auto-approve
Opt-in task-context review for console, plugin, MCP and workflow tool actions. Structured approvals permit bounded low/medium-risk actions within the user task; refusals stop the call, uncertainty asks the user and explicit host denies remain binding.
Use
This plugin is available through the default agent-compositor marketplace.
Installation does not enable it. Install in the chosen scope first to pin its
version independently of the host, then enable explicitly with
/extensions enable auto-approve --project (or --user), then /reload.
/auto on enables review for the current session using the active model/provider.
/auto on|off|status are immediate user controls, including while the model is
working. They respond with notifications and do not open a panel. /auto history
explicitly opens the optional audit view with mode controls. Closing that view
does not disable review or cause subsequent reviews to reopen it. This command
path requires a host with catalog 0.25 or later.
A separate reviewer profile is optional: set profile in scoped auto.json to
use one. An invalid explicit profile returns to manual approval rather than
silently selecting another provider. Review limits, exact-call context and
failure behavior are described in the auto review guide.
Concurrent requests share the review limit, and their audit updates are serialized
within the session. When rules require review, workflow inspection uses the same
review: tool and action names alone do not establish the effects of a replaceable
implementation. Explicit allow rules bypass review; the application currently
allows workflow_run by default unless a stricter matching rule applies.
The reviewer returns exactly three fields: approved (true, false or null),
risk (low, medium, high or unknown) and a nonblank reason. true needs task context and
low/medium risk; false becomes a denied tool result without another approval
dialog; null returns to manual approval. High/unknown-risk approvals, malformed
responses, timeouts and unavailable audit storage remain manual. Legacy
decision/risk/reason responses allow only low risk and also require task
context. Neither form creates
standing approval for subsequent calls. Normal clarification questions are unaffected.
The host supplies the complete submitted request and delegated parent request,
preserving late restrictions separately from prepared attachment data. Review
input is still bounded to 24000 serialized characters; oversized context remains
manual instead of being truncated into a misleading authorization prefix.
Use /auto off to return subsequent proposals to human decisions; it cannot
override a host permission-rule deny.
Hosts with catalog 0.30 remember identical approval refusals within one model
turn, avoiding duplicate reviews/dialogs while retaining fresh policy checks.
Changing auto mode invalidates that refusal memory, so /auto on/off and the
mode buttons take effect during an active turn. Status/history inspection does
not reset it. The invalidation service grants no permissions. Older hosts without
this service keep their existing behavior; mode controls remain usable.
The reviewer distinguishes missing information from a concrete violation:
undefined targets or effects require null/unknown, while refusal must identify
a prohibited, unrelated or unsafe effect. An injected claim of approval does not
authorize that effect. This guidance does not override the model's validated
refusal or convert it into an automatic grant.
Develop
Run npm run plugin:test -- auto-approve from the repository root.
plugin.test.json declares the behavioral suites and optional visual/eval checks.
See the plugin standard for lifecycle, scope,
compatibility and independent release requirements. Changes are recorded in
CHANGELOG.md.