v0.1.19

[0.1.19] - 2026-09-29

Versions 0.1.17 and 0.1.18 were numbered in the repository but never published;
everything they held is here.

Added

• networks.create_workspace() now accepts devices, cloud_accounts, vcenters and
  omissions, matching the full POST /networks/{networkId}/workspaces request body.
  Previously only name and retention_days were exposed, so every workspace forked the
  parent network's entire device set -- there was no way to scope a workspace to the
  handful of devices actually in a change window, which is the feature's whole point.
• client.collectors.create_collector_account(body={"collectorName": ...}) and
  .delete_collector_account(collector_name=...), unpublished: registers/removes a
  collector account for the org -- the same call Settings > Collectors > Add collector
  makes in the UI. There was previously no way to provision a collector from the SDK at
  all; a first-time registration required a human in the UI. The response's
  authorizationKey is the one-time install key (username:password when the org is on
  the V2 auth-key format, a base64 protobuf blob on V3) and cannot be fetched again after
  creation.
• client.collectors.upgrade_collector(collector_name=...) and
  .upgrade_collectors(), unpublished: start a remote upgrade of one collector, or of
  every online collector in the org that supports remote upgrade and auto-start (the
  latter returns the names it queued). Needs MANAGE_COLLECTORS. The single form
  refuses a collector it cannot upgrade -- 400 when it is offline, never connected, not
  remotely upgradeable or already current, 409 when an upgrade is already running --
  while the org-wide form skips those and upgrades the rest.
• client.client_software.download(destination, type="HEADLESS_LINUX") and
  get_client_package(type), unpublished: download a Forward client package -- the
  headless collector above all -- as the signed-in user, so a network-scoped API token
  can fetch its own collector instead of a person clicking through Software Central.
  download saves to a file (or into a directory under the server's file name, e.g.
  fwd-unix-26.9.0-18.tar.gz), writes through a .part file so a failed download never
  leaves a truncated archive, and returns a ClientPackage with the name, size and
  SHA-256 of what it saved. It reports the release it got rather than checking a pinned
  one: the package is whatever the server runs. fwd.app serves every type; an on-prem
  appserver serves only the Windows and Linux installers and 400s the headless ones.
• snapshots.invalidate(snapshot_id, reprocess=...) and snapshots.reprocess(snapshot_id),
  unpublished: rebuild a snapshot's model in place from its collected data under the
  org's current settings, the way a parser or modelling change (a new data-model
  feature, a parsing mode) reaches data already collected. The snapshot keeps its id
  and cannot answer queries until processing finishes; both return a
  SnapshotStateChange and do not wait, so follow with wait_until_processed().
• nqe.repo.bundle(entry, commit_id=..., overrides=...) and the pure
  forward_sdk.nqe.bundle_sources(): run a library query with some of its imported
  modules replaced, before committing anything. Every other module is read at the
  pinned commit and the whole closure is merged into one self-contained inline
  query, with each module's top-level names prefixed so private helpers that share a
  name stay distinct. Inline source could not do this before: Forward resolves an
  inline query's imports against head (sync) or not at all (async), with no way to
  pin them. Anything the token-level rewrite cannot rename safely raises
  NqeBundleError. Found when a hierarchy fix three imports deep had to be
  performance-tested against its baseline without committing it; a bundle of the
  unchanged commit returned rows byte-identical to that commit run by ID.
• busy_timeout on ai.start(), ai.ask() and a conversation's ask() /
  ask_and_wait(). Forward answers one question per user at a time and refuses the
  next with a 429 ("Please wait until 'conversation-212' is done processing"); the
  question was never accepted, so it is safe to repeat, but the transport retries no
  POST. With busy_timeout the SDK waits its turn (every five seconds, or after
  Retry-After) instead of raising ForwardRateLimitError. Default 0: unchanged.
  Found when a webhook's remediation chat landed while a research chat was thinking.

Fixed

• FORWARD_VERIFY_TLS now also takes the path to a CA bundle file or directory,
  which is what a self-hosted deployment with a private CA needs; it was read as a
  boolean only, so a path silently meant "verify against the system trust store"
  and failed on the private root. true/false (and 1/0, yes/no,
  on/off) keep their meaning; any other value must be an existing path, or
  from_env() raises ForwardConfigurationError instead of guessing. A CA path
  passed as verify= is checked for existence at construction too.
• A CA path in verify is handed to httpx as an ssl.SSLContext, so it no longer
  raises httpx 0.28's `verify=<str>` is deprecated warning.