fix(compose): bind output volume to host path and pre-create it

cli/src/run.rs

@@ -253,6 +253,20 @@ fn run_compose(
         return Ok(());
     }
 
+    // `services.yaml`'s `output` volume binds to `./output/{benchmark}/{task}`
+    // (compose/RULES.md rule 18) via a `driver_opts.device:` path — unlike a
+    // short-syntax host bind, that form does not auto-create the directory,
+    // so pre-create it here (as the invoking user, so the agent's uid-1002
+    // process can still write into it — Docker would otherwise make it
+    // root-owned on first mount).
+    let task_id = envs
+        .iter()
+        .find(|(k, _)| *k == "EVAL_TASK_ID")
+        .map(|(_, v)| v.as_str())
+        .unwrap_or("0");
+    std::fs::create_dir_all(format!("output/{benchmark}/{task_id}"))
+        .map_err(|e| format!("failed to create host output dir: {e}"))?;
+
     let mut cmd = Command::new("docker");
     cmd.arg("compose").arg("-f").arg(&compose_ref);
     // `-y`: a published `oci://` stack prompts to confirm (and echoes) the

containers/compose/services.yaml

@@ -55,4 +55,13 @@ networks:
   upstream: {}
 
 volumes:
+  # Bind to the host so results land at ./output/{benchmark}/{task-id}/
+  # (compose/RULES.md rule 18) instead of an anonymous managed volume with no
+  # host-visible path. `${PWD}` resolves to the CLI's invocation directory,
+  # not this file's directory, regardless of `include:` nesting.
   output:
+    driver: local
+    driver_opts:
+      type: none
+      o: bind
+      device: ${PWD}/output/${EVAL_BENCHMARK:-aime}/${EVAL_TASK_ID:-0}