1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
# Cursor CLI, headless. Installed separately from the Cursor editor, which
# ships no agent binary of its own:
#
# curl https://cursor.com/install -fsS | bash
#
# The installer symlinks both `agent` and `cursor-agent` to the same binary.
# This preset asks for the longer one: `agent` is a common enough word to be
# something else on a PATH, and being wrong about that is silent.
#
# Models come from your Cursor subscription rather than from a key you hold, so
# the roster changes without spar knowing, and what your plan covers is not the
# same as what it lists. `cursor-agent models` is the source of truth.
#
# There is no effort flag. Effort is part of the model name instead, as a
# `-low`, `-high` or `-max` suffix, which is why the template has no {effort}
# group and why `effort` in an agent block does nothing here.
#
# No structured output flag, so spar asks for JSON in the prompt and parses it
# back out. That is the one thing this preset gives up against codex and claude,
# and it is why cursor is a better backup than it is a primary.
= [
"cursor-agent", "--print",
["--model", "{model}"],
# Headless, in a throwaway worktree: nothing is going to answer a prompt.
# Without --force the agent proposes edits and applies none of them, which
# looks exactly like a model that did no work.
"--force",
"--trust",
"--output-format", "stream-json",
"{prompt}",
]
= "jsonl"
= "result"
= [
"~/.local/bin",
"~/.cursor/bin",
"/opt/homebrew/bin",
"/usr/local/bin",
]
# Hints for `spar init`. Never validated against, only written as comments.
#
# Kimi leads because it shares no training with Claude or with Codex, and an
# uncorrelated backup is the whole reason the pair is two models rather than
# one. A backup that thinks like the agent it stands in for is a spare tyre
# with the same puncture.
#
# It is metered rather than covered by a Cursor plan, so it needs a spend limit
# set or an allowance that has not run out. Without one it returns "You've hit
# your usage limit" for every call, which is the failure the primary already
# had. composer-2.5 and auto are covered by the plan itself and are the answer
# when being certain of a reply matters more than being certain of a second
# opinion.
#
# Cursor also serves Claude and GPT, and neither is listed here on purpose. A
# fallback running the same family as the OTHER agent turns one CLI failing
# into two of the same model reviewing each other, which is the single thing
# the pair exists to prevent, and nothing catches it: `correlation_warning`
# compares the two configured agents and never their fallbacks, and it matches
# on resolved binary and model string, so Claude through cursor-agent and
# Claude through claude do not look alike to it. Naming one here would be
# recommending it.
= ["kimi-k3-max", "composer-2.5", "auto"]
= "Run 'cursor-agent models' for the set your account actually has: it is long, it changes, and a model being listed is not the same as your plan covering it. Effort is a suffix on the model name (-low, -high, -max), not a separate setting."
# The terminal event carries the whole answer. The assistant events carry the
# same text as it streams, and matching those instead would hand spar the
# answer several times over.
[]
= "result"