Customization

Configuration Files

Swival supports persistent settings through TOML config files at two levels: a user-global file and a project-local file. Both are optional.

The global config lives at ~/.config/swival/config.toml (or $XDG_CONFIG_HOME/swival/config.toml if you set that variable). The project config is swival.toml in the base directory.

When both files exist, project settings override global settings. CLI flags override everything. The full precedence order is:

CLI flags > project config > global config > hardcoded defaults

To generate a starter config with all settings commented out, run one of:

swival --init-config            # global config
swival --init-config --project  # project config in current directory

Settings mostly use the same names as CLI flags, with a few exceptions: the config keys allowed_dirs and allowed_dirs_ro correspond to the CLI flags --add-dir and --add-dir-ro. Lists use TOML arrays instead of comma-separated strings.

provider = "openrouter"
model = "qwen/qwen3-coder-next"
max_turns = 250
allowed_commands = ["ls", "git", "python3"]
allowed_dirs = ["/tmp"]
allowed_dirs_ro = ["/opt/zig/lib/std"]
proactive_summaries = true
# no_continue = false
quiet = false
extra_body = { chat_template_kwargs = { enable_thinking = false } }
reasoning_effort = "high"
cache = true
# cache_dir = ".swival"
# color = true           # true = force color, false = force no-color, absent = auto
# no_mcp = false         # disable MCP server connections
# no_a2a = false         # disable A2A agent connections

# Reviewer settings
reviewer = "swival --reviewer-mode"
# self_review = true               # shorthand: mirrors provider/model into reviewer
review_prompt = "Focus on correctness and test coverage"
max_review_rounds = 15
objective = "objective.md"
verify = "verification/working.md"

# A2A agents (see docs.md/a2a.md)
[a2a_servers.research-agent]
url = "https://research.example.com"
auth_type = "bearer"
auth_token = "sk-..."

For the chatgpt provider (ChatGPT Plus/Pro subscriptions), no API key is needed since authentication is handled through OAuth:

provider = "chatgpt"
model = "gpt-5.4"

Relative paths in allowed_dirs, allowed_dirs_ro, skills_dir, cache_dir, objective, and verify resolve against the config file's parent directory, not the working directory. Tilde paths like ~/projects expand to the home directory.

The reviewer value is shell-split; only path-like first tokens (./, ../, ~) are resolved against the config directory, while bare command names like swival are left for PATH lookup at runtime. The same resolution applies to model when provider = "command".

If a project config contains api_key inside a git repository, Swival prints a warning because the key could be committed accidentally. Prefer environment variables for credentials.

The --system-prompt and no_system_prompt settings are mutually exclusive in config files, just as they are on the command line.

The library API (Session class) does not auto-load config files. If you want config file support in library code, call load_config() and config_to_session_kwargs() explicitly.

Instruction Files

Swival loads instruction files during startup and appends them to the system prompt. CLAUDE.md provides project rules (<project-instructions>), while AGENTS.md provides agent workflow commands and conventions (<agent-instructions>). Use /init in the REPL to auto-generate a project-level AGENTS.md.

CLAUDE.md

Loaded from the project base directory only. Capped at 10,000 characters.

AGENTS.md

Loaded from up to three locations, in this order:

  1. User-level: ~/.config/swival/AGENTS.md (or $XDG_CONFIG_HOME/swival/AGENTS.md)
  2. Global cross-agent: ~/.agents/AGENTS.md
  3. Project-level: <base-dir>/AGENTS.md

All are optional. Content is concatenated inside a single <agent-instructions> block sharing a combined 10,000 character budget. Files are read in the order shown; earlier files get budget priority.

The user-level file is for swival-specific personal conventions. The global file (~/.agents/AGENTS.md) follows the cross-agent standard shared with OpenCode, OpenHands, and similar tools — put conventions here that should apply regardless of which agent you're using. The project-level file is for project-specific rules.

This is a Go project using Chi for routing. Tests use testify.
Always run `go test ./...` after making changes.
Don't add dependencies without asking.

Disabling instructions

Use --no-instructions to skip all instruction files (CLAUDE.md and AGENTS.md at all levels, including ~/.agents/AGENTS.md).

swival --no-instructions "task"

Use --no-memory to skip loading auto-memory from .swival/memory/.

swival --no-memory "task"

Use --memory-full to inject the entire MEMORY.md file into the prompt instead of the default budgeted retrieval. This is the legacy behavior and serves as a fallback if the retrieval-based injection misses entries you need.

swival --memory-full "task"

If you set --system-prompt, instruction files are also skipped because you are providing the full prompt text directly.

System Prompt Control

The built-in prompt is stored in swival/system_prompt.txt and defines default behavior, tool policy, and coding expectations.

You can replace it completely with --system-prompt.

swival --system-prompt "You are a security auditor. Only report vulnerabilities." "Audit src/"

You can also remove the system message entirely with --no-system-prompt.

swival --no-system-prompt "Just answer: what is 2+2?"

--system-prompt and --no-system-prompt are mutually exclusive.

When a system message is present, Swival appends the current local date and time to that system content.

Sampling And Reproducibility

--temperature and --top-p control response sampling.

swival --temperature 0.3 --top-p 0.9 "task"

If you do not set --temperature, provider defaults apply. --top-p defaults to 1.0.

--seed passes a deterministic seed when the provider supports it.

swival --seed 42 "task"

Seeded runs are usually more stable, but identical output is still not guaranteed across all providers, model versions, and hardware environments.

Extra LLM Parameters

Some models accept parameters beyond the standard OpenAI chat completions API. The extra_body setting passes an arbitrary dictionary through to the underlying API call, so you can use provider-specific or model-specific options without Swival needing dedicated flags for each one.

On the command line, pass a JSON object:

swival --extra-body '{"chat_template_kwargs": {"enable_thinking": false}}' "task"

In a config file, use TOML inline table syntax:

extra_body = { chat_template_kwargs = { enable_thinking = false } }

In the library API:

session = Session(extra_body={"chat_template_kwargs": {"enable_thinking": False}})

The dictionary is forwarded as extra_body to the LiteLLM completion call, which passes it through to the provider's API. Refer to your model's documentation for supported parameters.

Reasoning Effort

Some models support a tunable reasoning level that controls how much effort the model puts into thinking before responding. This is a first-class parameter in Swival, separate from extra_body.

On the command line:

swival --provider chatgpt --model gpt-5.4 --reasoning-effort high "task"

In a config file:

reasoning_effort = "high"

In the library API:

session = Session(reasoning_effort="high")

Valid levels are none, minimal, low, medium, high, xhigh, and default. Not all models support this parameter — when used with a model that doesn't support it, the behavior depends on the provider (it may be ignored or cause an error). Only set it when you know your model supports it.

Thinking Tag Sanitization

Some open-weight models leak hidden-reasoning markers like <think> and </think> into their responses. This is especially common when using vLLM as the inference engine, where models may emit these tags even when thinking mode is disabled. The sanitize_thinking option strips these leaked tags from assistant content before returning it to the user.

This is off by default for all providers. Enable it in config if you're seeing leaked thinking tags:

sanitize_thinking = true

In the library API:

session = Session(sanitize_thinking=True)

The sanitizer strips <think>...</think> blocks, standalone <think> / </think> lines, and special tokens like <|start_header_id|>. Inline mentions of these tags in code examples or backtick-quoted text are preserved.

Outbound LLM Filter

Swival can run a user-defined script before every outbound LLM request to redact or block sensitive content. See Outbound LLM Filter for the script contract, configuration, and examples.

Secret Encryption

Swival can transparently encrypt recognized credential tokens before they leave your machine. See Secret Encryption for the full documentation, including built-in token patterns, custom patterns, key management, and threat model.

Lifecycle Hooks

Swival can run a user-defined command at startup and exit, useful for syncing .swival/ state to and from remote storage. See Lifecycle Hooks for the full documentation, including environment variables, execution ordering, failure semantics, and a complete Hugging Face Buckets example.

Turn And Token Limits

--max-turns limits how many agent-loop iterations are allowed.

swival --max-turns 10 "quick task"

The default turn limit is 100. If the loop reaches this limit without a final answer, Swival exits with code 2.

--max-output-tokens limits tokens generated per model call.

swival --max-output-tokens 16384 "task"

The default is 32768. If prompt size and context constraints require it, Swival clamps output budget downward automatically.

--max-context-tokens sets requested context length.

swival --max-context-tokens 65536 "task"

For LM Studio, this can trigger a model reload. When both --max-context-tokens and --max-output-tokens are set, --max-output-tokens must be less than or equal to context length.