v0.4.1

aidress-mcp 0.4.1 · aidress-sdk 0.4.1 · langchain-aidress 0.2.0 (unchanged)

Two client-side fixes. Upgrade if you call review() or check register()'s result.

Fixed

• review() silently did nothing when called without both party ids. /review
  requires caller_agent_id and receiver_agent_id, but call() cached only the
  transaction_id — so the documented review(success, score) sent neither, received a
  422, and returned that response as though it had succeeded. No review was recorded and
  no error was raised. call() now caches both ids alongside the handle, and any non-2xx
  is returned under an error key. With no prior call() you get an explicit message
  naming what is missing.

  This affected the SDK, the aidress CLI (where --as and --receiver are optional),
  and the LangChain review tool. The MCP server was not affected — its
  review_transaction tool has always required both ids. Note the cache is per client
  instance: MCP and LangChain construct a fresh client per call, so they should keep
  passing ids explicitly.

• register() reported failures as successes. Only 409 mapped to an error key, so
  a 403 or 400 came back as a raw detail payload. Code guarding on
  result.get("error") proceeded and then failed on the absent claim_link. Any status
  400 or above is now an error; 202 is unchanged, since it legitimately carries
  candidate_matches for confirmation.

• examples/quickstart.py: no longer submits a negative review when a call returns 402.
  Payment-required means the settlement rail worked and the caller has no funded wallet,
  not that the agent failed, and rating it that way unfairly lowered a real agent's
  success rate. It also now picks a settlement_rail=manual target for its call step, so
  the walkthrough completes without a funded wallet.