diff --git a/AGENTS.md b/AGENTS.md
index 4211ce8..23fd700 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -424,8 +424,9 @@ When a deploy fails, the deploy manager diagnoses it from the logs of that
 deploy, which workflow code gives it. `fetchDeployLogs()` reads them in the
 time range of the deploy, with no type filter. Thus it gets the build logs,
 the output of the pre-deploy command, and the logs of the new instance, and
-no line of an earlier deploy. A type filter can lose the pre-deploy output,
-because the Render documentation does not give its type. Each round reads
+no line of an earlier deploy. Do not add a type filter. The logs API gives
+the output of the pre-deploy command as `app` logs, not `build` logs, so a
+`build` filter loses it. With no filter, the API gives both. Each round reads
 the deploy that failed in that round. Do not give an agent `list_logs` in
 place of this read: an agent can read the logs of each service in the
 workspace, and the factory cannot redact them first.
diff --git a/app/render.ts b/app/render.ts
index 85ef8dd..70643c8 100644
--- a/app/render.ts
+++ b/app/render.ts
@@ -202,7 +202,8 @@ async function latestDeploy(
  * the caller must redact them before they go into a task input.
  *
  * The read has no type filter, so it gets the build, the pre-deploy command,
- * and the new instance. The time range of the deploy keeps out the lines of
+ * and the new instance. The API gives the pre-deploy output as app logs, so a
+ * build filter loses it. The time range of the deploy keeps out the lines of
  * an earlier deploy. It starts at createdAt: the API does not document
  * startedAt. The logs are only diagnostic, so a read that cannot finish gives
  * no logs and does not fail the run.