halo copilot
AI-assisted policy operations: the model drafts, Halos verifies, a human approves (ADR-0035)
The model proposes; Halos’ validator, plan, impact preview, bypass-token check and internal/stats verify; a human approves. Nothing is written without –apply, and –apply only commits to a new local branch: pushing, the PR and merging are human steps. Each model call logs its token counts to stderr; –offline needs no model.
halo copilotNone.
Global flags
Section titled “Global flags”| Flag | Shorthand | Type | Default | Description |
|---|---|---|---|---|
--output |
string | text | output format: text|json |
Commands
Section titled “Commands”| Command | Description |
|---|---|
halo copilot draft |
Turn an intent into a verified policy diff (validate, plan, impact); –apply commits it to a local branch |
halo copilot experiment |
Design an experiment: arms and metrics proposed, sample size and duration computed by internal/stats |
halo copilot explain |
Explain a release, a policy diff or an experiment in plain language, citing only Halos outputs |
halo copilot next |
Suggest next steps from experiment verdicts, eval scorecards, ROI and budgets, each with evidence and the command |