threshold: 8.0
+      evidence_bounded_section_quality_max_improvements: 2
+      evidence_bounded_section_quality_review_retry: 2
+      evidence_bounded_section_quality_revise_retry: 2
+      reviewer_max_suggestions: 5
+      section_revision_RAG_topk: 30
+      section_revise_temperature: 0.5
+      section_revise_retry: 3
+      no_suggestion_run_each_iteration: 2
+      max_review_revise_iterations: 3
+      revise_section_in_parallel: 4
+      draft_refinement_in_parts: true
+      refine_in_parts_mode: 2
+      draft_refinement_max_retry: 7
+      review_and_revise_whole_survey_in_refinement: true
+      agentic_refine_survey: true
+      review_and_revise_whole_survey_max_iteration: 5
+      valid_title_min_similarity: 0.5
+    # Optional post-save visual companion. It never replaces survey.md or
+    # blocks the survey pipeline: provider/QC failures simply omit a figure.
+    SurveyVisualization:
+      enabled: true
+      output_mode: companion_markdown
+      write_back_to_survey_md: false
+      # Final image files live directly beside survey.md, not in a subfolder.
+      image_output_location: survey_output_directory
+      image_filename_template: "fig_{index:02d}_{slug}.png"
+      min_figures: 3
+      max_figures: 5
+      candidates_per_figure: 2
+      allowed_figure_types:
+        - overview_framework
+        - mechanism
+        - causal_pathway
+        - evidence_to_inference
+        - method_comparison
+        - multiscale_synthesis
+        - conceptual_workflow
+        - research_landscape
+        - future_roadmap
+      # The Qwen image role names resolve through the top-level
+      # image_generation.role_models configuration.
+      draft_image_role: draft
+      final_image_role: academic_figure
+      optional_text_rich_image_role: text_rich_figure
+      final_image_size: "2048x1536"
+      # Used automatically when the configured Qwen image model does not
+      # declare 4K support (for example qwen-image-3.0-pro).
+      compatible_image_size: "2048x1536"
+      # All reader-facing image prose is English. The image prompt itself asks
+      # Qwen not to render prose; captions are inserted in Markdown instead.
+      reader_facing_language: en
+      caption_language: en
+      alt_text_language: en
+      overlay_label_language: en
+      prompt_language: en
+      # Image-model text is unreliable; concise English concept labels are
+      # overlaid locally when the VisualBrief supplies them.
+      append_english_label_strip: true
+      # One LLM-generated palette is locked and reused by all figures in a run.
+      generate_article_style_profile: true
+      palette_source: llm
+      lock_palette_per_survey: true
+      palette_min_colors: 4
+      palette_max_colors: 6
+      # A section excerpt, never the whole survey, is the scientific input to
+      # each figure. Data charts remain out of scope for image generation.
+      planner_section_max_chars: 6000
+      planner_temperature: 0.2
+      require_evidence_anchor: true
+      require_visual_brief: true
+      # A rejected LLM visual brief gets one section-local JSON repair pass
+      # before the figure is omitted. Scientific evidence gates remain strict.
+      visual_brief_repair_enabled: true
+      visual_brief_repair_max_output_tokens: 2400
+      allow_unsupported_claims: false
+      allow_generated_numeric_charts: false
+      show_explicit_evidence_gaps: true
+      visual_qc_enabled: true
+      # Retry only the affected figure when QC reports a locked-palette mismatch.
+      regenerate_on_palette_mismatch: true
+      max_regeneration_attempts: 1
+      # Visual enhancement is non-blocking by design.
+      fail_open: true
+    Judge:
+      skip_evaluation: false
+      use_different_api_for_judge: false
+      provider: "${oc.env:SURVEY_JUDGE_PROVIDER,''}"
+      model: qwen3.7-plus
+      judge_llm_api_key: ""
+      judge_llm_api_base_url: ""
+      nli_temperature: 0.1
+      citation_quality_threshold: 0.5
+      remove_failed_citation_in_eval: true
+      rubrics_eval_10_dimensions: true
+      rubrics_eval_4_dimensions: false
+      rubrics_deep_survey_eval: false
+      explanation_in_rubric: false
+      citation_eval: false
+    CodeAnalysis:
+      PseudoCreator:
+        planner_temperature: 0.5
+        planner_max_retry: 3
+
+
+# ============================================
+# Idea Agent Configuration
+# ============================================
+idea:
+  enabled: true
+  input: ""
+  topic: ""
+  mature_idea: ""
+  mature_ideas: []
+  refinement_scope: ""
+
+  run:
+    LigAgent-Pro: true
+    route_matrix_enabled: true
+    idea_route_matrix_max_workers: 8
+    max_mature_ideas: 12
+    allow_problem_reframing: true
+    allow_unanchored_seed: true
+    allow_high_risk_seed: true
+    input: "${idea.input}"
+    topic: "${idea.topic}"
+    mature_idea: "${idea.mature_idea}"
+    mature_ideas: "${idea.mature_ideas}"
+    refinement_scope: "${idea.refinement_scope}"
+    survey_manifest: ""
+    # Optional directory containing exactly one ablation-results JSON file.
+    ablation_results_path: ""
+    input_interpreter_model: "${idea.agent.model}"
+    output_root: "runs"
+    console_logs: true
+    rag_config: "src/config/default.yaml"
+    openai_api_key: "${api.openai.api_key}"
+    openai_base_url: "${api.openai.base_url}"
+
+  agent:
+    model: "${oc.env:IDEA_LLM_MODEL,''}"
+    paper_keynote_keep_top_k: 10
+    workflow_models:
+      default: "${idea.agent.model}"
+      main:
+        knowledge_aquisition: "${idea.agent.model}"
+        advanced_analysis: "${idea.agent.model}"
+        re_analysis_replan: "${idea.agent.model}"
+        idea_generation: "${idea.agent.model}"
+        mature_idea_portfolio_generation: "${idea.agent.model}"
+        mature_idea_adjudication: "${idea.agent.model}"
+        route_context_preparation: "${idea.agent.model}"
+        diversity_adjudication: "${idea.agent.model}"
+        cross_seed_debate: "${idea.agent.model}"
+        portfolio_synthesis: "${idea.agent.model}"
+        primary_idea_materialization: "${idea.agent.model}"
+
+      # Nested knowledge acquisition workflow
+      knowledge_acquisition:
+        ka_query_generation: "${idea.agent.model}"
+        ka_paper_triage: "${idea.agent.model}"
+
+      # Optional flat overrides by stage name if a stage needs to diverge
+      # from its workflow bucket later.
+      stage_overrides:
+        algorithm_structuring: "${idea.agent.model}"
+        algorithm_alignment: "${idea.agent.model}"
+    chat_max_retries: 3
+    chat_retry_backoff: 2.0
+    advanced_analysis_contract_retries: 2
+    introduction_max_output_tokens: 25600
+    introduction_json_repair_attempts: 2
+    paper_enrichment_timeout_sec: 7200
+    experiment_findings_extraction:
+      model: "${idea.agent.model}"
+      temperature: 0.1
+      max_output_tokens: 65536
+
+  mcts:
+    max_iterations: 16
+    max_depth: 3
+    branching_factor: 3
+    exploration_constant: 1.2
+    idea_taste_mode: "moonshot_inventor"
+    # Prompt mode for MCTS generation / retrieval / evaluation.
+    # Options:
+    #   - default
+    #   - conceptual_surprise
+    prompt_mode: "conceptual_surprise"
+    generation_model: "${oc.env:IDEA_GENERATION_MODEL,''}"
+    evaluation_model: "${oc.env:IDEA_EVALUATION_MODEL,''}"
+    generation_temperature: 0.7
+    evaluation_temperature: 0.001
+    generation_max_tokens: 16384
+    evaluation_max_tokens: 16384
+    component_novelty_model_path: "${repo_root}/models/bge-m3"
+    component_novelty_index_dir: null
+    component_novelty_retrieval_top_k: 50
+    component_novelty_evidence_top_k: 5
+    component_novelty_eval_model: ""
+    component_novelty_eval_temperature: 0.001
+    component_novelty_eval_max_tokens: 16384
+    mechanism_commit_retrieval_top_k: 3
+    mechanism_commit_similarity_threshold: 0.6
+    theory_transfer_retrieval_top_k: 3
+    theory_transfer_similarity_threshold: 0.6
+    enable_vector_memory: true
+    enable_symbolic_memory: true
+    max_parallel_seeds: 4
+    max_parallel_routes: 5
+    exploration_budget: 24
+    high_risk_budget: 32
+
+  portfolio:
+    min_independent_ideas: 2
+    max_candidates_per_seed: 5
+    max_route_expansions_per_seed: 5
+    primary_selection_policy: "scientific_maturity_diversity_validation"
+
+  diversity:
+    enabled: true
+    min_route_distance: 0.35
+    max_same_route_ratio: 0.60
+    regenerate_collapsed_routes: false
+    preserve_high_risk_unique_candidates: true
+
+    min_confidence_for_memory: 0.6
+    pareto_top_k: 5
+    symbolic_memory_path: "output/symbolic_memory"
+    skill_prior_success_threshold: 0.7
+
+  debate:
+    enabled: true
+    internal_max_rounds: 2
+    max_parallel_internal: 6
+    cross_seed_max_rounds: 1
+    max_parallel_cross_seed: 6
+    prompt_char_limit: 120000
+    internal_debate_prompt_limit: 80000
+    cross_seed_debate_prompt_limit: 60000
+
+  fusion:
+    enabled: true
+    only_when_ligagent_pro: true
+    model: ""
+    temperature: 0.7
+    max_tokens: 16384
+    min_candidates: 2
+    repair_max_steps: 10
+    repair_patience: 5
+    repair_epsilon: 0.02
+
+  memory:
+    vector_store:
+      model_path: "${repo_root}/models/all-MiniLM-L6-v2"
+      llm_name: ""
+      llm_backend: "openai"
+      llm_provider: ""
+
+# ============================================
+# Experiment Agent Configuration
+# ============================================
+experiment:
+  enabled: true
+  backend: "openharness"
+
+  openharness:
+    provider: "${oc.env:EXPERIMENT_LLM_PROVIDER,''}"
+    default_model: "${oc.env:EXPERIMENT_LLM_MODEL,''}"
+    api_key: ""
+    base_url: ""
+    role_models:
+      planner: "${oc.env:EXPERIMENT_PLANNER_MODEL,''}"
+      worker: "${oc.env:EXPERIMENT_WORKER_MODEL,''}"
+      reviewer: "${oc.env:EXPERIMENT_REVIEWER_MODEL,''}"
+      master: "${oc.env:EXPERIMENT_MASTER_MODEL,''}"
+    timeout_seconds: 7200
+    request_max_retries: 6
+    request_retry_base_delay: 1.0
+    request_retry_max_delay: 60.0
+    structured_output_max_hook_blocks: 6
+    max_tokens: 16384
+    max_turns: 0
+    runtime_dir_name: ".openharness_runtime"
+    max_budget_usd: 0
+
+  external_tools:
+    huggingface_endpoint: "https://hf-mirror.com"
+    github_ai_token: "${oc.env:GITHUB_AI_TOKEN,''}"
+    serper_api_key: "${oc.env:SERPER_API_KEY,''}"
+    jina_api_key: "${oc.env:JINA_API_KEY,''}"
+
+  # Execution configuration
+  execution:
+    max_iterations: 20
+    delegate_max_children: 1
+    planner_max_turns: 0
+    worker_max_turns: 0
+    prepare_review_feedback_rounds: 4
+    code_review_feedback_rounds: 2
+    science_review_feedback_rounds: 2
+    bash_timeout_seconds: 600000
+    mcp_timeout_seconds: 120
+    slurm:
+      partition: "gpu_llm"
+      gpus: 8
+      gpu_type: "NVIDIA A800-80G"
+      cpu_cores: 60
+      memory: "128G"
+      time_limit: "24:00:00"
+      log_dir_template: "{workspace}/logs/slurm"
+      job_name_template: "{experiment_condition}_{pde_type}_{timestamp}"
+
+  # Memory configuration
+  memory:
+    enabled: true
+    shared_dir: "~/.researchagent/shared_memory"
+    embedding_model_path: "${repo_root}/models/all-MiniLM-L6-v2"
+    llm_provider: "${oc.env:MEMORY_LLM_PROVIDER,''}"
+    llm_name: "${oc.env:MEMORY_LLM_MODEL,''}"
+    llm_backend: "openai"
+    query_method: "embedding"
+    writeback_enabled: true
+    tool_logs_enabled: false
+    prompt_injection_enabled: true
+    max_slots_per_task: 100
+
+  # Workspace configuration
+  workspace:
+    root: "${workspace.root}"
+    prepare_clone_depth: 1
+    model_candidate_seed: "~/.researchagent/model_seed/"
+    tavily_enabled: true
+    tavily_api_key: "${oc.env:TAVILY_API_KEY,''}"
+    tavily_remote_url_template: "https://mcp.tavily.com/mcp/?tavilyApiKey={api_key}"
+
+# ============================================
+# ExperimentDesign Agent Configuration
+# ============================================
+experiment_design:
+  enabled: true
+  provider: "${oc.env:EXPERIMENT_DESIGN_LLM_PROVIDER,'qwen'}"
+  model: "${oc.env:EXPERIMENT_DESIGN_LLM_MODEL,'qwen3.8-flash'}"
+  execution:
+    allow_digital_execution: false
+  retrieval:
+    # Each slot has at most two concise query variants. Keep provider result
+    # pages small; the global candidate budget below bounds all subsequent LLM work.
+    max_results_per_query: 6
+    max_fulltext_papers: 15
+    cache:
+      # Immutable content-addressed snapshots plus per-run manifests. Set
+      # mode=read_only for offline replay; a cache miss then degrades only the
+      # affected query, paper, or card instead of contacting an external service.
+      enabled: true
+      root: ".science/cache/experiment_design/v1"
+      mode: "read_write"
+    paper_screening:
+      fulltext_budget: 15
+      max_candidates_before_llm: 32
+      parallel_workers: 8
+    evidence_card_extraction:
+      # Card extraction may use bounded LLM concurrency; full-text acquisition
+      # and MinerU parsing remain serial in the collector.
+      parallel_workers: 3
+    fulltext:
+      enabled: true
+      cache_dir: "workspace/experiment_design/fulltext_cache"
+      max_papers: 15
+      max_candidates_per_paper: 12
+      max_pdf_bytes: 50000000
+      timeout_seconds: 30
+      per_host_concurrency: 2
+      parser_backend: "survey_mineru"
+      parser_batch_size: 1
+      enable_landing_page_recovery: true
+      enable_unpaywall_resolution: false
+      max_landing_page_bytes: 1500000
+      max_declared_pdf_links_per_landing_page: 8
+      max_redirects: 5
+      failure_non_pdf_ttl_seconds: 21600
+      failure_access_denied_ttl_seconds: 1800
+      failure_rate_limited_ttl_seconds: 900
+    openalex:
+      api_key: "${oc.env:OPENALEX_API_KEY,''}"
+      email: "${oc.env:OPENALEX_EMAIL,''}"
+      base_url: "https://api.openalex.org"
+    semantic_scholar:
+      api_key: "${api.semantic_scholar.api_key}"
+      base_url: "https://api.semanticscholar.org/graph/v1"
+
+# ============================================
+# Research Plan Author Configuration
+# ============================================
+research_plan_author:
+  enabled: true
+  # Empty preserves per-run co-location; set a path to centralize Author artifacts.
+  output_root: ""
+  provider: "${oc.env:RESEARCH_PLAN_AUTHOR_LLM_PROVIDER,'qwen'}"
+  model: "${oc.env:RESEARCH_PLAN_AUTHOR_LLM_MODEL,'qwen3.8-flash'}"
+  authoring:
+    require_survey_manifest: true
+    proposal_without_observed_results: true
+    unsupported_claims_forbidden: true
+    # 0 disables generic repair; 1 permits one bounded repair for each Author section and Blueprint assignment.
+    max_contract_repairs_per_section: 1
+    composer_concurrency: 5
+    section_cache:
+      enabled: true
+      mode: "read_write"
+      root: ".science/cache/research_plan_author/v1"
+    temperature: 0.5
+    require_survey_binding: false
+  document_quality:
+    enabled: true
+    model: "${oc.env:RESEARCH_PLAN_AUTHOR_QUALITY_MODEL,'qwen3.7-plus'}"
+    judge_temperature: 0.0
+    revision_temperature: 0.5
+    judge_max_retries: 3
+    max_iterations: 2
+    score_concurrency: 3
+    max_revision_block_edits: 14
+    special_score_weight: 0.25
+  idea_evolution:
+    default_mode: "auto"
+    max_iterations: 3
+    allow_raw_internal_reasoning: false
+    require_source_anchors: true
+  rendering:
+    # The source template is treated as read-only. Each Author run copies it
+    # into its own timestamped output project before any insertion occurs.
+    template_dir: "${repo_root}/Conference-LaTeX-template_10-17-19"
+    template_profile: "ieee_conference_v1"
+    main_tex: ""
+    engine: "${oc.env:SCIENCE_LATEX_ENGINE,'pdflatex'}"
+    bibtex: "${oc.env:SCIENCE_BIBTEX,'bibtex'}"
+    pdf_renderer: "${oc.env:SCIENCE_PDF_RENDERER,'pdftoppm'}"
+    compile_timeout_seconds: 180
+    pdf_validation_timeout_seconds: 60
+    allow_shell_escape: false
+
+# ============================================
+# Blog Agent Configuration
+# ============================================
+blog:
+  source_workspace_root: "${workspace.root}"
+  provider: "${oc.env:BLOG_LLM_PROVIDER,''}"
+
+  # Literature retrieval automatically prefers a usable local graph, then
+  # Semantic Scholar, then evidence already present in the experiment workspace.
+  # Set BLOG_RETRIEVAL_MODE to graph, semantic, or workspace to express a
+  # preference; an unavailable dependency always falls back safely.
+  retrieval:
+    mode: "${oc.env:BLOG_RETRIEVAL_MODE,'auto'}"
+    graph_db_path: "${oc.env:BLOG_GRAPH_DB_PATH,'data/processed/graph.db'}"
+
+  # OpenAI API
+  openai:
+    api_key: ""
+    base_url: ""
+
+  # Google Gemini API (optional)
+  gemini:
+    api_key: ""
+    base_url: "https://generativelanguage.googleapis.com/v1beta"
+
+  # GenGraph (image generation for figures)
+  gengraph:
+    provider: "${oc.env:IMAGE_GENERATION_PROVIDER,'qwen'}"
+    api_key: "${oc.env:DASHSCOPE_API_KEY,''}"
+    providers:  # Default configuration for each provider, can be customized
+      qwen:
+        base_url: "${oc.env:DASHSCOPE_IMAGE_BASE_URL,'https://dashscope.aliyuncs.com/api/v1'}"
+        model: "${oc.env:IMAGE_ACADEMIC_FIGURE_MODEL,'wan2.7-image-pro'}"
+      dashscope:
+        base_url: "${oc.env:DASHSCOPE_IMAGE_BASE_URL,'https://dashscope.aliyuncs.com/api/v1'}"
+        model: "${oc.env:IMAGE_ACADEMIC_FIGURE_MODEL,'wan2.7-image-pro'}"
+      openai:
+        base_url: "https://api.openai.com/v1"
+        model: "gpt-image-2"
+      openrouter:
+        base_url: "https://openrouter.ai/api/v1"
+        model: "google/gemini-3-pro-image-preview"
+      gemini:
+        base_url: "https://generativelanguage.googleapis.com/v1beta"
+        model: "gemini-3-pro-image-preview"
+
+  # DeepEraser (text removal)
+  deeperaser:
+    use_cuda: false  # Whether to use GPU
+
+  # Semantic Scholar API (for paper abstract retrieval and PDF download)
+  semantic_scholar:
+    api_key: "${api.semantic_scholar.api_key}"  # Optional: your Semantic Scholar API key
+
+  # Default Model settings (It is only used when you do not pass the model parameter when creating an agent.)
+  model: "${oc.env:BLOG_LLM_MODEL,''}"
+
+  # Blog Illustrate settings
+  illustrate:
+    only_gen_img: true  # true: only generate images, no OCR; false: generate + OCR cleanup
+
+# ============================================
+# Pipeline Configuration
+# ============================================
+pipeline:
+  # Blank uses the pipeline research start time; set a name explicitly to resume it.
+  name: ""
+
+  # State file for resume
+  state_file: "pipeline.yaml"
+
+  # Resume settings
+  resume_enabled: true
+
+  # Skip survey if already done
+  skip_survey: false
+
+  # Iteration settings
+  iterate:
+    max_iterations: 5
+
+    # Stop conditions
+    stop_conditions:
+      min_ideas_generated: 5
+      max_consecutive_failures: 3
+
+    # Feedback from experiment to idea agent
+    feedback:
+      include_metrics: true
+      include_errors: true
+      max_feedback_length: 8192
+
+  # Output settings
+  output:
+    root: "pipeline_runs"
+    idea_result_filename: "idea_result.json"
+    experiment_result_filename: "experiment_result.json"
+    resume_from_iteration: null
+
+  symbolic_memory_path: "idea_skill_priors"
+  default_macro_roles:
+    - "generator"
+    - "encoder"
+    - "decoder"
+    - "loss"
+    - "regularizer"
+    - "optimizer"
+    - "architecture"
+    - "attention"
+    - "embedding"
+    - "representation"
+    - "constraint"
+    - "retrieval"
+    - "augmentation"