Halios CLI and Agent Skill

The halios command is the machine-facing client used by the Halios Agent Skill and CI. It owns repository setup, evaluation-suite authoring, fresh scenario execution, evidence inspection, and prompt-optimization handoffs. Applications export their runtime telemetry with stock OpenTelemetry; the CLI is not an application tracing SDK.

uv tool install 'haliosai-cli>=2.0.0'
halios --version

Use pipx install 'haliosai-cli>=2.0.0' when uv is unavailable. Install the cross-harness skill separately:

npx skills add HaliosAI/halios --skill halios

The stable project workflow is:

halios auth login
halios project init --agent support-assistant --command "python tests/halios_adapter.py"
halios eval review --json
halios project configure --json
halios project check --json
halios eval run -k 3 --json

Prompt optimization is driven by the coding agent and the normal immutable eval workflow:

halios optimize start --baseline-run <run-id> --prompt-file path/to/system-prompt.txt --json
# make one bounded prompt edit from the returned guidance, then run the unchanged suite
halios eval run -k 3 --json
halios optimize record <optimization-run-id> \
  --evaluation-run <candidate-eval-run-id> \
  --prompt-file path/to/system-prompt.txt \
  --json

Credentials are stored outside the repository. The repository contains only .halios/config.toml, .halios/eval.yml, and .halios/scenarios.yml.

In non-interactive CI, use HALIOS_API_KEY for control-plane commands and the separate agent-scoped HALIOS_OTLP_TOKEN for adapter trace export. Only trusted default-branch jobs should receive HALIOS_CI_PUBLISH_TOKEN. The jsonl adapter is used by local/CI simulation only; deployed staging and production applications export their own stock OpenTelemetry directly to Halios.

See the repository's skills/halios workflow for coding-agent setup and OTEL_SCHEMA.md for the telemetry accepted by Halios.

Development

python -m pip install -e '.[dev]'
python -m pytest -q

See CONTRIBUTING.md for the contribution and release process.