sif.agent 2026.8.26.7

sif agent

<p align="center"> <img src="assets/sif.jpg" alt="Sif" width="720"> </p>

sif is a lightweight BYOK coding assistant written in C#. It works with OpenAI-compatible APIs, local model servers, and a small set of built-in tools for reading, editing, searching, serving, and inspecting local context.

C# is a first-class citizen — the built-in Roslyn tools let the agent analyze solutions and projects directly using the .NET Compiler Platform.

Highlights

• Interactive - run sif to start a coding session.
• Local model friendly - use vLLM, Ollama, LM Studio, LiteLLM, OpenAI, or any compatible endpoint.
• Tool calling - enable shell, file, context, sleep, static server, diagnostics, and Roslyn tools.
• C# as a first-class citizen - the built-in Roslyn tools let the agent analyze C# solutions directly — find symbols across projects, and inspect compiler diagnostics — without leaving the chat.
• MCP servers - connect to Model Context Protocol servers over stdio, HTTP, streamable HTTP, or SSE.
• Context store - large tool outputs are stored out-of-band and can be searched or read back by handle.
• History compaction - long chats are summarized automatically using the model's advertised context window when available.
• Crash recovery and resume - chat context is saved after every conversation update; restart and resume an interrupted session, or use /resume to pick a prior chat without loading every history first.
• VS Code context - the companion extension exposes active editor, cursor, and selection context to sif.
• Skills - reusable markdown instructions can be loaded from project or user skill folders.
• Persistent config - store model, endpoint, tool, reasoning, and compaction settings in ~/.sif/sif-agent.json.
• Model profiles - save multiple named endpoint/model presets and switch between them with sif models switch or the /model chat command.
• Secure API key storage - API keys can be stored in OS credential stores (Windows Credential Manager, macOS Keychain, Linux libsecret) instead of plaintext config files. Use sif secure migrate to move keys to secure storage.

Requirements

• .NET 10 SDK
• An OpenAI-compatible API endpoint
• Unix-like shells: curl, tar, and zip
• Windows PowerShell

Install

Install the released .NET tool from nuget.org:

dotnet tool install --global sif.agent

Update an existing install:

dotnet tool update --global sif.agent

The sif command will be available in your PATH after the .NET global tools directory is on your shell path.

sif update also uses the NuGet feed by default, and AGENT_AUTO_UPDATE_SOURCE can point updates at another NuGet feed URL or a local folder containing sif.agent packages.

Releases are published from Git tags like v1.2.3 via GitHub Actions. The publish workflow uses NuGet trusted publishing, so nuget.org must have a trusted publisher policy for this repository and workflow.

Install the companion VS Code extension:

sif vscode install

The first-run setup wizard also offers to install the extension. If VS Code is already running, restart it after installing the extension.

The source installers remain available for installing from the latest main branch. Linux/macOS:

curl -fsSL https://raw.githubusercontent.com/larsholm/sif/main/install.sh | bash

Windows PowerShell:

irm https://raw.githubusercontent.com/larsholm/sif/main/install.ps1 | iex

These download the latest main archive, build the agent, install it as a global .NET tool, and install the companion VS Code extension into ~/.vscode/extensions. Set VSCODE_EXTENSIONS=/path/to/extensions before running sif vscode install or the source installer to target another VS Code-compatible extensions directory.

On first launch, sif opens a setup wizard for the API base URL, optional API key, model, tools, and thinking/reasoning display. The wizard fetches available models from the configured endpoint when possible and falls back to manual model entry if the endpoint cannot be queried. Run sif setup any time to update those values. Chat startup also probes model metadata for context-window size when the endpoint exposes it.

For local development from a checked-out repository:

./sif.agent/build.sh install

On Windows, run .\install.ps1 from the repository root.

Quick Start

# Start an interactive chat
sif

# Use a specific endpoint and model
sif -u http://localhost:11434/v1 -m llama3.2

# Run one prompt and exit
sif complete "Explain async/await in C#"

# Pipe a prompt from stdin
cat prompt.txt | sif complete -

Tools

Tools can be enabled with --tools, AGENT_TOOLS, or persistent config. The CLI flag has highest priority, then AGENT_TOOLS, then ~/.sif/sif-agent.json.

The default tool set is:

bash,read,edit,write,sleep,serve,context

Available tools:

Tool	Description
bash	Execute allowed shell commands. Uses Bash on Unix-like systems and PowerShell on Windows. Includes macOS commands such as uname, sw_vers, and system_profiler; unknown commands ask for approval
read	Read file contents
edit	Edit files by replacing exact text
write	Create or overwrite files
sleep	Pause briefly before continuing or retrying
serve	Start a local static HTTP server for a directory
context	Enable ctx_index, ctx_search, ctx_read, ctx_summarize, and ctx_stats
diagnostics	Inspect sif configuration, AGENT_ environment variables, and chat history
roslyn	C# analysis via Roslyn — roslyn_find_symbols and roslyn_get_diagnostics (requires Roslyn Analyzers)

# Use the default tool set explicitly
sif --tools bash,read,edit,write,sleep,serve,context

# Add runtime diagnostics for this session
sif --tools bash,read,edit,write,sleep,serve,context,diagnostics

# Enable tools through the environment
export AGENT_TOOLS=bash,read,edit,write,context,diagnostics
sif

# Store tools persistently
sif config --set TOOLS=bash,read,edit,write,context,diagnostics
sif

When context is enabled, large local and MCP tool results are stored under ~/.sif/context/ and replaced in model context with a compact handle such as ctx_abc123. The model can then call ctx_search for focused snippets or ctx_read for a specific stored result. Defaults are sized for large local-model contexts: read returns up to 1,000 lines by default, shell output returns up to 24,000 characters, ctx_read returns up to 32,000 characters, and automatic context storage starts above 60,000 characters.

When chat history grows past the compaction threshold, sif summarizes older messages, stores the raw compacted history and summary in the context store, and keeps the system prompt plus the most recent messages in active history. On startup, sif probes the configured /models endpoint for context-window metadata such as context_length, context_window, max_model_len, or n_ctx. If the provider reports a context window and the threshold is still the built-in default, compaction starts at about 60% of that window. Set AGENT_COMPACTION_THRESHOLD or sif config --set COMPACTION_THRESHOLD=<tokens> to override it, or set it to 0 to disable compaction.

Saved Conversations

Interactive chats are saved under ~/.sif/conversations/ after every context update, including user messages, assistant responses, tool results, clears, system-prompt changes, and compaction. Their referenced large-context blobs are retained under ~/.sif/context/, so ctx_* handles continue to work after a resume. If sif exits unexpectedly, the next chat startup in the same working directory offers to resume the most recent unfinished session.

Use /resume in an interactive chat to choose a saved session with the arrow keys. That menu reads only lightweight session metadata; the message history is loaded only after you select one. /resume <id> also works with full IDs and unique ID prefixes. A normally exited session remains available for later resume as well.

The diagnostics tool is for inspecting sif's runtime state only. It is not a debugger and does not launch, attach to, or manage .NET debug adapter sessions. There is also a legacy debug tool alias for the same diagnostics behavior.

Sif streams model output, including thinking/reasoning deltas exposed by compatible providers, while accumulating any tool calls before executing them. Thinking and reasoning display works for OpenAI o-series models and Qwen3.x models via vLLM.

C# Development with Roslyn

Because sif is written in C#, it has deep native support for C# development through the Roslyn tool set. When enabled, the agent can inspect your C# solutions and projects using the .NET Compiler Platform (Roslyn) directly — no external tools or MCP servers needed.

Enable the Roslyn tools alongside your default set:

sif --tools bash,read,edit,write,sleep,serve,context,roslyn

# Or persistently
sif config --set TOOLS=bash,read,edit,write,sleep,serve,context,roslyn

The Roslyn tool set provides two tools:

Tool	Description
roslyn_find_symbols	Search for declarations of a symbol (class, method, property, etc.) across all projects in a .sln file. Returns the symbol name, kind, and source location.
roslyn_get_diagnostics	Compile a .csproj or .sln and surface all compiler diagnostics — errors, warnings, and information messages — with their severity, code, and source location.

Example usage

In a chat session, you can ask the agent to:

• Find where a class is defined: "Find the Program class in MyApp.sln"
• Check for compile errors: "Run diagnostics on MyApp/MyApp.csproj"
• Locate a method across projects: "Find all declarations of ConnectAsync in sif.sln"

The Roslyn tools use MSBuildWorkspace to load real project references and compilation contexts, so they respect your actual project configuration, targets, and imports — the same way Visual Studio or dotnet build would.

Secure API Key Storage

By default, API keys are stored in plaintext in ~/.sif/sif-agent.json. For better security, sif supports storing API keys in OS-native credential stores:

• Windows: Windows Credential Manager (DPAPI)
• macOS: Keychain (via security command)
• Linux: libsecret (via secret-tool) with encrypted file fallback

Commands

# Check secure storage status
sif secure

# Migrate API key from plaintext config to secure storage
sif secure migrate

# Restore API key from secure storage back to plaintext config (for troubleshooting)
sif secure restore

# Clear API key from secure storage
sif secure clear

After running sif secure migrate, the API key is removed from the plaintext config file and stored securely. The UseSecureApiKeyStorage: true setting is added to your config to indicate that the key should be loaded from secure storage.

Note: The current implementation secures the default API key. API keys stored in model profiles are not yet migrated to secure storage.

MCP Servers

sif can connect to Model Context Protocol servers configured in ~/.sif/sif-agent.json. MCP tools are listed at startup and exposed to the model alongside local tools.

{
  "McpServers": {
    "filesystem": {
      "Type": "stdio",
      "Command": "npx",
      "Args": ["-y", "@modelcontextprotocol/server-filesystem", "/home/lars/source"]
    },
    "remote-tools": {
      "Type": "streamableHttp",
      "Url": "https://example.com/mcp",
      "Headers": {
        "Authorization": "Bearer TOKEN"
      }
    }
  }
}

Supported Type values are stdio, http, streamableHttp, streamable-http, and sse. Stdio servers use Command, Args, and optional Env; HTTP-based servers use Url and optional Headers. Set Disabled to true to keep a server in the config without connecting to it.

Configuration

# Show current configuration
sif config

# Run the setup wizard
sif setup

# Uninstall the global tool and companion VS Code extension
sif uninstall

# Set persistent values
sif config --set BASE_URL=http://localhost:11434/v1
sif config --set MODEL=llama3.2
sif config --set TOOLS=bash,read,edit,write,context
sif config --set THINKING_ENABLED=true
sif config --set COMPACTION_THRESHOLD=60000
sif config --set MODEL_TIMEOUT_SECONDS=300

Configuration is loaded from ~/.sif/sif-agent.json, then overridden by environment variables, then overridden by CLI flags.

Variable	Description	Default
AGENT_BASE_URL	OpenAI-compatible API base URL	http://localhost:1234/v1
AGENT_API_KEY	API key, optional for local models	-
AGENT_MODEL	Model name	qwen3.6-27b-autoround
AGENT_TOOLS	Comma-separated list of tools	default tool set
AGENT_MAX_TOKENS	Maximum output tokens	model default
AGENT_TEMPERATURE	Sampling temperature	model default
AGENT_MODEL_TIMEOUT_SECONDS	Model request network timeout in seconds	SDK default
AGENT_THINKING_ENABLED	Enable model thinking or reasoning display	true
AGENT_COMPACTION_THRESHOLD	Chat-history token threshold for automatic compaction; 0 disables it	auto from model context window when available, otherwise 60000

AGENT_COMPACT_THRESHOLD is accepted as a backwards-compatible alias for AGENT_COMPACTION_THRESHOLD.

Model Profiles

sif can store multiple named model profiles in ~/.sif/sif-agent.json, each with its own base URL, model, optional API key, and optional compaction threshold. Switch between them without re-running setup.

# List configured profiles (current one is marked)
sif models

# Add a profile
sif models add ollama --url http://localhost:11434/v1 --model llama3.2
sif models add local --url http://localhost:8020/v1 --model qwen3.6-27b-autoround --compact 60000 --timeout 300

# Switch the active profile
sif models switch ollama

# Remove a profile
sif models remove ollama

sif models add accepts --url/-u, --model/-m, --key/-k, --compact (per-profile compaction threshold; falls back to the global setting when omitted), and --timeout (per-profile model request timeout in seconds).

During a chat session you can list and switch profiles inline. Switching clears the current conversation:

/model              # select a profile interactively
/model list         # list profiles and show the current one
/model switch local # switch to the "local" profile
/model local        # shorthand for switching

CLI Reference

Flag	Description
-m, --model	Model name
-u, --base-url	API base URL
-k, --api-key	API key
-s, --system	System prompt
-t, --temperature	Sampling temperature
-max, --max-tokens	Maximum output tokens
--timeout	Model request network timeout in seconds
-n, --no-stream	Disable streaming output
--tools	Enable tools, comma-separated
--thinking	Enable model thinking or reasoning display

Chat Commands

During an interactive chat session:

Command	Description
/q, /quit, /exit	Exit the chat session
/clear	Clear conversation history, keeping the system prompt
/sys <prompt>	Change the system prompt
/model	Select a model profile interactively
/model list	List model profiles and show the current one
/model <name>	Switch to a model profile (clears conversation)
/context	Show the last model request and persisted-state summary
/context full	Show the complete last model request, including tool schemas
/context history	Show persisted conversation history
/context list	List stored context entries
/context search <query>	Search stored context entries
/context read <id> [query]	Read a stored entry, optionally focused by query
/context delete <id>	Delete a stored context entry
/context drop <count>	Remove recent non-system chat messages
/context clear, /context clear history	Clear conversation history, keeping the system prompt
/context clear-history	Clear conversation history, keeping the system prompt
/context clear-store	Delete stored context entries for this session
/context clear all	Clear both chat history and stored context
btw <comment>	While the agent is working, queue a steering comment for the next tool call
/vscode	Show detected VS Code terminal/editor context
/debug	Show recent errors (saved to a log, survives /clear)
/debug latest	Show full details of the most recent error
/debug list	Show recent error entries
/help	Show help and options

In chat input, press Alt+Enter to insert a newline. Long input lines wrap in place and can be edited with the arrow, Home, End, Backspace, and Delete keys.

While the agent is working, type btw <comment> (or /btw <comment>) and press Enter to queue a steering comment for the next tool call. Any other non-empty input interrupts the current request, keeps the original task, and resumes with that input as a new instruction. If no further tool call occurs, queued comments run immediately after the current turn.

VS Code Context

When sif runs inside a VS Code integrated terminal, it detects that environment automatically. VS Code does not expose the active editor or selection to terminal child processes by default, so the companion extension writes live editor context to a small JSON file and exposes it through SIF_VSCODE_CONTEXT_FILE for new integrated terminals.

Variable	Description
SIF_VSCODE_CONTEXT_FILE	JSON file containing live editor context
SIF_VSCODE_FILE	Active editor file path or file:// URI
SIF_VSCODE_LINE	Active cursor line number
SIF_VSCODE_COLUMN	Active cursor column number
SIF_VSCODE_SELECTED_TEXT	Current selected text
SIF_VSCODE_SELECTED_TEXT_B64	Current selected text as UTF-8 base64

Use /vscode in chat to inspect what sif can see.

The VS Code extension lives in sif.vscode/. For local development, open the repository in VS Code and run the extension host with that folder as the extension development path. The sif: Start Chat With Editor Context command opens a sif terminal and keeps the context file updated as the active editor, cursor, or selection changes. Regular integrated terminals opened after the extension activates also receive SIF_VSCODE_CONTEXT_FILE.

Skills

sif loads skill files at startup and appends them to the system prompt. Use skills for reusable instructions that should apply when a request matches the skill description.

Supported locations:

./.sif/skills/
../.sif/skills/        # parent directories are checked up to filesystem root
~/.sif/skills/

Supported file layouts:

.sif/skills/my-skill.md
.sif/skills/my-skill/SKILL.md

Skill files are plain markdown. Frontmatter such as name and description is allowed and is passed through to the model as part of the skill content.

Examples

Local Model

sif config --set BASE_URL=http://100.118.58.55:8020/v1
sif config --set MODEL=qwen3.6-27b-autoround
sif config --set TOOLS=bash,read,edit,write,context
sif

Ollama

sif -u http://localhost:11434/v1 -m llama3.2

export AGENT_BASE_URL=http://localhost:11434/v1
export AGENT_MODEL=llama3.2
sif

LM Studio

export AGENT_BASE_URL=http://localhost:1234/v1
export AGENT_MODEL=local-model
sif

LiteLLM Proxy

export AGENT_BASE_URL=http://localhost:4000
export AGENT_API_KEY=your-key
export AGENT_MODEL=gpt-4o
sif

Thinking / Reasoning

# Enable thinking display for Qwen3.x via vLLM
sif --thinking true -u http://100.118.58.55:8020/v1 -m qwen3.6-27b-autoround

# Enable thinking display for OpenAI o-series
sif --thinking true -m o3-mini

# Store the setting
sif config --set THINKING_ENABLED=true

For Qwen3.x models, thinking is enabled by default on the server. The --thinking flag enables live display of streamed reasoning output when the provider exposes reasoning or reasoning_content deltas. Use --no-stream to request and display the complete reasoning response instead.

License

MIT