1
00:00:00,280 --> 00:00:05,020
Right, so you've got your coffee, your
terminal's open, and Claude's about to chew

2
00:00:05,060 --> 00:00:06,400
through a big refactor.

3
00:00:07,220 --> 00:00:11,560
This is The Claude Code Changelog, brought
to you by Jellypod AI.

4
00:00:12,320 --> 00:00:12,820
I'm Lachlan.

5
00:00:13,670 --> 00:00:14,270
And I'm James.

6
00:00:14,790 --> 00:00:17,970
Today is the advisor tool, the /advisor
command.

7
00:00:18,390 --> 00:00:21,950
It's experimental, and it's the cleanest
answer yet to a question we've circled

8
00:00:21,990 --> 00:00:26,670
before: how do you get Opus level thinking
without paying Opus prices for every

9
00:00:26,730 --> 00:00:27,050
grep?

10
00:00:28,225 --> 00:00:31,825
Yeah, this builds on the cost themes from
a couple of earlier episodes.

11
00:00:32,325 --> 00:00:37,925
Pinning Haiku for the auto mode safety
checks, and reining in runaway Opus threads.

12
00:00:38,625 --> 00:00:43,645
The advisor is the hybrid version of both
ideas, for your everyday CLI work.

13
00:00:44,840 --> 00:00:45,440
Here's what it is.

14
00:00:45,940 --> 00:00:51,840
A cheaper executor model, say Sonnet 5.5
or Haiku 5.5, does the actual work.

15
00:00:52,300 --> 00:00:54,920
When it hits something tricky, it consults
a stronger model,

16
00:00:55,160 --> 00:00:56,980
like Opus 5.5 or Fable.

17
00:00:57,560 --> 00:01:00,820
The advisor sees the full conversation,
every tool call and result,

18
00:01:01,200 --> 00:01:02,240
and sends back guidance.

19
00:01:02,740 --> 00:01:05,800
The advisor never calls tools and never
writes to your transcript.

20
00:01:06,100 --> 00:01:07,040
It only steers.

21
00:01:08,480 --> 00:01:09,880
Now, what changed recently.

22
00:01:10,540 --> 00:01:17,460
Version 2.1.300 added /advisor on Claude
Platform on AWS, and

23
00:01:17,520 --> 00:01:23,560
added CLAUDE_CODE_DISABLE_ADVISOR_TOOL set
to 1, which removes the tool entirely and

24
00:01:23,620 --> 00:01:25,880
ignores any advisor settings you've
configured.

25
00:01:27,020 --> 00:01:33,440
Earlier point releases, 2.1.293 and
2.1.287, brought Haiku

26
00:01:33,800 --> 00:01:38,540
5.5 and Sonnet 5.5 into the accepted
advisor lineup.

27
00:01:39,262 --> 00:01:40,422
Setup has four routes.

28
00:01:40,842 --> 00:01:44,302
Type /advisor on its own and you get an
interactive model picker.

29
00:01:44,662 --> 00:01:48,582
Type /advisor opus and it saves to your
user settings across sessions.

30
00:01:48,582 --> 00:01:50,762
/advisor off turns it off.

31
00:01:51,382 --> 00:01:56,142
For a persistent default, put advisorModel
set to opus in your global or project

32
00:01:56,162 --> 00:01:56,742
settings file.

33
00:01:57,222 --> 00:02:01,162
And for a single session, pass the advisor
flag at launch with opus after it.

34
00:02:01,870 --> 00:02:07,550
Quick note, per the 2.1.300 notes: no
advisor runs until you pick one.

35
00:02:07,990 --> 00:02:08,690
It's opt in.

36
00:02:09,432 --> 00:02:09,872
Why bother?

37
00:02:10,352 --> 00:02:11,092
Two comparisons.

38
00:02:11,452 --> 00:02:12,772
First, opusplan.

39
00:02:13,092 --> 00:02:16,952
There, Opus plans upfront, then the
session drops to Sonnet and Opus never speaks

40
00:02:16,992 --> 00:02:17,272
again.

41
00:02:17,832 --> 00:02:19,872
The advisor stays available the whole way
through.

42
00:02:20,492 --> 00:02:21,492
Second, subagents.

43
00:02:21,912 --> 00:02:24,372
Those run isolated tasks in their own
context.

44
00:02:24,692 --> 00:02:28,712
The advisor reads your actual shared
context, so its advice is grounded in what's

45
00:02:28,772 --> 00:02:29,492
already happened.

46
00:02:30,295 --> 00:02:34,615
On numbers, a video walkthrough we looked
at reports that Sonnet with an Opus

47
00:02:34,675 --> 00:02:39,855
advisor scores roughly three percent
higher on SWE bench than Opus alone,

48
00:02:40,275 --> 00:02:42,595
with token costs about twelve percent
lower.

49
00:02:43,375 --> 00:02:47,575
That's the reported figure, so treat it as
a benchmark, not a promise for your repo.

50
00:02:48,355 --> 00:02:54,175
The logic is that file reads, searches,
and boilerplate edits get billed at executor

51
00:02:54,275 --> 00:02:54,775
prices.

52
00:02:55,360 --> 00:02:59,660
As for when Claude calls the advisor,
that's model driven, not rule based.

53
00:03:00,240 --> 00:03:05,380
It tends to consult on ambiguous
architecture tradeoffs, on test or build failures

54
00:03:05,440 --> 00:03:09,020
that keep recurring, and before declaring
a complex task done.

55
00:03:09,420 --> 00:03:10,420
But you can force it.

56
00:03:10,960 --> 00:03:14,400
Put something in your prompt like: consult
the advisor before finalizing the

57
00:03:14,440 --> 00:03:15,900
database migration schema.

58
00:03:16,280 --> 00:03:19,220
Or: check with the advisor before
rewriting the auth middleware.

59
00:03:19,862 --> 00:03:23,662
You'll see an Advising row in the terminal
while it's thinking.

60
00:03:24,342 --> 00:03:29,322
Afterwards it reads Reviewed, Declined
with a reason, or Unavailable with an error

61
00:03:29,382 --> 00:03:29,722
code.

62
00:03:30,442 --> 00:03:34,842
Press Ctrl+O on that row and you can
expand it to read exactly what the advisor

63
00:03:34,902 --> 00:03:35,222
said.

64
00:03:35,762 --> 00:03:38,982
Worth doing the first few times, just to
calibrate how useful it is.

65
00:03:39,857 --> 00:03:41,477
Caveats, and there are a few.

66
00:03:42,017 --> 00:03:44,597
The advisor has to rank at or above your
main model.

67
00:03:44,997 --> 00:03:48,857
Haiku 4.5 can't be an advisor, and Sonnet
can't advise Opus.

68
00:03:49,537 --> 00:03:53,617
If you configure a bad pairing, or your
organization's availableModels allowlist

69
00:03:53,657 --> 00:03:58,617
excludes it, Claude Code either warns at
startup or quietly runs without an advisor.

70
00:03:59,097 --> 00:04:01,837
So if you think it's on and nothing's
happening, check that.

71
00:04:02,587 --> 00:04:04,127
Provider support matters too.

72
00:04:04,527 --> 00:04:09,287
It needs the first party Anthropic API or
Claude Platform on AWS.

73
00:04:09,847 --> 00:04:14,427
Amazon Bedrock, Google Vertex AI, and
Microsoft Foundry aren't supported.

74
00:04:14,917 --> 00:04:16,257
One more, on caching.

75
00:04:17,037 --> 00:04:21,317
Changing /advisor in the middle of a
session preserves your main conversation's

76
00:04:21,357 --> 00:04:25,337
prompt cache prefix, unlike /model, which
invalidates it.

77
00:04:25,997 --> 00:04:29,637
But the advisor's own read of the
transcript isn't cached between calls,

78
00:04:30,057 --> 00:04:32,417
so every consultation reprocesses the
history.

79
00:04:33,037 --> 00:04:35,317
Long sessions with frequent consults add
up.

80
00:04:36,237 --> 00:04:38,757
And Claude treats advice as advice.

81
00:04:39,097 --> 00:04:43,417
If your files or command output contradict
it, it surfaces the conflict instead of

82
00:04:43,457 --> 00:04:44,277
blindly following.

83
00:04:44,883 --> 00:04:50,103
A few smaller wins from 2.1.301 to
2.1.303.

84
00:04:50,883 --> 00:04:54,723
Vim mode got the percent motion to jump to
a matching bracket,

85
00:04:55,123 --> 00:04:57,663
alone or with d, c, or y.

86
00:04:58,383 --> 00:05:04,463
In 2.1.301, switching to a model with a
smaller context window now has the previous

87
00:05:04,583 --> 00:05:06,023
model summarize first.

88
00:05:06,643 --> 00:05:13,603
In 2.1.302, the Read tool refuses image
sources over 128 megabytes instead

89
00:05:13,643 --> 00:05:18,623
of loading them into memory, there's an
Idle compaction row in /config,

90
00:05:19,063 --> 00:05:21,983
and hooks are capped at twelve running at
once per event.

91
00:05:22,803 --> 00:05:26,263
2.1.303 is general reliability fixes.

92
00:05:27,035 --> 00:05:32,675
So your homework: run /advisor opus on a
Sonnet session, give it a messy migration,

93
00:05:32,835 --> 00:05:34,975
and hit Ctrl+O on the advice.

94
00:05:35,635 --> 00:05:36,115
See you next time.