* fix(cli): don't recover benign typos into destructive commands
CLI did-you-mean ranked purely by Levenshtein, so `orca worktree move`
sole-suggested `orca worktree remove` (distance 2) — an alias of the
destructive `worktree rm`. Suggestions also flow into --json
error.data.nextSteps, the agent recovery channel, so a blind retry could
delete a clean worktree.
Make destructiveness a declared property of the command instead of a verb
heuristic: add `destructive?: true` to CommandSpec and mark the
irreversible commands (worktree rm, environment rm, automations remove,
project setup-delete, tab profile delete, cookie delete, storage
local/session clear). The suggestion ranker excludes destructive
candidates unless the input token is itself a near-miss (distance <=1) of a
destructive verb, so `worktree remov` still recovers `rm`/`remove` while
`worktree move` no longer does. The guard tracks the registry, so it
also covers destructive verbs outside the delete family (e.g. kill).
Fixes#6303
Co-authored-by: Orca <help@stably.ai>
* fix(cli): use Array.at(-1) to satisfy oxlint prefer-at
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Orca <help@stably.ai>
* feat(cli): make the CLI self-correcting and self-describing for agents
Agents build a generalized model of how CLIs work and apply it to every
tool. When orca diverged — `rm` where git uses `remove` — a reasonable
first guess (`orca worktree remove`) dead-ended on a bare "Unknown
command" with no path forward. This makes the CLI degrade gracefully when
the orca-cli skill isn't loaded in context.
- First-class CommandSpec.aliases, resolved to the canonical path before
dispatch (no new handler registrations). `worktree remove`/`delete` now
resolve to `rm`; the ad-hoc `terminal focus` duplicate spec/handler is
migrated onto the mechanism.
- Did-you-mean suggestions on unknown commands and unknown flags, ranked
by edit distance over the live registry, surfaced in both stderr and
--json error.data (reusing the existing nextSteps channel).
- `orca agent-context [--json]`: a versioned, machine-readable dump of the
command schema. Pure local read (no RPC), so it works over SSH and when
the app isn't running.
- CI guards: specs<->handlers parity, and a vocabulary policy that fails
on new off-policy deletion/read verbs (existing ones grandfathered).
* Address PR review feedback (#6303)
- agent-context now emits each command's effective flag set (globals +
conditional --page), not just allowedFlags, so the schema no longer
under-reports --json/--help. Shared as effectiveAllowedFlags() between
validation and the schema.
- Collision check now covers alias paths too, so a duplicate alias that
would silently shadow a real command fails the build.
* fix(cli): harden agent recovery and introspection
Co-authored-by: Orca <help@stably.ai>
---------
Co-authored-by: Jinwoo-H <jinwoo0825@gmail.com>
Co-authored-by: Orca <help@stably.ai>
JinjingandOrca
