diff --git a/TODO.md b/TODO.md
index ab249fe..1f9dcde 100644
--- a/TODO.md
+++ b/TODO.md
@@ -5,19 +5,6 @@ one in place when it turns out to need a decision.
 
 ## App — transcript
 
-- [ ] Text inside code blocks does not highlight when selected. **Measured, and
-      it does** — the selection is drawn, but over the near-black surface a code
-      block and a tool's output sit on, Material's default 40%-alpha tint
-      composites to a barely-there smudge, much weaker than the same selection
-      over a reply. The app now states its own selection colours
-      (`AiAppSelectionColors`), which took the fill from #5B4C73 to #776394 on
-      that surface. Worth confirming this was the complaint rather than a
-      selection that draws *nothing* on the phone.
-- [ ] Text inside an opened peer message or memory note cannot be selected at
-      all — the heading of the same card can, and so can a tool call's output,
-      so it is the markdown text specifically. Pre-existing (measured against
-      the build before this session's changes, by stashing them). It
-      contradicts AGENTS.md's "all transcript text is selectable".
 - [ ] Messages received from other agents are inconsistent — sometimes they
       appear, sometimes they don't. **Needs a rig.** Read the code rather than
       measured: a live Claude session only learns of a peer message from the
@@ -32,10 +19,17 @@ one in place when it turns out to need a decision.
 ## Session settings
 
 - [ ] Autocompact belongs in session settings; empty disables it, which is the
-      default. **Needs a decision before building** — nothing called autocompact
-      exists yet on either side. `PLAN.md` has it only as a planned pi-driver
-      feature (`set_auto_compaction`), and Claude Code runs its own. So this is
-      a new server feature, and the open questions are what the empty-or-not
-      value *is* (a token count? a percentage of the context window?) and which
-      drivers it applies to.
+      default. Iris chose "hand it to the driver" — only where a driver has
+      auto-compaction of its own. **That option was offered on a false premise
+      and is not buildable yet.** It named pi's `set_auto_compaction`, but pi
+      was never built as a driver here: `session/llama.rs` talks to
+      `llama-server`'s OpenAI-compatible endpoint directly, and its `compact()`
+      refuses outright. Claude Code's auto-compaction is the CLI's own and
+      nothing in the stream-json control protocol this app uses configures it.
+      So the setting would be stored, passed to a driver, refused by every one
+      of them, and the field would never appear on any session. What is needed
+      first is either a driver that can take it, or a different rule — the
+      server watching `contextTokens` and running `/compact` itself is the one
+      that would work today, for Claude sessions, and it is the option that was
+      not chosen.
 
diff --git a/app/androidApp/src/main/kotlin/com/example/aiapp/Markdown.kt b/app/androidApp/src/main/kotlin/com/example/aiapp/Markdown.kt
index 82ad370..20a1b80 100644
--- a/app/androidApp/src/main/kotlin/com/example/aiapp/Markdown.kt
+++ b/app/androidApp/src/main/kotlin/com/example/aiapp/Markdown.kt
@@ -368,7 +368,15 @@ private fun MarkdownRoot(
                 // exactly a card's own fill, so a fenced block inside a tool call had no
                 // background at all and one in a reply read as a step *up* out of the page.
                 codeBackground = rawSurface,
-                inlineCodeBackground = rawSurface,
+                // The same colour, but let through. An inline span's background is part of the
+                // *text's* own drawing and the selection rectangle is drawn underneath it, so an
+                // opaque chip hides the selection completely: selecting a sentence highlighted
+                // every word of it except the ones in backticks, which is a difference in
+                // appearance the reader has no way to account for. Translucent, the selection
+                // shows through and the chip still reads as one step down from the page -- there
+                // is no way to draw it over the selection instead, since the order is the text
+                // node's.
+                inlineCodeBackground = rawSurface.copy(alpha = INLINE_CODE_ALPHA),
                 // The same tint a code block gets, rather than the renderer's 2%-alpha default:
                 // two adjacent tints that differ by a fiftieth read as one flat block on a phone,
                 // so the table would have had a border-less grid and nothing saying where it began.
@@ -711,3 +719,12 @@ class ParsedReplies {
         ready.clear()
     }
 }
+
+/**
+ * How much of the inline-code chip's fill is its own colour, the rest being whatever it sits on.
+ *
+ * High enough that the chip is still a clear step down from the page, low enough that a selection
+ * under it changes what the chip looks like. Both halves are the point: at 1.0 the chip was the
+ * only part of a selected sentence that did not look selected.
+ */
+private const val INLINE_CODE_ALPHA = 0.6f