selfware 0.6.0

# Selfware configuration
# Copy this file to selfware.toml and adjust values for your setup.
#
# UNATTENDED RUNS: selfware does NOT impose a token/wall-clock/cost cap by
# default. For any long or unattended run, pass hard limits on the CLI:
#   selfware --max-turns 200 --max-wall-secs 3600 --max-budget-tokens 2000000 ...
# (These are the only hard run-bounding levers; there is no TOML equivalent and
# they reset on resume, so re-pass them each launch.)

# Set this to your OpenAI-compatible API endpoint (e.g., local vLLM/sglang server, ngrok tunnel, etc.)
endpoint = "http://127.0.0.1:8000/v1"
model = "/media/thread/trebuchet6/qwen35/models/Qwen3.5-122B-A10B-NVFP4-yarn-1010k"
max_tokens = 98304
context_length = 1010000

[extra_body]
# Disable thinking for tool-heavy workflows
chat_template_kwargs = { enable_thinking = false }

[safety]
# Scope filesystem access to the project directory by default. Only widen this
# (e.g. add "~/**") if you genuinely need it — a wide scope lets an autonomous
# run read/modify files far outside your project.
allowed_paths = ["./**"]
denied_paths = ["**/.env", "**/secrets/**", "**/.ssh/**", "**/target/**"]

[agent]
# Keep the iteration cap modest. Raise deliberately for large tasks rather than
# shipping a five-figure default.
max_iterations = 100
step_timeout_secs = 600
# Use XML-based tool calling for SGLang
native_function_calling = false

[continuous_work]
enabled = true
checkpoint_interval_tools = 10
checkpoint_interval_secs = 300
auto_recovery = true
# Bound automatic recovery so a wedged task cannot retry indefinitely.
max_recovery_attempts = 3

[retry]
max_retries = 5
base_delay_ms = 1000
max_delay_ms = 60000