Feedback
One command files a structured issue — for humans and agents alike.
[object Object]
Hit a bug? File it without leaving the terminal:
hara feedback "write_file lost its path parameter and looped"The command collects your hara version, OS, Node version and provider:model label (never keys),
aggressively redacts credentials, and files a structured issue to
hara-cli/hara via the gh CLI — or prints a
copy-paste body if gh isn't available.
--session— append a redacted tail of your most recent session. Off by default, because issues are public.--dry-run— print the body without filing.
Agents use the same door: an agent that hits a hara bug can run hara feedback itself — the
issue lands in the same queue with the same structure. That is deliberate, and it is the whole
of hara's agent-to-agent surface:
Our security stance
hara agents exchange structured data — they never accept task instructions from untrusted parties.
We deliberately do not build "agents assigning tasks to strangers' agents". Untrusted natural language flowing into an agent that holds your credentials and can execute code is a prompt- injection highway, and no sandbox prevents a legitimately-authorized credential from being talked into misuse. Reports, issues, structured data: yes. Cross-trust-boundary instructions: never.