Skip to content

feat(tools): per-call approval predicates + tool-part lifecycle states (inspired by CF agents-starter) #26

Description

@drewstone

From reviewing cloudflare/agents-starter#154: the AI SDK models human-in-the-loop approval as states of the persisted tool-call part itself (input-available → approval-requested → output-available | output-denied | output-error) with a per-tool needsApproval: async (input) => boolean predicate.

Two adoptable pieces for agent-app, done compliance-grade (theirs is UX-grade — client-trusted approval, no audit trail):

  1. Per-call approval predicates. Today taxonomy.regulatedTypes is a static set. Add an optional needsApproval?: (args, ctx) => boolean | Promise<boolean> seam per proposal type so products can gate by cost threshold, environment, or first-use rather than always/never. Keep fail-closed: predicate errors → approval required.
  2. Normalized tool-part lifecycle states in LoopEvents + persisted messages. Emit tool-call state transitions (input-available, approval-requested, output-available, output-denied, output-error) so UIs render one switch over part state and approvals survive reload because they live in message history — no separate pending-approvals store to reconcile.

Unlike the starter, approval responses must be recorded server-side (who/when/decision in the proposals table) and the predicate re-checked at execution time, never trusted from the client.

🤖 Generated with Claude Code

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions