aegiscode 7.0.14 → 7.0.15
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/SOURCE-OFFER.txt +9 -0
- package/package.json +3 -2
- package/scripts/check.mjs +1 -52
- package/scripts/demo.mjs +1 -94
- package/scripts/predist.mjs +1 -75
- package/scripts/runtests.mjs +1 -56
- package/scripts/seal-gate.mjs +1 -74
- package/src/agents.js +1 -139
- package/src/app.js +1 -2932
- package/src/art.js +1 -289
- package/src/chatflow.js +1 -3020
- package/src/checkpoint.js +1 -90
- package/src/clipboard.js +1 -424
- package/src/cloudsync.js +1 -528
- package/src/commands.js +1 -4003
- package/src/config.js +1 -581
- package/src/connect.js +1 -349
- package/src/credentials.js +1 -33
- package/src/deps.js +1 -232
- package/src/devrun.js +1 -187
- package/src/diff.js +1 -333
- package/src/diffview.js +1 -296
- package/src/engine.js +1 -297
- package/src/entitlement.js +1 -355
- package/src/events.js +1 -707
- package/src/export.js +1 -67
- package/src/fingerprint.js +1 -592
- package/src/format.js +1 -259
- package/src/fuzzy.js +1 -119
- package/src/history.js +1 -401
- package/src/init.js +1 -175
- package/src/input.js +1 -139
- package/src/keys.js +1 -183
- package/src/markdown.js +1 -306
- package/src/models.js +1 -487
- package/src/naming.js +1 -229
- package/src/overlays.js +1 -461
- package/src/panels.js +1 -1418
- package/src/ping.js +1 -226
- package/src/presence.js +1 -238
- package/src/render.js +1 -515
- package/src/screen.js +1 -518
- package/src/screens.js +1 -896
- package/src/secret.js +1 -59
- package/src/securefs.js +1 -359
- package/src/settings.js +1 -155
- package/src/shared.js +1 -52
- package/src/sharedpaths.js +1 -53
- package/src/signup.js +1 -156
- package/src/summarize.js +1 -95
- package/src/sysdiag.js +1 -585
- package/src/system.js +1 -40
- package/src/theme.js +1 -570
- package/src/title.js +1 -220
- package/src/todos.js +1 -115
- package/src/tokens.js +1 -263
- package/src/topup.js +1 -210
- package/src/update.js +1 -181
- package/vendor/client/aegis.js +1 -2179
- package/vendor/client/attest.js +1 -442
- package/vendor/client/authorship-key.js +1 -410
- package/vendor/client/auto-provision.js +1 -134
- package/vendor/client/brain-catalog.js +1 -268
- package/vendor/client/byok-catalog.js +1 -230
- package/vendor/client/credentials.js +1 -494
- package/vendor/client/device-fingerprint.js +1 -141
- package/vendor/client/drive-connect.js +1 -110
- package/vendor/client/drive-upload.js +1 -174
- package/vendor/client/env-file.js +1 -1273
- package/vendor/client/fingerprint.js +1 -1064
- package/vendor/client/foreign-memory.js +1 -669
- package/vendor/client/google-signin.js +1 -133
- package/vendor/client/hooks.js +1 -430
- package/vendor/client/model-discovery.js +1 -373
- package/vendor/client/platform.js +1 -937
- package/vendor/client/prefs.js +1 -351
- package/vendor/client/presence-constants.js +1 -107
- package/vendor/client/presence.js +1 -438
- package/vendor/client/record-vault.js +3 -1055
- package/vendor/client/session-store.js +1 -548
- package/vendor/client/topup.js +1 -297
- package/vendor/client/touch.js +1 -458
- package/vendor/client/update.js +1 -704
- package/vendor/client/user-models.js +1 -381
- package/vendor/desktop/lib/local/agents.js +1 -105
- package/vendor/desktop/lib/local/autonomous.js +1 -1045
- package/vendor/desktop/lib/local/claude-pro.js +1 -993
- package/vendor/desktop/lib/local/context.js +1 -95
- package/vendor/desktop/lib/local/engine.js +1 -3301
- package/vendor/desktop/lib/local/entitlement.js +1 -660
- package/vendor/desktop/lib/local/git-scope.js +1 -228
- package/vendor/desktop/lib/local/local.js +1 -2135
- package/vendor/desktop/lib/local/prompt.js +1 -275
- package/vendor/desktop/lib/local/queue.js +1 -560
- package/vendor/desktop/lib/local/session-rounds.js +1 -208
- package/vendor/desktop/lib/local/shell.js +1 -680
- package/vendor/desktop/lib/local/tools.js +1 -1238
- package/vendor/desktop/lib/local/turn-guard.js +1 -300
- package/vendor/desktop/lib/local/worktree-lock.js +1 -235
- package/vendor/desktop/lib/settings.js +1 -890
- package/vendor/desktop/renderer/usage.js +1 -1239
- package/vendor/hosts/install.js +1 -638
- package/vendor/hosts/openai-shim.js +1 -515
- package/vendor/hosts/spec.js +1 -164
- package/vendor/hosts/targets.js +1 -346
- package/vendor/mcp/server.js +1 -221
- package/vendor/mcp/tools.js +1 -408
- package/vendor/one/src/child.js +1 -285
- package/vendor/one/src/permissions.js +1 -369
|
@@ -1,1241 +1,3 @@
|
|
|
1
1
|
// SPDX-License-Identifier: AGPL-3.0-or-later
|
|
2
2
|
// Copyright (C) 2026 Niklas Borneklint — AEGIS Code (aegiscode) / Aegiscloud.
|
|
3
|
-
|
|
4
|
-
// Conveying a modified copy, or running one as a network service, owes the
|
|
5
|
-
// corresponding source under this licence (§5/§13) — see LICENSING.md.
|
|
6
|
-
'use strict';
|
|
7
|
-
|
|
8
|
-
/**
|
|
9
|
-
* Pure token-usage → display-number mapping, and the cost half of it.
|
|
10
|
-
*
|
|
11
|
-
* Standalone from app.js (same reason as budget.js/stream-policy.js): it is
|
|
12
|
-
* requireable from a plain Node test without window.aegis. app.js only calls
|
|
13
|
-
* into it.
|
|
14
|
-
*
|
|
15
|
-
* Why this exists at all: the two wire formats spell the same quantity
|
|
16
|
-
* differently, and the renderer used to read exactly one of the spellings.
|
|
17
|
-
*
|
|
18
|
-
* OpenAI-compatible { prompt_tokens, completion_tokens, total_tokens }
|
|
19
|
-
* Anthropic-compatible { input_tokens, output_tokens } ← no total
|
|
20
|
-
*
|
|
21
|
-
* Reading only `total_tokens` — the field OpenAI happens to provide — meant
|
|
22
|
-
* every Anthropic-compatible model rendered with no token count at all, while
|
|
23
|
-
* the call was silently being billed. Accepting both spellings, and deriving
|
|
24
|
-
* the total when the provider doesn't state one, is what makes the spend
|
|
25
|
-
* visible regardless of which endpoint answered.
|
|
26
|
-
*
|
|
27
|
-
* ── The cost half, on the CLI's principle ──────────────────────────────────
|
|
28
|
-
*
|
|
29
|
-
* The renderer printed a token count and nothing else, so the desktop had no
|
|
30
|
-
* meter to compare against `aegiscodex` on the same engine and the same
|
|
31
|
-
* prompt — the comparison that started this whole line of work. The CLI's
|
|
32
|
-
* rule (cli/src/tokens.js, ported here) has three parts, and all three matter:
|
|
33
|
-
*
|
|
34
|
-
* 1. A pooled turn's bill is settled SERVER-side and carries a margin and a
|
|
35
|
-
* prompt-cache discount the client cannot see. When the response states
|
|
36
|
-
* `costUsd`, that figure wins verbatim — it is the truth about the
|
|
37
|
-
* charge, not an estimate of it.
|
|
38
|
-
* 2. Only in the absence of a settled charge does the local rate table
|
|
39
|
-
* apply, and the result is labeled an estimate so nobody reads a guess as
|
|
40
|
-
* a bill.
|
|
41
|
-
* 3. The rate table is resolved by exact id, then longest prefix. The
|
|
42
|
-
* DeepSeek row is load-bearing: without it every direct DeepSeek turn
|
|
43
|
-
* fell through to Sonnet's $3.00/$15.00 per M against a real
|
|
44
|
-
* $0.14/$0.28 — 21x the input rate and 54x the output rate — which made
|
|
45
|
-
* the *meter* the largest single contributor to the apparent cost gap
|
|
46
|
-
* between two surfaces running identical code.
|
|
47
|
-
*/
|
|
48
|
-
|
|
49
|
-
/**
|
|
50
|
-
* Number of tokens to show for a completed turn, or `null` when the provider
|
|
51
|
-
* reported none (an unknown count must render as nothing, never as `0`).
|
|
52
|
-
*
|
|
53
|
-
* @param {{total_tokens?: number, prompt_tokens?: number, completion_tokens?: number,
|
|
54
|
-
* input_tokens?: number, output_tokens?: number,
|
|
55
|
-
* input?: number, output?: number}|null|undefined} usage
|
|
56
|
-
* @returns {number|null}
|
|
57
|
-
*/
|
|
58
|
-
function usageTokens(usage) {
|
|
59
|
-
if (!usage || typeof usage !== 'object') return null;
|
|
60
|
-
if (typeof usage.total_tokens === 'number') return usage.total_tokens;
|
|
61
|
-
// Three spellings of the same quantity: OpenAI's, Anthropic's, and this
|
|
62
|
-
// file's own bucket shape — which is also the shape a LEDGER row carries
|
|
63
|
-
// (cli/src/history.js writes `{input, output, cacheRead, cacheWrite}` into
|
|
64
|
-
// sessions.json, and the CLI's demo path writes `{input, output}`). Reading
|
|
65
|
-
// only the two wire spellings left every stored row uncountable, so a
|
|
66
|
-
// resumed session's rolling total started at zero even though its ledger
|
|
67
|
-
// said otherwise.
|
|
68
|
-
const input = usage.input_tokens ?? usage.prompt_tokens ?? usage.input;
|
|
69
|
-
const output = usage.output_tokens ?? usage.completion_tokens ?? usage.output;
|
|
70
|
-
if (typeof input !== 'number' && typeof output !== 'number') return null;
|
|
71
|
-
return (input || 0) + (output || 0);
|
|
72
|
-
}
|
|
73
|
-
|
|
74
|
-
/**
|
|
75
|
-
* Rough token count for text the wire never measured — a direct port of the
|
|
76
|
-
* CLI's own estimator (the upstream engine's estimateTokens), and the
|
|
77
|
-
* reason its session total is MONOTONIC.
|
|
78
|
-
*
|
|
79
|
-
* This is the whole difference being fixed. The CLI's `appendHistory`
|
|
80
|
-
* (the upstream engine's history module) writes a `tokens` object for EVERY finished
|
|
81
|
-
* exchange: `{input, output, cacheRead, cacheWrite, real: true}` when the wire
|
|
82
|
-
* reported usage, and `{input: estimateTokens(prompt), output:
|
|
83
|
-
* estimateTokens(reply), real: false}` when it did not. `real` is a FLAG, not a
|
|
84
|
-
* gate — `aggregateSessionUsage` adds `t.input || 0` for every row regardless,
|
|
85
|
-
* so a turn without usage still moves the total and only marks it as partly
|
|
86
|
-
* estimated.
|
|
87
|
-
*
|
|
88
|
-
* The desktop rolled only reported usage and dropped everything else, so its
|
|
89
|
-
* total was flat across any turn the pool did not report on — the meter looked
|
|
90
|
-
* dead while the conversation was being billed. ~4 characters per token, the
|
|
91
|
-
* CLI's heuristic, kept identical so the two surfaces estimate the same turn
|
|
92
|
-
* the same way. Empty text is 0, never the `Math.max(1, …)` floor.
|
|
93
|
-
*/
|
|
94
|
-
function estimateTokens(text) {
|
|
95
|
-
if (!text) return 0;
|
|
96
|
-
return Math.max(1, Math.ceil([...String(text)].length / 4));
|
|
97
|
-
}
|
|
98
|
-
|
|
99
|
-
/**
|
|
100
|
-
* The bucket shape an exchange gets when the wire reported no usage: the CLI's
|
|
101
|
-
* `{input: estimateTokens(prompt), output: estimateTokens(reply)}`.
|
|
102
|
-
*
|
|
103
|
-
* `null` when there is no text at all to estimate from, so a dispatch that
|
|
104
|
-
* genuinely reported nothing (no usage, no prompt, no reply) still lands in
|
|
105
|
-
* `unknown` instead of being handed a fabricated zero.
|
|
106
|
-
*
|
|
107
|
-
* `reasoning` is the extended-thinking trace, and it is OUTPUT: a reasoning
|
|
108
|
-
* model bills its chain of thought as output tokens, which is why the wire's
|
|
109
|
-
* own `output_tokens` already includes it. The desktop renders that trace in
|
|
110
|
-
* its own element (app.js `reasoningText`) rather than in the reply body, so
|
|
111
|
-
* the two streams had to be added back together here — an estimate measured off
|
|
112
|
-
* the visible answer alone sat still for the entire think phase, which on a
|
|
113
|
-
* reasoning model is both the longest and the most expensive part of the turn.
|
|
114
|
-
*
|
|
115
|
-
* @param {string} [prompt]
|
|
116
|
-
* @param {string} [reply]
|
|
117
|
-
* @param {string} [reasoning]
|
|
118
|
-
* @returns {{input: number, output: number, cacheRead: number, cacheWrite: number}|null}
|
|
119
|
-
*/
|
|
120
|
-
function estimatedBuckets(prompt, reply, reasoning) {
|
|
121
|
-
const hasPrompt = typeof prompt === 'string' && prompt.length > 0;
|
|
122
|
-
const out =
|
|
123
|
-
(typeof reply === 'string' ? reply : '') +
|
|
124
|
-
(typeof reasoning === 'string' ? reasoning : '');
|
|
125
|
-
const hasReply = out.length > 0;
|
|
126
|
-
if (!hasPrompt && !hasReply) return null;
|
|
127
|
-
return {
|
|
128
|
-
input: estimateTokens(prompt),
|
|
129
|
-
output: estimateTokens(out),
|
|
130
|
-
cacheRead: 0,
|
|
131
|
-
cacheWrite: 0,
|
|
132
|
-
real: false,
|
|
133
|
-
};
|
|
134
|
-
}
|
|
135
|
-
|
|
136
|
-
// Per-million-token USD rates. Cache-read/write matter for long sessions.
|
|
137
|
-
//
|
|
138
|
-
// These are the PROVIDER's rates, for the fallback path where a turn has no
|
|
139
|
-
// server-settled charge (a byok relay turn reports the provider rate it
|
|
140
|
-
// relays against). A pooled turn reports the ledger figure instead — see
|
|
141
|
-
// turnAccounting's `costUsd` — because the pool's bill carries a margin and a
|
|
142
|
-
// prompt-cache discount this table cannot see.
|
|
143
|
-
//
|
|
144
|
-
// Keys are matched by exact id first, then by prefix (ratesFor below), so a
|
|
145
|
-
// family row covers every dated variant of it.
|
|
146
|
-
const RATES = {
|
|
147
|
-
// Sonnet, Opus, Fable, Haiku — the four FAMILY rows, matched by prefix, and
|
|
148
|
-
// deliberately NOT re-pointed at any one generation's price.
|
|
149
|
-
//
|
|
150
|
-
// The tempting change is to price this row as Claude Sonnet 5's published
|
|
151
|
-
// $2.00/$10.00, on the theory that `sonnet` names Sonnet 5. It does not: this
|
|
152
|
-
// repo's own alias tables carry no such id. `cli/src/models.js` lists pooled
|
|
153
|
-
// seats by name and `cli/src/config.js` records that the platform "does not
|
|
154
|
-
// advertise" `sonnet` at all — a pin on it was a phantom that made the pool
|
|
155
|
-
// answer from its own default with no error. So a bare `sonnet` is a LABEL,
|
|
156
|
-
// not a route: `cli/src/tokens.js` uses it as the default stand-in when no
|
|
157
|
-
// model is named, and the family row here is what a non-`claude-`-prefixed
|
|
158
|
-
// spelling resolves to by prefix.
|
|
159
|
-
//
|
|
160
|
-
// $3.00/$15.00 is therefore the honest figure for it, and it is the SAME
|
|
161
|
-
// figure as `RATES.default` below — which is the point: an id that names no
|
|
162
|
-
// real vendored route prices exactly like an unpriced one, never like a
|
|
163
|
-
// current-generation model. Every id that DOES name a real model has an exact
|
|
164
|
-
// row in the Claude 5 block below, and an exact id out-ranks a prefix match,
|
|
165
|
-
// so correcting those is what fixes the money — re-pointing this family row
|
|
166
|
-
// only moves the number an admitted placeholder reports, and silently
|
|
167
|
-
// re-prices every dated spelling that resolves here by prefix
|
|
168
|
-
// (`sonnet-4-5-20250929` is pinned at $3.00 in test/turn-accounting.test.mjs
|
|
169
|
-
// for exactly that reason).
|
|
170
|
-
sonnet: { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },
|
|
171
|
-
default: { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },
|
|
172
|
-
// Fable — the alias for **claude-fable-5**, at Fable's published
|
|
173
|
-
// $10.00/$50.00. The $5.00/$25.00 this row used to carry is the Opus-class
|
|
174
|
-
// figure; Fable has never been priced there, and the error ran in the
|
|
175
|
-
// under-reporting direction on the most expensive model in the lineup.
|
|
176
|
-
fable: { input: 10.00, output: 50.00, cacheRead: 1.00, cacheWrite: 12.50 },
|
|
177
|
-
opus: { input: 5.00, output: 25.00, cacheRead: 0.50, cacheWrite: 6.25 },
|
|
178
|
-
// Haiku — the same class of hole as the DeepSeek row below, on the provider
|
|
179
|
-
// whose prices we already thought we knew. The three rows above price Sonnet,
|
|
180
|
-
// Opus and Fable; a Haiku turn matched no exact key, no prefix
|
|
181
|
-
// ("claude-haiku-4-5" starts with none of them) and no family name, so it
|
|
182
|
-
// fell to the Sonnet-class default at 3x its real input and output rate.
|
|
183
|
-
//
|
|
184
|
-
// $1.00/$5.00 per M (Anthropic's published rate). cacheRead/cacheWrite
|
|
185
|
-
// follow the convention the three rows above already encode — 0.1x and 1.25x
|
|
186
|
-
// input — rather than being quoted independently, so the Anthropic rows
|
|
187
|
-
// cannot drift apart from each other.
|
|
188
|
-
haiku: { input: 1.00, output: 5.00, cacheRead: 0.10, cacheWrite: 1.25 },
|
|
189
|
-
// `claude-haiku-4-5` names its family after the vendor prefix, so the
|
|
190
|
-
// family-scan fallback (raw.includes(key)) would find it — but only after the
|
|
191
|
-
// prefix loop, and the prefix `claude` is not a key. Listed explicitly so the
|
|
192
|
-
// match is EXACT and cannot be reordered away.
|
|
193
|
-
'claude-haiku': { input: 1.00, output: 5.00, cacheRead: 0.10, cacheWrite: 1.25 },
|
|
194
|
-
|
|
195
|
-
// ── the Claude 5 line, one row per ID ────────────────────────────────────
|
|
196
|
-
//
|
|
197
|
-
// Every figure here is Anthropic's own published rate (docs.claude.com
|
|
198
|
-
// pricing table, fetched 2026-10-11), not a fraction derived from a sibling
|
|
199
|
-
// row. They are EXACT rows because the family fallback prices this
|
|
200
|
-
// generation wrong in both directions: `claude-sonnet-5-5` family-matched
|
|
201
|
-
// `sonnet` and `claude-fable-5-1` family-matched `fable`, so the very ids
|
|
202
|
-
// the picker offers and the BYOK relay forwards were billed at rates the
|
|
203
|
-
// vendor does not charge for them.
|
|
204
|
-
//
|
|
205
|
-
// cacheRead is NOT the 0.1x-input convention the 4.x rows above use.
|
|
206
|
-
// Anthropic changed the cache-hit multiplier on this generation and quotes
|
|
207
|
-
// it per model — 5% of input on Opus 5.5 and Sonnet 5.5 ($0.20 / $0.10),
|
|
208
|
-
// 2.5% on Fable 5.1 ($0.25), 10% on Fable 5 and Haiku 5.5 ($1.00 / $0.01)
|
|
209
|
-
// — so each row carries the published number rather than a multiplier this
|
|
210
|
-
// file computed. cacheWrite stays 1.25x input, which the same table
|
|
211
|
-
// confirms for all of them.
|
|
212
|
-
//
|
|
213
|
-
// Haiku 5.5 is priced by PROMPT LENGTH: the vendor charges 5x input and 5x
|
|
214
|
-
// output once a request's prompt passes 100k tokens ($0.50/$2.50). This
|
|
215
|
-
// schema has no prompt-size field — the only tiering it models is
|
|
216
|
-
// DeepSeek's time-of-day peak/offPeak — so the row carries the up-to-100k
|
|
217
|
-
// tier and a long-context Haiku turn UNDER-reports. Said out loud rather
|
|
218
|
-
// than left implicit: the alternative is 5x-ing every ordinary Haiku turn,
|
|
219
|
-
// which is the louder error.
|
|
220
|
-
'claude-opus-5': { input: 5.00, output: 25.00, cacheRead: 0.50, cacheWrite: 6.25 },
|
|
221
|
-
'claude-opus-5-5': { input: 4.00, output: 20.00, cacheRead: 0.20, cacheWrite: 5.00 },
|
|
222
|
-
'claude-sonnet-5': { input: 2.00, output: 10.00, cacheRead: 0.20, cacheWrite: 2.50 },
|
|
223
|
-
'claude-sonnet-5-5': { input: 2.00, output: 10.00, cacheRead: 0.10, cacheWrite: 2.50 },
|
|
224
|
-
'claude-haiku-5-5': { input: 0.10, output: 0.50, cacheRead: 0.01, cacheWrite: 0.125 },
|
|
225
|
-
'claude-fable-5': { input: 10.00, output: 50.00, cacheRead: 1.00, cacheWrite: 12.50 },
|
|
226
|
-
'claude-fable-5-1': { input: 10.00, output: 50.00, cacheRead: 0.25, cacheWrite: 12.50 },
|
|
227
|
-
// The Sonnet 4.x spellings keep their own line items now that `sonnet`
|
|
228
|
-
// above carries the 5-line price: without these, a Claude Sonnet 4.x turn —
|
|
229
|
-
// the id a config saved before this change still holds, and the wire model
|
|
230
|
-
// the dated form names — would family-match the alias row and be billed
|
|
231
|
-
// $2.00/$10.00 against the vendor's $3.00/$15.00.
|
|
232
|
-
'claude-sonnet-4-6': { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },
|
|
233
|
-
'claude-sonnet-4-5': { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },
|
|
234
|
-
|
|
235
|
-
// DeepSeek — read off the vendor's own pricing table
|
|
236
|
-
// (api-docs.deepseek.com/quick_start/pricing), not from a secondary blog and
|
|
237
|
-
// not from a mirror of our own catalog.
|
|
238
|
-
//
|
|
239
|
-
// The row this replaces said $0.14/$0.28, sourced from aegis1's
|
|
240
|
-
// catalog.py (cost_per_1k_input=0.00014) and referenced in pricing.py. That
|
|
241
|
-
// figure is the retired **V4-Flash** tier, and the vendor's own footnote (1)
|
|
242
|
-
// retires it explicitly: "The legacy names deepseek-v4-flash and
|
|
243
|
-
// deepseek-v4-flash-vision-exp are still accepted, but the corresponding
|
|
244
|
-
// models have been retired, their requests are served by the
|
|
245
|
-
// DeepSeek-V4.1-Flash model and billed at the Flash price." So every legacy
|
|
246
|
-
// id — which is what the picker advertises and what a user pins — bills at
|
|
247
|
-
// the V4.1 figures below, ~2.1x the old row.
|
|
248
|
-
//
|
|
249
|
-
// The rate is TIME-DEPENDENT (peak 01:00-04:00 and 06:00-10:00 UTC Mon-Fri),
|
|
250
|
-
// so one flat row cannot state the price — it can only state one of the two.
|
|
251
|
-
// `offPeak` carries the other and ratesFor() picks by the wall clock.
|
|
252
|
-
//
|
|
253
|
-
// cacheRead is the CACHE HIT price the vendor quotes directly ($0.006 peak),
|
|
254
|
-
// not a fraction of input: it is 50x cheaper than input, which is what makes
|
|
255
|
-
// the cache-hit rate worth measuring at all. Deriving it as 0.1x input — the
|
|
256
|
-
// convention aegis1/pricing.py used — overstated it 2.3x.
|
|
257
|
-
// cacheWrite is the input rate: DeepSeek bills a cache write as ordinary
|
|
258
|
-
// input tokens with no write premium, unlike Anthropic's 1.25x.
|
|
259
|
-
deepseek: {
|
|
260
|
-
input: 0.30, output: 1.20, cacheRead: 0.006, cacheWrite: 0.30,
|
|
261
|
-
offPeak: { input: 0.15, output: 0.60, cacheRead: 0.003, cacheWrite: 0.15 },
|
|
262
|
-
},
|
|
263
|
-
// DeepSeek-V4-Pro-0813 — a different, pricier model. Longest-prefix matching
|
|
264
|
-
// keeps it from being shadowed by the `deepseek` family row above.
|
|
265
|
-
'deepseek-v4-pro': {
|
|
266
|
-
input: 1.32, output: 3.96, cacheRead: 0.044, cacheWrite: 1.32,
|
|
267
|
-
offPeak: { input: 0.66, output: 1.98, cacheRead: 0.022, cacheWrite: 0.66 },
|
|
268
|
-
},
|
|
269
|
-
// `deepseek-v4.1-pro` is a spelling the plugin's OWN regex accepts
|
|
270
|
-
// (DEEPSEEK_REASONING_MODEL_RE: `v4(\.\d+)?-(flash|pro)`), so a config can
|
|
271
|
-
// carry it and the picker can advertise it. It is listed explicitly because
|
|
272
|
-
// the prefix loop cannot reach the row above — "deepseek-v4.1-pro" does not
|
|
273
|
-
// start with "deepseek-v4-pro" — so it matched the `deepseek` KEY by PREFIX
|
|
274
|
-
// (the prefix branch, not the family scan, though that key is also the row
|
|
275
|
-
// the family scan lands on) and a Pro-tier turn was billed at Flash rates.
|
|
276
|
-
//
|
|
277
|
-
// Figures are the Pro tier's ($1.32/$3.96 per M), the same tier the tail of
|
|
278
|
-
// the id names. This is deliberately the DEARER reading: an id that names
|
|
279
|
-
// "pro" and prices at Flash under-reports spend, which is the direction that
|
|
280
|
-
// cannot be recovered after the fact. Sources disagree on V4 Pro's current
|
|
281
|
-
// rate (one registry still carries an August 12 check at $0.435/$0.87), so
|
|
282
|
-
// treat this row as the one figure here most worth confirming against an
|
|
283
|
-
// invoice.
|
|
284
|
-
'deepseek-v4.1-pro': {
|
|
285
|
-
input: 1.32, output: 3.96, cacheRead: 0.044, cacheWrite: 1.32,
|
|
286
|
-
offPeak: { input: 0.66, output: 1.98, cacheRead: 0.022, cacheWrite: 0.66 },
|
|
287
|
-
},
|
|
288
|
-
|
|
289
|
-
// The three rows below back the BYOK additions in client/byok-catalog.js —
|
|
290
|
-
// without a row here each would resolve to RATE_BASIS.DEFAULT, the same
|
|
291
|
-
// "priced at Sonnet's $3/$15 and nobody can tell" hole this file exists to
|
|
292
|
-
// close for DeepSeek and Haiku. None of the three publish a cache-WRITE
|
|
293
|
-
// premium (unlike Anthropic's 1.25x): all three bill a cache write as
|
|
294
|
-
// ordinary input, so cacheWrite repeats the input rate, the same convention
|
|
295
|
-
// the DeepSeek rows above use for the same reason.
|
|
296
|
-
|
|
297
|
-
// Grok 4.7 (api.x.ai/v1, flagship since 2026-09-21): $2.00/$6.00 per M,
|
|
298
|
-
// $0.50 cached input, standard (<200k token) tier. xAI doubles every rate
|
|
299
|
-
// above a 200k-token prompt, a prompt-SIZE threshold this table has no field
|
|
300
|
-
// for (the only tiering this schema models is DeepSeek's time-of-day
|
|
301
|
-
// peak/offPeak) — left at the standard tier rather than inventing one, so a
|
|
302
|
-
// long-context Grok turn under-reports rather than carrying a fabricated
|
|
303
|
-
// split.
|
|
304
|
-
'grok-4.7': { input: 2.00, output: 6.00, cacheRead: 0.50, cacheWrite: 2.00 },
|
|
305
|
-
|
|
306
|
-
// Gemini 3.8 Flash (ai.google.dev, GA 2026-09-02): $0.75/$3.75 per M,
|
|
307
|
-
// $0.075 cached input — the introductory rate published through
|
|
308
|
-
// 2026-12-31 (scheduled to rise to $1.50/$7.50 on 2027-01-01, not modeled
|
|
309
|
-
// here for the same reason the Grok long-context tier above is not: no
|
|
310
|
-
// field in this table carries a future-dated rate change).
|
|
311
|
-
'gemini-3.8-flash': { input: 0.75, output: 3.75, cacheRead: 0.075, cacheWrite: 0.75 },
|
|
312
|
-
|
|
313
|
-
// GPT-5.5 (api.openai.com/v1, released 2026-04-23): $5.00/$30.00 per M,
|
|
314
|
-
// $0.50 cached input, standard (<272k token) tier. Like Grok above, OpenAI
|
|
315
|
-
// doubles input and 1.5x's output past that prompt size; left at the
|
|
316
|
-
// standard tier for the same reason.
|
|
317
|
-
'gpt-5.5': { input: 5.00, output: 30.00, cacheRead: 0.50, cacheWrite: 5.00 },
|
|
318
|
-
};
|
|
319
|
-
|
|
320
|
-
/**
|
|
321
|
-
* Is `at` inside DeepSeek's peak window? 01:00-04:00 and 06:00-10:00 UTC,
|
|
322
|
-
* Monday-Friday. Everything else, weekends included, is off-peak at half rate.
|
|
323
|
-
*
|
|
324
|
-
* The vendor also excludes Chinese public holidays, which we do NOT model: the
|
|
325
|
-
* published table lists them by date, so honouring them would mean shipping a
|
|
326
|
-
* calendar that goes stale. The consequence is bounded and one-directional —
|
|
327
|
-
* on a holiday we price at peak (the dearer reading) instead of off-peak — and
|
|
328
|
-
* over-reporting a spend is recoverable where under-reporting is not.
|
|
329
|
-
*/
|
|
330
|
-
function isPeak(provider, at = new Date()) {
|
|
331
|
-
if (String(provider || '').toLowerCase().indexOf('deepseek') !== 0) return true;
|
|
332
|
-
const day = at.getUTCDay();
|
|
333
|
-
if (day === 0 || day === 6) return false;
|
|
334
|
-
const hour = at.getUTCHours();
|
|
335
|
-
return (hour >= 1 && hour < 4) || (hour >= 6 && hour < 10);
|
|
336
|
-
}
|
|
337
|
-
|
|
338
|
-
/** How a model id was matched to a rate row — see ratesFor. */
|
|
339
|
-
const RATE_BASIS = Object.freeze({
|
|
340
|
-
EXACT: 'exact',
|
|
341
|
-
PREFIX: 'prefix',
|
|
342
|
-
FAMILY: 'family',
|
|
343
|
-
DEFAULT: 'default',
|
|
344
|
-
});
|
|
345
|
-
|
|
346
|
-
/**
|
|
347
|
-
* The rate rows a model id may name as a whole word rather than as a prefix.
|
|
348
|
-
* Anthropic spells its family *inside* the id — "claude-opus-4" matches no
|
|
349
|
-
* prefix — so prefix matching alone cannot see it.
|
|
350
|
-
*/
|
|
351
|
-
const RATE_FAMILIES = Object.keys(RATES).filter((k) => k !== 'default');
|
|
352
|
-
|
|
353
|
-
/**
|
|
354
|
-
* Ids this module can be handed that deliberately have NO row, with the reason.
|
|
355
|
-
*
|
|
356
|
-
* `test/rate-coverage.test.mjs` fails the build when an id the plugin's own
|
|
357
|
-
* source routes to resolves to RATE_BASIS.DEFAULT without appearing here. That
|
|
358
|
-
* is the mechanism that was missing on the day the picker started advertising
|
|
359
|
-
* DeepSeek: nothing compared "ids we route to" against "ids we have priced", so
|
|
360
|
-
* the gap had no surface on which to show up as a failure.
|
|
361
|
-
*
|
|
362
|
-
* Two honest reasons to be in this list, and no others:
|
|
363
|
-
*
|
|
364
|
-
* - POOLED: aegis1 bills the pool call and reports the settled figure, which
|
|
365
|
-
* `turnAccounting` already prefers over anything computed here (its
|
|
366
|
-
* `costUsd` branch). A local guess would be a second, worse number printed
|
|
367
|
-
* beside the real charge — and for the brain tiers it would also miss the
|
|
368
|
-
* fan-out factor (see ID_NOTES in cli/src/models.js: four billed provider
|
|
369
|
-
* calls per turn at effort=high, sized by the ask).
|
|
370
|
-
* - PROVIDER: a bare provider (`anthropic`, `groq`) is a ROUTING choice — "let
|
|
371
|
-
* the pool pick" — not a model. It names no rate because it names no model.
|
|
372
|
-
*
|
|
373
|
-
* Anything else belongs in RATES. An entry here is a claim that we have no
|
|
374
|
-
* business pricing it, not a parking space for a rate we have not looked up.
|
|
375
|
-
*/
|
|
376
|
-
const UNPRICED_BY_DESIGN = Object.freeze({
|
|
377
|
-
// Pooled seats — the ledger carries the charge.
|
|
378
|
-
'nexus-brain': 'pooled: multi-call fan-out, aegis1 settles the charge',
|
|
379
|
-
'aegis-brain': 'pooled alias of nexus-brain',
|
|
380
|
-
'nexus-brain-smart': 'pooled alias of nexus-brain',
|
|
381
|
-
'nexus-brain-neo': 'pooled alias of nexus-brain',
|
|
382
|
-
'aegis-brain-smart': 'pooled alias of nexus-brain',
|
|
383
|
-
'aegis-brain-neo': 'pooled alias of nexus-brain',
|
|
384
|
-
'openai-gpt4o-mini': 'pooled seat, aegis1 settles the charge',
|
|
385
|
-
// Bare provider names — a routing choice, not a model.
|
|
386
|
-
anthropic: 'provider name, not a model id',
|
|
387
|
-
openai: 'provider name, not a model id',
|
|
388
|
-
groq: 'provider name, not a model id',
|
|
389
|
-
google: 'provider name, not a model id',
|
|
390
|
-
xai: 'provider name, not a model id',
|
|
391
|
-
});
|
|
392
|
-
|
|
393
|
-
/**
|
|
394
|
-
* The rate row for a model id, plus HOW it was matched.
|
|
395
|
-
*
|
|
396
|
-
* `basis` is the whole point of the return shape. This used to be
|
|
397
|
-
* `return family || RATES.default` — a bare row — so an id nobody had priced
|
|
398
|
-
* (DeepSeek, until this change; Haiku too) was silently priced at Sonnet's
|
|
399
|
-
* $3/$15 and rendered as though it were a measurement. A wrong number that
|
|
400
|
-
* announces itself is a bug report; a wrong number that looks like every other
|
|
401
|
-
* number is a fabrication, and it survived because nothing downstream could
|
|
402
|
-
* tell the two apart.
|
|
403
|
-
*
|
|
404
|
-
* So the fallback still exists — a caller needs *a* number to render — but it
|
|
405
|
-
* is now labeled: `basis: 'default'` means "we do not know this model's price
|
|
406
|
-
* and this figure is a Sonnet-class placeholder". Callers that display money
|
|
407
|
-
* are expected to say so, and test/rate-coverage.test.mjs fails the build when
|
|
408
|
-
* a model the plugin can actually route to lands here.
|
|
409
|
-
*
|
|
410
|
-
* @param {string} model a model id, provider name, alias, or "provider:model"
|
|
411
|
-
* @param {{at?: Date}} [opts] the instant to price against, for peak/off-peak
|
|
412
|
-
* @returns {object} the row ({input,output,cacheRead,cacheWrite}) plus `basis`,
|
|
413
|
-
* `peak`, and the `model` it resolved from.
|
|
414
|
-
*/
|
|
415
|
-
function ratesFor(model, { at } = {}) {
|
|
416
|
-
const raw = String(model || '').toLowerCase();
|
|
417
|
-
const finish = (row, basis) => {
|
|
418
|
-
const provider = raw.indexOf('deepseek') === 0 || raw.indexOf(':deepseek') > -1
|
|
419
|
-
? 'deepseek'
|
|
420
|
-
: raw;
|
|
421
|
-
const peak = isPeak(provider, at || new Date());
|
|
422
|
-
// Off-peak rows are half of peak by construction, but read them from the
|
|
423
|
-
// table rather than dividing, so a future discount that is NOT exactly half
|
|
424
|
-
// is expressed where the numbers live instead of in arithmetic here.
|
|
425
|
-
const use = !peak && row.offPeak ? { ...row, ...row.offPeak } : row;
|
|
426
|
-
return { ...use, basis, peak, model: raw };
|
|
427
|
-
};
|
|
428
|
-
if (!raw) return finish(RATES.default, RATE_BASIS.DEFAULT);
|
|
429
|
-
|
|
430
|
-
// A byok model id carries its provider as a routing label:
|
|
431
|
-
// "anthropic:claude-opus-4", "deepseek:deepseek-v4-flash". The label is not
|
|
432
|
-
// a rate, and rating it as written fell through every prefix to
|
|
433
|
-
// RATES.default — so an Opus BYOK turn was priced at Sonnet's $3/$15 against
|
|
434
|
-
// the vendor's real $5/$25. 15x, on the one class whose figure can only ever
|
|
435
|
-
// be an estimate, because there the vendor bills the caller and AEGIS only
|
|
436
|
-
// relays.
|
|
437
|
-
//
|
|
438
|
-
// Precedence is by MATCH QUALITY first and candidate second, not
|
|
439
|
-
// candidate-first: matching candidate-by-candidate let the label `deepseek`
|
|
440
|
-
// prefix-match the family row before the tail was ever considered, so a
|
|
441
|
-
// `deepseek:deepseek-v4-pro` turn was priced at Flash rates — the routing
|
|
442
|
-
// label beating the model it labels. So: try every candidate as an exact id,
|
|
443
|
-
// then every candidate as a prefix.
|
|
444
|
-
const ids = [raw];
|
|
445
|
-
// The routing label is separated by ':' in a byok id ("anthropic:claude-opus-4")
|
|
446
|
-
// and by '/' in the ids the PLATFORM advertises ("deepseek/deepseek-v4-flash"
|
|
447
|
-
// — see test/cli-tools.test.mjs, test/mcp-balance.test.mjs). Splitting only on
|
|
448
|
-
// ':' left the slash form resolving its whole string as one id, so
|
|
449
|
-
// "deepseek/deepseek-v4-pro" never reached the tail and prefix-matched the
|
|
450
|
-
// `deepseek` FAMILY row — the pro model priced at Flash rates, the same
|
|
451
|
-
// label-shadowing bug the ':' form had, in the spelling that is actually
|
|
452
|
-
// shipped.
|
|
453
|
-
const sep = Math.max(raw.lastIndexOf(':'), raw.lastIndexOf('/'));
|
|
454
|
-
const tail = raw.slice(sep + 1);
|
|
455
|
-
if (tail && tail !== raw) ids.push(tail);
|
|
456
|
-
for (const id of ids) {
|
|
457
|
-
if (RATES[id]) return finish(RATES[id], RATE_BASIS.EXACT);
|
|
458
|
-
}
|
|
459
|
-
let best = null;
|
|
460
|
-
let bestLen = 0;
|
|
461
|
-
for (const id of ids) {
|
|
462
|
-
for (const [prefix, rates] of Object.entries(RATES)) {
|
|
463
|
-
if (prefix !== 'default' && id.startsWith(prefix) && prefix.length > bestLen) {
|
|
464
|
-
best = rates;
|
|
465
|
-
bestLen = prefix.length;
|
|
466
|
-
}
|
|
467
|
-
}
|
|
468
|
-
}
|
|
469
|
-
if (best) return finish(best, RATE_BASIS.PREFIX);
|
|
470
|
-
// Longest family-key hit wins, so an overlapping pair resolves the same way
|
|
471
|
-
// twice running rather than by object key order.
|
|
472
|
-
let family = null;
|
|
473
|
-
let familyLen = 0;
|
|
474
|
-
for (const key of RATE_FAMILIES) {
|
|
475
|
-
if (raw.includes(key) && key.length > familyLen) {
|
|
476
|
-
family = RATES[key];
|
|
477
|
-
familyLen = key.length;
|
|
478
|
-
}
|
|
479
|
-
}
|
|
480
|
-
return finish(family || RATES.default, family ? RATE_BASIS.FAMILY : RATE_BASIS.DEFAULT);
|
|
481
|
-
}
|
|
482
|
-
|
|
483
|
-
/**
|
|
484
|
-
* The four billable buckets from a wire usage object, accepting every
|
|
485
|
-
* provider's spelling of them.
|
|
486
|
-
*
|
|
487
|
-
* ── The one thing the providers genuinely disagree about ───────────────────
|
|
488
|
-
*
|
|
489
|
-
* Cache fields are not merely spelled differently, they are MEASURED
|
|
490
|
-
* differently, and the difference is silent:
|
|
491
|
-
*
|
|
492
|
-
* Anthropic `input_tokens` EXCLUDES `cache_read_input_tokens` and
|
|
493
|
-
* `cache_creation_input_tokens` — the buckets are DISJOINT.
|
|
494
|
-
* OpenAI `prompt_tokens` INCLUDES
|
|
495
|
-
* `prompt_tokens_details.cached_tokens`.
|
|
496
|
-
* DeepSeek `prompt_tokens` INCLUDES `prompt_cache_hit_tokens` — the
|
|
497
|
-
* cached count is a SUBSET of the prompt count.
|
|
498
|
-
*
|
|
499
|
-
* Reading a subset pair as though it were disjoint double-bills every cached
|
|
500
|
-
* token: once at the input rate and again at the cache rate. That is the exact
|
|
501
|
-
* inverse of the saving caching exists to produce, and it is invisible in the
|
|
502
|
-
* output — the figure simply comes out larger, and never looks wrong. So the
|
|
503
|
-
* convention is detected from which field supplied the prompt count, and only
|
|
504
|
-
* the subset spellings subtract.
|
|
505
|
-
*
|
|
506
|
-
* Earlier revisions deferred this to `desktop/lib/local/providers.js`, a file
|
|
507
|
-
* that does not exist — so nothing normalised these fields and every DeepSeek
|
|
508
|
-
* turn reported `cacheRead: 0`, pricing its whole prompt at the miss rate.
|
|
509
|
-
* That is why the convention is resolved here, in the one module both hosts
|
|
510
|
-
* resolve (cli/src/deps.js), rather than in a transport either could drop.
|
|
511
|
-
*
|
|
512
|
-
* @param {object|null|undefined} usage
|
|
513
|
-
* @returns {{input: number, output: number, cacheRead: number, cacheWrite: number}}
|
|
514
|
-
*/
|
|
515
|
-
function usageBuckets(usage) {
|
|
516
|
-
const u = usage && typeof usage === 'object' ? usage : {};
|
|
517
|
-
const num = (...candidates) => {
|
|
518
|
-
for (const c of candidates) if (typeof c === 'number') return c;
|
|
519
|
-
return 0;
|
|
520
|
-
};
|
|
521
|
-
const opt = (v) => (typeof v === 'number' ? v : undefined);
|
|
522
|
-
|
|
523
|
-
const anthropicIn = opt(u.input_tokens);
|
|
524
|
-
const openaiIn = opt(u.prompt_tokens);
|
|
525
|
-
const details = u.prompt_tokens_details;
|
|
526
|
-
const nested = details && typeof details === 'object' ? opt(details.cached_tokens) : undefined;
|
|
527
|
-
// DeepSeek's `prompt_cache_hit_tokens`, OpenAI's nested `cached_tokens`, and
|
|
528
|
-
// the bare `cached_tokens` some OpenAI-compatible gateways emit. All three
|
|
529
|
-
// are SUBSETS of the prompt count they arrive beside.
|
|
530
|
-
const subsetHit = num(u.prompt_cache_hit_tokens, nested, u.cached_tokens);
|
|
531
|
-
|
|
532
|
-
// Only the OpenAI/DeepSeek spelling names the cache INSIDE the prompt stat.
|
|
533
|
-
// The Anthropic spelling, and this file's own already-bucketed `{input,...}`
|
|
534
|
-
// shape (which a ledger row carries), are disjoint — subtracting there would
|
|
535
|
-
// understate the bill.
|
|
536
|
-
const subset = anthropicIn === undefined && openaiIn !== undefined;
|
|
537
|
-
const promptCount = subset ? openaiIn : num(anthropicIn, openaiIn, u.input);
|
|
538
|
-
|
|
539
|
-
const cacheReadRaw = num(u.cache_read_input_tokens, u.cacheRead, subsetHit);
|
|
540
|
-
// Clamp to the prompt it belongs to: a provider over-reporting its hit count
|
|
541
|
-
// must not be able to produce negative input tokens.
|
|
542
|
-
const cacheRead = subset ? Math.min(Math.max(0, promptCount), cacheReadRaw) : cacheReadRaw;
|
|
543
|
-
const cacheWrite = num(u.cache_creation_input_tokens, u.cacheWrite);
|
|
544
|
-
const output = num(u.output_tokens, u.completion_tokens, u.output);
|
|
545
|
-
const input = subset ? Math.max(0, promptCount - cacheRead) : promptCount;
|
|
546
|
-
|
|
547
|
-
// A stated total larger than the split means the provider counted tokens the
|
|
548
|
-
// split does not name (thinking, cached reads). Attribute the remainder to
|
|
549
|
-
// input rather than dropping it: dropping it would understate the bill.
|
|
550
|
-
const total = num(u.total_tokens);
|
|
551
|
-
const accounted = input + output + cacheRead + cacheWrite;
|
|
552
|
-
// ── Did the source report a cache at all? ─────────────────────────────────
|
|
553
|
-
//
|
|
554
|
-
// Every `num(...)` above coerces an absent field to 0, which is right for
|
|
555
|
-
// arithmetic and WRONG for a claim. A provider that reports only
|
|
556
|
-
// `prompt_tokens` and no cache field of any spelling is not reporting a cold
|
|
557
|
-
// cache — it is not reporting a cache — and the two must not render the same.
|
|
558
|
-
// Aider does exactly this (it prints "Tokens: N sent" and no hit count), and
|
|
559
|
-
// reading its silence as a 0 made its prompt total the lowest of five
|
|
560
|
-
// harnesses when it was really a partial number. See `cacheHitRate`.
|
|
561
|
-
const cacheReported =
|
|
562
|
-
typeof u.cache_read_input_tokens === 'number'
|
|
563
|
-
|| typeof u.cacheRead === 'number'
|
|
564
|
-
|| typeof u.cache_creation_input_tokens === 'number'
|
|
565
|
-
|| typeof u.cacheWrite === 'number'
|
|
566
|
-
|| typeof u.prompt_cache_hit_tokens === 'number'
|
|
567
|
-
|| typeof u.cached_tokens === 'number'
|
|
568
|
-
|| typeof nested === 'number';
|
|
569
|
-
return {
|
|
570
|
-
input: total > accounted ? input + (total - accounted) : input,
|
|
571
|
-
output,
|
|
572
|
-
cacheRead,
|
|
573
|
-
cacheWrite,
|
|
574
|
-
cacheReported,
|
|
575
|
-
};
|
|
576
|
-
}
|
|
577
|
-
|
|
578
|
-
/**
|
|
579
|
-
* Cache hit rate for a usage record: the share of prompt tokens the provider
|
|
580
|
-
* actually served from its cache rather than billing at miss price.
|
|
581
|
-
*
|
|
582
|
-
* Ported from the upstream engine reference. It reads through `usageBuckets`
|
|
583
|
-
* rather than off the raw object, which is the one deliberate difference: the
|
|
584
|
-
* engine's copy is handed an already-normalised record, while this module is
|
|
585
|
-
* also handed WIRE usage (a provider's raw response body) and ledger rows. A
|
|
586
|
-
* raw DeepSeek body spells the hit count `prompt_cache_hit_tokens` inside
|
|
587
|
-
* `prompt_tokens`; reading `usage.cacheRead` off that shape finds nothing and
|
|
588
|
-
* reports `pct: null` — "no cache activity" — for a turn that was almost
|
|
589
|
-
* entirely cached. Going through the normaliser is what makes the number the
|
|
590
|
-
* same on both spellings.
|
|
591
|
-
*
|
|
592
|
-
* The denominator is `input + cacheRead`, and `input` means the MISS half
|
|
593
|
-
* everywhere in this codebase — `usageBuckets` derives it as
|
|
594
|
-
* `prompt_tokens - prompt_cache_hit_tokens`, and the Anthropic spelling
|
|
595
|
-
* `input_tokens` already excludes cached reads by definition. Adding
|
|
596
|
-
* `cacheWrite` to the denominator would be wrong twice over: on Anthropic a
|
|
597
|
-
* written block is read back on the next turn (the same tokens counted again),
|
|
598
|
-
* and on DeepSeek there is no write field at all.
|
|
599
|
-
*
|
|
600
|
-
* `pct` is null — not 0 — when nothing has been billed yet, so a caller can
|
|
601
|
-
* tell "no cache activity recorded" apart from "everything missed". A
|
|
602
|
-
* hardcoded 0 in a parser once made the benchmark's "0 cache reads" column a
|
|
603
|
-
* measurement of our own code rather than of the cache.
|
|
604
|
-
*
|
|
605
|
-
* `reported` closes the other half of that hole, which this module shares with
|
|
606
|
-
* the engine and had to fix in the same commit or the two hosts would disagree.
|
|
607
|
-
* `usageBuckets` coerces every absent field to 0 for arithmetic, so a provider
|
|
608
|
-
* that reports prompt tokens and NO cache spelling of any kind came out
|
|
609
|
-
* `pct: 0` — "everything missed" — when the honest reading is "never said".
|
|
610
|
-
* Aider is that shape: it prints a sent-token count and no hit count, and its
|
|
611
|
-
* silence priced it as the most cache-efficient harness in the benchmark. So
|
|
612
|
-
* `pct` is null unless a cache field was actually present, and a caller that
|
|
613
|
-
* prints the rate renders an unreported one as "—".
|
|
614
|
-
*
|
|
615
|
-
* @returns {{read: number, miss: number, prompt: number, pct: number|null,
|
|
616
|
-
* reported: boolean}} `pct` is 0-100, or null when the source reported no
|
|
617
|
-
* cache field, or when no prompt tokens have been accounted.
|
|
618
|
-
*/
|
|
619
|
-
function cacheHitRate(usage) {
|
|
620
|
-
const u = usageBuckets(usage);
|
|
621
|
-
const read = Math.max(0, u.cacheRead || 0);
|
|
622
|
-
const miss = Math.max(0, u.input || 0);
|
|
623
|
-
const prompt = read + miss;
|
|
624
|
-
return {
|
|
625
|
-
read,
|
|
626
|
-
miss,
|
|
627
|
-
prompt,
|
|
628
|
-
pct: u.cacheReported && prompt > 0 ? (read / prompt) * 100 : null,
|
|
629
|
-
reported: u.cacheReported,
|
|
630
|
-
};
|
|
631
|
-
}
|
|
632
|
-
|
|
633
|
-
/**
|
|
634
|
-
* Dollar cost of a usage record at the given model's rates, with the
|
|
635
|
-
* provenance attached.
|
|
636
|
-
*
|
|
637
|
-
* `basis === 'default'` means the model was not priced and `cost` is a
|
|
638
|
-
* Sonnet-class placeholder — the one state the old code could not express.
|
|
639
|
-
*
|
|
640
|
-
* `model` deliberately has NO `= 'sonnet'` default parameter, which is a
|
|
641
|
-
* behaviour change worth stating: the old signature was `model = 'sonnet'`, so
|
|
642
|
-
* `costBreakdown(u, undefined)` substituted the alias and reported
|
|
643
|
-
* `basis: 'exact'` — "this model's real rate" — while `costBreakdown(u, '')`
|
|
644
|
-
* (and `ratesFor(undefined)`) reported `basis: 'default'` — "we do not know
|
|
645
|
-
* this model". The two spellings produced the SAME money, because
|
|
646
|
-
* RATES.default and RATES.sonnet are numerically identical, but two different
|
|
647
|
-
* labels for the same number is the exact ambiguity this return shape exists to
|
|
648
|
-
* remove. So an unspecified model now flows to the default row and says so.
|
|
649
|
-
* The money is unchanged; only the claim about it is.
|
|
650
|
-
*
|
|
651
|
-
* @param {object} usage a wire usage object or a stored bucket row
|
|
652
|
-
* @param {string} model the model id that answered
|
|
653
|
-
* @param {{at?: Date}} [opts] the instant to price against (peak/off-peak)
|
|
654
|
-
* @returns {{cost: number, basis: string, peak: boolean, model: string, priced: boolean}}
|
|
655
|
-
*/
|
|
656
|
-
function costBreakdown(usage, model, opts = {}) {
|
|
657
|
-
const r = ratesFor(model, opts);
|
|
658
|
-
const u = usageBuckets(usage);
|
|
659
|
-
const toD = (n, rate) => (Math.max(0, Number(n) || 0) / 1_000_000) * rate;
|
|
660
|
-
const cost = toD(u.input, r.input)
|
|
661
|
-
+ toD(u.output, r.output)
|
|
662
|
-
+ toD(u.cacheRead, r.cacheRead)
|
|
663
|
-
+ toD(u.cacheWrite, r.cacheWrite);
|
|
664
|
-
return { cost, basis: r.basis, peak: r.peak, model: r.model, priced: r.basis !== RATE_BASIS.DEFAULT };
|
|
665
|
-
}
|
|
666
|
-
|
|
667
|
-
/**
|
|
668
|
-
* Dollar cost of a usage record at the given model's rates (USD, estimate).
|
|
669
|
-
* See `costBreakdown` for why an unnamed model reports `default` rather than
|
|
670
|
-
* silently claiming Sonnet's row as an exact match.
|
|
671
|
-
*/
|
|
672
|
-
function usageCost(usage, model, opts = {}) {
|
|
673
|
-
return costBreakdown(usage, model, opts).cost;
|
|
674
|
-
}
|
|
675
|
-
|
|
676
|
-
/**
|
|
677
|
-
* What one turn cost, as the CLI reports it: the server's settled charge when
|
|
678
|
-
* the response carries one, otherwise the rate table's estimate, always with
|
|
679
|
-
* the distinction preserved so it can be labeled.
|
|
680
|
-
*
|
|
681
|
-
* @param {object|null|undefined} usage the response's `usage` object
|
|
682
|
-
* @param {string} model the model id that answered
|
|
683
|
-
* @param {{costUsd?: number}} [opts] the server-settled charge, if any
|
|
684
|
-
* @returns {{tokens: number|null, output: number|null, cost: number|null,
|
|
685
|
-
* real: boolean, estimated: boolean}}
|
|
686
|
-
* `real` is true only when `cost` is the settled charge. `cost` is
|
|
687
|
-
* `null` when nothing was reported and no model was named — an
|
|
688
|
-
* unpriced turn must render as nothing, never as $0.0000.
|
|
689
|
-
* `tokens` is the wire's TOTAL (input + output) and is what the ledger
|
|
690
|
-
* and the rolling fold need; `output` is the half a person reads, and
|
|
691
|
-
* it is what every per-turn surface shows (see `displayTokens`).
|
|
692
|
-
*/
|
|
693
|
-
function turnAccounting(usage, model, opts = {}) {
|
|
694
|
-
const tokens = usageTokens(usage);
|
|
695
|
-
const u = usage && typeof usage === 'object' ? usage : {};
|
|
696
|
-
// The half the turn PRODUCED. Kept beside the total rather than replacing it:
|
|
697
|
-
// `tokens` still has to fold the ledger, and a rebuild from stored buckets
|
|
698
|
-
// must agree with the live fold to the unit.
|
|
699
|
-
const output =
|
|
700
|
-
usage && typeof usage === 'object' ? usageBuckets(usage).output : null;
|
|
701
|
-
// The settled charge can arrive either on the response or folded into the
|
|
702
|
-
// usage object — the upstream engine does the latter
|
|
703
|
-
// (`{ ...result.usage, costUsd: result.costUsd }`), so both are accepted.
|
|
704
|
-
const settled = typeof opts.costUsd === 'number'
|
|
705
|
-
? opts.costUsd
|
|
706
|
-
: (typeof u.costUsd === 'number' ? u.costUsd : undefined);
|
|
707
|
-
if (typeof settled === 'number') {
|
|
708
|
-
// The ledger's own figure. `basis: 'settled'` rather than a table match,
|
|
709
|
-
// because a settled charge is a fact and a table match is a guess — and the
|
|
710
|
-
// pool's bill carries a margin and a prompt-cache discount this table
|
|
711
|
-
// cannot see, so the two must never be averaged or confused.
|
|
712
|
-
return {
|
|
713
|
-
tokens, output, cost: settled, real: true, estimated: false, basis: 'settled', priced: true,
|
|
714
|
-
};
|
|
715
|
-
}
|
|
716
|
-
if (tokens == null) {
|
|
717
|
-
return {
|
|
718
|
-
tokens, output, cost: null, real: false, estimated: false, basis: null, priced: false,
|
|
719
|
-
};
|
|
720
|
-
}
|
|
721
|
-
const b = costBreakdown(u, model, opts);
|
|
722
|
-
// A model the table cannot place still yields a number (ratesFor never
|
|
723
|
-
// returns nothing), so this is always an estimate — but `priced: false` says
|
|
724
|
-
// the figure is a placeholder rather than this model's rate, so the caller
|
|
725
|
-
// can label it instead of printing it as though it had been measured.
|
|
726
|
-
return {
|
|
727
|
-
tokens, output, cost: b.cost, real: false, estimated: true, basis: b.basis, priced: b.priced,
|
|
728
|
-
};
|
|
729
|
-
}
|
|
730
|
-
|
|
731
|
-
/**
|
|
732
|
-
* ── The rolling session tallies, on the CLI's rule ─────────────────────────
|
|
733
|
-
*
|
|
734
|
-
* The two surfaces did not merely print different numbers — they counted
|
|
735
|
-
* differently. The CLI never shows a turn's tokens in isolation: `recordTurn`
|
|
736
|
-
* (cli/src/app.js) folds each finished turn's usage into one `session` object
|
|
737
|
-
* — `tokens`, `inputTokens`, `outputTokens`, `calls` — and what the user reads
|
|
738
|
-
* back is that RUNNING TOTAL. The status bar prints `state.tokens`
|
|
739
|
-
* (renderStatus), and `ctrl+t` prints the tallies in one line
|
|
740
|
-
* (`tokenSummary`: `12,400 tokens (10,100 in / 2,300 out) · 4 calls · €0.03`).
|
|
741
|
-
* The tallies are reproduced; that one abbreviation is not — `fmtRoll` explains
|
|
742
|
-
* why it drops the parenthetical rather than carrying the CLI's line verbatim.
|
|
743
|
-
*
|
|
744
|
-
* The desktop counted per turn only. Every meta row was a fresh count that
|
|
745
|
-
* reset at the next call, so "what has this session spent" was answerable only
|
|
746
|
-
* by adding the rows up by eye across a scrollback — which is most of why the
|
|
747
|
-
* desktop looked like it accounted differently from the CLI on identical
|
|
748
|
-
* engine code and an identical prompt.
|
|
749
|
-
*
|
|
750
|
-
* Three properties of the CLI's fold are load-bearing and are reproduced here
|
|
751
|
-
* exactly, because dropping any one of them reintroduces a specific lie:
|
|
752
|
-
*
|
|
753
|
-
* 1. `turns` and `calls` are incremented BEFORE the "did usage come back?"
|
|
754
|
-
* gate (recordTurn counts first, then tests `tokens != null`). A turn
|
|
755
|
-
* that reported nothing still happened; a tally that skipped it would
|
|
756
|
-
* report the session as shorter and cheaper than it was.
|
|
757
|
-
* 2. `tokens`, `input` and `output` accumulate — never reset. A rolling
|
|
758
|
-
* total that resets per turn is the per-turn count it replaced.
|
|
759
|
-
* 3. A `null` count contributes nothing and is counted in `unknown`, so
|
|
760
|
-
* `tokens: 0` is only ever read as a real zero. Nothing is ever added as
|
|
761
|
-
* a fabricated 0 to make the arithmetic look complete.
|
|
762
|
-
*
|
|
763
|
-
* The money split is the CLI's too: a settled charge (the pool's ledger
|
|
764
|
-
* figure) rolls into `cost`, and a locally priced turn rolls into `estimate`.
|
|
765
|
-
* They are kept apart rather than summed so a `~`-estimate can never be read
|
|
766
|
-
* as part of the bill — the distinction `fmtCost` marks on a single turn, held
|
|
767
|
-
* across the session.
|
|
768
|
-
*/
|
|
769
|
-
|
|
770
|
-
/** A session with nothing accounted for yet. */
|
|
771
|
-
function emptyRoll() {
|
|
772
|
-
return {
|
|
773
|
-
turns: 0,
|
|
774
|
-
calls: 0,
|
|
775
|
-
tokens: 0,
|
|
776
|
-
input: 0,
|
|
777
|
-
output: 0,
|
|
778
|
-
cacheRead: 0,
|
|
779
|
-
cacheWrite: 0,
|
|
780
|
-
/** Dispatches that reported no usage at all — the honest gap in `tokens`. */
|
|
781
|
-
unknown: 0,
|
|
782
|
-
/**
|
|
783
|
-
* Exchanges folded from TEXT rather than reported usage, on the CLI's
|
|
784
|
-
* `real: false` rule. They are counted in `tokens` — that is the point —
|
|
785
|
-
* and named here so an estimated figure is never read as a measured one.
|
|
786
|
-
*/
|
|
787
|
-
estimated: 0,
|
|
788
|
-
/** Settled charges (server-settled `costUsd`), rolled. */
|
|
789
|
-
cost: 0,
|
|
790
|
-
/** Locally priced turns, rolled. Never mixed into `cost`. */
|
|
791
|
-
estimate: 0,
|
|
792
|
-
};
|
|
793
|
-
}
|
|
794
|
-
|
|
795
|
-
/**
|
|
796
|
-
* Fold one completed dispatch into a session's rolling tallies. Pure: it
|
|
797
|
-
* returns a NEW roll and never mutates the one it was handed, so a half-applied
|
|
798
|
-
* fold cannot exist.
|
|
799
|
-
*
|
|
800
|
-
* @param {object} [roll] the roll so far (emptyRoll() when omitted)
|
|
801
|
-
* @param {object} [usage] the response's `usage` object
|
|
802
|
-
* @param {{model?: string, costUsd?: number, calls?: number, turns?: number,
|
|
803
|
-
* prompt?: string, reply?: string, estimated?: boolean}} [opts]
|
|
804
|
-
* `calls` defaults to 1; a pooled turn may report how many provider
|
|
805
|
-
* calls it actually made. `turns: 0` folds a dispatch that is not a
|
|
806
|
-
* turn of its own — the discovery-lane card, which bills like any other
|
|
807
|
-
* call but is not something the user asked for. `prompt`/`reply` are the
|
|
808
|
-
* turn's text, used ONLY when the wire reported no usage, so the turn is
|
|
809
|
-
* estimated rather than dropped (the CLI's `appendHistory` rule).
|
|
810
|
-
* @returns {object} the new roll
|
|
811
|
-
*/
|
|
812
|
-
function rollTurn(roll, usage, opts = {}) {
|
|
813
|
-
const next = Object.assign(emptyRoll(), roll || {});
|
|
814
|
-
next.turns += opts.turns === undefined ? 1 : Number(opts.turns) || 0;
|
|
815
|
-
next.calls += opts.calls === undefined ? 1 : Number(opts.calls) || 0;
|
|
816
|
-
const turn = turnAccounting(usage, opts.model, { costUsd: opts.costUsd });
|
|
817
|
-
// A row folded from text rather than from the wire: either this turn's own
|
|
818
|
-
// fallback below, or a stored ledger row carrying `real: false` — the shape
|
|
819
|
-
// the CLI's `appendHistory` writes for an exchange the provider did not
|
|
820
|
-
// report on.
|
|
821
|
-
const estimated = opts.estimated === true || (usage && usage.real === false);
|
|
822
|
-
if (turn.tokens == null) {
|
|
823
|
-
// No reported usage. The CLI does not let such a turn vanish from the
|
|
824
|
-
// total: `appendHistory` estimates it from the text and marks it
|
|
825
|
-
// `real: false` (the upstream engine's history module), and
|
|
826
|
-
// `aggregateSessionUsage` then adds it like any other row. Folding the same
|
|
827
|
-
// estimate here is what makes the live roll and the rebuilt roll the same
|
|
828
|
-
// number, and what stops the total standing still on exactly the turns the
|
|
829
|
-
// pool declined to report on — the symptom this fallback exists to remove.
|
|
830
|
-
const est = estimatedBuckets(opts.prompt, opts.reply, opts.reasoning);
|
|
831
|
-
// Thinking tokens the provider MEASURED, not estimated. On a subscription
|
|
832
|
-
// turn the deliberation text is withheld entirely (every real
|
|
833
|
-
// `thinking_delta` carries `thinking: ""`), so `estimatedBuckets` above
|
|
834
|
-
// cannot see the longest, priciest part of the turn and the live preview
|
|
835
|
-
// read 0 for the whole thinking phase. Added verbatim — it is already a
|
|
836
|
-
// token count, and it is billed output exactly as the reply is.
|
|
837
|
-
const measured =
|
|
838
|
-
typeof opts.measuredReasoningTokens === 'number' && opts.measuredReasoningTokens > 0
|
|
839
|
-
? opts.measuredReasoningTokens
|
|
840
|
-
: 0;
|
|
841
|
-
if (!est) {
|
|
842
|
-
// A subscription turn with nothing written yet but thinking under way is
|
|
843
|
-
// NOT the `unknown` case: the provider measured it, so a zero would be a
|
|
844
|
-
// fabricated reading rather than an absent one.
|
|
845
|
-
if (!measured) {
|
|
846
|
-
// Nothing reported AND no text to estimate from. The single case that
|
|
847
|
-
// stays uncounted: `unknown` names the gap, where a fabricated zero
|
|
848
|
-
// would read as a measurement.
|
|
849
|
-
next.unknown += 1;
|
|
850
|
-
return next;
|
|
851
|
-
}
|
|
852
|
-
next.tokens += measured;
|
|
853
|
-
next.output += measured;
|
|
854
|
-
next.estimated += 1;
|
|
855
|
-
return next;
|
|
856
|
-
}
|
|
857
|
-
next.tokens += est.input + est.output + measured;
|
|
858
|
-
next.input += est.input;
|
|
859
|
-
next.output += est.output + measured;
|
|
860
|
-
next.estimated += 1;
|
|
861
|
-
return next;
|
|
862
|
-
}
|
|
863
|
-
const b = usageBuckets(usage);
|
|
864
|
-
next.tokens += turn.tokens;
|
|
865
|
-
next.input += b.input;
|
|
866
|
-
next.output += b.output;
|
|
867
|
-
next.cacheRead += b.cacheRead;
|
|
868
|
-
next.cacheWrite += b.cacheWrite;
|
|
869
|
-
if (estimated) {
|
|
870
|
-
// Money stays out of an estimated exchange, deliberately and in both
|
|
871
|
-
// directions: the CLI writes no `costUsd` for one, so pricing it here would
|
|
872
|
-
// make the live roll drift from the rebuilt one — and pricing tokens that
|
|
873
|
-
// were themselves guessed would stack one guess on another.
|
|
874
|
-
next.estimated += 1;
|
|
875
|
-
return next;
|
|
876
|
-
}
|
|
877
|
-
if (turn.real) next.cost += turn.cost;
|
|
878
|
-
else if (turn.cost != null) next.estimate += turn.cost;
|
|
879
|
-
return next;
|
|
880
|
-
}
|
|
881
|
-
|
|
882
|
-
/**
|
|
883
|
-
* Rebuild a session's rolling total from stored exchanges — the desktop's
|
|
884
|
-
* counterpart of the CLI's `aggregateSessionUsage`, which sums history.jsonl so
|
|
885
|
-
* a resumed session (and a compacted one) still reports everything it spent.
|
|
886
|
-
*
|
|
887
|
-
* Reads the shapes the shared store writes: an assistant message carrying
|
|
888
|
-
* `tokens: {input, output, cacheRead, cacheWrite}` and, when the pool settled
|
|
889
|
-
* the turn, `costUsd` (cli/src/history.js → session-store.recordExchange).
|
|
890
|
-
* Messages the desktop itself appended carry no `tokens` and fold as `unknown`
|
|
891
|
-
* — a resumed thread states what is known and does not invent the rest.
|
|
892
|
-
*
|
|
893
|
-
* @param {Array<{role?: string, tokens?: object, costUsd?: number, model?: string}>} [messages]
|
|
894
|
-
* @returns {object} the roll
|
|
895
|
-
*/
|
|
896
|
-
function rollMessages(messages) {
|
|
897
|
-
let roll = emptyRoll();
|
|
898
|
-
for (const m of Array.isArray(messages) ? messages : []) {
|
|
899
|
-
if (!m || m.role !== 'assistant') continue;
|
|
900
|
-
roll = rollTurn(roll, m.tokens, {
|
|
901
|
-
model: m.model,
|
|
902
|
-
costUsd: typeof m.costUsd === 'number' ? m.costUsd : undefined,
|
|
903
|
-
calls: m.calls,
|
|
904
|
-
});
|
|
905
|
-
}
|
|
906
|
-
return roll;
|
|
907
|
-
}
|
|
908
|
-
|
|
909
|
-
/**
|
|
910
|
-
* The ledger row one finished dispatch must carry into the shared session
|
|
911
|
-
* store, so the rolling total can be REBUILT when the thread is reopened —
|
|
912
|
-
* the desktop's counterpart of the CLI's `appendHistory`
|
|
913
|
-
* (the upstream engine's history module).
|
|
914
|
-
*
|
|
915
|
-
* One authority for the row, on purpose. The CLI writes a `tokens` object for
|
|
916
|
-
* EVERY exchange and skips none:
|
|
917
|
-
*
|
|
918
|
-
* tokens: usage
|
|
919
|
-
* ? { input, output, cacheRead, cacheWrite, real: true }
|
|
920
|
-
* : { input: estimateTokens(prompt), output: estimateTokens(reply), real: false }
|
|
921
|
-
*
|
|
922
|
-
* …and `aggregateSessionUsage` then sums `t.input || 0` over every entry with
|
|
923
|
-
* `real` as a FLAG, not a gate. Mirroring that shape here is what makes the
|
|
924
|
-
* live roll and the rebuilt roll the same number: `rollTurn` is handed this
|
|
925
|
-
* exact object on reopen, so an estimated exchange adds the same buckets it
|
|
926
|
-
* added live and is counted under `estimated` in both.
|
|
927
|
-
*
|
|
928
|
-
* Returns `null` when there is neither reported usage nor text to estimate
|
|
929
|
-
* from — the one case that must stay unrecorded, because a fabricated
|
|
930
|
-
* `{input: 0, output: 0}` row would read as a measured zero forever after.
|
|
931
|
-
*
|
|
932
|
-
* @param {object} [usage] the response's `usage` object
|
|
933
|
-
* @param {object} [turn] its accounting (turnAccounting), when already done
|
|
934
|
-
* @param {{model?: string, costUsd?: number, calls?: number, prompt?: string,
|
|
935
|
-
* reply?: string}} [opts]
|
|
936
|
-
* @returns {object|null} the row's ledger fields
|
|
937
|
-
*/
|
|
938
|
-
function ledgerRow(usage, turn, opts = {}) {
|
|
939
|
-
const t = turn || turnAccounting(usage, opts.model, { costUsd: opts.costUsd });
|
|
940
|
-
const calls = Number.isFinite(opts.calls) && opts.calls > 0 ? opts.calls : undefined;
|
|
941
|
-
if (t.tokens != null) {
|
|
942
|
-
const row = { tokens: usageBuckets(usage) };
|
|
943
|
-
// Only a SETTLED charge is persisted. The CLI writes `costUsd` for a real
|
|
944
|
-
// charge only, and the rebuild prices an unpriced row from the same rate
|
|
945
|
-
// table this window used — persisting a local guess would let a stale
|
|
946
|
-
// table outlive the change that wrote it.
|
|
947
|
-
if (t.real && t.cost != null) row.costUsd = t.cost;
|
|
948
|
-
if (calls !== undefined) row.calls = calls;
|
|
949
|
-
return row;
|
|
950
|
-
}
|
|
951
|
-
// `reasoning` is part of the billed output, so it is estimated with the
|
|
952
|
-
// reply — a stored row for a thinking-heavy turn must not persist a count
|
|
953
|
-
// that excludes the longest thing the model wrote.
|
|
954
|
-
const est = estimatedBuckets(opts.prompt, opts.reply, opts.reasoning);
|
|
955
|
-
if (!est) return null;
|
|
956
|
-
const row = { tokens: est };
|
|
957
|
-
if (calls !== undefined) row.calls = calls;
|
|
958
|
-
return row;
|
|
959
|
-
}
|
|
960
|
-
|
|
961
|
-
/**
|
|
962
|
-
* The CLI's rendering of a session tally — `cli/src/format.js fmtTokens`, which
|
|
963
|
-
* is the one `tokenSummary` actually imports (`cli/src/app.js:50`). NOT the
|
|
964
|
-
* `1.5k`/`12.3k` form in `cli/src/tokens.js`: that one belongs to the /cost
|
|
965
|
-
* panels, and using it here would print `12.4k` on the very total the CLI
|
|
966
|
-
* prints as `12,400` — a rendering difference stacked on top of the accounting
|
|
967
|
-
* difference this change exists to remove.
|
|
968
|
-
*
|
|
969
|
-
* Comma-grouped integer. Anything not finite and positive renders `0`, which is
|
|
970
|
-
* the CLI's rule and keeps a stray NaN from reaching the topbar.
|
|
971
|
-
*/
|
|
972
|
-
function fmtTokens(n) {
|
|
973
|
-
const v = Number(n);
|
|
974
|
-
if (!Number.isFinite(v) || v <= 0) return '0';
|
|
975
|
-
return Math.round(v)
|
|
976
|
-
.toString()
|
|
977
|
-
.replace(/\B(?=(\d{3})+(?!\d))/g, ',');
|
|
978
|
-
}
|
|
979
|
-
|
|
980
|
-
/**
|
|
981
|
-
* The token figure a person is SHOWN for a finished turn or a session roll.
|
|
982
|
-
*
|
|
983
|
-
* One authority, because there are two shapes that reach it and they had
|
|
984
|
-
* drifted apart: `turnAccounting` (a single dispatch, `{tokens, output}`) and
|
|
985
|
-
* `emptyRoll`/`rollTurn` (a session tally, same two fields). Every per-turn row
|
|
986
|
-
* and the topbar both route through here, so no surface can print a different
|
|
987
|
-
* quantity than another.
|
|
988
|
-
*
|
|
989
|
-
* What it returns is the OUTPUT — what the turn produced. The wire's total is
|
|
990
|
-
* input + output, and on any cached or long-context turn the input half
|
|
991
|
-
* dominates it: a 3,145-token call whose prompt was 3,000 tokens is a turn that
|
|
992
|
-
* wrote 145. Showing 3,145 beside a `session:` line that also reads 3,145 told
|
|
993
|
-
* the reader the model had emitted ten times what it did, and there is no way
|
|
994
|
-
* to tell from the row which of the two numbers is the prompt restated. The
|
|
995
|
-
* CLI settled this the same way (`fmtUsageSplit`, cli/src/format.js): the figure
|
|
996
|
-
* is the output, and the input half is not shown unlabelled beside it.
|
|
997
|
-
*
|
|
998
|
-
* It does NOT fall back to the total. Fall back to `tokens` and the input half
|
|
999
|
-
* is back on the screen with no label: a wire that priced a turn but did not
|
|
1000
|
-
* itemise it would print `3,145` for a turn that wrote 145. Returns `null`
|
|
1001
|
-
* rather than `0` in that case — `0 tokens` is a figure no provider sent,
|
|
1002
|
-
* and this file refuses that fabrication everywhere else (`fmtRoll`).
|
|
1003
|
-
*
|
|
1004
|
-
* @param {{tokens?: number|null, output?: number|null}} [source]
|
|
1005
|
-
* @returns {number|null}
|
|
1006
|
-
*/
|
|
1007
|
-
function displayTokens(source) {
|
|
1008
|
-
if (!source || typeof source !== 'object') return null;
|
|
1009
|
-
// The half the turn PRODUCED, from either shape this file is handed: an
|
|
1010
|
-
// accounting or roll record (`output`), or a stored token bucket
|
|
1011
|
-
// (`tokens: {output}` — what `rollMessages` reads back out of the shared
|
|
1012
|
-
// store, and what a rebuilt session presents).
|
|
1013
|
-
const nested = source.tokens;
|
|
1014
|
-
const candidates = [source.output];
|
|
1015
|
-
if (nested && typeof nested === 'object') {
|
|
1016
|
-
candidates.push(nested.output, nested.output_tokens, nested.completion_tokens);
|
|
1017
|
-
} else {
|
|
1018
|
-
candidates.push(source.output_tokens, source.completion_tokens);
|
|
1019
|
-
}
|
|
1020
|
-
for (const c of candidates) {
|
|
1021
|
-
if (typeof c === 'number' && isFinite(c) && c > 0) return c;
|
|
1022
|
-
}
|
|
1023
|
-
// The wire's TOTAL is deliberately NOT a fallback here. `tokens` is
|
|
1024
|
-
// input + output, and on a cached turn the input half is the larger one, so
|
|
1025
|
-
// falling back to it put back the exact reading this rule exists to remove:
|
|
1026
|
-
// `tokens: 3,145` on a turn that wrote 145. This was the last surviving
|
|
1027
|
-
// place the input half reached the screen — the split had been dropped, and
|
|
1028
|
-
// the total it fell back to still carried it.
|
|
1029
|
-
//
|
|
1030
|
-
// Stated once more, because it is the whole point: a turn the wire priced but
|
|
1031
|
-
// did not itemise has NO output figure, and inventing one from the total
|
|
1032
|
-
// would be a number no provider sent. That turn's row states its own text
|
|
1033
|
-
// estimate instead (`~N tokens`), which is at least a figure about what was
|
|
1034
|
-
// written.
|
|
1035
|
-
return null;
|
|
1036
|
-
}
|
|
1037
|
-
|
|
1038
|
-
/**
|
|
1039
|
-
* The rolling session total as one line — the desktop's counterpart of the
|
|
1040
|
-
* CLI's `tokenSummary`. Empty string when nothing has been accounted for, so
|
|
1041
|
-
* an untouched session adds no noise to a turn's meta line.
|
|
1042
|
-
*
|
|
1043
|
-
* @param {object} [roll]
|
|
1044
|
-
* @returns {string} e.g. `2,300 tokens · 4 calls · $0.0310` — the CLI's fields in
|
|
1045
|
-
* the CLI's order, minus the input/output parenthetical (see below). The
|
|
1046
|
-
* figure is the OUTPUT (see `displayTokens`).
|
|
1047
|
-
*/
|
|
1048
|
-
function fmtRoll(roll, opts = {}) {
|
|
1049
|
-
const r = roll || emptyRoll();
|
|
1050
|
-
if (!r.turns && !r.calls) return '';
|
|
1051
|
-
// The fields are `tokenSummary`'s (cli/src/app.js:1413) in its order — tokens,
|
|
1052
|
-
// calls, money — with one dropped: the `(10,100 in / 2,300 out)` split. `in`
|
|
1053
|
-
// and `out` are a terminal status-line shorthand, legible to someone already
|
|
1054
|
-
// reading that status line and to nobody else, which is what an unlabelled
|
|
1055
|
-
// abbreviation in a GUI topbar turns into.
|
|
1056
|
-
//
|
|
1057
|
-
// …and the figure that remains is the OUTPUT, not the wire's total. This line
|
|
1058
|
-
// used to print `r.tokens`, which is input + output: on a cached turn the
|
|
1059
|
-
// prompt restated is the larger half, so a session that wrote 2,300 tokens on
|
|
1060
|
-
// a 10,100-token prompt read `12,400 tokens` and looked like a model emitting
|
|
1061
|
-
// five times what it did. The parenthetical was not the only place the input
|
|
1062
|
-
// half showed up — it was in the total too, and dropping the split alone left
|
|
1063
|
-
// the reader comparing two numbers that were both mostly their own prompt.
|
|
1064
|
-
// `displayTokens` holds the rule, the same one the CLI's `fmtUsageSplit` uses.
|
|
1065
|
-
// `r.input` is still folded onto the roll (see `rollTurn`) for any surface
|
|
1066
|
-
// that wants to show the split with real labels.
|
|
1067
|
-
const bits = [];
|
|
1068
|
-
// The token half only when something was actually counted. `rollTurn` counts
|
|
1069
|
-
// turns and calls BEFORE it looks at the token count, so a dispatch that
|
|
1070
|
-
// reported nothing still reaches here — and printing an empty tally would put
|
|
1071
|
-
// `0 tokens` on the topbar, a figure the meter never took, which reads as a
|
|
1072
|
-
// counter that does not move. `displayTokens` returns null rather than 0 for
|
|
1073
|
-
// exactly that case, and falls back to the total only when no split arrived.
|
|
1074
|
-
// The turn count is still stated, because that much is true.
|
|
1075
|
-
const produced = displayTokens(r);
|
|
1076
|
-
if (produced != null) {
|
|
1077
|
-
bits.push(`${fmtTokens(produced)} tokens`);
|
|
1078
|
-
}
|
|
1079
|
-
bits.push(`${r.calls} call${r.calls === 1 ? '' : 's'}`);
|
|
1080
|
-
// Money is where this line departs from `tokenSummary`, deliberately: that
|
|
1081
|
-
// one sums a single `session.cost` in EUR via fmtEur, while this surface keeps
|
|
1082
|
-
// a settled charge and a local estimate apart so a `~`-estimate can never be
|
|
1083
|
-
// read as part of the bill. Desktop's pre-existing fmtCost renders both, and
|
|
1084
|
-
// its `$` convention is left exactly as it was.
|
|
1085
|
-
// Money is omitted entirely on a tier AEGIS does not bill a fee to
|
|
1086
|
-
// (`opts.omitMoney` — see `feeWaived`). Not zeroed, not marked `~`: a `$0.0000`
|
|
1087
|
-
// states a price, and this account has no price on this turn.
|
|
1088
|
-
if (!opts.omitMoney && (r.cost > 0 || r.estimate > 0)) {
|
|
1089
|
-
const money = [];
|
|
1090
|
-
if (r.cost > 0) money.push(fmtCost(r.cost, true));
|
|
1091
|
-
if (r.estimate > 0) money.push(fmtCost(r.estimate, false));
|
|
1092
|
-
bits.push(money.join(' + '));
|
|
1093
|
-
}
|
|
1094
|
-
// No counterpart in `tokenSummary`, which folds a usage-less turn silently and
|
|
1095
|
-
// so reports a total short of the truth without saying so. Named here instead:
|
|
1096
|
-
// the count appears only when something went unreported, and it never changes
|
|
1097
|
-
// a number — it only says the number is not the whole story.
|
|
1098
|
-
if (r.unknown) bits.push(`${r.unknown} unrpt`);
|
|
1099
|
-
// The other half of the same honesty: a total that includes estimated
|
|
1100
|
-
// exchanges says so, because the CLI marks the same rows `real: false` and a
|
|
1101
|
-
// reader is entitled to know which figure they are looking at. It changes no
|
|
1102
|
-
// number — it says the number is partly inferred.
|
|
1103
|
-
if (r.estimated) bits.push(`${r.estimated} est`);
|
|
1104
|
-
return bits.join(' · ');
|
|
1105
|
-
}
|
|
1106
|
-
|
|
1107
|
-
/**
|
|
1108
|
-
* A cost for display. Estimates are marked with `~` so an estimate is never
|
|
1109
|
-
* mistaken for a settled charge.
|
|
1110
|
-
*
|
|
1111
|
-
* @param {number} cost
|
|
1112
|
-
* @param {boolean} [real]
|
|
1113
|
-
* @returns {string}
|
|
1114
|
-
*/
|
|
1115
|
-
function fmtCost(cost, real) {
|
|
1116
|
-
if (typeof cost !== 'number' || !isFinite(cost)) return '';
|
|
1117
|
-
// The `$` is not decoration: its loss was a regression, not a decision.
|
|
1118
|
-
// `4d5d1e9` introduced this function with it (`~$18.0000`); `6d48470` — a
|
|
1119
|
-
// commit about the token TOTAL, unrelated to money — dropped it as a stray
|
|
1120
|
-
// hunk. Every desktop turn row then painted a bare `18.0000`, a number with no
|
|
1121
|
-
// currency on it, where the CLI paints `€0.03` for the same figure.
|
|
1122
|
-
return `${real ? '' : '~'}$${cost.toFixed(4)}`;
|
|
1123
|
-
}
|
|
1124
|
-
|
|
1125
|
-
/**
|
|
1126
|
-
* True when this account's tier means AEGIS bills no per-turn fee, so no money
|
|
1127
|
-
* figure belongs on a turn row or a session roll.
|
|
1128
|
-
*
|
|
1129
|
-
* The fee in question is the BYOK handling fee — the only charge AEGIS adds on
|
|
1130
|
-
* top of a caller's own provider bill — and aegis1 waives it on Unlimited:
|
|
1131
|
-
* `lib/local/entitlement.js` states the rule in as many words ("the throughput
|
|
1132
|
-
* allowance and the BYOK fee waiver are `is_unlimited()`-gated"), and the tier
|
|
1133
|
-
* a welcome trial grants IS Unlimited (a 30-day `unlimited` row in
|
|
1134
|
-
* `subscriptions`, which `register_device` writes in the same request that mints
|
|
1135
|
-
* the key). So a trial account resolves here exactly like a paying one, which is
|
|
1136
|
-
* what "trial unlimited" means — the trial is not a separate tier with its own
|
|
1137
|
-
* money rule, it is Unlimited for 30 days.
|
|
1138
|
-
*
|
|
1139
|
-
* Matching is on the resolved tier rather than `pro`: `pro` is "the PLAN is paid
|
|
1140
|
-
* for" and is also true of the €9 tier's aliases, while the waiver is gated on
|
|
1141
|
-
* `unlimited` specifically. A funded bank on a free plan is not a tier at all
|
|
1142
|
-
* (`entitlement.js` refuses to name one) and is therefore NOT waived — that
|
|
1143
|
-
* account pays the fee from the bank, and hiding the number would leave it with
|
|
1144
|
-
* no meter for the only money it spends.
|
|
1145
|
-
*
|
|
1146
|
-
* @param {{tier?: string|null, plan?: string|null, planRaw?: string|null}} [entitlement]
|
|
1147
|
-
* @returns {boolean}
|
|
1148
|
-
*/
|
|
1149
|
-
function feeWaived(entitlement) {
|
|
1150
|
-
if (!entitlement || typeof entitlement !== 'object') return false;
|
|
1151
|
-
const named = [
|
|
1152
|
-
entitlement.tier,
|
|
1153
|
-
entitlement.plan,
|
|
1154
|
-
entitlement.planRaw,
|
|
1155
|
-
entitlement.subscription,
|
|
1156
|
-
entitlement.subscriptionTier,
|
|
1157
|
-
entitlement.product,
|
|
1158
|
-
// The server's own two fields, as `register_device` writes them: a welcome
|
|
1159
|
-
// trial is a 30-day `unlimited` row in `subscriptions`, and `item`/`sku` is
|
|
1160
|
-
// that row's product name. Naming them here is what keeps a trial account
|
|
1161
|
-
// from resolving to "no tier" and being billed a fee it does not owe.
|
|
1162
|
-
entitlement.item,
|
|
1163
|
-
entitlement.sku,
|
|
1164
|
-
]
|
|
1165
|
-
.filter((v) => typeof v === 'string')
|
|
1166
|
-
.map((v) => v.trim().toLowerCase());
|
|
1167
|
-
// An explicit boolean outranks every spelling below: aegis1's `is_unlimited()`
|
|
1168
|
-
// verdict is the fact, and the strings are only what it is written as on the
|
|
1169
|
-
// wire. A probe that answers with the bit and not the tier is still an answer.
|
|
1170
|
-
if (entitlement.unlimited === true) return true;
|
|
1171
|
-
if (named.some((v) => v === 'unlimited' || v === 'trial')) return true;
|
|
1172
|
-
// A tier string that merely CONTAINS the word — `unlimited_trial`,
|
|
1173
|
-
// `plan.unlimited`, `Unlimited (trial)` — so a server that decorates the name
|
|
1174
|
-
// cannot put the fee back on an account whose tier is Unlimited.
|
|
1175
|
-
return named.some((v) => /(^|[^a-z])unlimited([^a-z]|$)/.test(v));
|
|
1176
|
-
}
|
|
1177
|
-
|
|
1178
|
-
/**
|
|
1179
|
-
* Whether a money figure belongs on this account's turn row or session roll.
|
|
1180
|
-
*
|
|
1181
|
-
* The CLI's exact rule (`cli/src/format.js` → `feeBillable`), and it is the same
|
|
1182
|
-
* rule on purpose — the two hosts render the same turn from the same engine, so
|
|
1183
|
-
* a row that shows a fee on one and waives it on the other is a disagreement
|
|
1184
|
-
* about what the user owes.
|
|
1185
|
-
*
|
|
1186
|
-
* A price is printed UNLESS AEGIS either waives it or has not proven it owed:
|
|
1187
|
-
*
|
|
1188
|
-
* - a tier AEGIS waives (`feeWaived`) → no price. Unlimited, paid or on the
|
|
1189
|
-
* 30-day welcome trial, has no per-turn charge and no bank draw;
|
|
1190
|
-
* - a snapshot asserting a plan it cannot prove (`degraded`) → no price. The
|
|
1191
|
-
* entitlement module STARTS at exactly that: `makeFreeSnapshot()` reads
|
|
1192
|
-
* `plan: 'free'`, `tier: null`, `source: 'unresolved'`, `degraded: true`,
|
|
1193
|
-
* which says "no verdict yet", not "this account is free". Reading the
|
|
1194
|
-
* placeholder as a verdict is what put `~$0.0005` and an "AEGIS handling fee"
|
|
1195
|
-
* line on the turns of trial accounts whose probe had not answered — a fee
|
|
1196
|
-
* charged in writing to the one account that does not pay it;
|
|
1197
|
-
* - everything else → the price stands, including when no snapshot exists at
|
|
1198
|
-
* all. A cost figure is a measurement of the turn, and dropping a real number
|
|
1199
|
-
* because no tier was published would leave a self-hosted or BYO-key run with
|
|
1200
|
-
* no meter for money it actually spent.
|
|
1201
|
-
*
|
|
1202
|
-
* The token figure is never gated by this, on any branch: what a turn produced
|
|
1203
|
-
* is not a price, and hiding it would leave an Unlimited account with no meter.
|
|
1204
|
-
*
|
|
1205
|
-
* @param {{tier?: string|null, plan?: string|null, planRaw?: string|null,
|
|
1206
|
-
* degraded?: boolean, source?: string|null, unlimited?: boolean}} [entitlement]
|
|
1207
|
-
* @returns {boolean}
|
|
1208
|
-
*/
|
|
1209
|
-
function feeBillable(entitlement) {
|
|
1210
|
-
if (!entitlement || typeof entitlement !== 'object') return true;
|
|
1211
|
-
if (feeWaived(entitlement)) return false;
|
|
1212
|
-
return entitlement.degraded !== true;
|
|
1213
|
-
}
|
|
1214
|
-
|
|
1215
|
-
if (typeof module !== 'undefined' && module.exports) {
|
|
1216
|
-
module.exports = {
|
|
1217
|
-
feeWaived,
|
|
1218
|
-
feeBillable,
|
|
1219
|
-
RATES,
|
|
1220
|
-
RATE_BASIS,
|
|
1221
|
-
UNPRICED_BY_DESIGN,
|
|
1222
|
-
isPeak,
|
|
1223
|
-
usageTokens,
|
|
1224
|
-
ratesFor,
|
|
1225
|
-
usageBuckets,
|
|
1226
|
-
cacheHitRate,
|
|
1227
|
-
usageCost,
|
|
1228
|
-
costBreakdown,
|
|
1229
|
-
turnAccounting,
|
|
1230
|
-
displayTokens,
|
|
1231
|
-
estimateTokens,
|
|
1232
|
-
estimatedBuckets,
|
|
1233
|
-
emptyRoll,
|
|
1234
|
-
rollTurn,
|
|
1235
|
-
rollMessages,
|
|
1236
|
-
ledgerRow,
|
|
1237
|
-
fmtTokens,
|
|
1238
|
-
fmtRoll,
|
|
1239
|
-
fmtCost,
|
|
1240
|
-
};
|
|
1241
|
-
}
|
|
3
|
+
"use strict";function usageTokens(e){if(!e||"object"!=typeof e)return null;if("number"==typeof e.total_tokens)return e.total_tokens;const t=e.input_tokens??e.prompt_tokens??e.input,o=e.output_tokens??e.completion_tokens??e.output;return"number"!=typeof t&&"number"!=typeof o?null:(t||0)+(o||0)}function estimateTokens(e){return e?Math.max(1,Math.ceil([...String(e)].length/4)):0}function estimatedBuckets(e,t,o){const n=("string"==typeof t?t:"")+("string"==typeof o?o:"");return"string"==typeof e&&e.length>0||n.length>0?{input:estimateTokens(e),output:estimateTokens(n),cacheRead:0,cacheWrite:0,real:!1}:null}const RATES={sonnet:{input:3,output:15,cacheRead:.3,cacheWrite:3.75},default:{input:3,output:15,cacheRead:.3,cacheWrite:3.75},fable:{input:10,output:50,cacheRead:1,cacheWrite:12.5},opus:{input:5,output:25,cacheRead:.5,cacheWrite:6.25},haiku:{input:1,output:5,cacheRead:.1,cacheWrite:1.25},"claude-haiku":{input:1,output:5,cacheRead:.1,cacheWrite:1.25},"claude-opus-5":{input:5,output:25,cacheRead:.5,cacheWrite:6.25},"claude-opus-5-5":{input:4,output:20,cacheRead:.2,cacheWrite:5},"claude-sonnet-5":{input:2,output:10,cacheRead:.2,cacheWrite:2.5},"claude-sonnet-5-5":{input:2,output:10,cacheRead:.1,cacheWrite:2.5},"claude-haiku-5-5":{input:.1,output:.5,cacheRead:.01,cacheWrite:.125},"claude-fable-5":{input:10,output:50,cacheRead:1,cacheWrite:12.5},"claude-fable-5-1":{input:10,output:50,cacheRead:.25,cacheWrite:12.5},"claude-sonnet-4-6":{input:3,output:15,cacheRead:.3,cacheWrite:3.75},"claude-sonnet-4-5":{input:3,output:15,cacheRead:.3,cacheWrite:3.75},deepseek:{input:.3,output:1.2,cacheRead:.006,cacheWrite:.3,offPeak:{input:.15,output:.6,cacheRead:.003,cacheWrite:.15}},"deepseek-v4-pro":{input:1.32,output:3.96,cacheRead:.044,cacheWrite:1.32,offPeak:{input:.66,output:1.98,cacheRead:.022,cacheWrite:.66}},"deepseek-v4.1-pro":{input:1.32,output:3.96,cacheRead:.044,cacheWrite:1.32,offPeak:{input:.66,output:1.98,cacheRead:.022,cacheWrite:.66}},"grok-4.7":{input:2,output:6,cacheRead:.5,cacheWrite:2},"gemini-3.8-flash":{input:.75,output:3.75,cacheRead:.075,cacheWrite:.75},"gpt-5.5":{input:5,output:30,cacheRead:.5,cacheWrite:5}};function isPeak(e,t=new Date){if(0!==String(e||"").toLowerCase().indexOf("deepseek"))return!0;const o=t.getUTCDay();if(0===o||6===o)return!1;const n=t.getUTCHours();return n>=1&&n<4||n>=6&&n<10}const RATE_BASIS=Object.freeze({EXACT:"exact",PREFIX:"prefix",FAMILY:"family",DEFAULT:"default"}),RATE_FAMILIES=Object.keys(RATES).filter((e=>"default"!==e)),UNPRICED_BY_DESIGN=Object.freeze({"nexus-brain":"pooled: multi-call fan-out, aegis1 settles the charge","aegis-brain":"pooled alias of nexus-brain","nexus-brain-smart":"pooled alias of nexus-brain","nexus-brain-neo":"pooled alias of nexus-brain","aegis-brain-smart":"pooled alias of nexus-brain","aegis-brain-neo":"pooled alias of nexus-brain","openai-gpt4o-mini":"pooled seat, aegis1 settles the charge",anthropic:"provider name, not a model id",openai:"provider name, not a model id",groq:"provider name, not a model id",google:"provider name, not a model id",xai:"provider name, not a model id"});function ratesFor(e,{at:t}={}){const o=String(e||"").toLowerCase(),n=(e,n)=>{const s=isPeak(0===o.indexOf("deepseek")||o.indexOf(":deepseek")>-1?"deepseek":o,t||new Date);return{...!s&&e.offPeak?{...e,...e.offPeak}:e,basis:n,peak:s,model:o}};if(!o)return n(RATES.default,RATE_BASIS.DEFAULT);const s=[o],a=Math.max(o.lastIndexOf(":"),o.lastIndexOf("/")),u=o.slice(a+1);u&&u!==o&&s.push(u);for(const e of s)if(RATES[e])return n(RATES[e],RATE_BASIS.EXACT);let c=null,i=0;for(const e of s)for(const[t,o]of Object.entries(RATES))"default"!==t&&e.startsWith(t)&&t.length>i&&(c=o,i=t.length);if(c)return n(c,RATE_BASIS.PREFIX);let r=null,p=0;for(const e of RATE_FAMILIES)o.includes(e)&&e.length>p&&(r=RATES[e],p=e.length);return n(r||RATES.default,r?RATE_BASIS.FAMILY:RATE_BASIS.DEFAULT)}function usageBuckets(e){const t=e&&"object"==typeof e?e:{},o=(...e)=>{for(const t of e)if("number"==typeof t)return t;return 0},n=e=>"number"==typeof e?e:void 0,s=n(t.input_tokens),a=n(t.prompt_tokens),u=t.prompt_tokens_details,c=u&&"object"==typeof u?n(u.cached_tokens):void 0,i=o(t.prompt_cache_hit_tokens,c,t.cached_tokens),r=void 0===s&&void 0!==a,p=r?a:o(s,a,t.input),l=o(t.cache_read_input_tokens,t.cacheRead,i),d=r?Math.min(Math.max(0,p),l):l,f=o(t.cache_creation_input_tokens,t.cacheWrite),m=o(t.output_tokens,t.completion_tokens,t.output),h=r?Math.max(0,p-d):p,k=o(t.total_tokens),R=h+m+d+f;return{input:k>R?h+(k-R):h,output:m,cacheRead:d,cacheWrite:f,cacheReported:"number"==typeof t.cache_read_input_tokens||"number"==typeof t.cacheRead||"number"==typeof t.cache_creation_input_tokens||"number"==typeof t.cacheWrite||"number"==typeof t.prompt_cache_hit_tokens||"number"==typeof t.cached_tokens||"number"==typeof c}}function cacheHitRate(e){const t=usageBuckets(e),o=Math.max(0,t.cacheRead||0),n=Math.max(0,t.input||0),s=o+n;return{read:o,miss:n,prompt:s,pct:t.cacheReported&&s>0?o/s*100:null,reported:t.cacheReported}}function costBreakdown(e,t,o={}){const n=ratesFor(t,o),s=usageBuckets(e),a=(e,t)=>Math.max(0,Number(e)||0)/1e6*t;return{cost:a(s.input,n.input)+a(s.output,n.output)+a(s.cacheRead,n.cacheRead)+a(s.cacheWrite,n.cacheWrite),basis:n.basis,peak:n.peak,model:n.model,priced:n.basis!==RATE_BASIS.DEFAULT}}function usageCost(e,t,o={}){return costBreakdown(e,t,o).cost}function turnAccounting(e,t,o={}){const n=usageTokens(e),s=e&&"object"==typeof e?e:{},a=e&&"object"==typeof e?usageBuckets(e).output:null,u="number"==typeof o.costUsd?o.costUsd:"number"==typeof s.costUsd?s.costUsd:void 0;if("number"==typeof u)return{tokens:n,output:a,cost:u,real:!0,estimated:!1,basis:"settled",priced:!0};if(null==n)return{tokens:n,output:a,cost:null,real:!1,estimated:!1,basis:null,priced:!1};const c=costBreakdown(s,t,o);return{tokens:n,output:a,cost:c.cost,real:!1,estimated:!0,basis:c.basis,priced:c.priced}}function emptyRoll(){return{turns:0,calls:0,tokens:0,input:0,output:0,cacheRead:0,cacheWrite:0,unknown:0,estimated:0,cost:0,estimate:0}}function rollTurn(e,t,o={}){const n=Object.assign(emptyRoll(),e||{});n.turns+=void 0===o.turns?1:Number(o.turns)||0,n.calls+=void 0===o.calls?1:Number(o.calls)||0;const s=turnAccounting(t,o.model,{costUsd:o.costUsd}),a=!0===o.estimated||t&&!1===t.real;if(null==s.tokens){const e=estimatedBuckets(o.prompt,o.reply,o.reasoning),t="number"==typeof o.measuredReasoningTokens&&o.measuredReasoningTokens>0?o.measuredReasoningTokens:0;return e?(n.tokens+=e.input+e.output+t,n.input+=e.input,n.output+=e.output+t,n.estimated+=1,n):t?(n.tokens+=t,n.output+=t,n.estimated+=1,n):(n.unknown+=1,n)}const u=usageBuckets(t);return n.tokens+=s.tokens,n.input+=u.input,n.output+=u.output,n.cacheRead+=u.cacheRead,n.cacheWrite+=u.cacheWrite,a?(n.estimated+=1,n):(s.real?n.cost+=s.cost:null!=s.cost&&(n.estimate+=s.cost),n)}function rollMessages(e){let t=emptyRoll();for(const o of Array.isArray(e)?e:[])o&&"assistant"===o.role&&(t=rollTurn(t,o.tokens,{model:o.model,costUsd:"number"==typeof o.costUsd?o.costUsd:void 0,calls:o.calls}));return t}function ledgerRow(e,t,o={}){const n=t||turnAccounting(e,o.model,{costUsd:o.costUsd}),s=Number.isFinite(o.calls)&&o.calls>0?o.calls:void 0;if(null!=n.tokens){const t={tokens:usageBuckets(e)};return n.real&&null!=n.cost&&(t.costUsd=n.cost),void 0!==s&&(t.calls=s),t}const a=estimatedBuckets(o.prompt,o.reply,o.reasoning);if(!a)return null;const u={tokens:a};return void 0!==s&&(u.calls=s),u}function fmtTokens(e){const t=Number(e);return!Number.isFinite(t)||t<=0?"0":Math.round(t).toString().replace(/\B(?=(\d{3})+(?!\d))/g,",")}function displayTokens(e){if(!e||"object"!=typeof e)return null;const t=e.tokens,o=[e.output];t&&"object"==typeof t?o.push(t.output,t.output_tokens,t.completion_tokens):o.push(e.output_tokens,e.completion_tokens);for(const e of o)if("number"==typeof e&&isFinite(e)&&e>0)return e;return null}function fmtRoll(e,t={}){const o=e||emptyRoll();if(!o.turns&&!o.calls)return"";const n=[],s=displayTokens(o);if(null!=s&&n.push(`${fmtTokens(s)} tokens`),n.push(`${o.calls} call${1===o.calls?"":"s"}`),!t.omitMoney&&(o.cost>0||o.estimate>0)){const e=[];o.cost>0&&e.push(fmtCost(o.cost,!0)),o.estimate>0&&e.push(fmtCost(o.estimate,!1)),n.push(e.join(" + "))}return o.unknown&&n.push(`${o.unknown} unrpt`),o.estimated&&n.push(`${o.estimated} est`),n.join(" · ")}function fmtCost(e,t){return"number"==typeof e&&isFinite(e)?`${t?"":"~"}$${e.toFixed(4)}`:""}function feeWaived(e){if(!e||"object"!=typeof e)return!1;const t=[e.tier,e.plan,e.planRaw,e.subscription,e.subscriptionTier,e.product,e.item,e.sku].filter((e=>"string"==typeof e)).map((e=>e.trim().toLowerCase()));return!0===e.unlimited||!!t.some((e=>"unlimited"===e||"trial"===e))||t.some((e=>/(^|[^a-z])unlimited([^a-z]|$)/.test(e)))}function feeBillable(e){return!e||"object"!=typeof e||!feeWaived(e)&&!0!==e.degraded}"undefined"!=typeof module&&module.exports&&(module.exports={feeWaived:feeWaived,feeBillable:feeBillable,RATES:RATES,RATE_BASIS:RATE_BASIS,UNPRICED_BY_DESIGN:UNPRICED_BY_DESIGN,isPeak:isPeak,usageTokens:usageTokens,ratesFor:ratesFor,usageBuckets:usageBuckets,cacheHitRate:cacheHitRate,usageCost:usageCost,costBreakdown:costBreakdown,turnAccounting:turnAccounting,displayTokens:displayTokens,estimateTokens:estimateTokens,estimatedBuckets:estimatedBuckets,emptyRoll:emptyRoll,rollTurn:rollTurn,rollMessages:rollMessages,ledgerRow:ledgerRow,fmtTokens:fmtTokens,fmtRoll:fmtRoll,fmtCost:fmtCost});
|