diff --git a/lib/stack/ai/client.py b/lib/stack/ai/client.py
index d539f16..52bd570 100644
--- a/lib/stack/ai/client.py
+++ b/lib/stack/ai/client.py
@@ -190,15 +190,24 @@ def from_env(cls, *, namespace: str | None = None,
                 "No AI endpoint configured — set up AI with 'stack up ai'"
             )
         key = os.environ.get("OPENAI_KEY", "") or "not-needed"
-        # The SDK default timeout is 600s — a hung local endpoint would
-        # freeze a caller for ten minutes. 300s is the compromise: a
-        # 30K-token prefill on a local model (a scanned multi-page PDF
-        # going into a vision classify) needs several minutes before the
-        # first token, and 120s cancelled those mid-prefill; past 300s
-        # it is a stuck server and should fail loudly.
+        # This is a read timeout, and the call is not streamed, so it
+        # covers the entire silent wait while the model reads a document
+        # and generates an answer. Prefill alone, measured against local
+        # oMLX (Qwen3.5-9B-MLX-4bit) on a dense 14-page PDF:
+        #
+        #     1,000 tokens ->   7.6s      15,000 tokens ->  58.4s
+        #     3,000 tokens ->  20.4s      28,400 tokens -> 163.6s
+        #     7,500 tokens ->  36.7s
+        #
+        # That machine had RAM to spare. A 16GB Mac running photos and
+        # documents alongside is several times slower, so 120s and then
+        # 300s were both cancelling healthy work partway through a big
+        # scan and discarding every minute already spent on it. 900s
+        # covers ~158K tokens here, past any context this serves, and
+        # still bounds a genuinely dead endpoint.
         client = AsyncOpenAI(
             base_url=url, api_key=key, max_retries=max_retries,
-            timeout=300.0,
+            timeout=900.0,
         )
         return cls(client, namespace=namespace, capabilities=capabilities)
 
diff --git a/stacklets/docs/bot/archivist.py b/stacklets/docs/bot/archivist.py
index 943f840..f7c11c3 100644
--- a/stacklets/docs/bot/archivist.py
+++ b/stacklets/docs/bot/archivist.py
@@ -253,15 +253,23 @@ class ArchivistBot(MicroBot):
 
     # Filing is the slowest thing any famstack bot does, and the framework
     # default (180s) is a chat-bot budget, not a document-pipeline one. One
-    # archived email costs, in sequence: the Paperless upload + OCR wait
+    # archived document costs, in sequence: the Paperless upload + OCR wait
     # (up to 120s), a vision classify pass, a layout reformat pass, and the
-    # entity-enrichment pass, each a full LLM call whose prefill on a local
-    # model can run into minutes for a long scan. Cancelling at 180s threw
-    # away work that was still progressing. Eight minutes is a stuck-handler
-    # guard, not a latency target: the pipeline's own per-step timeouts bound
-    # normal slowness, and a handler that blows this budget now fails
-    # visibly (❌ + a threaded notice) and can be retried with 🔁.
-    HANDLER_TIMEOUT_SECONDS = 480
+    # entity-enrichment pass — three separate LLM calls, each paying its own
+    # prefill before it emits a token.
+    #
+    # Those prefills are the whole budget. Measured on a dense 14-page PDF
+    # against local oMLX, one 28K-token pass took 163s on a machine with RAM
+    # to spare; a 16GB Mac running photos and documents alongside is several
+    # times slower. Three such calls plus the OCR wait clears eight minutes
+    # comfortably, which is why 480s still cancelled long documents and threw
+    # the work away.
+    #
+    # 30 minutes is deliberately far past any healthy filing. It is a
+    # stuck-handler guard, not a latency target: the pipeline's own per-step
+    # timeouts bound normal slowness, and a handler that does blow this now
+    # fails visibly (❌ + a threaded notice) and can be retried with 🔁.
+    HANDLER_TIMEOUT_SECONDS = 1800
 
     def __init__(self, homeserver, user_id, password, session_dir, **settings):
         super().__init__(homeserver, user_id, password, session_dir, **settings)