1.8.9b2 — every help page says what the command does

An opt-in preview on the 1.8.9 fix line (#1722). Stable is 1.8.8.

Help an agent can act on

For an agent, a co help page is the prompt that describes the tool. We audited all 300 commands (#1721). Each one opened its help without error and none wrote a file doing it. But only 14 had what #1643 asks of every page: an example, what the command changes, and the way back to its parent.

• Every command outside co wiki (264 of them) now has an example whose flags exist, and a first line that says what it changes: Read-only, Sends, Deletes, Charges and so on. Each of those statements was written from the command's code, not guessed from its name.
• Every page ends with the way back (Back: co trust --help). It is generated from the command tree, so commands added later get it as well.
• CI enforces it (#1657). tests/unit/test_cli_help_contract.py checks every page offline, with no exceptions. A new command fails CI until its help meets the contract.
• Discovery fixes. A model that saw only the help pages it asked for found the right command for 9 of 14 goals. After these fixes it finds 13 of 13 single-step goals:
  • co status says it shows your credit balance.
  • co trust says its lists decide who may call your agent.
  • co env set says it writes the ~/.co/keys.env that every project reads.
  • co skills and co sub say which one is for someone else's published skills.

Help that was wrong, now corrected

• co auth feishu and co auth lark create a Feishu application you own. The page now says so.
• co keys --ssh --write writes to ~/.co/ssh/, not ~/.ssh/ as its option help said.

Install

python -m pip install --upgrade 'connectonion==1.8.9b2'
co --help

Known limits

• co wiki keeps its own reviewed help pages (#1656). 27 of its 36 pages still lack an example or a statement of what they change. They are being rewritten in #1667.
• The discovery test is opt-in (pytest -m real_api tests/e2e/real_api/test_cli_discovery_journeys.py). Two multi-step goals, writing a new skill and adding a published skill to a project, are expected failures: the test accepts a single command.