zevairouter 1.1.23 → 1.1.25

package/cli/app/CHANGELOG.md

@@ -1,5 +1,18 @@
 # Unreleased
 
+# v1.1.25 (2026-09-18)
+
+## Fixes
+- **Antigravity CLI shows every model tier again while MITM is active.** MITM forced a single hardcoded `ideVersion`/User-Agent (`1.28.0`) on every forwarded request, and CLI 1.2.5 hides whole tiers when it believes it is an older IDE. Measured against a live account: 14 models with the override off, 8 with it on — the 3.8 and 3.7 tiers disappeared entirely, which is why they were visible without MITM but missing with it. The catalog response is byte-identical either way, so this was not a catalog problem: the same `tieredModelIds` payload is simply rendered differently by the client. The override is disabled; re-enabling it requires a per-client version the installed CLI actually reports, verified against the rendered model list rather than the catalog payload.
+  This also corrects the reasoning recorded for v1.1.23 and v1.1.24: `gemini-3.8-flash-*` was never dropped by the account. It moved to the `tieredModelIds` scheme, which earlier catalog inspection did not check.
+
+# v1.1.24 (2026-09-18)
+
+## Fixes
+- **Antigravity CLI no longer crashes mid-answer when a request is routed.** The router speaks OpenAI-style SSE, so it appends a `[DONE]` sentinel and may emit a bare `{"error":...}` chunk. Antigravity's language server dereferences `.response` on every chunk without a nil check, so either one segfaulted the CLI (`SIGSEGV` in `getStreamingTextCompletion`) right after the answer began streaming. MITM now forwards only chunks carrying a `response` object and ends the stream cleanly otherwise, which is how Google signals completion and failure on this endpoint. Forwarded chunks are passed through verbatim so opaque fields such as `thoughtSignature` are not perturbed.
+- **A cleanly finished Antigravity stream is no longer accounted as a failure.** Same-format passthrough only recognised the OpenAI terminal marker `choices[].finish_reason`, but a Cloud Code stream carries it as `response.candidates[].finishReason`. Successful answers were therefore recorded as `STREAM_MISSING_TERMINAL` (`FAILED 502`) and fed provider cooldown. Both shapes are now recognised, including a `promptFeedback.blockReason` terminal; a stream that genuinely ends without any terminal still fails.
+- **Reverted catalog injection for models the upstream account no longer offers.** v1.1.23 advertised every mapped model on `:fetchAvailableModels` so dropped tiers stayed selectable. Making an unavailable model selectable does not make it usable: the CLI sends the injected model's backend enum and Google answers `NOT_FOUND (404)` before any chat request is made, which is worse than the model being absent. The catalog is forwarded untouched again. Models the account cannot reach remain available through the router directly (for example `ag/gemini-3.8-flash-high`).
+
 # v1.1.23 (2026-09-18)
 
 ## Fixes

package/cli/app/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zevairouter",
-  "version": "1.1.23",
+  "version": "1.1.25",
   "description": "Self-hosted AI router: many providers, one OpenAI-compatible /v1 endpoint and dashboard",
   "bin": {
     "zevai": "cli/cli.js"