Skip to content

python-sdk-v0.12.0

Choose a tag to compare

@github-actions github-actions released this 22 Aug 13:56
· 60 commits to main since this release

0.12.0 — 2026-08-22

Enforcement and exactly-once correctness. The ADK's governance knobs were
enforceable on paper but reachable around in practice; this closes that, and
fixes three defaults that quietly meant the opposite of what they read.

Fixed

  • Tool policy and approval are now enforced on ADK agent dispatch. blocked_tools
    and require_approval_for were evaluated on ordinary tool nodes but not on the
    agent's own dispatch node, so an ADK agent's tool calls bypassed both. An approval
    is bound to a content hash of (name, arguments), so it cannot be replayed against
    a different payload.
  • A spent budget refuses to fire. Previously a run that had already exhausted its
    budget could still make one more model call.
  • max_cost_usd no longer defaults to 1.0. The default doubled as the "unset"
    sentinel, so asking for a $1.00 ceiling was indistinguishable from asking for
    nothing and was silently dropped. The default is now None — no implicit ceiling —
    and the governance knobs track explicitly-set values separately.
  • The model allowlist globs in-process, matching the durable engine. ["*"] — the
    natural way to write "allow everything" — denied every model in-process while being
    accepted durable-side, and a provider-only entry like ["anthropic"] was the mirror
    image. Both transports now share one matcher, pinned by a parity table.
  • Team no longer overrides a governance value a sub-agent set explicitly. It
    compared values rather than tracking what was set, so a sub-agent that deliberately
    chose the coordinator's default had that choice overwritten.

Added

  • External tool effects carry an idempotency key. complete_work_item accepts and
    echoes the key the claim hands out, so the engine records the result and a re-run
    replays it instead of firing the tool a second time. Without it nothing was recorded
    and every replay re-fired — on the transport ADK agents actually take, since
    python_tool nodes run with no in-process worker.