tokenusage-cli 0.1.0 → 0.2.1
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 +46 -19
- package/LICENSE +22 -22
- package/README.md +123 -94
- package/THIRD_PARTY_NOTICES.md +140 -140
- package/bin/tokenusage.js +69 -69
- package/package.json +59 -59
- package/src/cli.js +146 -134
- package/src/commands/device-login.js +205 -205
- package/src/commands/diagnostics.js +38 -38
- package/src/commands/doctor.js +110 -110
- package/src/commands/init.js +2484 -2436
- package/src/commands/pause.js +77 -0
- package/src/commands/sessions.js +29 -29
- package/src/commands/status.js +1758 -1747
- package/src/commands/sync.js +8048 -7934
- package/src/commands/uninstall.js +322 -312
- package/src/lib/account-usage-dedup.js +267 -267
- package/src/lib/antigravity-paths.js +48 -48
- package/src/lib/ark-agent-plan-limits.js +438 -438
- package/src/lib/ark-coding-plan-limits.js +537 -537
- package/src/lib/async-queue-reader.js +138 -138
- package/src/lib/atomic-json.js +27 -27
- package/src/lib/bearer-token.js +11 -11
- package/src/lib/categorizer-utils.js +232 -232
- package/src/lib/claude-categorizer.js +1343 -1343
- package/src/lib/claude-config.js +279 -279
- package/src/lib/claude-user-identity.js +56 -56
- package/src/lib/cli-ui.js +179 -179
- package/src/lib/cloud-account.js +541 -541
- package/src/lib/cloud-device-token.js +141 -141
- package/src/lib/cloud-sync-prefs.js +14 -14
- package/src/lib/cloud-usage-reset.js +115 -115
- package/src/lib/codex-config.js +595 -595
- package/src/lib/codex-context-breakdown.js +806 -806
- package/src/lib/codex-model-attribution.js +128 -128
- package/src/lib/codex-rollout-parser.js +1471 -1471
- package/src/lib/codex-service-tier.js +49 -49
- package/src/lib/codex-token-refresh.js +140 -140
- package/src/lib/codex-token-usage.js +195 -195
- package/src/lib/codex-usage-record.js +195 -195
- package/src/lib/collection-pause.js +77 -0
- package/src/lib/command-runner.js +313 -313
- package/src/lib/commandcode-limits.js +317 -317
- package/src/lib/context-health.js +124 -124
- package/src/lib/cursor-config.js +605 -605
- package/src/lib/cursor-store.js +1918 -1918
- package/src/lib/debug-flags.js +7 -7
- package/src/lib/devin-limits.js +249 -249
- package/src/lib/devin-usage.js +212 -212
- package/src/lib/diagnostics.js +383 -383
- package/src/lib/doctor.js +385 -385
- package/src/lib/dumate.js +95 -95
- package/src/lib/fs.js +455 -455
- package/src/lib/function-url.js +14 -14
- package/src/lib/gemini-config.js +291 -291
- package/src/lib/git-outcomes.js +243 -243
- package/src/lib/grok-context-breakdown.js +890 -890
- package/src/lib/grok-hook.js +360 -360
- package/src/lib/grok-limits.js +865 -865
- package/src/lib/grok-usage.js +98 -98
- package/src/lib/heatmap-compact.js +114 -114
- package/src/lib/ingest-sanitize.js +41 -41
- package/src/lib/init-flow.js +42 -42
- package/src/lib/install-resolver.js +170 -170
- package/src/lib/jsonl-lines.js +134 -134
- package/src/lib/kimi-model-repair.js +173 -173
- package/src/lib/kimi-profile.js +61 -61
- package/src/lib/kimi-session.js +216 -216
- package/src/lib/local-api.js +3588 -3588
- package/src/lib/machine-id.js +238 -238
- package/src/lib/multi-install-parser.js +113 -113
- package/src/lib/notify-command.js +38 -38
- package/src/lib/omp-hook.js +589 -589
- package/src/lib/openclaw-cli.js +104 -104
- package/src/lib/openclaw-hook.js +375 -375
- package/src/lib/openclaw-session-plugin.js +560 -560
- package/src/lib/opencode-config.js +121 -121
- package/src/lib/opencode-go-limits.js +807 -807
- package/src/lib/outcomes-engine.js +253 -253
- package/src/lib/passive-mode.js +219 -219
- package/src/lib/pet-packages.js +722 -722
- package/src/lib/pricing/curated-overrides.json +822 -822
- package/src/lib/pricing/index.js +379 -379
- package/src/lib/pricing/litellm-fetcher.js +172 -172
- package/src/lib/pricing/matcher.js +456 -456
- package/src/lib/pricing/seed-snapshot.json +1 -1
- package/src/lib/progress.js +76 -76
- package/src/lib/project-usage-purge.js +117 -117
- package/src/lib/prompt.js +24 -24
- package/src/lib/provider-status.js +104 -104
- package/src/lib/proxy-env.js +489 -489
- package/src/lib/proxy-settings.js +163 -163
- package/src/lib/qoder-cn-cli-limits.js +119 -119
- package/src/lib/qoder-limits.js +880 -880
- package/src/lib/quiet-warnings.js +26 -26
- package/src/lib/rollout.js +26005 -25966
- package/src/lib/runtime-config.js +9 -4
- package/src/lib/scan-roots.js +277 -277
- package/src/lib/scheduled-task.js +145 -0
- package/src/lib/sentry.js +231 -231
- package/src/lib/session-analytics.js +2972 -2972
- package/src/lib/session-performance.js +237 -237
- package/src/lib/session-upload.js +379 -379
- package/src/lib/skill-usage.js +267 -267
- package/src/lib/skills-manager.js +1591 -1591
- package/src/lib/source-metadata.js +48 -48
- package/src/lib/sqlite-reader.js +215 -215
- package/src/lib/star-cta.js +73 -73
- package/src/lib/subscription-manager.js +293 -293
- package/src/lib/subscriptions.js +508 -508
- package/src/lib/sync-report.js +138 -0
- package/src/lib/telemetry.js +193 -193
- package/src/lib/tracker-paths.js +16 -16
- package/src/lib/trae-cn-config.js +548 -548
- package/src/lib/trae-db.js +127 -127
- package/src/lib/trae-snapshot.js +230 -230
- package/src/lib/trae-sqlite.js +143 -143
- package/src/lib/trae-usage.js +121 -121
- package/src/lib/upload-queue.js +134 -134
- package/src/lib/upload-throttle.js +171 -171
- package/src/lib/usage-limits.js +4552 -4552
- package/src/lib/wrapped-aggregator.js +236 -236
- package/src/lib/wsl-probe.js +281 -281
- package/src/lib/zcode-limits.js +1172 -1172
- package/src/lib/zcode-reset-credits.js +66 -66
- package/test/audit-fixes.test.js +200 -0
- package/test/cli-smoke.test.js +53 -53
- package/test/collection-controls.test.js +210 -0
- package/test/ingest-sanitize.test.js +162 -162
- package/test/quiet-warnings.test.js +53 -53
- package/test/session-upload.test.js +427 -427
|
@@ -1,456 +1,456 @@
|
|
|
1
|
-
// Pure pricing-lookup logic. No I/O, no async. Tested in isolation.
|
|
2
|
-
//
|
|
3
|
-
// Resolve order:
|
|
4
|
-
// 0. CURATED source exact match (source-specific pricing)
|
|
5
|
-
// 1. CURATED exact match (self-defined aliases like kiro-*, hy3-*)
|
|
6
|
-
// 2. LiteLLM exact match (mainstream claude/gpt-5/gemini)
|
|
7
|
-
// 3. CURATED alias (e.g. "auto" -> "composer-1")
|
|
8
|
-
// 4. CURATED fuzzy substring (e.g. "kiro-future-xyz" matches via "kiro")
|
|
9
|
-
// 5. LiteLLM suffix-strip (gpt-5-codex-high-fast -> gpt-5-codex)
|
|
10
|
-
// 5b. LiteLLM provider-prefix strip (mimo-v2.5-pro -> openrouter/xiaomi/mimo-v2.5-pro)
|
|
11
|
-
// 6. LiteLLM reverse substring (longest-key first)
|
|
12
|
-
// 7. null (caller decides what to do — typically zero-pricing + negative cache)
|
|
13
|
-
|
|
14
|
-
const SUFFIX_STRIP_PATTERNS = [
|
|
15
|
-
/-xhigh-fast$/,
|
|
16
|
-
/-high-fast$/,
|
|
17
|
-
/-medium-fast$/,
|
|
18
|
-
/-low-fast$/,
|
|
19
|
-
/-xhigh$/,
|
|
20
|
-
/-high$/,
|
|
21
|
-
/-medium$/,
|
|
22
|
-
/-low$/,
|
|
23
|
-
/-fast$/,
|
|
24
|
-
];
|
|
25
|
-
|
|
26
|
-
function stripReasoningSuffix(model) {
|
|
27
|
-
for (const re of SUFFIX_STRIP_PATTERNS) {
|
|
28
|
-
if (re.test(model)) return model.replace(re, "");
|
|
29
|
-
}
|
|
30
|
-
return model;
|
|
31
|
-
}
|
|
32
|
-
|
|
33
|
-
function normalizeAntigravityModel(model) {
|
|
34
|
-
if (!model || typeof model !== "string") return model;
|
|
35
|
-
let lower = model
|
|
36
|
-
.trim()
|
|
37
|
-
.replace(/\([^)]*\)/g, " ")
|
|
38
|
-
.replace(/\b(thinking|xhigh|high|medium|low|fast)\b/gi, " ")
|
|
39
|
-
.toLowerCase()
|
|
40
|
-
.replace(/[^a-z0-9.]+/g, "-")
|
|
41
|
-
.replace(/^-+|-+$/g, "")
|
|
42
|
-
.replace(/-{2,}/g, "-");
|
|
43
|
-
lower = stripReasoningSuffix(lower);
|
|
44
|
-
|
|
45
|
-
if (lower.startsWith("gemini-claude-") || lower.startsWith("gemini-gpt-")) {
|
|
46
|
-
lower = lower.substring(7);
|
|
47
|
-
}
|
|
48
|
-
|
|
49
|
-
if (/^gemini-3\.\d+-flash-lite/.test(lower)) return "gemini-2.5-flash-lite";
|
|
50
|
-
if (/^gemini-3\.\d+-flash/.test(lower)) return "gemini-2.5-flash";
|
|
51
|
-
if (/^gemini-3\.\d+-pro/.test(lower)) return "gemini-2.5-pro";
|
|
52
|
-
if (/^claude-(sonnet|opus|haiku)-4\.\d+/.test(lower)) {
|
|
53
|
-
return lower.replace(/^claude-(sonnet|opus|haiku)-4\.(\d+)/, "claude-$1-4-$2");
|
|
54
|
-
}
|
|
55
|
-
if (lower.startsWith("gpt-oss-120b")) return "antigravity-gpt-oss-120b";
|
|
56
|
-
|
|
57
|
-
return lower;
|
|
58
|
-
}
|
|
59
|
-
|
|
60
|
-
// Zed stores model names inconsistently across versions and providers: a mix
|
|
61
|
-
// of canonical ids (`gpt-5.5`, `claude-opus-4.8`) and human display names
|
|
62
|
-
// (`Claude Sonnet 4`, `GPT-5 (Preview)`, `Gemini 3 Pro (Preview)`). Map both to
|
|
63
|
-
// the pricing engine's keys for cost lookup ONLY — the raw name is still what
|
|
64
|
-
// gets stored/displayed. Unlike normalizeAntigravityModel we must NOT strip the
|
|
65
|
-
// word "fast" (it is part of `grok-code-fast-1`) and we keep dotted GPT minors
|
|
66
|
-
// (`gpt-5.2`) while hyphenating Claude minors (`claude-opus-4.8` ->
|
|
67
|
-
// `claude-opus-4-8`) to match each family's LiteLLM/curated key style.
|
|
68
|
-
function normalizeZedModel(model) {
|
|
69
|
-
if (!model || typeof model !== "string") return model;
|
|
70
|
-
let m = model
|
|
71
|
-
.trim()
|
|
72
|
-
.replace(/\([^)]*\)/g, " ") // drop "(Preview)" and similar qualifiers
|
|
73
|
-
.toLowerCase()
|
|
74
|
-
.replace(/[^a-z0-9./]+/g, "-") // spaces/underscores -> hyphen; keep . and /
|
|
75
|
-
.replace(/^-+|-+$/g, "")
|
|
76
|
-
.replace(/-{2,}/g, "-");
|
|
77
|
-
if (/^claude-(sonnet|opus|haiku)-\d+\.\d+/.test(m)) {
|
|
78
|
-
m = m.replace(/^(claude-(?:sonnet|opus|haiku)-\d+)\.(\d+)/, "$1-$2");
|
|
79
|
-
}
|
|
80
|
-
return m;
|
|
81
|
-
}
|
|
82
|
-
|
|
83
|
-
// Claude desktop/CLI sometimes reports display-style names such as
|
|
84
|
-
// `claude-opus-4.8` or `Claude Opus 4.8`, while curated keys use Anthropic's
|
|
85
|
-
// historical dash style (`claude-opus-4-8`). Relay/gateway backends (OpenRouter
|
|
86
|
-
// and API proxies) add a provider path prefix and may invert tier/version order
|
|
87
|
-
// — e.g. `anthropic/claude-4.6-opus-20260205` — which previously rendered $0
|
|
88
|
-
// because the prefix slash got flattened to a dash and the order never matched.
|
|
89
|
-
function normalizeClaudeModel(model) {
|
|
90
|
-
if (!model || typeof model !== "string") return model;
|
|
91
|
-
// Drop the provider path prefix relays prepend ("anthropic/...",
|
|
92
|
-
// "openrouter/anthropic/...") so only the bare Claude id reaches the cleanup
|
|
93
|
-
// and lookups below. Without this the "/" is flattened to "-" by the regex
|
|
94
|
-
// step, breaking both the exact and provider-prefix-strip matches.
|
|
95
|
-
const base = model.includes("/") ? model.slice(model.lastIndexOf("/") + 1) : model;
|
|
96
|
-
let m = base
|
|
97
|
-
.trim()
|
|
98
|
-
.replace(/\([^)]*\)/g, " ")
|
|
99
|
-
.toLowerCase()
|
|
100
|
-
.replace(/[^a-z0-9.]+/g, "-")
|
|
101
|
-
.replace(/^-+|-+$/g, "")
|
|
102
|
-
.replace(/-{2,}/g, "-");
|
|
103
|
-
|
|
104
|
-
if (/^claude-(sonnet|opus|haiku)-\d+\.\d+/.test(m)) {
|
|
105
|
-
return m.replace(/^(claude-(?:sonnet|opus|haiku)-\d+)\.(\d+)/, "$1-$2");
|
|
106
|
-
}
|
|
107
|
-
if (/^(sonnet|opus|haiku)-\d+[.-]\d+/.test(m)) {
|
|
108
|
-
return m
|
|
109
|
-
.replace(/^(sonnet|opus|haiku)-/, "claude-$1-")
|
|
110
|
-
.replace(/^(claude-(?:sonnet|opus|haiku)-\d+)\.(\d+)/, "$1-$2");
|
|
111
|
-
}
|
|
112
|
-
// Some relays invert tier/version order (`claude-4.6-opus` instead of the
|
|
113
|
-
// canonical `claude-opus-4-6`). Restore it ONLY for major>=4 — Claude 3.x is
|
|
114
|
-
// genuinely version-first (`claude-3-5-sonnet`, `claude-3-opus`) and must stay
|
|
115
|
-
// untouched.
|
|
116
|
-
if (/^claude-(?:[4-9]|\d{2,})[.-]\d+-(?:sonnet|opus|haiku)/.test(m)) {
|
|
117
|
-
return m.replace(/^claude-(\d+)[.-](\d+)-(sonnet|opus|haiku)/, "claude-$3-$1-$2");
|
|
118
|
-
}
|
|
119
|
-
|
|
120
|
-
return m;
|
|
121
|
-
}
|
|
122
|
-
|
|
123
|
-
// Cursor decorates model ids with reasoning effort and sometimes places the
|
|
124
|
-
// Claude version before the tier. Normalize those lookup-only aliases while
|
|
125
|
-
// preserving Cursor Grok Fast SKUs (4.5/4.6/4.7 Fast rates are not the base
|
|
126
|
-
// rates). Gemini 4 Argon effort suffixes (-high/-medium/-low/-minimal) are
|
|
127
|
-
// stripped with the shared decorations set below.
|
|
128
|
-
function normalizeCursorModel(model) {
|
|
129
|
-
if (!model || typeof model !== "string") return model;
|
|
130
|
-
let m = model
|
|
131
|
-
.trim()
|
|
132
|
-
.replace(/\([^)]*\)/g, " ")
|
|
133
|
-
.toLowerCase()
|
|
134
|
-
.replace(/[^a-z0-9.]+/g, "-")
|
|
135
|
-
.replace(/^-+|-+$/g, "")
|
|
136
|
-
.replace(/-{2,}/g, "-");
|
|
137
|
-
const parts = m.split("-");
|
|
138
|
-
const isFast = parts.includes("fast");
|
|
139
|
-
|
|
140
|
-
if (/^(?:cursor-)?grok-4[.-]7(?:-|$)/.test(m)) {
|
|
141
|
-
return isFast ? "cursor-grok-4.7-fast" : "cursor-grok-4.7";
|
|
142
|
-
}
|
|
143
|
-
if (/^(?:cursor-)?grok-4[.-]6(?:-|$)/.test(m)) {
|
|
144
|
-
return isFast ? "cursor-grok-4.6-fast" : "cursor-grok-4.6";
|
|
145
|
-
}
|
|
146
|
-
if (/^(?:cursor-)?grok-4[.-]5(?:-|$)/.test(m)) {
|
|
147
|
-
return isFast ? "cursor-grok-4.5-fast" : "cursor-grok-4.5";
|
|
148
|
-
}
|
|
149
|
-
|
|
150
|
-
// `fast` can be a separately priced SKU (for example composer-2-fast and
|
|
151
|
-
// claude-opus-5-fast). Preserve it for the exact lookups below; when it is
|
|
152
|
-
// merely a decoration, lookupPricing's later suffix-strip still falls back
|
|
153
|
-
// to the base LiteLLM model. Gemini Argon uses "minimal" as an effort level.
|
|
154
|
-
const decorations = new Set(["thinking", "xhigh", "high", "medium", "low", "minimal"]);
|
|
155
|
-
m = parts.filter((part) => !decorations.has(part)).join("-");
|
|
156
|
-
if (m.startsWith("claude-")) return normalizeClaudeModel(m);
|
|
157
|
-
return m;
|
|
158
|
-
}
|
|
159
|
-
|
|
160
|
-
// WorkBuddy's auto-router records the model as the literal "auto" and never
|
|
161
|
-
// exposes the underlying model it picked. That collides with Cursor's curated
|
|
162
|
-
// alias ("auto" -> "composer-1"), which would misprice WorkBuddy usage as
|
|
163
|
-
// Cursor's Composer. Map it instead to hy3-preview-agent — WorkBuddy's default
|
|
164
|
-
// Tencent Hunyuan model — mirroring how Cursor's "auto" maps to its own default
|
|
165
|
-
// (composer-1). The auto-router can pick pricier models, so this can slightly
|
|
166
|
-
// under-count, but it tracks the token cost of WorkBuddy's representative model
|
|
167
|
-
// rather than an unrelated vendor's. (The raw "auto" string is still
|
|
168
|
-
// stored/displayed; only the pricing lookup is remapped.)
|
|
169
|
-
function normalizeWorkbuddyModel(model) {
|
|
170
|
-
if (typeof model === "string" && model.trim().toLowerCase() === "auto") {
|
|
171
|
-
return "hy3-preview-agent";
|
|
172
|
-
}
|
|
173
|
-
return model;
|
|
174
|
-
}
|
|
175
|
-
|
|
176
|
-
function normalizeIFlytekMaasModel(model) {
|
|
177
|
-
if (!model || typeof model !== "string") return model;
|
|
178
|
-
const trimmed = model.trim();
|
|
179
|
-
const lower = trimmed.toLowerCase();
|
|
180
|
-
// The default Agent slug maps to the Spark X2 service in the supplied price list.
|
|
181
|
-
if (lower === "xsparkx2agent") return "xsparkx2";
|
|
182
|
-
return trimmed;
|
|
183
|
-
}
|
|
184
|
-
|
|
185
|
-
// Unsloth Studio can mix local engines and paid API providers in one database.
|
|
186
|
-
// The parser qualifies metered models with their provider and marks local,
|
|
187
|
-
// subscription-backed, or ambiguous custom routes as unpriced. Returning a
|
|
188
|
-
// sentinel here prevents the generic reverse-substring matcher from mistaking
|
|
189
|
-
// a locally hosted model name for the similarly named cloud SKU.
|
|
190
|
-
function normalizeUnslothModel(model) {
|
|
191
|
-
if (typeof model !== "string") return model;
|
|
192
|
-
const lower = model.trim().toLowerCase();
|
|
193
|
-
if (lower.startsWith("local/") || lower.startsWith("unpriced/")) {
|
|
194
|
-
return "__tokenusage_unpriced_unsloth_model__";
|
|
195
|
-
}
|
|
196
|
-
return model;
|
|
197
|
-
}
|
|
198
|
-
|
|
199
|
-
// Per-source model-name normalizers, applied at pricing-lookup time only (the
|
|
200
|
-
// raw model name is preserved for storage/display). Add a source here when its
|
|
201
|
-
// model strings don't match the LiteLLM/curated keys verbatim.
|
|
202
|
-
const SOURCE_MODEL_NORMALIZERS = {
|
|
203
|
-
acode: normalizeIFlytekMaasModel,
|
|
204
|
-
antigravity: normalizeAntigravityModel,
|
|
205
|
-
claude: normalizeClaudeModel,
|
|
206
|
-
cursor: normalizeCursorModel,
|
|
207
|
-
"pi-anthropic": normalizeClaudeModel,
|
|
208
|
-
"prime-agent-anthropic": normalizeClaudeModel,
|
|
209
|
-
zed: normalizeZedModel,
|
|
210
|
-
unsloth: normalizeUnslothModel,
|
|
211
|
-
workbuddy: normalizeWorkbuddyModel,
|
|
212
|
-
};
|
|
213
|
-
|
|
214
|
-
// Memoise the sorted-by-length LiteLLM key list. Reverse-substring scan walks
|
|
215
|
-
// this once per uncached model; ~2k keys × negligible per-iteration cost, but
|
|
216
|
-
// computing the sort on every call would add up across a sync.
|
|
217
|
-
const sortedKeysCache = new WeakMap();
|
|
218
|
-
function getSortedKeys(litellm) {
|
|
219
|
-
let cached = sortedKeysCache.get(litellm);
|
|
220
|
-
if (!cached) {
|
|
221
|
-
cached = Object.keys(litellm).sort((a, b) => b.length - a.length);
|
|
222
|
-
sortedKeysCache.set(litellm, cached);
|
|
223
|
-
}
|
|
224
|
-
return cached;
|
|
225
|
-
}
|
|
226
|
-
|
|
227
|
-
function buildDotRestoredModel(model) {
|
|
228
|
-
if (typeof model !== "string") return "";
|
|
229
|
-
const lower = model.toLowerCase();
|
|
230
|
-
const restored = lower.replace(/(\d+)-(\d+)/g, "$1.$2");
|
|
231
|
-
return restored === lower ? "" : restored;
|
|
232
|
-
}
|
|
233
|
-
|
|
234
|
-
function lookupExactCaseInsensitive(table, model) {
|
|
235
|
-
if (!table || !model) return null;
|
|
236
|
-
if (table[model]) return table[model];
|
|
237
|
-
const lower = model.toLowerCase();
|
|
238
|
-
for (const key of Object.keys(table)) {
|
|
239
|
-
if (key.toLowerCase() === lower) return table[key];
|
|
240
|
-
}
|
|
241
|
-
return null;
|
|
242
|
-
}
|
|
243
|
-
|
|
244
|
-
function lookupContainedExactCaseInsensitive(table, model) {
|
|
245
|
-
if (!table || !model) return null;
|
|
246
|
-
const lower = model.toLowerCase();
|
|
247
|
-
const keys = Object.keys(table).sort((a, b) => b.length - a.length);
|
|
248
|
-
for (const key of keys) {
|
|
249
|
-
const start = lower.indexOf(key.toLowerCase());
|
|
250
|
-
if (start < 0) continue;
|
|
251
|
-
// A model generation must not match a later numeric version (Opus 5.5
|
|
252
|
-
// is a separate SKU from Opus 5). Dated and provider suffixes still work.
|
|
253
|
-
const suffix = lower.slice(start + key.length);
|
|
254
|
-
if (/^[.\d]/.test(suffix)) continue;
|
|
255
|
-
return table[key];
|
|
256
|
-
}
|
|
257
|
-
return null;
|
|
258
|
-
}
|
|
259
|
-
|
|
260
|
-
function lookupPricing(model, { curated, litellm, source } = {}) {
|
|
261
|
-
if (!model || typeof model !== "string") {
|
|
262
|
-
return { hit: false, source: "empty", value: null };
|
|
263
|
-
}
|
|
264
|
-
const normalize =
|
|
265
|
-
typeof source === "string" ? SOURCE_MODEL_NORMALIZERS[source.toLowerCase()] : null;
|
|
266
|
-
const lookupModel = normalize ? normalize(model) : model;
|
|
267
|
-
const lower = lookupModel.toLowerCase();
|
|
268
|
-
const dotForm = buildDotRestoredModel(lookupModel);
|
|
269
|
-
|
|
270
|
-
// 0. CURATED source exact. Source-specific prices apply only to their source,
|
|
271
|
-
// preventing collisions with public prices for same-named models from other CLIs.
|
|
272
|
-
const sourceKey = typeof source === "string" ? source.toLowerCase() : "";
|
|
273
|
-
if (sourceKey === "workbuddy-ai" && lower.trim() === "auto") {
|
|
274
|
-
return { hit: false, source: "miss", value: null };
|
|
275
|
-
}
|
|
276
|
-
if (sourceKey === "cline" && lower.endsWith(":free")) {
|
|
277
|
-
return {
|
|
278
|
-
hit: true,
|
|
279
|
-
source: "curated:cline-free-suffix",
|
|
280
|
-
value: { input: 0, output: 0, cache_read: 0, cache_write: 0 },
|
|
281
|
-
};
|
|
282
|
-
}
|
|
283
|
-
// OpenCode Zen free models carry a `-free` suffix on the paid model id
|
|
284
|
-
// (`deepseek-v4.1-flash-free`); the base-name fuzzy match must not bill them.
|
|
285
|
-
if (sourceKey === "opencode" && lower.endsWith("-free")) {
|
|
286
|
-
return {
|
|
287
|
-
hit: true,
|
|
288
|
-
source: "curated:opencode-free-suffix",
|
|
289
|
-
value: { input: 0, output: 0, cache_read: 0, cache_write: 0 },
|
|
290
|
-
};
|
|
291
|
-
}
|
|
292
|
-
// AStudio does not disclose its routed model. Stop before generic aliases
|
|
293
|
-
// and fuzzy matching can turn an unresolved router into a priced model.
|
|
294
|
-
if (sourceKey === "acode" && (lower === "auto" || lower.endsWith("-auto"))) {
|
|
295
|
-
return { hit: false, source: "miss", value: null };
|
|
296
|
-
}
|
|
297
|
-
const sourceExact = curated.source_exact?.[sourceKey];
|
|
298
|
-
if (sourceExact) {
|
|
299
|
-
const sourceValue = lookupExactCaseInsensitive(sourceExact, lookupModel);
|
|
300
|
-
if (sourceValue) {
|
|
301
|
-
return { hit: true, source: "curated:source-exact", value: sourceValue };
|
|
302
|
-
}
|
|
303
|
-
}
|
|
304
|
-
|
|
305
|
-
// 1. CURATED exact
|
|
306
|
-
// Qwen Flash's provider-qualified and dated IDs share the reference SKU.
|
|
307
|
-
// Keep the boundary strict: Flash-Next and Max have independent prices.
|
|
308
|
-
if (/(?:^|\/)qwen3[.-]8-flash(?:-\d{4}-\d{2}-\d{2})?$/.test(lower.trim()) && curated.exact?.["qwen3.8-flash"]) {
|
|
309
|
-
return { hit: true, source: "curated:exact", value: curated.exact["qwen3.8-flash"] };
|
|
310
|
-
}
|
|
311
|
-
if (curated.exact && curated.exact[lookupModel]) {
|
|
312
|
-
return { hit: true, source: "curated:exact", value: curated.exact[lookupModel] };
|
|
313
|
-
}
|
|
314
|
-
const curatedDotExact = lookupExactCaseInsensitive(curated.exact, dotForm);
|
|
315
|
-
if (curatedDotExact) {
|
|
316
|
-
return { hit: true, source: "curated:exact-dot", value: curatedDotExact };
|
|
317
|
-
}
|
|
318
|
-
|
|
319
|
-
// 2. LiteLLM exact
|
|
320
|
-
if (litellm && litellm[lookupModel]) {
|
|
321
|
-
return { hit: true, source: "litellm:exact", value: litellm[lookupModel] };
|
|
322
|
-
}
|
|
323
|
-
const litellmDotExact = lookupExactCaseInsensitive(litellm, dotForm);
|
|
324
|
-
if (litellmDotExact) {
|
|
325
|
-
return { hit: true, source: "litellm:exact-dot", value: litellmDotExact };
|
|
326
|
-
}
|
|
327
|
-
const curatedDotContainedExact = lookupContainedExactCaseInsensitive(curated.exact, dotForm);
|
|
328
|
-
if (curatedDotContainedExact) {
|
|
329
|
-
return { hit: true, source: "curated:exact-dot", value: curatedDotContainedExact };
|
|
330
|
-
}
|
|
331
|
-
|
|
332
|
-
// 3. CURATED alias (literal mapping like "auto" -> "composer-1")
|
|
333
|
-
if (curated.alias && curated.alias[lookupModel] && curated.exact[curated.alias[lookupModel]]) {
|
|
334
|
-
return {
|
|
335
|
-
hit: true,
|
|
336
|
-
source: "curated:alias",
|
|
337
|
-
value: curated.exact[curated.alias[lookupModel]],
|
|
338
|
-
};
|
|
339
|
-
}
|
|
340
|
-
|
|
341
|
-
// 4. CURATED fuzzy substring. Also try a dot-restored variant of the input
|
|
342
|
-
// (digits separated by `-` rejoined as `.`) so providers that dash-normalize
|
|
343
|
-
// numeric segments — Droid emits `glm-5-1-0` for upstream `GLM-5.1` — still
|
|
344
|
-
// resolve against dot-keyed curated entries like `glm-5.1`, `glm-4.6`, etc.
|
|
345
|
-
// The regex only fires on digit-dash-digit, so `claude-3-7-sonnet`,
|
|
346
|
-
// `gpt-5-codex`, `gemini-2-5-pro` are unaffected (no digit-pair to rejoin or
|
|
347
|
-
// no matching curated key).
|
|
348
|
-
if (Array.isArray(curated.fuzzy)) {
|
|
349
|
-
for (const { match, ref } of curated.fuzzy) {
|
|
350
|
-
if (!match || !ref) continue;
|
|
351
|
-
const needle = match.toLowerCase();
|
|
352
|
-
if (!curated.exact[ref]) continue;
|
|
353
|
-
if (lower.includes(needle)) {
|
|
354
|
-
return { hit: true, source: "curated:fuzzy", value: curated.exact[ref] };
|
|
355
|
-
}
|
|
356
|
-
if (dotForm && dotForm.includes(needle)) {
|
|
357
|
-
return { hit: true, source: "curated:fuzzy", value: curated.exact[ref] };
|
|
358
|
-
}
|
|
359
|
-
}
|
|
360
|
-
}
|
|
361
|
-
|
|
362
|
-
// 5. LiteLLM suffix-strip
|
|
363
|
-
if (litellm) {
|
|
364
|
-
const stripped = stripReasoningSuffix(lookupModel);
|
|
365
|
-
if (stripped !== lookupModel && litellm[stripped]) {
|
|
366
|
-
return { hit: true, source: "litellm:strip", value: litellm[stripped] };
|
|
367
|
-
}
|
|
368
|
-
}
|
|
369
|
-
|
|
370
|
-
// 5b. LiteLLM provider-prefix strip. Queue rows store the bare model name
|
|
371
|
-
// (e.g. "mimo-v2.5-pro"), but LiteLLM keys are provider-qualified (e.g.
|
|
372
|
-
// "openrouter/xiaomi/mimo-v2.5-pro"), so the exact lookups above miss. Match
|
|
373
|
-
// any key whose path suffix equals the bare model. When several providers
|
|
374
|
-
// expose the same model, pick the lexicographically smallest key so the
|
|
375
|
-
// resolved price is deterministic and independent of JSON ordering. Runs
|
|
376
|
-
// AFTER curated alias/fuzzy so e.g. Cursor's "auto" still resolves to
|
|
377
|
-
// composer-1 rather than a LiteLLM "*/auto" entry.
|
|
378
|
-
if (litellm) {
|
|
379
|
-
const suffix = "/" + lower;
|
|
380
|
-
let best = null;
|
|
381
|
-
for (const key of Object.keys(litellm)) {
|
|
382
|
-
if (key.length > suffix.length && key.toLowerCase().endsWith(suffix)) {
|
|
383
|
-
if (best === null || key < best) best = key;
|
|
384
|
-
}
|
|
385
|
-
}
|
|
386
|
-
if (best) return { hit: true, source: "litellm:prefix-strip", value: litellm[best] };
|
|
387
|
-
}
|
|
388
|
-
|
|
389
|
-
// 6. LiteLLM reverse substring (longest-key first)
|
|
390
|
-
if (litellm) {
|
|
391
|
-
const sorted = getSortedKeys(litellm);
|
|
392
|
-
for (const key of sorted) {
|
|
393
|
-
const keyLower = key.toLowerCase();
|
|
394
|
-
// Only accept if model is a superset of key (model contains key), to
|
|
395
|
-
// avoid e.g. "gpt-5" matching "gpt-5-pro" in the wrong direction.
|
|
396
|
-
if (lower.includes(keyLower) || (dotForm && dotForm.includes(keyLower))) {
|
|
397
|
-
return { hit: true, source: "litellm:fuzzy", value: litellm[key] };
|
|
398
|
-
}
|
|
399
|
-
}
|
|
400
|
-
}
|
|
401
|
-
|
|
402
|
-
return { hit: false, source: "miss", value: null };
|
|
403
|
-
}
|
|
404
|
-
|
|
405
|
-
// Convert one LiteLLM entry (per-token) to internal per-million USD shape.
|
|
406
|
-
// Missing fields stay missing — callers default with `(pricing.x || 0)`.
|
|
407
|
-
//
|
|
408
|
-
// Why the round: floating-point math means 1e-7 * 1e6 = 0.09999999999999999.
|
|
409
|
-
// Rounding to 10 significant decimals ($0.0000000001 / MTok) is well below
|
|
410
|
-
// any realistic price step but cleans up the printed/asserted numbers.
|
|
411
|
-
function roundToTenDecimals(n) {
|
|
412
|
-
return Math.round(n * 1e10) / 1e10;
|
|
413
|
-
}
|
|
414
|
-
|
|
415
|
-
function convertLitellmEntry(entry) {
|
|
416
|
-
if (!entry || typeof entry !== "object") return null;
|
|
417
|
-
const out = {};
|
|
418
|
-
if (typeof entry.input_cost_per_token === "number") {
|
|
419
|
-
out.input = roundToTenDecimals(entry.input_cost_per_token * 1_000_000);
|
|
420
|
-
}
|
|
421
|
-
if (typeof entry.output_cost_per_token === "number") {
|
|
422
|
-
out.output = roundToTenDecimals(entry.output_cost_per_token * 1_000_000);
|
|
423
|
-
}
|
|
424
|
-
if (typeof entry.cache_read_input_token_cost === "number") {
|
|
425
|
-
out.cache_read = roundToTenDecimals(entry.cache_read_input_token_cost * 1_000_000);
|
|
426
|
-
}
|
|
427
|
-
if (typeof entry.cache_creation_input_token_cost === "number") {
|
|
428
|
-
out.cache_write = roundToTenDecimals(entry.cache_creation_input_token_cost * 1_000_000);
|
|
429
|
-
}
|
|
430
|
-
return Object.keys(out).length ? out : null;
|
|
431
|
-
}
|
|
432
|
-
|
|
433
|
-
// Build a per-million-USD map from a LiteLLM raw map (or seed snapshot which
|
|
434
|
-
// uses the same field names). Skips meta keys starting with "_".
|
|
435
|
-
function buildLitellmPerMillionMap(rawData) {
|
|
436
|
-
if (!rawData || typeof rawData !== "object") return {};
|
|
437
|
-
const out = {};
|
|
438
|
-
for (const [name, entry] of Object.entries(rawData)) {
|
|
439
|
-
if (name.startsWith("_")) continue;
|
|
440
|
-
const converted = convertLitellmEntry(entry);
|
|
441
|
-
if (converted) out[name] = converted;
|
|
442
|
-
}
|
|
443
|
-
return out;
|
|
444
|
-
}
|
|
445
|
-
|
|
446
|
-
module.exports = {
|
|
447
|
-
lookupPricing,
|
|
448
|
-
stripReasoningSuffix,
|
|
449
|
-
normalizeAntigravityModel,
|
|
450
|
-
normalizeIFlytekMaasModel,
|
|
451
|
-
normalizeClaudeModel,
|
|
452
|
-
normalizeCursorModel,
|
|
453
|
-
normalizeZedModel,
|
|
454
|
-
convertLitellmEntry,
|
|
455
|
-
buildLitellmPerMillionMap,
|
|
456
|
-
};
|
|
1
|
+
// Pure pricing-lookup logic. No I/O, no async. Tested in isolation.
|
|
2
|
+
//
|
|
3
|
+
// Resolve order:
|
|
4
|
+
// 0. CURATED source exact match (source-specific pricing)
|
|
5
|
+
// 1. CURATED exact match (self-defined aliases like kiro-*, hy3-*)
|
|
6
|
+
// 2. LiteLLM exact match (mainstream claude/gpt-5/gemini)
|
|
7
|
+
// 3. CURATED alias (e.g. "auto" -> "composer-1")
|
|
8
|
+
// 4. CURATED fuzzy substring (e.g. "kiro-future-xyz" matches via "kiro")
|
|
9
|
+
// 5. LiteLLM suffix-strip (gpt-5-codex-high-fast -> gpt-5-codex)
|
|
10
|
+
// 5b. LiteLLM provider-prefix strip (mimo-v2.5-pro -> openrouter/xiaomi/mimo-v2.5-pro)
|
|
11
|
+
// 6. LiteLLM reverse substring (longest-key first)
|
|
12
|
+
// 7. null (caller decides what to do — typically zero-pricing + negative cache)
|
|
13
|
+
|
|
14
|
+
const SUFFIX_STRIP_PATTERNS = [
|
|
15
|
+
/-xhigh-fast$/,
|
|
16
|
+
/-high-fast$/,
|
|
17
|
+
/-medium-fast$/,
|
|
18
|
+
/-low-fast$/,
|
|
19
|
+
/-xhigh$/,
|
|
20
|
+
/-high$/,
|
|
21
|
+
/-medium$/,
|
|
22
|
+
/-low$/,
|
|
23
|
+
/-fast$/,
|
|
24
|
+
];
|
|
25
|
+
|
|
26
|
+
function stripReasoningSuffix(model) {
|
|
27
|
+
for (const re of SUFFIX_STRIP_PATTERNS) {
|
|
28
|
+
if (re.test(model)) return model.replace(re, "");
|
|
29
|
+
}
|
|
30
|
+
return model;
|
|
31
|
+
}
|
|
32
|
+
|
|
33
|
+
function normalizeAntigravityModel(model) {
|
|
34
|
+
if (!model || typeof model !== "string") return model;
|
|
35
|
+
let lower = model
|
|
36
|
+
.trim()
|
|
37
|
+
.replace(/\([^)]*\)/g, " ")
|
|
38
|
+
.replace(/\b(thinking|xhigh|high|medium|low|fast)\b/gi, " ")
|
|
39
|
+
.toLowerCase()
|
|
40
|
+
.replace(/[^a-z0-9.]+/g, "-")
|
|
41
|
+
.replace(/^-+|-+$/g, "")
|
|
42
|
+
.replace(/-{2,}/g, "-");
|
|
43
|
+
lower = stripReasoningSuffix(lower);
|
|
44
|
+
|
|
45
|
+
if (lower.startsWith("gemini-claude-") || lower.startsWith("gemini-gpt-")) {
|
|
46
|
+
lower = lower.substring(7);
|
|
47
|
+
}
|
|
48
|
+
|
|
49
|
+
if (/^gemini-3\.\d+-flash-lite/.test(lower)) return "gemini-2.5-flash-lite";
|
|
50
|
+
if (/^gemini-3\.\d+-flash/.test(lower)) return "gemini-2.5-flash";
|
|
51
|
+
if (/^gemini-3\.\d+-pro/.test(lower)) return "gemini-2.5-pro";
|
|
52
|
+
if (/^claude-(sonnet|opus|haiku)-4\.\d+/.test(lower)) {
|
|
53
|
+
return lower.replace(/^claude-(sonnet|opus|haiku)-4\.(\d+)/, "claude-$1-4-$2");
|
|
54
|
+
}
|
|
55
|
+
if (lower.startsWith("gpt-oss-120b")) return "antigravity-gpt-oss-120b";
|
|
56
|
+
|
|
57
|
+
return lower;
|
|
58
|
+
}
|
|
59
|
+
|
|
60
|
+
// Zed stores model names inconsistently across versions and providers: a mix
|
|
61
|
+
// of canonical ids (`gpt-5.5`, `claude-opus-4.8`) and human display names
|
|
62
|
+
// (`Claude Sonnet 4`, `GPT-5 (Preview)`, `Gemini 3 Pro (Preview)`). Map both to
|
|
63
|
+
// the pricing engine's keys for cost lookup ONLY — the raw name is still what
|
|
64
|
+
// gets stored/displayed. Unlike normalizeAntigravityModel we must NOT strip the
|
|
65
|
+
// word "fast" (it is part of `grok-code-fast-1`) and we keep dotted GPT minors
|
|
66
|
+
// (`gpt-5.2`) while hyphenating Claude minors (`claude-opus-4.8` ->
|
|
67
|
+
// `claude-opus-4-8`) to match each family's LiteLLM/curated key style.
|
|
68
|
+
function normalizeZedModel(model) {
|
|
69
|
+
if (!model || typeof model !== "string") return model;
|
|
70
|
+
let m = model
|
|
71
|
+
.trim()
|
|
72
|
+
.replace(/\([^)]*\)/g, " ") // drop "(Preview)" and similar qualifiers
|
|
73
|
+
.toLowerCase()
|
|
74
|
+
.replace(/[^a-z0-9./]+/g, "-") // spaces/underscores -> hyphen; keep . and /
|
|
75
|
+
.replace(/^-+|-+$/g, "")
|
|
76
|
+
.replace(/-{2,}/g, "-");
|
|
77
|
+
if (/^claude-(sonnet|opus|haiku)-\d+\.\d+/.test(m)) {
|
|
78
|
+
m = m.replace(/^(claude-(?:sonnet|opus|haiku)-\d+)\.(\d+)/, "$1-$2");
|
|
79
|
+
}
|
|
80
|
+
return m;
|
|
81
|
+
}
|
|
82
|
+
|
|
83
|
+
// Claude desktop/CLI sometimes reports display-style names such as
|
|
84
|
+
// `claude-opus-4.8` or `Claude Opus 4.8`, while curated keys use Anthropic's
|
|
85
|
+
// historical dash style (`claude-opus-4-8`). Relay/gateway backends (OpenRouter
|
|
86
|
+
// and API proxies) add a provider path prefix and may invert tier/version order
|
|
87
|
+
// — e.g. `anthropic/claude-4.6-opus-20260205` — which previously rendered $0
|
|
88
|
+
// because the prefix slash got flattened to a dash and the order never matched.
|
|
89
|
+
function normalizeClaudeModel(model) {
|
|
90
|
+
if (!model || typeof model !== "string") return model;
|
|
91
|
+
// Drop the provider path prefix relays prepend ("anthropic/...",
|
|
92
|
+
// "openrouter/anthropic/...") so only the bare Claude id reaches the cleanup
|
|
93
|
+
// and lookups below. Without this the "/" is flattened to "-" by the regex
|
|
94
|
+
// step, breaking both the exact and provider-prefix-strip matches.
|
|
95
|
+
const base = model.includes("/") ? model.slice(model.lastIndexOf("/") + 1) : model;
|
|
96
|
+
let m = base
|
|
97
|
+
.trim()
|
|
98
|
+
.replace(/\([^)]*\)/g, " ")
|
|
99
|
+
.toLowerCase()
|
|
100
|
+
.replace(/[^a-z0-9.]+/g, "-")
|
|
101
|
+
.replace(/^-+|-+$/g, "")
|
|
102
|
+
.replace(/-{2,}/g, "-");
|
|
103
|
+
|
|
104
|
+
if (/^claude-(sonnet|opus|haiku)-\d+\.\d+/.test(m)) {
|
|
105
|
+
return m.replace(/^(claude-(?:sonnet|opus|haiku)-\d+)\.(\d+)/, "$1-$2");
|
|
106
|
+
}
|
|
107
|
+
if (/^(sonnet|opus|haiku)-\d+[.-]\d+/.test(m)) {
|
|
108
|
+
return m
|
|
109
|
+
.replace(/^(sonnet|opus|haiku)-/, "claude-$1-")
|
|
110
|
+
.replace(/^(claude-(?:sonnet|opus|haiku)-\d+)\.(\d+)/, "$1-$2");
|
|
111
|
+
}
|
|
112
|
+
// Some relays invert tier/version order (`claude-4.6-opus` instead of the
|
|
113
|
+
// canonical `claude-opus-4-6`). Restore it ONLY for major>=4 — Claude 3.x is
|
|
114
|
+
// genuinely version-first (`claude-3-5-sonnet`, `claude-3-opus`) and must stay
|
|
115
|
+
// untouched.
|
|
116
|
+
if (/^claude-(?:[4-9]|\d{2,})[.-]\d+-(?:sonnet|opus|haiku)/.test(m)) {
|
|
117
|
+
return m.replace(/^claude-(\d+)[.-](\d+)-(sonnet|opus|haiku)/, "claude-$3-$1-$2");
|
|
118
|
+
}
|
|
119
|
+
|
|
120
|
+
return m;
|
|
121
|
+
}
|
|
122
|
+
|
|
123
|
+
// Cursor decorates model ids with reasoning effort and sometimes places the
|
|
124
|
+
// Claude version before the tier. Normalize those lookup-only aliases while
|
|
125
|
+
// preserving Cursor Grok Fast SKUs (4.5/4.6/4.7 Fast rates are not the base
|
|
126
|
+
// rates). Gemini 4 Argon effort suffixes (-high/-medium/-low/-minimal) are
|
|
127
|
+
// stripped with the shared decorations set below.
|
|
128
|
+
function normalizeCursorModel(model) {
|
|
129
|
+
if (!model || typeof model !== "string") return model;
|
|
130
|
+
let m = model
|
|
131
|
+
.trim()
|
|
132
|
+
.replace(/\([^)]*\)/g, " ")
|
|
133
|
+
.toLowerCase()
|
|
134
|
+
.replace(/[^a-z0-9.]+/g, "-")
|
|
135
|
+
.replace(/^-+|-+$/g, "")
|
|
136
|
+
.replace(/-{2,}/g, "-");
|
|
137
|
+
const parts = m.split("-");
|
|
138
|
+
const isFast = parts.includes("fast");
|
|
139
|
+
|
|
140
|
+
if (/^(?:cursor-)?grok-4[.-]7(?:-|$)/.test(m)) {
|
|
141
|
+
return isFast ? "cursor-grok-4.7-fast" : "cursor-grok-4.7";
|
|
142
|
+
}
|
|
143
|
+
if (/^(?:cursor-)?grok-4[.-]6(?:-|$)/.test(m)) {
|
|
144
|
+
return isFast ? "cursor-grok-4.6-fast" : "cursor-grok-4.6";
|
|
145
|
+
}
|
|
146
|
+
if (/^(?:cursor-)?grok-4[.-]5(?:-|$)/.test(m)) {
|
|
147
|
+
return isFast ? "cursor-grok-4.5-fast" : "cursor-grok-4.5";
|
|
148
|
+
}
|
|
149
|
+
|
|
150
|
+
// `fast` can be a separately priced SKU (for example composer-2-fast and
|
|
151
|
+
// claude-opus-5-fast). Preserve it for the exact lookups below; when it is
|
|
152
|
+
// merely a decoration, lookupPricing's later suffix-strip still falls back
|
|
153
|
+
// to the base LiteLLM model. Gemini Argon uses "minimal" as an effort level.
|
|
154
|
+
const decorations = new Set(["thinking", "xhigh", "high", "medium", "low", "minimal"]);
|
|
155
|
+
m = parts.filter((part) => !decorations.has(part)).join("-");
|
|
156
|
+
if (m.startsWith("claude-")) return normalizeClaudeModel(m);
|
|
157
|
+
return m;
|
|
158
|
+
}
|
|
159
|
+
|
|
160
|
+
// WorkBuddy's auto-router records the model as the literal "auto" and never
|
|
161
|
+
// exposes the underlying model it picked. That collides with Cursor's curated
|
|
162
|
+
// alias ("auto" -> "composer-1"), which would misprice WorkBuddy usage as
|
|
163
|
+
// Cursor's Composer. Map it instead to hy3-preview-agent — WorkBuddy's default
|
|
164
|
+
// Tencent Hunyuan model — mirroring how Cursor's "auto" maps to its own default
|
|
165
|
+
// (composer-1). The auto-router can pick pricier models, so this can slightly
|
|
166
|
+
// under-count, but it tracks the token cost of WorkBuddy's representative model
|
|
167
|
+
// rather than an unrelated vendor's. (The raw "auto" string is still
|
|
168
|
+
// stored/displayed; only the pricing lookup is remapped.)
|
|
169
|
+
function normalizeWorkbuddyModel(model) {
|
|
170
|
+
if (typeof model === "string" && model.trim().toLowerCase() === "auto") {
|
|
171
|
+
return "hy3-preview-agent";
|
|
172
|
+
}
|
|
173
|
+
return model;
|
|
174
|
+
}
|
|
175
|
+
|
|
176
|
+
function normalizeIFlytekMaasModel(model) {
|
|
177
|
+
if (!model || typeof model !== "string") return model;
|
|
178
|
+
const trimmed = model.trim();
|
|
179
|
+
const lower = trimmed.toLowerCase();
|
|
180
|
+
// The default Agent slug maps to the Spark X2 service in the supplied price list.
|
|
181
|
+
if (lower === "xsparkx2agent") return "xsparkx2";
|
|
182
|
+
return trimmed;
|
|
183
|
+
}
|
|
184
|
+
|
|
185
|
+
// Unsloth Studio can mix local engines and paid API providers in one database.
|
|
186
|
+
// The parser qualifies metered models with their provider and marks local,
|
|
187
|
+
// subscription-backed, or ambiguous custom routes as unpriced. Returning a
|
|
188
|
+
// sentinel here prevents the generic reverse-substring matcher from mistaking
|
|
189
|
+
// a locally hosted model name for the similarly named cloud SKU.
|
|
190
|
+
function normalizeUnslothModel(model) {
|
|
191
|
+
if (typeof model !== "string") return model;
|
|
192
|
+
const lower = model.trim().toLowerCase();
|
|
193
|
+
if (lower.startsWith("local/") || lower.startsWith("unpriced/")) {
|
|
194
|
+
return "__tokenusage_unpriced_unsloth_model__";
|
|
195
|
+
}
|
|
196
|
+
return model;
|
|
197
|
+
}
|
|
198
|
+
|
|
199
|
+
// Per-source model-name normalizers, applied at pricing-lookup time only (the
|
|
200
|
+
// raw model name is preserved for storage/display). Add a source here when its
|
|
201
|
+
// model strings don't match the LiteLLM/curated keys verbatim.
|
|
202
|
+
const SOURCE_MODEL_NORMALIZERS = {
|
|
203
|
+
acode: normalizeIFlytekMaasModel,
|
|
204
|
+
antigravity: normalizeAntigravityModel,
|
|
205
|
+
claude: normalizeClaudeModel,
|
|
206
|
+
cursor: normalizeCursorModel,
|
|
207
|
+
"pi-anthropic": normalizeClaudeModel,
|
|
208
|
+
"prime-agent-anthropic": normalizeClaudeModel,
|
|
209
|
+
zed: normalizeZedModel,
|
|
210
|
+
unsloth: normalizeUnslothModel,
|
|
211
|
+
workbuddy: normalizeWorkbuddyModel,
|
|
212
|
+
};
|
|
213
|
+
|
|
214
|
+
// Memoise the sorted-by-length LiteLLM key list. Reverse-substring scan walks
|
|
215
|
+
// this once per uncached model; ~2k keys × negligible per-iteration cost, but
|
|
216
|
+
// computing the sort on every call would add up across a sync.
|
|
217
|
+
const sortedKeysCache = new WeakMap();
|
|
218
|
+
function getSortedKeys(litellm) {
|
|
219
|
+
let cached = sortedKeysCache.get(litellm);
|
|
220
|
+
if (!cached) {
|
|
221
|
+
cached = Object.keys(litellm).sort((a, b) => b.length - a.length);
|
|
222
|
+
sortedKeysCache.set(litellm, cached);
|
|
223
|
+
}
|
|
224
|
+
return cached;
|
|
225
|
+
}
|
|
226
|
+
|
|
227
|
+
function buildDotRestoredModel(model) {
|
|
228
|
+
if (typeof model !== "string") return "";
|
|
229
|
+
const lower = model.toLowerCase();
|
|
230
|
+
const restored = lower.replace(/(\d+)-(\d+)/g, "$1.$2");
|
|
231
|
+
return restored === lower ? "" : restored;
|
|
232
|
+
}
|
|
233
|
+
|
|
234
|
+
function lookupExactCaseInsensitive(table, model) {
|
|
235
|
+
if (!table || !model) return null;
|
|
236
|
+
if (table[model]) return table[model];
|
|
237
|
+
const lower = model.toLowerCase();
|
|
238
|
+
for (const key of Object.keys(table)) {
|
|
239
|
+
if (key.toLowerCase() === lower) return table[key];
|
|
240
|
+
}
|
|
241
|
+
return null;
|
|
242
|
+
}
|
|
243
|
+
|
|
244
|
+
function lookupContainedExactCaseInsensitive(table, model) {
|
|
245
|
+
if (!table || !model) return null;
|
|
246
|
+
const lower = model.toLowerCase();
|
|
247
|
+
const keys = Object.keys(table).sort((a, b) => b.length - a.length);
|
|
248
|
+
for (const key of keys) {
|
|
249
|
+
const start = lower.indexOf(key.toLowerCase());
|
|
250
|
+
if (start < 0) continue;
|
|
251
|
+
// A model generation must not match a later numeric version (Opus 5.5
|
|
252
|
+
// is a separate SKU from Opus 5). Dated and provider suffixes still work.
|
|
253
|
+
const suffix = lower.slice(start + key.length);
|
|
254
|
+
if (/^[.\d]/.test(suffix)) continue;
|
|
255
|
+
return table[key];
|
|
256
|
+
}
|
|
257
|
+
return null;
|
|
258
|
+
}
|
|
259
|
+
|
|
260
|
+
function lookupPricing(model, { curated, litellm, source } = {}) {
|
|
261
|
+
if (!model || typeof model !== "string") {
|
|
262
|
+
return { hit: false, source: "empty", value: null };
|
|
263
|
+
}
|
|
264
|
+
const normalize =
|
|
265
|
+
typeof source === "string" ? SOURCE_MODEL_NORMALIZERS[source.toLowerCase()] : null;
|
|
266
|
+
const lookupModel = normalize ? normalize(model) : model;
|
|
267
|
+
const lower = lookupModel.toLowerCase();
|
|
268
|
+
const dotForm = buildDotRestoredModel(lookupModel);
|
|
269
|
+
|
|
270
|
+
// 0. CURATED source exact. Source-specific prices apply only to their source,
|
|
271
|
+
// preventing collisions with public prices for same-named models from other CLIs.
|
|
272
|
+
const sourceKey = typeof source === "string" ? source.toLowerCase() : "";
|
|
273
|
+
if (sourceKey === "workbuddy-ai" && lower.trim() === "auto") {
|
|
274
|
+
return { hit: false, source: "miss", value: null };
|
|
275
|
+
}
|
|
276
|
+
if (sourceKey === "cline" && lower.endsWith(":free")) {
|
|
277
|
+
return {
|
|
278
|
+
hit: true,
|
|
279
|
+
source: "curated:cline-free-suffix",
|
|
280
|
+
value: { input: 0, output: 0, cache_read: 0, cache_write: 0 },
|
|
281
|
+
};
|
|
282
|
+
}
|
|
283
|
+
// OpenCode Zen free models carry a `-free` suffix on the paid model id
|
|
284
|
+
// (`deepseek-v4.1-flash-free`); the base-name fuzzy match must not bill them.
|
|
285
|
+
if (sourceKey === "opencode" && lower.endsWith("-free")) {
|
|
286
|
+
return {
|
|
287
|
+
hit: true,
|
|
288
|
+
source: "curated:opencode-free-suffix",
|
|
289
|
+
value: { input: 0, output: 0, cache_read: 0, cache_write: 0 },
|
|
290
|
+
};
|
|
291
|
+
}
|
|
292
|
+
// AStudio does not disclose its routed model. Stop before generic aliases
|
|
293
|
+
// and fuzzy matching can turn an unresolved router into a priced model.
|
|
294
|
+
if (sourceKey === "acode" && (lower === "auto" || lower.endsWith("-auto"))) {
|
|
295
|
+
return { hit: false, source: "miss", value: null };
|
|
296
|
+
}
|
|
297
|
+
const sourceExact = curated.source_exact?.[sourceKey];
|
|
298
|
+
if (sourceExact) {
|
|
299
|
+
const sourceValue = lookupExactCaseInsensitive(sourceExact, lookupModel);
|
|
300
|
+
if (sourceValue) {
|
|
301
|
+
return { hit: true, source: "curated:source-exact", value: sourceValue };
|
|
302
|
+
}
|
|
303
|
+
}
|
|
304
|
+
|
|
305
|
+
// 1. CURATED exact
|
|
306
|
+
// Qwen Flash's provider-qualified and dated IDs share the reference SKU.
|
|
307
|
+
// Keep the boundary strict: Flash-Next and Max have independent prices.
|
|
308
|
+
if (/(?:^|\/)qwen3[.-]8-flash(?:-\d{4}-\d{2}-\d{2})?$/.test(lower.trim()) && curated.exact?.["qwen3.8-flash"]) {
|
|
309
|
+
return { hit: true, source: "curated:exact", value: curated.exact["qwen3.8-flash"] };
|
|
310
|
+
}
|
|
311
|
+
if (curated.exact && curated.exact[lookupModel]) {
|
|
312
|
+
return { hit: true, source: "curated:exact", value: curated.exact[lookupModel] };
|
|
313
|
+
}
|
|
314
|
+
const curatedDotExact = lookupExactCaseInsensitive(curated.exact, dotForm);
|
|
315
|
+
if (curatedDotExact) {
|
|
316
|
+
return { hit: true, source: "curated:exact-dot", value: curatedDotExact };
|
|
317
|
+
}
|
|
318
|
+
|
|
319
|
+
// 2. LiteLLM exact
|
|
320
|
+
if (litellm && litellm[lookupModel]) {
|
|
321
|
+
return { hit: true, source: "litellm:exact", value: litellm[lookupModel] };
|
|
322
|
+
}
|
|
323
|
+
const litellmDotExact = lookupExactCaseInsensitive(litellm, dotForm);
|
|
324
|
+
if (litellmDotExact) {
|
|
325
|
+
return { hit: true, source: "litellm:exact-dot", value: litellmDotExact };
|
|
326
|
+
}
|
|
327
|
+
const curatedDotContainedExact = lookupContainedExactCaseInsensitive(curated.exact, dotForm);
|
|
328
|
+
if (curatedDotContainedExact) {
|
|
329
|
+
return { hit: true, source: "curated:exact-dot", value: curatedDotContainedExact };
|
|
330
|
+
}
|
|
331
|
+
|
|
332
|
+
// 3. CURATED alias (literal mapping like "auto" -> "composer-1")
|
|
333
|
+
if (curated.alias && curated.alias[lookupModel] && curated.exact[curated.alias[lookupModel]]) {
|
|
334
|
+
return {
|
|
335
|
+
hit: true,
|
|
336
|
+
source: "curated:alias",
|
|
337
|
+
value: curated.exact[curated.alias[lookupModel]],
|
|
338
|
+
};
|
|
339
|
+
}
|
|
340
|
+
|
|
341
|
+
// 4. CURATED fuzzy substring. Also try a dot-restored variant of the input
|
|
342
|
+
// (digits separated by `-` rejoined as `.`) so providers that dash-normalize
|
|
343
|
+
// numeric segments — Droid emits `glm-5-1-0` for upstream `GLM-5.1` — still
|
|
344
|
+
// resolve against dot-keyed curated entries like `glm-5.1`, `glm-4.6`, etc.
|
|
345
|
+
// The regex only fires on digit-dash-digit, so `claude-3-7-sonnet`,
|
|
346
|
+
// `gpt-5-codex`, `gemini-2-5-pro` are unaffected (no digit-pair to rejoin or
|
|
347
|
+
// no matching curated key).
|
|
348
|
+
if (Array.isArray(curated.fuzzy)) {
|
|
349
|
+
for (const { match, ref } of curated.fuzzy) {
|
|
350
|
+
if (!match || !ref) continue;
|
|
351
|
+
const needle = match.toLowerCase();
|
|
352
|
+
if (!curated.exact[ref]) continue;
|
|
353
|
+
if (lower.includes(needle)) {
|
|
354
|
+
return { hit: true, source: "curated:fuzzy", value: curated.exact[ref] };
|
|
355
|
+
}
|
|
356
|
+
if (dotForm && dotForm.includes(needle)) {
|
|
357
|
+
return { hit: true, source: "curated:fuzzy", value: curated.exact[ref] };
|
|
358
|
+
}
|
|
359
|
+
}
|
|
360
|
+
}
|
|
361
|
+
|
|
362
|
+
// 5. LiteLLM suffix-strip
|
|
363
|
+
if (litellm) {
|
|
364
|
+
const stripped = stripReasoningSuffix(lookupModel);
|
|
365
|
+
if (stripped !== lookupModel && litellm[stripped]) {
|
|
366
|
+
return { hit: true, source: "litellm:strip", value: litellm[stripped] };
|
|
367
|
+
}
|
|
368
|
+
}
|
|
369
|
+
|
|
370
|
+
// 5b. LiteLLM provider-prefix strip. Queue rows store the bare model name
|
|
371
|
+
// (e.g. "mimo-v2.5-pro"), but LiteLLM keys are provider-qualified (e.g.
|
|
372
|
+
// "openrouter/xiaomi/mimo-v2.5-pro"), so the exact lookups above miss. Match
|
|
373
|
+
// any key whose path suffix equals the bare model. When several providers
|
|
374
|
+
// expose the same model, pick the lexicographically smallest key so the
|
|
375
|
+
// resolved price is deterministic and independent of JSON ordering. Runs
|
|
376
|
+
// AFTER curated alias/fuzzy so e.g. Cursor's "auto" still resolves to
|
|
377
|
+
// composer-1 rather than a LiteLLM "*/auto" entry.
|
|
378
|
+
if (litellm) {
|
|
379
|
+
const suffix = "/" + lower;
|
|
380
|
+
let best = null;
|
|
381
|
+
for (const key of Object.keys(litellm)) {
|
|
382
|
+
if (key.length > suffix.length && key.toLowerCase().endsWith(suffix)) {
|
|
383
|
+
if (best === null || key < best) best = key;
|
|
384
|
+
}
|
|
385
|
+
}
|
|
386
|
+
if (best) return { hit: true, source: "litellm:prefix-strip", value: litellm[best] };
|
|
387
|
+
}
|
|
388
|
+
|
|
389
|
+
// 6. LiteLLM reverse substring (longest-key first)
|
|
390
|
+
if (litellm) {
|
|
391
|
+
const sorted = getSortedKeys(litellm);
|
|
392
|
+
for (const key of sorted) {
|
|
393
|
+
const keyLower = key.toLowerCase();
|
|
394
|
+
// Only accept if model is a superset of key (model contains key), to
|
|
395
|
+
// avoid e.g. "gpt-5" matching "gpt-5-pro" in the wrong direction.
|
|
396
|
+
if (lower.includes(keyLower) || (dotForm && dotForm.includes(keyLower))) {
|
|
397
|
+
return { hit: true, source: "litellm:fuzzy", value: litellm[key] };
|
|
398
|
+
}
|
|
399
|
+
}
|
|
400
|
+
}
|
|
401
|
+
|
|
402
|
+
return { hit: false, source: "miss", value: null };
|
|
403
|
+
}
|
|
404
|
+
|
|
405
|
+
// Convert one LiteLLM entry (per-token) to internal per-million USD shape.
|
|
406
|
+
// Missing fields stay missing — callers default with `(pricing.x || 0)`.
|
|
407
|
+
//
|
|
408
|
+
// Why the round: floating-point math means 1e-7 * 1e6 = 0.09999999999999999.
|
|
409
|
+
// Rounding to 10 significant decimals ($0.0000000001 / MTok) is well below
|
|
410
|
+
// any realistic price step but cleans up the printed/asserted numbers.
|
|
411
|
+
function roundToTenDecimals(n) {
|
|
412
|
+
return Math.round(n * 1e10) / 1e10;
|
|
413
|
+
}
|
|
414
|
+
|
|
415
|
+
function convertLitellmEntry(entry) {
|
|
416
|
+
if (!entry || typeof entry !== "object") return null;
|
|
417
|
+
const out = {};
|
|
418
|
+
if (typeof entry.input_cost_per_token === "number") {
|
|
419
|
+
out.input = roundToTenDecimals(entry.input_cost_per_token * 1_000_000);
|
|
420
|
+
}
|
|
421
|
+
if (typeof entry.output_cost_per_token === "number") {
|
|
422
|
+
out.output = roundToTenDecimals(entry.output_cost_per_token * 1_000_000);
|
|
423
|
+
}
|
|
424
|
+
if (typeof entry.cache_read_input_token_cost === "number") {
|
|
425
|
+
out.cache_read = roundToTenDecimals(entry.cache_read_input_token_cost * 1_000_000);
|
|
426
|
+
}
|
|
427
|
+
if (typeof entry.cache_creation_input_token_cost === "number") {
|
|
428
|
+
out.cache_write = roundToTenDecimals(entry.cache_creation_input_token_cost * 1_000_000);
|
|
429
|
+
}
|
|
430
|
+
return Object.keys(out).length ? out : null;
|
|
431
|
+
}
|
|
432
|
+
|
|
433
|
+
// Build a per-million-USD map from a LiteLLM raw map (or seed snapshot which
|
|
434
|
+
// uses the same field names). Skips meta keys starting with "_".
|
|
435
|
+
function buildLitellmPerMillionMap(rawData) {
|
|
436
|
+
if (!rawData || typeof rawData !== "object") return {};
|
|
437
|
+
const out = {};
|
|
438
|
+
for (const [name, entry] of Object.entries(rawData)) {
|
|
439
|
+
if (name.startsWith("_")) continue;
|
|
440
|
+
const converted = convertLitellmEntry(entry);
|
|
441
|
+
if (converted) out[name] = converted;
|
|
442
|
+
}
|
|
443
|
+
return out;
|
|
444
|
+
}
|
|
445
|
+
|
|
446
|
+
module.exports = {
|
|
447
|
+
lookupPricing,
|
|
448
|
+
stripReasoningSuffix,
|
|
449
|
+
normalizeAntigravityModel,
|
|
450
|
+
normalizeIFlytekMaasModel,
|
|
451
|
+
normalizeClaudeModel,
|
|
452
|
+
normalizeCursorModel,
|
|
453
|
+
normalizeZedModel,
|
|
454
|
+
convertLitellmEntry,
|
|
455
|
+
buildLitellmPerMillionMap,
|
|
456
|
+
};
|