@@ -54,7 +54,7 @@ categories = ["network-programming", "web-programming::http-server"]
 
 # 二进制走 CalVer，crate 走 SemVer —— 两者是两回事，卖给不同的人。
 # 这里是 crate 的版本。
-version = "0.75.0"
+version = "0.76.0"
 
 [workspace.dependencies]
 # ── 内部 crate ───────────────────────────────────────────
 
@@ -0,0 +1,16 @@
+This release makes `/live` report how fast running requests are generating right now, so a client can show a live speed during a long request, and names a broken answer body instead of reporting a decoding error.
+
+**Upgrade notes**
+
+- The control-plane protocol version is now 49. ThinkWatch Lite must be upgraded together with core: a Lite release with protocol 48 does not connect to 0.76.0, and the other way round.
+- The request store's schema and the configuration are unchanged; nothing is cleared.
+- For crates that embed the gateway: `Ending::streaming` takes the gateway's `pace` table, `AppState` has a `pace` field, `tw_observe::Meter` and `EventBus::meter` are new, and `tw_engine::TextTokens` is public. The public API of `tw_guard` is unchanged.
+
+**Live generation speed.** `/live` used to give the speed of requests that finished in the last minute, so it had nothing to show while a long request ran.
+- The gateway keeps reading a successful stream after its first token and estimates the output as it arrives: text, reasoning text and tool arguments, with the same rule as the input estimate.
+- The estimate is multiplied by a ratio per upstream and model. Each streaming request that ends with reported usage updates that ratio (moving average, bounded between 0.5 and 2; skipped when reasoning was hidden). Gemini's cumulative usage in the stream is used when it is higher.
+- `LiveView.tokens_per_sec` is the total of the running requests over the last 3 seconds: null when nothing runs, 0 when requests run but none is generating yet.
+- `RunningView` has `tokens_per_sec` and `output_tokens` for each running request, null before the first token and for non-streaming requests.
+- The speed recorded for a finished request is unchanged and still uses the usage the upstream reports.
+
+**Broken answer bodies.** An upstream that hung up in the middle of an answer was reported as "error decoding response body". Such failures now have their own code, `gw.upstream.body_broken`, with the causes spelled out, and `gw.upstream.forward_failed` carries every cause, with the URL reduced to its origin.