ModelDock v0.3.84

Full Changelog: v0.3.83...v0.3.84

ModelDock v0.3.83

Full Changelog: v0.3.82...v0.3.83

ModelDock v0.3.82

Full Changelog: v0.3.81...v0.3.82

ModelDock v0.3.81

Full Changelog: v0.3.80...v0.3.81

ModelDock v0.3.80

Full Changelog: v0.3.79...v0.3.80

ModelDock v0.3.79

Full Changelog: v0.3.78...v0.3.79

ModelDock v0.3.78

Full Changelog: v0.3.77...v0.3.78

ModelDock v0.3.77

Parallel tool results with promoted images

• Keep a complete parallel tool-result group together before its promoted image
  messages. Promoted images are now held until the current contiguous run of tool
  results ends, so every result of a group stays directly after its calls.
• Previously each image-bearing result inserted its promoted image message
  immediately, which could place a message between two results of the same parallel
  group. Providers that require every call in a group to be answered before the next
  message rejected the request instead of completing it.
• The change introduces no new call/result matcher: canonical normalization already
  places each group's results contiguously, so the promoted messages only wait for
  that run to end.

Verification

• Real provider A/B: the previous ordering returned HTTP 400 No tool output found;
  the fixed ordering returned HTTP 200 with response.completed in about 3 seconds
  (20,702 input tokens, 288 output tokens).
• The A/B case is a sanitized historical fixture plus a reconstructed transition, not
  a capture of the complete original incident.
• Regression coverage in this release exercises mixed function/custom groups, an error
  result inside a group, image histories, and the routed OpenCode Go full-fixture
  request, asserting that no message splits a parallel group from its results.

GitHub CI, the manual install gate, and post-release Windows installation verification
passed. All 12 checksummed release files were downloaded and verified. The local full
suite rerun was stopped at the user's request; release gates ran on GitHub.

The fix ships in the gateway bundle; installing or updating ModelDock applies it.

ModelDock v0.3.76

Vision capabilities

• Declare DeepSeek V4.1 Flash as vision-capable when discovered through OpenCode Go or DeepSeek, and recognize Command Code's V4.1 Flash vision model.
• Make the Models page Vision column editable for remote provider models. Corrections are saved by provider-qualified model identity and survive refreshes, restarts, and upgrades.
• Apply saved corrections to the shared model entries used by the catalog, vision picker, and gateway routing. Native and local capabilities remain controlled by their runtime.
• Prevent disabling vision on the currently selected vision helper until another helper is selected.

Restart Codex after changing a capability to reload its model declarations.

Validation for this release runs in the existing GitHub CI and release-install workflows. No new tests or local test/preview runs were performed for this change.

ModelDock v0.3.75

OpenCode Go automatic model discovery

• Fix a model-directory lookup failure that prevented newly published OpenCode Go models from appearing in ModelDock and the Codex catalog.
• Provider-wide discovery now resolves the Go directory without requiring a model argument. Model-specific Zen/free-tier routing is unchanged.
• Retain automatic discovery at startup and every 24 hours, with no inference probes or per-model whitelist additions.

Verification

• A real authenticated Go directory request returned HTTP 200 and discovered deepseek-v4.1-flash.
• The regression fails against the original v0.3.74 bundle and passes against the corrected bundle, covering discovery, the dashboard picker, the Codex catalog, and full-history routing to the discovered model.
• Local full regression suite: 1,146 passed, 7 skipped, 0 failed.

After updating ModelDock, restart Codex to reload its model catalog.