v0.4.0-beta.4

What's Changed

• fix: use global endpoint for Kimi API Platform by @jepgambardella in #176
• fix: coerce whole-number tool arguments for Codex integer schemas by @Zzy-min in #178
• fix: preserve account native model catalog by @wousp112 in #169
• fix: normalize tool schemas for strict upstreams and stop reporting 401s as rate limits by @duolahypercho in #187
• feat: split the mainland China Moonshot platform into its own provider by @duolahypercho in #189
• fix: make the Dynamic Island opt-in and stop duplicate overlays by @duolahypercho in #188
• fix: resolve and spawn commands the way Windows requires by @duolahypercho in #186
• fix: make the missing-Codex test actually exercise the missing case by @duolahypercho in #190
• feat: surface tool-result aging savings and extend aging to native GPT by @duolahypercho in #185
• Fix observed token rate calculation by @duolahypercho in #192
• feat: Simplified Chinese, and a tray language setting to choose it by @duolahypercho in #194
• fix: measure output speed after the first token, not after the headers by @duolahypercho in #195
• fix: stop one broken sample from carrying a model's speed, add per-model rates, default compaction off by @duolahypercho in #196
• feat: catch Command Code's registry up, and fix one dead Messages route by @duolahypercho in #199
• fix: make gpt-5.6-luna usable as a v2 subagent without a mode switch by @Guojiz in #197
• feat: publish routed models into DeepSeek Harness as a second target by @duolahypercho in #193
• feat: operate DeepSeek Harness from the tray, and let it use the Codex session by @duolahypercho in #198
• docs: say that native models publish, because they do now by @duolahypercho in #203
• feat: ship GLM-5.3, and split Z.ai's metered platform from its plan by @duolahypercho in #201
• fix: release the empty-completion hold when the upstream proves it is generating by @duolahypercho in #205
• feat: give Model Router an app icon, and open it before Codex needs it by @duolahypercho in #204
• feat: add anonymous free model gateways by @Yuxin-Qiao in #184
• docs: record the empty-completion hold release in the changelog by @duolahypercho in #218
• docs: check bin scripts by language by @jepgambardella in #215
• chore: remove dead catalog helper by @jepgambardella in #214
• refactor: share private JSON writes by @jepgambardella in #213
• perf: start forwarders in parallel by @jepgambardella in #212
• perf: speed up desktop snapshot by @jepgambardella in #211
• force a stable HTTP/1.1 upstream transport by @IndelibleVivi in #217
• preserve user tables inside managed config blocks by @IndelibleVivi in #216
• feat: unify desktop app icons by @ctxnn in #206
• feat: add LM Studio local provider by @jepgambardella in #210
• fix(desktop): make Windows tray control spawns and task status work by @richardotomislav in #209
• feat: group a vendor's providers, and let the tray restart itself by @duolahypercho in #220
• fix: close finished subagents instead of leaving them Working by @duolahypercho in #221
• fix: address reviewer findings from the Aug 14 PR sweep by @duolahypercho in #223
• fix: repair a nullable object root on every tool, not just namespaced ones by @duolahypercho in #222
• feat: research any enabled model for the v2 subagent role automatically by @duolahypercho in #225
• fix: escape interpolated Windows paths in config-manager TOML fixtures by @duolahypercho in #226
• fix: opencode's DeepSeek models never receive a forced tool_choice by @duolahypercho in #227
• fix: a transient probe failure defers verification instead of condemning the model by @duolahypercho in #228
• fix: the tray's subagent rows honor machine-local proofs by @duolahypercho in #229
• fix: an entitlement refusal defers verification instead of condemning the model by @duolahypercho in #230
• fix: five more opencode models never receive a forced tool_choice by @duolahypercho in #231
• fix: an unreachable probe defers verification instead of condemning the model by @duolahypercho in #232
• feat: credential-free install mode via --no-provider and --no-discovery by @duolahypercho in #233
• fix: the account catalog is an account read, and a flaked config read is not an uninstall by @duolahypercho in #234
• feat: offer the Qwen3.8 family in the local Ollama catalog by @duolahypercho in #236
• fix(catalog): preserve native GPT metadata by @jepgambardella in #235
• feat: quota-reset rows with countdowns; graduate tool-result compaction by @duolahypercho in #237
• feat: tray Restart button rebuilds backend and tray, then starts them by @duolahypercho in #238
• Ship desktop companion and complete Windows tray parity by @duolahypercho in #202

New Contributors

• @Zzy-min made their first contribution in #178
• @wousp112 made their first contribution in #169
• @Yuxin-Qiao made their first contribution in #184
• @IndelibleVivi made their first contribution in #217
• @ctxnn made their first contribution in #206
• @richardotomislav made their first contribution in #209

Full Changelog: v0.4.0-beta.3...v0.4.0-beta.4