DuoPort Connector for OpenCode

Changelog

0.1.9

• Fixed: three Zen MiniMax models (minimax-m3, minimax-m2.7, minimax-m2.5) were allowlisted but blocked as unsupported, hiding them from the picker and failing generation. They match OpenCode’s published Zen chat/completions endpoint table and are now routed as chat.
• Changed: allowlisted models whose chat/completions family is not evidenced in OpenCode’s published endpoint/pricing tables are now recorded as verification-required and stay hidden until a chat/completions probe passes (Go: glm-5.1, glm-5, kimi-k2.5, mimo-v2-pro, mimo-v2-omni, hy3-preview). Default-visible counts are now Go: 11, Zen: 15.
• Removed: mimo-v2.5-free and deepseek-v4-flash-free left the Zen allowlist and free set — neither ID is present in the live Zen /models catalog anymore. big-pickle keeps its reviewed free label: it is billed Free in OpenCode’s published Zen pricing table (https://opencode.ai/docs/zen/#pricing).
• Fixed: free-model labels are now per-catalog, so a free-named ID in one catalog can never mislabel the other catalog’s picker.
• Fixed: network uninstall now sweeps the settings row and plugin transients on every subsite (capped), instead of only the current site.
• Improved: the reviewed-model counts in the documentation are now covered by a test that reads them from the registry, so they cannot drift again.

0.1.8

• Fixed: rotating a saved key did not clear its cached availability state.
  WordPress passes the option name as the THIRD argument to
  update_option_{$option} but as the FIRST to add_option_ and
  delete_option_. The cleanup hook read the first argument, so on a key
  CHANGE it received the previous key text, did not match, and cleared
  nothing. A rotated key could keep its last-known-good state for up to 30
  days. Removal and first-time setup were never affected.
• Fixed: an error naming the credential only in its message text, with no
  machine-readable code, could be read as the model was retired” and leave a
  revoked key marked valid.
• Fixed: a provider error naming PROBE_MODEL in its message text is now treated
  as telling us nothing about the key, instead of being recorded as a fresh
  result and overwriting what was known before. No model has been retired.
  PROBE_MODEL is still sent on every probe; this release only adds handling for
  a future retirement.
• Behaviour to be aware of: when the gateway answers with a response this plugin
  has no rule for, the dashboard keeps reporting the last successful check
  (connected”) instead of flipping to not connected”, so an unrecognised
  reply or outage does not disconnect a working key. That stored value is
  written only by a successful check and lasts up to 30 days from that check
  (each new success extends it), plus roughly five minutes of cached probe
  result before a new probe is sent, so a change can lag behind the gateway
  by that much. Only a recognised credential answer (an invalid or missing
  key) clears the stored value; a failure naming the probed model leaves it
  alone too, and a catalog that never checked successfully still reads not
  connected”. Changing or re-entering the key for that catalog on
  Settings → Connectors clears the stored state; otherwise wait for a later
  probe to return a definitive answer.

0.1.7

• Fixed: credential verification now detects the WordPress AI Client HTTP method enum correctly, so valid keys are no longer treated as unverifiable because of a dynamic SDK method.
• Improved: temporary server, transport, and rate-limit outcomes preserve the last known good configuration state; Zen free-tier limits are still recognized as a valid key.
• Fixed: malformed model-list entries with non-scalar IDs are skipped safely instead of causing a fatal error.
• Fixed: endpoint-family and catalog identifiers now use shared constants in the registry, metadata, and Model Radar paths.
• CI: release verification now runs the test suite on PHP 8.2, 8.3, 8.4, and 8.5.

0.1.6

• Added two real WordPress.org screenshots for the Connectors and DuoPort Settings screens.
• Expanded the listing with factual Go + Zen, free-model, verified-support, and WordPress AI Client sections.
• No model registry, transport, capability, or credential behavior changed.

0.1.5

• Centralized Go session and client User-Agent headers across text and image requests.
• Added an authoritative architecture baseline, class inventory, dependency graph, boundaries, scope, improvement rules, and single-item improvement queue.
• Hardened release documentation so post-0.1.4 additions carry accurate @since metadata.

0.1.4

• Added deepseek-v4.1-flash to the Go catalog.
• Fixed: the key validation probe now sends the required x-opencode-session header and probes an authentication-discriminating model, so valid Go/Zen keys are accepted instead of rejected.
• Added image-generation model path: image-capable models advertise the imageGeneration capability through new Go/Zen image models behind a per-catalog allowlist. Ships with no image IDs allowlisted yet, so the path stays inert and text generation is unaffected. Also added a Media Library saver with MIME, size, and capability guards.

0.1.3

• Updated Zen model catalog: removed hy3-free and laguna-s-2.1-free, which OpenCode retired from the API.

0.1.2

• Fixed: Go and Zen now use separate API key fields on the Connectors screen. The previous shared-key setup made WordPress reject valid keys on save (It was not possible to connect to the provider using this key”). If you connected 0.1.1, please re-enter your key in both fields after updating.

0.1.1

• Renamed plugin to DuoPort Connector for OpenCode; Go and Zen now share a single API key field on the Connectors screen; connecting the providers now requires a one-click approval consent in the Connectors screen.

0.1.0

• Initial release: Go and Zen providers, allowlisted models with free labels, probe availability, shared-key sync, and Show all models toggle.