Datus Files

New Feature

• Local image reading and multimodal conversations - The new filesystem_tools.read_image tool sends PNG, JPEG, and WebP images to the currently selected multimodal model and preserves image content or metadata across session restoration, context compaction, CLI rendering, and observability. It follows the existing filesystem permissions, corrects orientation automatically, and enforces image-dimension limits. (#1412) Docs: Image Reading

Enhancement

• Dosi/OSI metric attribution and dimension discovery - OSI Date, Time, DateTime, and DateTimeTz types now map automatically to time dimensions and remain intact during editing. AskMetrics can invoke Dosi-native term_wise, mix_shift, and factor_shapley attribution. The metric-dimensions API exposes grouping recommendations and their sources while retaining every queryable dimension. list_metrics(path=...) and Subject Tree now use the same Knowledge Base paths, and non-blocking Dosi validation warnings no longer make successful writes appear to have failed. (#1421 #1436 #1443 #1447 datus-semantic-adapter#100 #101 #104 #106 #107) Docs: Metrics and Dimensions, Dosi Semantic Adapter
• Unified model-call observability and cache usage accounting - CLI, API, Agent, and Gateway now share consistent logging configuration. Traces retain model-provider correlation IDs, complete tool definitions, tool-selection settings, normalized messages, and call results. API/SSE payloads, sessions, benchmarks, and tracing also report prompt cache-write tokens consistently, making call-chain diagnosis and cache-write cost accounting more reliable. (#1408 #1418) Docs: Agent Configuration, Observability
• Datasource configuration documentation for 21 databases - The documentation navigation now replaces the broad Services section with Datasources and provides dedicated configuration guidance for BigQuery, ClickHouse, ClickZetta, Doris, DuckDB, DWS, GaussDB, Greenplum, Hive, Hologres, MaxCompute, MySQL, Oracle, PostgreSQL, Redshift, Snowflake, Spark, SQLite, StarRocks, TiDB, and Trino, including fields, defaults, aliases, authentication, namespaces, and TLS behavior. (#1448) Docs: Datasource Configuration, Database Adapters
• More efficient metric discovery and querying - Metric catalogs now use compact summaries, reducing context usage when browsing large metric sets. The new get_metric tool returns a metric's queryable dimensions, grouping recommendations, time grains, and any derivation relationships or window requirements supplied by the semantic adapter in a single call. query_metrics now includes compiled SQL only for dry_run, uses a larger result-preview budget, prioritizes requested dimensions and metrics during compression, and keeps column metadata aligned with the returned preview data. The CLI also displays discovery details, result size, compression state, and validation warnings. Derivation expressions and required dimensions are returned only when the semantic adapter provides them. (#1444) Docs: Metrics and Dimensions, AskMetrics

BugFix

• More reliable model streaming and reasoning modes - Streamed LiteLLM responses no longer re-export the complete model-call span for every chunk, preventing tracing from monopolizing the event loop and stalling the backend instance. OpenAIModel also no longer disables a configured Kimi thinking mode by default and now accommodates current Kimi and DeepSeek parameter and output-budget behavior. (#1419 #1435)
• Concurrent file edits no longer lose updates silently - Concurrent edit_file operations against the same path are now serialized and version-checked, preventing multiple calls from reading the same stale content and overwriting one another while both report success. (#1429)
• Reliable hosted sessions and custom node configuration - Base Agent nodes now select templates from node_config.system_prompt; an unbound request-level datasource no longer aborts the entire chat turn; and context-compaction archives now honor the explicit session_dir, allowing deployments with read-only project snapshots and separately mounted writable session directories to operate correctly. (#1437 #1438 #1439)

Upgrade Notes

• Migrate SQL Policy configuration and request context - Every row_filter and metric_row_filter must define a stable, unique id. The global access_mode: unrestricted mode has been removed; use request-level access_mode: scoped, then set policy_modes.<policy_id>: unrestricted to bypass only the selected policy. (#1416 datus-sql-policies#9) Docs: SQL Policy
• Upgrade Dosi semantic dependencies - To use native attribution and dimension recommendations, install datus-semantic-core>=0.2.6 and datus-semantic-dosi>=0.1.14. The current datus-semantic-dosi 0.1.14 release pins dosi-engine 0.1.11. (#1421 #1443 datus-semantic-adapter#101 #107) Docs: Dosi Semantic Adapter
• Synchronize model SDK dependencies - Deployments that maintain their own dependency locks or model compatibility patches must update to openai>=2.45,<3, openai-agents[litellm]==0.18.1, litellm==1.100.1, and httpx==0.28.1. Image reading additionally requires Pillow>=10.4. (#1412 #1418) Docs: Image Reading