diff --git a/CHANGELOG.md b/CHANGELOG.md
index 245fdc9..c7f0916 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,13 @@
+## [Unreleased]
+### Fixed
+- Generation speed counts the time a model spends writing a tool call's
+  arguments. Those tokens were counted but their streaming time was not, so
+  a step that wrote a file showed several times its real speed (3,672 tokens
+  "at 162.9 tok/s" against the engine's 36.4), and the Session average was
+  inflated with it. Affects OpenCode's figures, not an engine's own. On
+  MTPLX, a turn that wrote a file now gives 36.4 tok/s from OpenCode's
+  figures against the engine's 36.1.
+
 ## [0.3.2] – 2026-09-25
 ### Fixed
 - A reply is recorded when it ends, not when OpenCode's execution does. A
diff --git a/tui.tsx b/tui.tsx
index 88f9595..73cc6bd 100644
--- a/tui.tsx
+++ b/tui.tsx
@@ -1094,7 +1094,17 @@ export default Plugin.define({
       off.push(
         ctx.data.on("session.step.streamed", (evt) => {
           const id = stepID(evt)
-          if (id) readStep(id, "step.streamed")
+          if (!id) return
+          if (HUD_DEBUG) {
+            const t = turns.get(id)
+            const now = Date.now()
+            dbg(
+              `  stream window ${id}: ${t?.firstAt !== undefined && t.lastAt !== undefined ? ((t.lastAt - t.firstAt) / 1000).toFixed(2) : "?"}s; ` +
+                `last mark ${t?.lastAt !== undefined ? now - t.lastAt : "?"}ms before step.streamed; tool deltas ${toolDeltas.get(id) ?? 0}`
+            )
+            toolDeltas.delete(id)
+          }
+          readStep(id, "step.streamed")
         })
       )
       off.push(
@@ -1105,6 +1115,26 @@ export default Plugin.define({
       )
       off.push(ctx.data.on("session.text.delta", mark))
       off.push(ctx.data.on("session.reasoning.delta", mark))
+      // A tool call's arguments are generated tokens too, counted in the
+      // step's output. Unwatched, a step that wrote a file had its tokens
+      // divided by only its text's streaming time (measured: 3,672 tokens "at
+      // 162.9 tok/s" on a `write` step, against the engine's 36.4 for the
+      // turn), inflating the session's speed. The argument deltas never
+      // reach a plugin (measured: 0 on every step of an 8-step turn that
+      // wrote a file); the start and end of the arguments do, and with them
+      // the window's last mark lands 0-1ms before the stream ends. Session
+      // 36.4 tok/s against the engine's 36.1, from 55.1 against 36.4. The
+      // delta subscription is kept in case a later OpenCode forwards them.
+      const toolDeltas = new Map<string, number>()
+      off.push(
+        ctx.data.on("session.tool.input.delta", (evt) => {
+          mark(evt)
+          const id = stepID(evt)
+          if (id) toolDeltas.set(id, (toolDeltas.get(id) ?? 0) + 1)
+        })
+      )
+      off.push(ctx.data.on("session.tool.input.started", mark))
+      off.push(ctx.data.on("session.tool.input.ended", mark))
 
       // Diagnostics only (OPENCODE_HUD_DEBUG): the per-step events, to design
       // reading the engine once per step instead of once per turn. On