diff --git a/server/src/session/claude.rs b/server/src/session/claude.rs
index b118d01..a459e0d 100644
--- a/server/src/session/claude.rs
+++ b/server/src/session/claude.rs
@@ -172,6 +172,25 @@ impl Queue {
 /// the attach path must agree about which file is which; if they drift,
 /// a reattached session reads a file nothing is writing and simply looks
 /// idle forever.
+/// The context size the CLI compacts at, rather than letting it choose.
+///
+/// Left to `auto` a phone session drifts: these run for hours, nobody
+/// closes them, and the transcript carries screenshots. One measured here
+/// reached 491,562 tokens before the CLI compacted it, and *every* API
+/// call re-reads the whole context -- so at that size a single tool call
+/// bills about 49k tokens before it does anything, and one ordinary
+/// request ("can you make it so you can rename a session?") cost 4.2
+/// million across its 130 calls.
+///
+/// 100k is the smallest window the CLI accepts and roughly the cheapest:
+/// per-call cost falls with the cap, while the compaction it forces costs
+/// about the same in total either way -- a smaller window compacts more
+/// often but each pass is proportionally smaller. What it trades is how
+/// much detail survives a compaction, which is a real cost to the work
+/// and the reason this is one named constant rather than a computed
+/// value. Raise it if sessions start losing the thread.
+const AUTOCOMPACT_WINDOW: &str = "100k";
+
 const STDIN_FIFO: &str = "stdin.fifo";
 const STDOUT_LOG: &str = "stdout.log";
 const STDERR_LOG: &str = "stderr.log";
@@ -327,6 +346,7 @@ impl ClaudeDriver {
         };
         push("--input-format", "stream-json");
         push("--output-format", "stream-json");
+        push("--autocompact", AUTOCOMPACT_WINDOW);
         // Hidden but load-bearing: without it the CLI resolves permissions
         // itself and nothing ever reaches the phone.
         push("--permission-prompt-tool", "stdio");