1
00:00:00,580 --> 00:00:05,280
This one's brought to you by Jellypod AI,
and it's about a small option with a big

2
00:00:05,400 --> 00:00:06,140
bill attached.

3
00:00:07,060 --> 00:00:14,020
In Claude Code 2.1.318, dollar sign dot
model dot fork now accepts an

4
00:00:14,080 --> 00:00:14,940
effort parameter.

5
00:00:15,937 --> 00:00:18,917
Right, and to see why that matters, you
need the old behavior.

6
00:00:19,537 --> 00:00:24,197
Fork reruns the main thread's last request
with one more user message and every tool

7
00:00:24,277 --> 00:00:24,677
denied.

8
00:00:25,177 --> 00:00:28,777
The API serves that prefix from the prompt
cache, so it's cheap on input.

9
00:00:29,377 --> 00:00:33,757
But before 2.1.318, the fork inherited
your session's reasoning effort.

10
00:00:34,377 --> 00:00:39,297
So if you'd dialed effort up to high or
max for real work, every little side job

11
00:00:39,357 --> 00:00:40,957
from a mod ran at that same level.

12
00:00:42,098 --> 00:00:44,778
So you're using a sledgehammer to hang a
picture frame.

13
00:00:45,378 --> 00:00:49,938
Things like a commit message, an export
summary, a handoff brief,

14
00:00:50,418 --> 00:00:51,678
or classifying an action.

15
00:00:52,198 --> 00:00:56,118
Tiny jobs, but they'd burn tokens and add
seconds of latency,

16
00:00:56,678 --> 00:01:00,718
especially if the result was feeding a
pane or a command the user was waiting on.

17
00:01:01,732 --> 00:01:03,592
Now you can say what you want per call.

18
00:01:04,112 --> 00:01:07,852
Inside any hook in your register function,
so a command run handler,

19
00:01:07,972 --> 00:01:11,552
a turn complete handler, or a UI action,
you write it like this.

20
00:01:12,372 --> 00:01:17,232
const result equals await dollar sign dot
model dot fork, open brace,

21
00:01:17,772 --> 00:01:22,212
prompt colon, summarize the open questions
and touched files in three bullet points,

22
00:01:22,632 --> 00:01:25,292
comma, effort colon low, close brace.

23
00:01:25,992 --> 00:01:26,292
That's it.

24
00:01:27,195 --> 00:01:29,295
And the result tells you what actually
happened.

25
00:01:29,815 --> 00:01:33,475
The returned object reports the effort
level that was sent and executed.

26
00:01:34,075 --> 00:01:39,375
So result dot text is your brief, and
result dot effort is your proof.

27
00:01:39,875 --> 00:01:44,095
That's handy if you want to log it, or
assert in a test that your side job really

28
00:01:44,155 --> 00:01:47,075
ran at low and not at whatever the user
had cranked up.

29
00:01:47,665 --> 00:01:51,805
Concretely, think about the ccseats
handoff pattern we covered before,

30
00:01:52,325 --> 00:01:55,305
where a fork writes a brief so another
tool can pick up the work.

31
00:01:55,985 --> 00:01:58,105
That's an extraction and formatting job.

32
00:01:58,545 --> 00:02:01,005
Low effort is plenty, and the pane opens
faster.

33
00:02:01,585 --> 00:02:04,665
Keep high effort for forks where the model
actually has to reason,

34
00:02:04,845 --> 00:02:06,485
like reviewing a design decision.

35
00:02:07,818 --> 00:02:08,738
Few caveats, though.

36
00:02:09,518 --> 00:02:15,418
Mods are early access, so check the types
in the dot claude-plugin slash types

37
00:02:15,538 --> 00:02:19,098
folder for your installed version, because
that's the exact interface.

38
00:02:19,998 --> 00:02:23,898
And caching only helps when you fork from
the current transcript state.

39
00:02:24,718 --> 00:02:29,538
If the cache is cold or expired, the fork
still pays full context ingestion.

40
00:02:30,287 --> 00:02:32,267
And remember mods aren't sandboxed.

41
00:02:32,507 --> 00:02:34,767
They run inside Claude Code as you.

42
00:02:35,367 --> 00:02:39,547
The fork itself has model tools denied by
default, but your mod code doesn't.

43
00:02:40,327 --> 00:02:43,867
Before shipping, run claude plugin
validate on your mod folder,

44
00:02:44,167 --> 00:02:47,387
and use claude dash dash debug to inspect
what's executing.

45
00:02:47,895 --> 00:02:51,115
Now the smaller stuff in 2.1.318.

46
00:02:51,775 --> 00:02:56,495
Claude apps gateway gets assume underscore
role on mantle upstreams,

47
00:02:56,915 --> 00:03:01,935
for cross account AWS IAM roles and per
developer cost attribution.

48
00:03:02,695 --> 00:03:07,955
And a fix worth knowing: claude dash p
could exit zero, a false success,

49
00:03:08,395 --> 00:03:13,595
if an API error cut a reply off mid stream
while a mod hooked turn dot step.

50
00:03:14,368 --> 00:03:18,588
Also, subagents running shell commands now
move processes to the background after

51
00:03:18,628 --> 00:03:21,148
three minutes, to avoid prompt cache
expiration.

52
00:03:21,928 --> 00:03:26,088
Cloud sessions on Anthropic hosted
environments can start with a No repository

53
00:03:26,168 --> 00:03:26,508
option.

54
00:03:27,548 --> 00:03:31,908
And starting Claude Code in an IDE
terminal no longer auto installs or updates the

55
00:03:31,948 --> 00:03:32,448
extension.

56
00:03:32,888 --> 00:03:35,088
Run slash ide yourself when you want it.

57
00:03:35,868 --> 00:03:37,148
So the takeaway is simple.

58
00:03:37,928 --> 00:03:42,388
Audit your mods, find the forks doing
light work, and add effort low.