@limitry/cli 1.0.0-rc.7 → 1.0.0-rc.9

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -2,6 +2,118 @@
2
2
 
3
3
  What changed in the Limitry CLI (`@limitry/cli`). Generated from the source on each release; newest first.
4
4
 
5
+ ## 1.0.0-rc.9 (2026-10-22)
6
+
7
+ ### Added
8
+
9
+ - `POST /v1/checks` can now answer `503` with the code `busy` when
10
+ Limitry is deciding too many checks for those subjects at once. Nothing
11
+ was charged. `guard` and `reserve` (TypeScript and Python) send the
12
+ check again with the same key, within the same 30 seconds they give
13
+ `moving`, before the limit's fail mode decides. Each retry of `busy` or
14
+ `moving` now waits the answer's `Retry-After` plus a random backoff that
15
+ grows with each try, so clients don't all come back at the same moment.
16
+ Older SDK versions treat `busy` as Limitry being unreachable after their
17
+ usual two quick retries, so the fail mode decides; upgrade to retry it.
18
+
19
+ ### Changed
20
+
21
+ - A very busy workspace's counts are now spread over more of Limitry's
22
+ servers as its traffic grows. While some of them move, which takes a few
23
+ seconds, checks, commits and releases on those subjects answer `503`
24
+ with the new error code `moving` and `Retry-After: 1`; nothing is
25
+ counted. `guard` and `reserve` (TypeScript and Python) and the CLI send
26
+ the same request again with the same Idempotency-Key for up to 30
27
+ seconds, each try with its own time budget, before a limit's fail mode
28
+ applies, so work is neither run uncounted nor left uncommitted.
29
+
30
+ ### Fixed
31
+
32
+ - A reservation (`mode: "reserve"`) now starts a subject's period, as a
33
+ plain check does: on a `billing_date` limit with no billing date set
34
+ yet, the first reservation's time becomes the subject's billing date,
35
+ and on a `first_use` limit it starts the period. Before, reservations
36
+ made before the subject's first commit could each count in a period of
37
+ their own, so a monthly limit could let more through than its amount.
38
+ A reservation you release still starts the period, with nothing used in
39
+ it.
40
+ - A `503 moving` answer with `Retry-After: 0` is now retried after at least
41
+ a second, not at once, and still for at most 30 seconds. A client made
42
+ with `maxRetries: 0` no longer sends the SDK's internal
43
+ `x-limitry-moving` header to the API.
44
+
45
+ ## 1.0.0-rc.8 (2026-10-21)
46
+
47
+ ### Added
48
+
49
+ - Limits have `counting`: `fast` answers the instant it decides, and
50
+ `guaranteed` stores every count safely before answering, so the limit
51
+ never goes over, even if a server fails. A `total` balance defaults to
52
+ `guaranteed`, every other window to `fast`. Set it when creating or
53
+ updating a limit (`counting`, `--counting`), or in `limitry.json`.
54
+
55
+ ### Changed
56
+
57
+ - - Creating a limit needs only who it is for, the amount and the window:
58
+ `limits.create({ subject: { kind: "user" }, amount: 100, window: "month" })`
59
+ (`limitry limits create --subject '{"kind":"user"}' --amount 100 --window month`).
60
+ `action` defaults to every action (`*`), `unit` to `units`, and `name`
61
+ to the sentence the limit reads as, such as "100 units per user per
62
+ month". If a limit already has that name, the API answers `409`: give
63
+ this one a `name`. `limitry.json` entries still need a `name`.
64
+ - A limit's input names who it is for as `subject: { kind, id }` (was
65
+ `subjectKind` and `subjectId`), the same shape the limit itself has.
66
+ - What a limit counts is fixed once it is created: its subject, action,
67
+ unit, window, `every`, `start` and `rolling`. Changing one answers
68
+ `400 invalid_request`: create a new limit and archive the old one (its
69
+ history stays). Name, amount, fail mode and enabled still change any
70
+ time. `limitry sync` refuses such a change until you rename the entry.
71
+ - A limit says when a check last reached it (`lastCheckedAt`), so you can
72
+ see a limit no check names.
73
+ - One name for each idea, everywhere: the API, both SDKs, the CLI, the MCP
74
+ tools, `limitry.json` and webhooks. Update your code when you upgrade:
75
+
76
+ - `hold()` is `reserve()`: it takes `cost` (was `estimate`), and the block
77
+ gets a `Reservation` (was `Held`; Python `Reservation` and
78
+ `AsyncReservation`), with the same `cost()` and `extend()`.
79
+ - A refusal throws `LimitReached` (was `LimitExceeded`), with `limitId`
80
+ (Python `limit_id`) and `reason` `limit_reached` or
81
+ `too_many_reservations`. When Limitry can't be reached and a limit fails
82
+ closed, you now get a separate `Unavailable` error. "Can't be reached"
83
+ also covers a `402 plan_limit_reached` and a `429` from Limitry's flood
84
+ guard, so each limit's fail mode decides instead of your code getting an
85
+ error.
86
+ - A check's answer lists its limits in `limits` (was `remaining`): read
87
+ `decision.limits[0].remaining`. Every answer that names a limit says
88
+ `limitId`: `reason.limitId`, `limits[].limitId`, `usage.standing`'s
89
+ `limits[].limitId`, and webhook payloads.
90
+ - The refusal code is `limit_reached` (was `limit_exceeded`), and the
91
+ webhook event is `limit.reached` (was `limit.exceeded`): subscribe your
92
+ endpoints to the new name.
93
+ - Usage says `refused` (was `denied`); `usage.subjects` no longer has
94
+ `over` (it is `refused > 0`).
95
+ - Credits: `limits.credits.create()` and `limits.credits.list()` at
96
+ `/v1/limits/{limitId}/credits` (were `limits.grants.*` at `…/grants`),
97
+ ids start `crd_` (were `grt_`), and the approval an agent asks for is
98
+ `limit.credit` (was `limit.grant`).
99
+ - A period that starts on the subject's billing date is
100
+ `start: "billing_date"` (was `"anchor"`), and `subjects.update()` takes
101
+ `billingDate` (was `anchor`); the subject says `billingDate` too.
102
+ - `usage.standing()` and `usage.query()` take one `subject`, written
103
+ `kind:id` (`user:usr_123`), instead of separate kind and id fields
104
+ (`limitry usage standing --subject user:usr_123`).
105
+ - A limit's id starts `lmt_` (was `lim_`); keys still start `lim_test_`
106
+ and `lim_live_`.
107
+ - `limitry sync --preview` (was `--dry-run`), and the API's `preview`
108
+ (was `dryRun`), in what you send and what comes back.
109
+
110
+ ### Removed
111
+
112
+ - Limits no longer have `counting` (Fast or Guaranteed), and the CLI no
113
+ longer has `--counting`: every count is now saved before Limitry answers,
114
+ for every limit. Remove `counting` from your code and from
115
+ `limitry.json`; Limitry ignores it.
116
+
5
117
  ## 1.0.0-rc.7 (2026-10-19)
6
118
 
7
119
  ### Added
package/README.md CHANGED
@@ -28,15 +28,15 @@ Commands act in **test** unless you say otherwise, so trying things
28
28
  never touches the limits your customers depend on.
29
29
 
30
30
  ```bash
31
- # 100 AI replies a day for each user
32
- limitry limits create --name "AI replies per day" --subject-kind user \
33
- --action ai-call --amount 100 --window day
31
+ # 100 AI replies a day for each user ("100 replies per user per day for ai-call")
32
+ limitry limits create --subject '{"kind":"user"}' --action ai-call \
33
+ --amount 100 --unit replies --window day
34
34
 
35
35
  # Check a user, as your code would
36
36
  limitry check user:usr_123 ai-call
37
37
 
38
38
  # Where a user stands, without counting anything
39
- limitry usage standing --kind user --id usr_123
39
+ limitry usage standing --subject user:usr_123
40
40
  ```
41
41
 
42
42
  When a limit is right, copy it to live from the **Limits** page in the