open-data-sci 0.1.0__tar.gz → 0.2.1__tar.gz

{open_data_sci-0.1.0 → open_data_sci-0.2.1}/.env.example

@@ -78,9 +78,8 @@ LLM_SERVER_BASE_URL=
 # (falls back to "EMPTY" when unset, which works for most local servers):
 # OPENAI_API_KEY=
 
-# ── Sampling & reasoning ───────────────────────────────
+# ── Sampling ───────────────────────────────────────────
 TEMPERATURE=
-THINKING_BUDGET=
 
 # ── Agent customization ────────────────────────────────
 NAME=

{open_data_sci-0.1.0 → open_data_sci-0.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: open-data-sci
-Version: 0.1.0
+Version: 0.2.1
 Summary: AI agent for data science and machine learning
 Project-URL: Homepage, https://github.com/f4roukb/open-data-sci
 Project-URL: Repository, https://github.com/f4roukb/open-data-sci
@@ -218,70 +218,71 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: <3.13,>=3.12
 Requires-Dist: catboost<2.0.0,>=1.2.0
 Requires-Dist: category-encoders<3.0.0,>=2.6.0
-Requires-Dist: connectorx<1.0.0,>=0.3.0
-Requires-Dist: duckduckgo-search<7.0.0,>=6.0.0
+Requires-Dist: duckdb<2.0.0,>=1.0.0
+Requires-Dist: duckduckgo-search<9.0.0,>=8.0.0
 Requires-Dist: httpx<1.0.0,>=0.27.0
 Requires-Dist: imbalanced-learn<1.0.0,>=0.12.0
-Requires-Dist: langchain-anthropic<1.0.0,>=0.2.0
-Requires-Dist: langchain-core<1.0.0,>=0.3.0
-Requires-Dist: langchain-openai<1.0.0,>=0.2.0
-Requires-Dist: langgraph<1.0.0,>=0.2.0
+Requires-Dist: langchain-anthropic<2.0.0,>=1.0.0
+Requires-Dist: langchain-core<2.0.0,>=1.0.0
+Requires-Dist: langchain-openai<2.0.0,>=1.0.0
+Requires-Dist: langgraph<2.0.0,>=1.0.0
 Requires-Dist: lightgbm<5.0.0,>=4.0.0
 Requires-Dist: lxml<7.0.0,>=6.1.0
 Requires-Dist: matplotlib<4.0.0,>=3.7.0
 Requires-Dist: networkx<4.0.0,>=3.0.0
-Requires-Dist: numpy<2.0.0,>=1.26.0
-Requires-Dist: optuna<4.0.0,>=3.6.0
-Requires-Dist: pandas[excel,feather,parquet]<3.0.0,>=2.0.0
-Requires-Dist: plotly<6.0.0,>=5.15.0
+Requires-Dist: numpy<2.5.0,>=2.0.0
+Requires-Dist: optuna<5.0.0,>=4.0.0
+Requires-Dist: pandas[excel,feather,parquet]<4.0.0,>=3.0.0
+Requires-Dist: plotly<7.0.0,>=6.0.0
 Requires-Dist: polars[calamine]<2.0.0,>=1.0.0
 Requires-Dist: prophet<2.0.0,>=1.1.0
 Requires-Dist: pydantic-settings<3.0.0,>=2.0.0
 Requires-Dist: pydantic<3.0.0,>=2.0.0
-Requires-Dist: pyod<3.0.0,>=2.0.0
+Requires-Dist: pyod<4.0.0,>=3.0.0
 Requires-Dist: python-dotenv<2.0.0,>=1.0.0
 Requires-Dist: pyyaml<7.0,>=6.0
-Requires-Dist: rich<14.0.0,>=13.0.0
+Requires-Dist: rich<16.0.0,>=15.0.0
 Requires-Dist: sandbox-runtime<1.0.0,>=0.2.0
 Requires-Dist: scikit-learn<2.0.0,>=1.5.0
 Requires-Dist: scipy<2.0.0,>=1.14.0
 Requires-Dist: seaborn<1.0.0,>=0.12.0
 Requires-Dist: shap<1.0.0,>=0.46.0
 Requires-Dist: statsmodels<1.0.0,>=0.14.0
-Requires-Dist: textual<0.48.0,>=0.47.0
+Requires-Dist: textual<9.0.0,>=8.0.0
 Requires-Dist: umap-learn<1.0.0,>=0.5.0
-Requires-Dist: xgboost<3.0.0,>=2.0.0
+Requires-Dist: xgboost<4.0.0,>=3.0.0
 Requires-Dist: xxhash<4.0.0,>=3.5.0
 Provides-Extra: aws
 Requires-Dist: boto3<2.0.0,>=1.34.0; extra == 'aws'
-Requires-Dist: langchain-aws<1.0.0,>=0.2.0; extra == 'aws'
+Requires-Dist: langchain-aws<2.0.0,>=1.0.0; extra == 'aws'
 Provides-Extra: azure
 Requires-Dist: azure-identity<2.0.0,>=1.15.0; extra == 'azure'
 Provides-Extra: dev
-Requires-Dist: kaggle<2.0.0,>=1.6.0; extra == 'dev'
-Requires-Dist: mypy<2.0.0,>=1.13.0; extra == 'dev'
-Requires-Dist: pandas-stubs<3.0.0,>=2.0.0; extra == 'dev'
-Requires-Dist: pytest-asyncio<1.0.0,>=0.23.0; extra == 'dev'
+Requires-Dist: kaggle<3.0.0,>=2.0.0; extra == 'dev'
+Requires-Dist: mypy<3.0.0,>=2.0.0; extra == 'dev'
+Requires-Dist: pandas-stubs<4.0.0,>=3.0.0; extra == 'dev'
+Requires-Dist: pyright<2.0.0,>=1.1.390; extra == 'dev'
+Requires-Dist: pytest-asyncio<2.0.0,>=0.23.0; extra == 'dev'
 Requires-Dist: pytest-cov<5.0.0,>=4.0.0; extra == 'dev'
 Requires-Dist: pytest-xdist<4.0.0,>=3.0.0; extra == 'dev'
-Requires-Dist: pytest<8.0.0,>=7.0.0; extra == 'dev'
+Requires-Dist: pytest<10.0.0,>=9.0.0; extra == 'dev'
 Requires-Dist: ruff<1.0.0,>=0.8.0; extra == 'dev'
 Requires-Dist: types-openpyxl<4.0.0,>=3.1.0; extra == 'dev'
 Requires-Dist: vulture<3.0.0,>=2.0.0; extra == 'dev'
 Provides-Extra: docs
 Requires-Dist: mkdocs-material<10.0.0,>=9.5.0; extra == 'docs'
 Requires-Dist: mkdocs<2.0.0,>=1.6.0; extra == 'docs'
-Requires-Dist: mkdocstrings[python]<1.0.0,>=0.26.0; extra == 'docs'
+Requires-Dist: mkdocstrings[python]<2.0.0,>=1.0.0; extra == 'docs'
 Provides-Extra: gcp
-Requires-Dist: langchain-google-vertexai<3.0.0,>=2.0.0; extra == 'gcp'
+Requires-Dist: langchain-google-vertexai<4.0.0,>=3.0.0; extra == 'gcp'
 Provides-Extra: gemini
-Requires-Dist: langchain-google-genai<3.0.0,>=2.0.0; extra == 'gemini'
+Requires-Dist: langchain-google-genai<5.0.0,>=4.0.0; extra == 'gemini'
 Provides-Extra: jax
 Requires-Dist: flax<1.0.0,>=0.8.0; extra == 'jax'
 Requires-Dist: jax[cpu]<1.0.0,>=0.4.0; extra == 'jax'
 Requires-Dist: optax<1.0.0,>=0.2.0; extra == 'jax'
 Provides-Extra: ollama
-Requires-Dist: langchain-ollama<1.0.0,>=0.2.0; extra == 'ollama'
+Requires-Dist: langchain-ollama<2.0.0,>=1.0.0; extra == 'ollama'
 Description-Content-Type: text/markdown
 
 # OpenDataSci
@@ -300,6 +301,7 @@ A production-grade AI agent for data science and machine learning. See the [proj
 - [Python SDK](#python-sdk)
 - [Models](#models)
 - [Configuration](#configuration)
+- [Custom Skills](#custom-skills)
 - [Environment Variables](#environment-variables)
 
 ---
@@ -384,7 +386,7 @@ To use a different provider, pass `--provider`:
 
 ```bash
 opendatasci data.csv --provider openai --api-key sk-...
-opendatasci data.csv --provider ollama --model llama3.2:3b   # local, no key needed
+opendatasci data.csv --provider ollama --model qwen3.5:9b   # local, no key needed
 ```
 
 ### Setup with a config file
@@ -394,11 +396,10 @@ For a reusable configuration across projects, create a YAML file and pass it wit
 ```yaml
 # datasci.yaml
 provider: anthropic
-model: claude-sonnet-4-6
+model: claude-sonnet-5
 secondary_provider: openai
-secondary_model: gpt-4o-mini
+secondary_model: gpt-5.6-luna
 temperature: 0.1
-thinking_budget: 8000
 ```
 
 ```bash
@@ -445,7 +446,6 @@ opendatasci PATH [OPTIONS]
 | `--secondary-model` | *(provider default)* | Secondary model name for lightweight tasks (summarisation, etc.) |
 | `--api-key` | *(env var)* | API key for the primary provider. Falls back to the standard env var for the selected provider |
 | `--theme` | `default` | Colour palette. Choices: `default`, `accessible`, `light`, `solarized`, `dracula`. Run `/themes` inside the TUI for descriptions |
-| `--debug` | `false` | Enable debug output — writes a detailed `opendatasci_debug.log` |
 | `--config` | *(none)* | Path to a YAML file containing `OpenDataSciConfig` fields; explicit TUI flags take precedence |
 | `--list-providers` | | Print all supported providers and their default models, then exit |
 | `--version` | | Print the installed version, then exit |
@@ -457,7 +457,7 @@ opendatasci PATH [OPTIONS]
 opendatasci data.xlsx
 
 # Switch provider and primary model
-opendatasci data.csv --provider openai --model gpt-4o
+opendatasci data.csv --provider openai --model gpt-5.6-sol
 
 # Bedrock with a region
 REGION=us-west-2 opendatasci ./project/ --provider bedrock
@@ -466,7 +466,7 @@ REGION=us-west-2 opendatasci ./project/ --provider bedrock
 opendatasci data.parquet --theme accessible
 
 # Mix providers — heavy model on one, lightweight secondary on another
-opendatasci data.csv --provider anthropic --secondary-provider openai --secondary-model gpt-5.4-mini
+opendatasci data.csv --provider anthropic --secondary-provider openai --secondary-model gpt-5.6-luna
 
 # See all available providers
 opendatasci --list-providers
@@ -480,6 +480,8 @@ Type `/` in the input box to trigger autocomplete. All commands are available at
 
 | Command | Description |
 |---------|-------------|
+| `/cancel-all-messages` | Cancel all messages queued while the agent was busy |
+| `/cancel-message` | Cancel the most recently queued message |
 | `/clear` | Clear conversation context (preserves session variables and loaded data) |
 | `/compact` | Summarise and compress conversation history to free up context |
 | `/help` | Show all available commands |
@@ -490,6 +492,8 @@ Type `/` in the input box to trigger autocomplete. All commands are available at
 | `/themes` | List available colour themes with descriptions |
 | `/exit` | Quit OpenDataSci |
 
+Sending a message while the agent is still working doesn't reject it — it's pinned above the input box as a queued message and run automatically, in order, once the agent finishes (unless the agent is waiting on your answer to a question). Use `/cancel-message` or `/cancel-all-messages` to discard queued messages instead of waiting for them to run.
+
 ---
 
 ## File Attachments
@@ -554,7 +558,7 @@ from opendatasci import OpenDataSciConfig, create_agent
 
 config = OpenDataSciConfig(
     provider="openai",
-    model="gpt-4o",
+    model="gpt-5.6-sol",
     openai_api_key="sk-...",
     temperature=0.2,
 )
@@ -580,7 +584,18 @@ async with create_agent("data.parquet", config=config) as agent:
 | `google_cloud_project` | GCP project ID for Vertex AI (env: `GOOGLE_CLOUD_PROJECT`) |
 | `google_cloud_location` | Vertex AI region (env: `GOOGLE_CLOUD_LOCATION`) |
 | `azure_endpoint` | Azure OpenAI resource endpoint URL (env: `AZURE_OPENAI_ENDPOINT`) |
+| `azure_api_version` | Azure OpenAI API version — defaults to `2025-01-01-preview` (env: `AZURE_OPENAI_API_VERSION`) |
 | `llm_server_base_url` | Custom API base URL — required for `ollama` and `openai_compatible_server` (env: `LLM_SERVER_BASE_URL`) |
+| `temperature` | Sampling temperature — not sent to Claude 4.6+ / Sonnet 5 models, which use adaptive thinking (env: `TEMPERATURE`) |
+| `name` | Display name for the agent — defaults to `"Sai"` (env: `NAME`) |
+| `mcp_servers` | List of MCP server URLs the agent may connect to (env: `MCP_SERVERS`) |
+| `extra_web_domains` | Additional hostnames the `fetch_url` tool may retrieve, on top of the built-in allowlist (env: `EXTRA_FETCH_DOMAINS`) |
+| `override_web_domains` | When set, replaces the built-in domain allowlist entirely — `extra_web_domains` is still applied on top |
+| `skills_directory` | Path to a directory of custom skill files loaded in addition to built-ins (env: `SKILLS_DIRECTORY`) |
+| `builtin_skills_directory` | Path to the built-in skills directory — override only to replace defaults entirely (env: `BUILTIN_SKILLS_DIRECTORY`) |
+| `worker_timeout_seconds` | Max seconds to wait for spawned workers to finish — `null` disables the timeout, default `300` (env: `WORKER_TIMEOUT_SECONDS`) |
+| `midturn_compaction_threshold` | Token count at which context is compacted mid-turn — default `96000` (env: `MIDTURN_COMPACTION_THRESHOLD`) |
+| `local_code_exec_timeout` | Max seconds for a single sandboxed code-execution run — default `1800` (env: `CODE_EXEC_TIMEOUT`) |
 
 ---
 
@@ -590,14 +605,14 @@ OpenDataSci supports every major LLM provider. Pass `--provider` to the TUI or s
 
 | Provider | Flag | Extra required | Default model |
 |----------|------|----------------|---------------|
-| Anthropic | `anthropic` | *(none — default)* | `claude-sonnet-4-6` |
-| OpenAI | `openai` | *(none)* | `gpt-5.5` |
-| OpenAI-compatible server (e.g. vLLM) | `openai_compatible_server` | *(none)* | `meta-llama/Llama-3.2-3B-Instruct` |
-| AWS Bedrock | `bedrock` | `open-data-sci[aws]` | `us.anthropic.claude-sonnet-4-6` |
-| Google Gemini | `gemini` | `open-data-sci[gemini]` | `gemini-2.5-pro` |
-| Google Vertex AI | `vertexai` | `open-data-sci[gcp]` | `gemini-2.5-pro` |
-| Azure OpenAI | `azure` | `open-data-sci[azure]` | `gpt-4o` |
-| Ollama | `ollama` | `open-data-sci[ollama]` | `llama3.2:3b` |
+| Anthropic | `anthropic` | *(none — default)* | `claude-sonnet-5` |
+| OpenAI | `openai` | *(none)* | `gpt-5.6-sol` |
+| OpenAI-compatible server (e.g. vLLM) | `openai_compatible_server` | *(none)* | `Qwen/Qwen3.5-4B` |
+| AWS Bedrock | `bedrock` | `open-data-sci[aws]` | `us.anthropic.claude-sonnet-5` |
+| Google Gemini | `gemini` | `open-data-sci[gemini]` | `gemini-3.5-flash` |
+| Google Vertex AI | `vertexai` | `open-data-sci[gcp]` | `gemini-3.5-flash` |
+| Azure OpenAI | `azure` | `open-data-sci[azure]` | `gpt-5.6-sol` |
+| Ollama | `ollama` | `open-data-sci[ollama]` | `qwen3.5:9b` |
 
 Pass `--list-providers` to print this table from the TUI at any time.
 
@@ -611,19 +626,94 @@ Place these files inside your workspace's `.opendatasci/` directory:
 
 | Path | Purpose |
 |------|---------|
-| `.opendatasci/mcp.json` | MCP server URLs — connects the agent to external tool servers |
+| `.opendatasci/mcp.json` | MCP server definitions — connects the agent to external tool servers |
 | `.opendatasci/plans/` | Persisted plan files — auto-managed, one file per planning session |
 
+`mcp.json` uses the same convention as Cursor:
+
+```json
+{
+  "mcpServers": {
+    "my-server":   { "url": "http://localhost:8080" },
+    "another":     { "url": "http://localhost:9000" }
+  }
+}
+```
+
+---
+
+## Custom Skills
+
+Skills are Markdown (or YAML/JSON) files that give the agent a specialised persona and instruction set. OpenDataSci ships several built-in skills; you can add your own at the workspace level or point the agent at any directory you choose.
+
+### Workspace skills (recommended)
+
+Place skill files inside `.opendatasci/skills/` in your workspace directory — they are picked up automatically, no configuration needed:
+
+```
+<workspace>/
+└── .opendatasci/
+    ├── skills/
+    │   ├── my_skill.md              # standalone skill named "my_skill"
+    │   └── my_domain/
+    │       └── specialist.md        # domain-scoped skill: "my_domain::specialist"
+    └── skill_domains/
+        └── my_domain/
+            └── manifest.md          # optional domain manifest
+```
+
+Subdirectories create *domain-scoped* skills. The agent refers to them with the `domain::skill` naming convention (e.g. `my_domain::specialist`).
+
+### File format
+
+Skill files must be `.md`. The filename stem becomes the skill name and the file body is the prompt content. Files with any other extension are silently ignored.
+
+```markdown
+<!-- .opendatasci/skills/forecasting.md -->
+You are a time-series forecasting specialist. When analysing data, always...
+```
+
+### Global skills directory
+
+To share skills across workspaces, set `SKILLS_DIRECTORY` in your environment or `.env` file:
+
+```bash
+SKILLS_DIRECTORY=/home/user/my-skills
+```
+
+This directory is scanned *in addition to* the workspace `.opendatasci/skills/` directory and the built-in skills. When two sources define a skill with the same name, the later source wins (built-ins → workspace → `SKILLS_DIRECTORY`).
+
+You can also pass `skills_directory` directly when using the Python SDK:
+
+```python
+from opendatasci import OpenDataSciConfig, create_agent
+
+config = OpenDataSciConfig(skills_directory="/home/user/my-skills")
+
+async with create_agent("data.csv", config=config) as agent:
+    ...
+```
+
 ### Environment variables
 
 | Variable | Description |
 |----------|-------------|
 | `ANTHROPIC_API_KEY` | API key for the Anthropic provider |
-| `OPENAI_API_KEY` | API key for the OpenAI provider |
-| `REGION` | Cloud provider region |
+| `OPENAI_API_KEY` | API key for the OpenAI / OpenAI-compatible server provider |
+| `GOOGLE_API_KEY` | API key for the Google Gemini provider |
+| `AZURE_OPENAI_API_KEY` | API key for the Azure OpenAI provider |
+| `REGION` | AWS region for Bedrock |
+| `GOOGLE_CLOUD_PROJECT` | GCP project ID for Vertex AI |
+| `GOOGLE_CLOUD_LOCATION` | Vertex AI region / location |
+| `AZURE_OPENAI_ENDPOINT` | Azure OpenAI resource endpoint URL |
+| `AZURE_OPENAI_API_VERSION` | Azure OpenAI API version (default: `2025-01-01-preview`) |
 | `LLM_SERVER_BASE_URL` | Custom API base URL — used by `ollama` and `openai_compatible_server` providers |
-| `SKILLS_DIRECTORY` | Path to a directory of user-defined skill files (overrides none by default) |
-| `BUILTIN_SKILLS_DIRECTORY` | Path to the built-in skills directory (defaults to the bundled `resources/skills`) |
+| `TEMPERATURE` | LLM sampling temperature |
+| `MCP_SERVERS` | Comma-separated list of MCP server URLs |
+| `SKILLS_DIRECTORY` | Path to a directory of user-defined skill files |
+| `BUILTIN_SKILLS_DIRECTORY` | Path to the built-in skills directory (defaults to the bundled skills) |
+| `WORKER_TIMEOUT_SECONDS` | Max seconds to wait for spawned workers (default: `300`) |
+| `MIDTURN_COMPACTION_THRESHOLD` | Token count at which context is compacted mid-turn (default: `96000`) |
 | `CODE_EXEC_TIMEOUT` | Max seconds for a single sandboxed code execution (default: `1800`) |
 
 A `.env` file in the working directory is loaded automatically at startup.

{open_data_sci-0.1.0 → open_data_sci-0.2.1}/README.md

@@ -14,6 +14,7 @@ A production-grade AI agent for data science and machine learning. See the [proj
 - [Python SDK](#python-sdk)
 - [Models](#models)
 - [Configuration](#configuration)
+- [Custom Skills](#custom-skills)
 - [Environment Variables](#environment-variables)
 
 ---
@@ -98,7 +99,7 @@ To use a different provider, pass `--provider`:
 
 ```bash
 opendatasci data.csv --provider openai --api-key sk-...
-opendatasci data.csv --provider ollama --model llama3.2:3b   # local, no key needed
+opendatasci data.csv --provider ollama --model qwen3.5:9b   # local, no key needed
 ```
 
 ### Setup with a config file
@@ -108,11 +109,10 @@ For a reusable configuration across projects, create a YAML file and pass it wit
 ```yaml
 # datasci.yaml
 provider: anthropic
-model: claude-sonnet-4-6
+model: claude-sonnet-5
 secondary_provider: openai
-secondary_model: gpt-4o-mini
+secondary_model: gpt-5.6-luna
 temperature: 0.1
-thinking_budget: 8000
 ```
 
 ```bash
@@ -159,7 +159,6 @@ opendatasci PATH [OPTIONS]
 | `--secondary-model` | *(provider default)* | Secondary model name for lightweight tasks (summarisation, etc.) |
 | `--api-key` | *(env var)* | API key for the primary provider. Falls back to the standard env var for the selected provider |
 | `--theme` | `default` | Colour palette. Choices: `default`, `accessible`, `light`, `solarized`, `dracula`. Run `/themes` inside the TUI for descriptions |
-| `--debug` | `false` | Enable debug output — writes a detailed `opendatasci_debug.log` |
 | `--config` | *(none)* | Path to a YAML file containing `OpenDataSciConfig` fields; explicit TUI flags take precedence |
 | `--list-providers` | | Print all supported providers and their default models, then exit |
 | `--version` | | Print the installed version, then exit |
@@ -171,7 +170,7 @@ opendatasci PATH [OPTIONS]
 opendatasci data.xlsx
 
 # Switch provider and primary model
-opendatasci data.csv --provider openai --model gpt-4o
+opendatasci data.csv --provider openai --model gpt-5.6-sol
 
 # Bedrock with a region
 REGION=us-west-2 opendatasci ./project/ --provider bedrock
@@ -180,7 +179,7 @@ REGION=us-west-2 opendatasci ./project/ --provider bedrock
 opendatasci data.parquet --theme accessible
 
 # Mix providers — heavy model on one, lightweight secondary on another
-opendatasci data.csv --provider anthropic --secondary-provider openai --secondary-model gpt-5.4-mini
+opendatasci data.csv --provider anthropic --secondary-provider openai --secondary-model gpt-5.6-luna
 
 # See all available providers
 opendatasci --list-providers
@@ -194,6 +193,8 @@ Type `/` in the input box to trigger autocomplete. All commands are available at
 
 | Command | Description |
 |---------|-------------|
+| `/cancel-all-messages` | Cancel all messages queued while the agent was busy |
+| `/cancel-message` | Cancel the most recently queued message |
 | `/clear` | Clear conversation context (preserves session variables and loaded data) |
 | `/compact` | Summarise and compress conversation history to free up context |
 | `/help` | Show all available commands |
@@ -204,6 +205,8 @@ Type `/` in the input box to trigger autocomplete. All commands are available at
 | `/themes` | List available colour themes with descriptions |
 | `/exit` | Quit OpenDataSci |
 
+Sending a message while the agent is still working doesn't reject it — it's pinned above the input box as a queued message and run automatically, in order, once the agent finishes (unless the agent is waiting on your answer to a question). Use `/cancel-message` or `/cancel-all-messages` to discard queued messages instead of waiting for them to run.
+
 ---
 
 ## File Attachments
@@ -268,7 +271,7 @@ from opendatasci import OpenDataSciConfig, create_agent
 
 config = OpenDataSciConfig(
     provider="openai",
-    model="gpt-4o",
+    model="gpt-5.6-sol",
     openai_api_key="sk-...",
     temperature=0.2,
 )
@@ -294,7 +297,18 @@ async with create_agent("data.parquet", config=config) as agent:
 | `google_cloud_project` | GCP project ID for Vertex AI (env: `GOOGLE_CLOUD_PROJECT`) |
 | `google_cloud_location` | Vertex AI region (env: `GOOGLE_CLOUD_LOCATION`) |
 | `azure_endpoint` | Azure OpenAI resource endpoint URL (env: `AZURE_OPENAI_ENDPOINT`) |
+| `azure_api_version` | Azure OpenAI API version — defaults to `2025-01-01-preview` (env: `AZURE_OPENAI_API_VERSION`) |
 | `llm_server_base_url` | Custom API base URL — required for `ollama` and `openai_compatible_server` (env: `LLM_SERVER_BASE_URL`) |
+| `temperature` | Sampling temperature — not sent to Claude 4.6+ / Sonnet 5 models, which use adaptive thinking (env: `TEMPERATURE`) |
+| `name` | Display name for the agent — defaults to `"Sai"` (env: `NAME`) |
+| `mcp_servers` | List of MCP server URLs the agent may connect to (env: `MCP_SERVERS`) |
+| `extra_web_domains` | Additional hostnames the `fetch_url` tool may retrieve, on top of the built-in allowlist (env: `EXTRA_FETCH_DOMAINS`) |
+| `override_web_domains` | When set, replaces the built-in domain allowlist entirely — `extra_web_domains` is still applied on top |
+| `skills_directory` | Path to a directory of custom skill files loaded in addition to built-ins (env: `SKILLS_DIRECTORY`) |
+| `builtin_skills_directory` | Path to the built-in skills directory — override only to replace defaults entirely (env: `BUILTIN_SKILLS_DIRECTORY`) |
+| `worker_timeout_seconds` | Max seconds to wait for spawned workers to finish — `null` disables the timeout, default `300` (env: `WORKER_TIMEOUT_SECONDS`) |
+| `midturn_compaction_threshold` | Token count at which context is compacted mid-turn — default `96000` (env: `MIDTURN_COMPACTION_THRESHOLD`) |
+| `local_code_exec_timeout` | Max seconds for a single sandboxed code-execution run — default `1800` (env: `CODE_EXEC_TIMEOUT`) |
 
 ---
 
@@ -304,14 +318,14 @@ OpenDataSci supports every major LLM provider. Pass `--provider` to the TUI or s
 
 | Provider | Flag | Extra required | Default model |
 |----------|------|----------------|---------------|
-| Anthropic | `anthropic` | *(none — default)* | `claude-sonnet-4-6` |
-| OpenAI | `openai` | *(none)* | `gpt-5.5` |
-| OpenAI-compatible server (e.g. vLLM) | `openai_compatible_server` | *(none)* | `meta-llama/Llama-3.2-3B-Instruct` |
-| AWS Bedrock | `bedrock` | `open-data-sci[aws]` | `us.anthropic.claude-sonnet-4-6` |
-| Google Gemini | `gemini` | `open-data-sci[gemini]` | `gemini-2.5-pro` |
-| Google Vertex AI | `vertexai` | `open-data-sci[gcp]` | `gemini-2.5-pro` |
-| Azure OpenAI | `azure` | `open-data-sci[azure]` | `gpt-4o` |
-| Ollama | `ollama` | `open-data-sci[ollama]` | `llama3.2:3b` |
+| Anthropic | `anthropic` | *(none — default)* | `claude-sonnet-5` |
+| OpenAI | `openai` | *(none)* | `gpt-5.6-sol` |
+| OpenAI-compatible server (e.g. vLLM) | `openai_compatible_server` | *(none)* | `Qwen/Qwen3.5-4B` |
+| AWS Bedrock | `bedrock` | `open-data-sci[aws]` | `us.anthropic.claude-sonnet-5` |
+| Google Gemini | `gemini` | `open-data-sci[gemini]` | `gemini-3.5-flash` |
+| Google Vertex AI | `vertexai` | `open-data-sci[gcp]` | `gemini-3.5-flash` |
+| Azure OpenAI | `azure` | `open-data-sci[azure]` | `gpt-5.6-sol` |
+| Ollama | `ollama` | `open-data-sci[ollama]` | `qwen3.5:9b` |
 
 Pass `--list-providers` to print this table from the TUI at any time.
 
@@ -325,19 +339,94 @@ Place these files inside your workspace's `.opendatasci/` directory:
 
 | Path | Purpose |
 |------|---------|
-| `.opendatasci/mcp.json` | MCP server URLs — connects the agent to external tool servers |
+| `.opendatasci/mcp.json` | MCP server definitions — connects the agent to external tool servers |
 | `.opendatasci/plans/` | Persisted plan files — auto-managed, one file per planning session |
 
+`mcp.json` uses the same convention as Cursor:
+
+```json
+{
+  "mcpServers": {
+    "my-server":   { "url": "http://localhost:8080" },
+    "another":     { "url": "http://localhost:9000" }
+  }
+}
+```
+
+---
+
+## Custom Skills
+
+Skills are Markdown (or YAML/JSON) files that give the agent a specialised persona and instruction set. OpenDataSci ships several built-in skills; you can add your own at the workspace level or point the agent at any directory you choose.
+
+### Workspace skills (recommended)
+
+Place skill files inside `.opendatasci/skills/` in your workspace directory — they are picked up automatically, no configuration needed:
+
+```
+<workspace>/
+└── .opendatasci/
+    ├── skills/
+    │   ├── my_skill.md              # standalone skill named "my_skill"
+    │   └── my_domain/
+    │       └── specialist.md        # domain-scoped skill: "my_domain::specialist"
+    └── skill_domains/
+        └── my_domain/
+            └── manifest.md          # optional domain manifest
+```
+
+Subdirectories create *domain-scoped* skills. The agent refers to them with the `domain::skill` naming convention (e.g. `my_domain::specialist`).
+
+### File format
+
+Skill files must be `.md`. The filename stem becomes the skill name and the file body is the prompt content. Files with any other extension are silently ignored.
+
+```markdown
+<!-- .opendatasci/skills/forecasting.md -->
+You are a time-series forecasting specialist. When analysing data, always...
+```
+
+### Global skills directory
+
+To share skills across workspaces, set `SKILLS_DIRECTORY` in your environment or `.env` file:
+
+```bash
+SKILLS_DIRECTORY=/home/user/my-skills
+```
+
+This directory is scanned *in addition to* the workspace `.opendatasci/skills/` directory and the built-in skills. When two sources define a skill with the same name, the later source wins (built-ins → workspace → `SKILLS_DIRECTORY`).
+
+You can also pass `skills_directory` directly when using the Python SDK:
+
+```python
+from opendatasci import OpenDataSciConfig, create_agent
+
+config = OpenDataSciConfig(skills_directory="/home/user/my-skills")
+
+async with create_agent("data.csv", config=config) as agent:
+    ...
+```
+
 ### Environment variables
 
 | Variable | Description |
 |----------|-------------|
 | `ANTHROPIC_API_KEY` | API key for the Anthropic provider |
-| `OPENAI_API_KEY` | API key for the OpenAI provider |
-| `REGION` | Cloud provider region |
+| `OPENAI_API_KEY` | API key for the OpenAI / OpenAI-compatible server provider |
+| `GOOGLE_API_KEY` | API key for the Google Gemini provider |
+| `AZURE_OPENAI_API_KEY` | API key for the Azure OpenAI provider |
+| `REGION` | AWS region for Bedrock |
+| `GOOGLE_CLOUD_PROJECT` | GCP project ID for Vertex AI |
+| `GOOGLE_CLOUD_LOCATION` | Vertex AI region / location |
+| `AZURE_OPENAI_ENDPOINT` | Azure OpenAI resource endpoint URL |
+| `AZURE_OPENAI_API_VERSION` | Azure OpenAI API version (default: `2025-01-01-preview`) |
 | `LLM_SERVER_BASE_URL` | Custom API base URL — used by `ollama` and `openai_compatible_server` providers |
-| `SKILLS_DIRECTORY` | Path to a directory of user-defined skill files (overrides none by default) |
-| `BUILTIN_SKILLS_DIRECTORY` | Path to the built-in skills directory (defaults to the bundled `resources/skills`) |
+| `TEMPERATURE` | LLM sampling temperature |
+| `MCP_SERVERS` | Comma-separated list of MCP server URLs |
+| `SKILLS_DIRECTORY` | Path to a directory of user-defined skill files |
+| `BUILTIN_SKILLS_DIRECTORY` | Path to the built-in skills directory (defaults to the bundled skills) |
+| `WORKER_TIMEOUT_SECONDS` | Max seconds to wait for spawned workers (default: `300`) |
+| `MIDTURN_COMPACTION_THRESHOLD` | Token count at which context is compacted mid-turn (default: `96000`) |
 | `CODE_EXEC_TIMEOUT` | Max seconds for a single sandboxed code execution (default: `1800`) |
 
 A `.env` file in the working directory is loaded automatically at startup.

{open_data_sci-0.1.0 → open_data_sci-0.2.1}/docs/api/agent.md

@@ -1,6 +1,6 @@
 # Agent
 
-`Agent` is the core conversational AI agent. It wraps a LangGraph state machine that orchestrates LLM calls, tool execution, concurrent workers, and rolling memory.
+`Agent` is the core conversational AI agent. It orchestrates LLM calls, tool execution, concurrent workers, and rolling memory automatically, so you only need to send queries and consume the resulting stream.
 
 ## Lifecycle
 
@@ -55,8 +55,8 @@ async for event in agent.astream(query):
 | Method | Description |
 |--------|-------------|
 | `clear_chat_history()` | Remove all messages and rolling memory summaries. Preserves sandbox state. |
-| `rewind_turn()` | Remove only the last turn (user message + agent response) from the graph state. |
-| `compact_chat_history()` | Use the LLM to summarise old turns, then discard them. Returns the summary text. Use this instead of `clear_chat_history` when you want to keep context across a long session without blowing up the context window. |
+| `rewind_turn()` | Remove only the last turn (user message + agent response) from the conversation. |
+| `compact_chat_history()` | Fold all turn summaries and any existing compaction into a single `ChatHistoryCompaction` record. Returns the compaction text. Use this instead of `clear_chat_history` when you want to preserve context across a long session. |
 
 ```python
 # After many turns, compact instead of clearing:
@@ -72,11 +72,49 @@ print("Compacted:", summary)
       show_source: false
       members:
         - astream
-        - resume_with_input
         - rewind_turn
         - clear_chat_history
         - compact_chat_history
-        - graph
+
+---
+
+## Session manager
+
+A session manager tracks the mapping from a session to its conversation threads in the graph checkpointer. Clearing the conversation (`clear_chat_history`) creates a new thread, so no prior state is visible to the LLM on the next turn.
+
+The default `LocalSessionManager` persists this mapping to `.opendatasci/session.json` inside the workspace. For cloud or multi-process deployments — where two agent instances must share a conversation — replace it with a custom `BaseSessionManager` backed by shared storage:
+
+```python
+import uuid
+from opendatasci.session.session_manager import BaseSessionManager
+
+class RedisSessionManager(BaseSessionManager):
+    def get_or_create_thread(self) -> uuid.UUID: ...
+    def create_thread(self) -> uuid.UUID: ...
+    def get_current_thread(self) -> uuid.UUID: ...
+
+from opendatasci.agents.agents import Agent
+from opendatasci import LocalWorkspace, OpenDataSciConfig
+
+async with Agent(
+    workspace=LocalWorkspace("data/"),
+    session_manager=RedisSessionManager(),
+    config=OpenDataSciConfig(),
+) as agent:
+    ...
+```
+
+::: opendatasci.session.session_manager.BaseSessionManager
+    options:
+      show_root_heading: true
+      show_source: false
+
+---
+
+::: opendatasci.session.session_manager.LocalSessionManager
+    options:
+      show_root_heading: true
+      show_source: false
 
 ---
 
@@ -89,4 +127,4 @@ print("Compacted:", summary)
       show_root_heading: true
       show_source: false
       members:
-        - run
+        - ainvoke