@alexeiled/claude-router 1.16.0 → 1.17.1

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/lib/display.mjs CHANGED
@@ -48,12 +48,21 @@ export function modeLever(config, view) {
48
48
  const UNPRICED_LABEL = { usage: 'no usage', model: 'no price', ttl: 'cache lifetime unknown', error: 'error' };
49
49
  export function unpricedLine(metrics) {
50
50
  const { unpriced } = readMetrics(metrics);
51
- const total = UNPRICED_CAUSES.reduce((n, cause) => n + unpriced[cause], 0);
52
- if (!total) return null;
53
- const causes = UNPRICED_CAUSES.filter((cause) => unpriced[cause] > 0).map(
54
- (cause) => `${UNPRICED_LABEL[cause]} ${unpriced[cause]}`,
55
- );
56
- return `Unpriced ${total} routed repl${total === 1 ? 'y' : 'ies'} left out of usage: ${causes.join(' · ')}`;
51
+ const missing = UNPRICED_CAUSES.filter((cause) => cause !== 'comparison');
52
+ const total = missing.reduce((n, cause) => n + unpriced[cause], 0);
53
+ const comparison = unpriced.comparison
54
+ ? `Comparison unavailable: ${unpriced.comparison} routed repl${unpriced.comparison === 1 ? 'y' : 'ies'}; observed costs are in activity stats`
55
+ : null;
56
+ if (!total) return comparison;
57
+ const causes = missing
58
+ .filter((cause) => unpriced[cause] > 0)
59
+ .map((cause) => `${UNPRICED_LABEL[cause]} ${unpriced[cause]}`);
60
+ return [
61
+ `Unpriced ${total} routed repl${total === 1 ? 'y' : 'ies'} left out of usage: ${causes.join(' · ')}`,
62
+ comparison,
63
+ ]
64
+ .filter(Boolean)
65
+ .join(' · ');
57
66
  }
58
67
 
59
68
  // Why routing starts off: the session's model is no tier's model. Set at session start, read by the band.
@@ -106,6 +115,7 @@ export const REASONS = {
106
115
  'cash-gate': 'estimated cold cache write exceeds the cap',
107
116
  'context-fit': 'a larger context window is needed',
108
117
  'context-unknown': 'context is not measured reliably',
118
+ 'context-full': 'the context is too full for any route, so Claude’s model is kept',
109
119
  'model-unavailable': 'the requested model is not available',
110
120
  pinned: 'one-turn model pin',
111
121
  manual: 'routing did not choose this turn, so Claude’s model is kept',
@@ -176,6 +186,7 @@ export const SHORT_REASONS = {
176
186
  'cash-gate': '= cash cap',
177
187
  'context-fit': '↑ context',
178
188
  'context-unknown': '= context unknown',
189
+ 'context-full': '! context full',
179
190
  'model-unavailable': '! model unavailable',
180
191
  'native-fallback': '! fallback',
181
192
  pinned: '⏵ pinned',
package/lib/facts.mjs CHANGED
@@ -7,6 +7,15 @@ const REMINDER = /<system-reminder>[\s\S]*?<\/system-reminder>/g;
7
7
  const CLIP_MARKER = ' […] ';
8
8
  const ANSI_ESCAPE = new RegExp(`${String.fromCharCode(0x1b)}\\[[0-?]*[ -/]*[@-~]`, 'g');
9
9
 
10
+ // Images can arrive inside a tool result, not just a user message.
11
+ export function hasImages(messages) {
12
+ const contains = (content) =>
13
+ blocks(content).some(
14
+ (block) => block.type === 'image' || (block.type === 'tool_result' && contains(block.content)),
15
+ );
16
+ return messages.some((message) => contains(message.content));
17
+ }
18
+
10
19
  export function extractFacts(body, memory, { recentTurns, maxTextChars }) {
11
20
  const messages = Array.isArray(body.messages) ? body.messages : [];
12
21
  // Claude Code puts hook output and tool additions in `system` messages after the user message, and keeps them in
@@ -35,6 +44,7 @@ export function extractFacts(body, memory, { recentTurns, maxTextChars }) {
35
44
  turns.push({ role: message.role, text: clip(text, maxTextChars) });
36
45
  });
37
46
  return {
47
+ hasImages: hasImages(messages),
38
48
  turns: turns.slice(-recentTurns),
39
49
  prompt: current === -1 ? '' : clip(textOf(lastBlocks), maxTextChars),
40
50
  continuation: last?.role === 'user' && lastBlocks.some((b) => b.type === 'tool_result'),
package/lib/panel.mjs CHANGED
@@ -4,7 +4,7 @@ import {
4
4
  ACTIVITIES,
5
5
  ACTIVITY_MODES,
6
6
  activeClassifier,
7
- DEFAULTS,
7
+ defaultsFor,
8
8
  effectiveActivities,
9
9
  effectiveRoutes,
10
10
  ROUTING_MODES,
@@ -114,6 +114,7 @@ const EFFORT_SHORT = { low: 'low', medium: 'med', high: 'high', xhigh: 'xh', max
114
114
  export function routeDraftOf(config, view) {
115
115
  if (view.routeDraft) return view.routeDraft;
116
116
  const base = structuredClone({
117
+ ...(config.provider && config.provider !== 'anthropic' ? { provider: config.provider } : {}),
117
118
  routes: config.routes,
118
119
  baselineTier: config.baselineTier,
119
120
  activities: config.activities,
@@ -185,6 +186,9 @@ export function renderPanel(
185
186
  flexDirection: 'column',
186
187
  children: [
187
188
  ...header(ui, config, view, tab, actions),
189
+ ...(config.provider && config.provider !== 'anthropic'
190
+ ? [ui.dim(`Provider: ${config.provider} · ${view.configPath ?? config.nativePath ?? ''}`)]
191
+ : []),
188
192
  ...body(ui, config, view, usage, actions, modelOptions, columns, now),
189
193
  ...(view.help ? helpLines(ui) : []),
190
194
  ...statusBar(ui, config, view, tab, actions),
@@ -634,7 +638,7 @@ function shortRoute(config, route, narrow) {
634
638
  return `${route.model}·${narrow ? (EFFORT_SHORT[effort] ?? effort) : effort}`;
635
639
  }
636
640
  const routeKey = (route) => `${route.model}@${route.effort ?? ''}`;
637
- const builtIn = (activity, tier) => DEFAULTS.activities[activity]?.[tier] ?? null;
641
+ const builtIn = (config, activity, tier) => defaultsFor(config.provider).activities[activity]?.[tier] ?? null;
638
642
  const sharesEffort = (config, route) => effortSharesCache(config, config.models[route.model]?.id);
639
643
  // The cache a route runs on: one per model on a model that keeps one cache across efforts, else one per route.
640
644
  const cacheOf = (config, route) => (sharesEffort(config, route) ? route.model : routeKey(route));
@@ -647,7 +651,9 @@ function overrideWarnings(config, uses, activity, tier) {
647
651
  return [
648
652
  above && compareRoutes(config, route, above) > 0 ? 'stronger than the tier above' : null,
649
653
  uses.get(cacheOf(config, route)) === 1
650
- ? `${sharesEffort(config, route) ? 'a model' : 'a model + effort'} only this cell uses: one more cache`
654
+ ? config.cache.mode === 'automatic'
655
+ ? 'a route only this cell uses; cache sharing is unknown'
656
+ : `${sharesEffort(config, route) ? 'a model' : 'a model + effort'} only this cell uses: one more cache`
651
657
  : null,
652
658
  ].filter(Boolean);
653
659
  }
@@ -770,7 +776,7 @@ function routesTab(ui, config, view, _usage, actions, modelOptions, columns) {
770
776
  const route = resolveRoute(draft, tier, activity);
771
777
  const edited = changes.cells.some(([a, t]) => a === activity && t === tier) ? '●' : '';
772
778
  if (sameRoute(route, draft.routes[tier])) return cellButton(activity, tier, `·${edited}`, true);
773
- const preset = builtIn(activity, tier);
779
+ const preset = builtIn(draft, activity, tier);
774
780
  const mark = preset && sameRoute(route, { model: preset.model, effort: preset.effort ?? null }) ? '°' : '';
775
781
  return cellButton(activity, tier, `${shortRoute(draft, route, narrow)}${mark}${edited}`, mode !== 'on');
776
782
  }),
@@ -781,7 +787,15 @@ function routesTab(ui, config, view, _usage, actions, modelOptions, columns) {
781
787
  ui.text(' '),
782
788
  ...cellEditor(ui, draft, changes, selected, cacheUses(draft, true).uses, actions, modelOptions),
783
789
  ui.text(' '),
784
- ...cacheLines(ui, uses.size, shared, mode, columns),
790
+ ...(config.cache.mode === 'automatic'
791
+ ? [
792
+ ui.dim(
793
+ 'Automatic prefix cache: hits are observed; lifetime and sharing across efforts are unknown. No cache-write premium.',
794
+ ),
795
+ ]
796
+ : config.cache.mode === 'disabled'
797
+ ? [ui.dim('Prompt caching disabled: input is priced at the uncached rate.')]
798
+ : cacheLines(ui, uses.size, shared, mode, columns)),
785
799
  ...LADDER.slice(0, -1)
786
800
  .filter((upper, i) => sameRoute(draft.routes[upper], draft.routes[LADDER[i + 1]]))
787
801
  .map((upper) =>
@@ -791,7 +805,7 @@ function routesTab(ui, config, view, _usage, actions, modelOptions, columns) {
791
805
  ),
792
806
  ),
793
807
  ui.Button({ key: 'reset-routes', label: 'Reset routes to defaults', hotkey: 'r', onPress: actions.resetRoutes }),
794
- ui.dim('A new model ID needs router.json → models: price, window, efforts.'),
808
+ ui.dim('A new model ID needs the active file → models: price, window, efforts.'),
795
809
  );
796
810
  return out;
797
811
  }
@@ -843,7 +857,7 @@ function cellEditor(ui, draft, changes, selected, uses, actions, modelOptions) {
843
857
  const activity = row;
844
858
  const route = resolveRoute(draft, tier, activity);
845
859
  const base = draft.routes[tier];
846
- const preset = builtIn(activity, tier);
860
+ const preset = builtIn(draft, activity, tier);
847
861
  const presetRoute = preset ? { model: preset.model, effort: preset.effort ?? null } : null;
848
862
  const edited = changes.cells.some(([a, t]) => a === activity && t === tier);
849
863
  const own = !sameRoute(route, base);
@@ -948,7 +962,8 @@ const MODE_HINT = {
948
962
  function policyTab(ui, config, view, _usage, actions) {
949
963
  const changes = routingChanges(config, view);
950
964
  const { draft, tuning, policy } = changes;
951
- const defaults = tuningOf(DEFAULTS);
965
+ const preset = defaultsFor(config.provider);
966
+ const defaults = tuningOf(preset);
952
967
  const label = classifierLabel(config);
953
968
  const pad = ''.padEnd(13);
954
969
  const tune = (key) => choice(ui, key, tuning[key], POLICY_FORMAT[key], (value) => actions.tune(key, value));
@@ -991,7 +1006,7 @@ function policyTab(ui, config, view, _usage, actions) {
991
1006
  : draft.baselineTier,
992
1007
  ' ',
993
1008
  marker(ui, changes.baseline),
994
- draft.baselineTier === DEFAULTS.baselineTier ? null : ui.dim(`default ${DEFAULTS.baselineTier}`),
1009
+ draft.baselineTier === preset.baselineTier ? null : ui.dim(`default ${preset.baselineTier}`),
995
1010
  ),
996
1011
  ui.dim(`${pad}when the session’s model is no tier’s model`),
997
1012
  ui.line(
@@ -0,0 +1,116 @@
1
+ // Execution presets, not classifier services. Prices are API list-price estimates, never a plan bill.
2
+ // Sources and the limits of these starting routes are in docs/configuration.md#provider-profiles.
3
+ const EFFORTS = ['low', 'high', 'max'];
4
+ const AUTOMATIC = {
5
+ mode: 'automatic',
6
+ effortScope: 'unknown',
7
+ writeMultiplier: {},
8
+ ttlMs: {},
9
+ warmMarginMs: 0,
10
+ };
11
+ const route = (model, effort) => ({ model, effort });
12
+ const agents = (light, heavy) => ({
13
+ types: { Explore: light, Plan: 'classify', 'general-purpose': 'classify' },
14
+ light,
15
+ standard: heavy,
16
+ heavy,
17
+ heavyMass: 0.8,
18
+ });
19
+
20
+ export const PROVIDER_DEFAULTS = {
21
+ deepseek: {
22
+ models: {
23
+ flash: {
24
+ id: 'deepseek-flash[1m]',
25
+ input: 0.3,
26
+ output: 1.2,
27
+ cacheRead: 0.006,
28
+ contextWindow: 1_000_000,
29
+ billing: 'credits',
30
+ efforts: EFFORTS,
31
+ strength: 1,
32
+ imageInput: true,
33
+ },
34
+ },
35
+ routes: {
36
+ micro: route('flash', 'low'),
37
+ low: route('flash', 'high'),
38
+ medium: route('flash', 'high'),
39
+ high: route('flash', 'max'),
40
+ },
41
+ activities: Object.fromEntries(
42
+ ['code', 'debug', 'plan', 'review'].map((a) => [a, { medium: route('flash', 'max') }]),
43
+ ),
44
+ subagents: agents('flash', 'flash'),
45
+ cache: AUTOMATIC,
46
+ },
47
+ zai: {
48
+ models: {
49
+ flash: {
50
+ id: 'glm-5.3-flash[1m]',
51
+ input: 0.15,
52
+ output: 0.5,
53
+ cacheRead: 0.03,
54
+ contextWindow: 1_000_000,
55
+ billing: 'credits',
56
+ efforts: EFFORTS,
57
+ strength: 1,
58
+ imageInput: true,
59
+ },
60
+ glm: {
61
+ id: 'glm-5.3[1m]',
62
+ input: 1.4,
63
+ output: 4.4,
64
+ cacheRead: 0.26,
65
+ contextWindow: 1_000_000,
66
+ billing: 'credits',
67
+ efforts: EFFORTS,
68
+ strength: 2,
69
+ imageInput: false,
70
+ },
71
+ },
72
+ routes: {
73
+ micro: route('flash', 'low'),
74
+ low: route('flash', 'high'),
75
+ medium: route('glm', 'high'),
76
+ high: route('glm', 'max'),
77
+ },
78
+ activities: {
79
+ ...Object.fromEntries(['code', 'debug', 'plan', 'review'].map((a) => [a, { medium: route('glm', 'max') }])),
80
+ ...Object.fromEntries(
81
+ ['ops', 'explore', 'docs'].map((a) => [a, { medium: route('flash', 'high'), high: route('flash', 'max') }]),
82
+ ),
83
+ },
84
+ subagents: agents('flash', 'glm'),
85
+ cache: AUTOMATIC,
86
+ },
87
+ };
88
+
89
+ // A provider is an identifier, never a path. Unknown providers may supply a complete standalone profile.
90
+ export function providerName(value) {
91
+ if (value == null || value === '') return 'anthropic';
92
+ if (typeof value !== 'string' || !/^[a-z][a-z0-9-]{0,63}$/.test(value))
93
+ throw new Error('invalid router provider identifier');
94
+ return value;
95
+ }
96
+
97
+ export function providerFile(value) {
98
+ const provider = providerName(value);
99
+ return provider === 'anthropic' ? 'router.json' : `${provider}-router.json`;
100
+ }
101
+
102
+ export const canonicalModel = (id) => (typeof id === 'string' ? id.replace(/\[1m\]$/, '') : id);
103
+
104
+ export function executionDefaults(provider, defaults) {
105
+ if (provider === 'anthropic') return defaults;
106
+ return {
107
+ ...defaults,
108
+ models: {},
109
+ routes: {},
110
+ activities: {},
111
+ subagentRouting: 'off',
112
+ subagents: { types: {}, light: '', standard: '', heavy: '', heavyMass: 0.8 },
113
+ cache: AUTOMATIC,
114
+ ...(Object.hasOwn(PROVIDER_DEFAULTS, provider) ? { subagentRouting: 'on', ...PROVIDER_DEFAULTS[provider] } : {}),
115
+ };
116
+ }
package/lib/route.mjs CHANGED
@@ -3,6 +3,7 @@ import { ACTIVITIES, rank, resolveRoute, routeSpec, sameRoute, TIERS } from './c
3
3
  import * as costs from './cost.mjs';
4
4
  import { clip, completedSince, extractFacts } from './facts.mjs';
5
5
  import { CONTEXT_FILL, decide, fitTier, initialState, sameTaskOf } from './policy.mjs';
6
+ import { canonicalModel } from './providers.mjs';
6
7
 
7
8
  // Exact observed API snapshot, not a rule that merges different dated versions.
8
9
  const SNAPSHOTS = { 'claude-haiku-4-5': 'claude-haiku-4-5-20251001' };
@@ -13,7 +14,9 @@ export const HISTORY_SHRINK = 0.8;
13
14
 
14
15
  // Whether the engine served the requested model, allowing for a dated snapshot of it.
15
16
  export function isSameModel(requested, served) {
16
- return requested === served || SNAPSHOTS[requested] === served;
17
+ if (typeof requested !== 'string' || typeof served !== 'string') return false;
18
+ const [a, b] = [canonicalModel(requested), canonicalModel(served)];
19
+ return a === b || SNAPSHOTS[a] === b;
17
20
  }
18
21
 
19
22
  // The engine may echo our own routed model on a continuation; any third model is its fallback.
@@ -120,14 +123,19 @@ export function nativeFacts(config, loop, { messages, prompt, effort, turnId, co
120
123
  };
121
124
  }
122
125
 
123
- export function isModelAllowed(id, available, nativeModel) {
124
- if (available === undefined) return true;
125
- if (!Array.isArray(available) || !available.every((v) => typeof v === 'string')) return false;
126
+ export function isModelAllowed(requested, allowed, native) {
127
+ if (allowed === undefined) return true;
128
+ if (!Array.isArray(allowed) || !allowed.every((v) => typeof v === 'string')) return false;
129
+ const id = canonicalModel(requested);
130
+ const nativeModel = canonicalModel(native);
131
+ const available = allowed.map(canonicalModel);
126
132
  if (available.includes('default') && id === nativeModel) return true;
127
133
  const family = /^claude-(opus|sonnet|haiku)-/.exec(id)?.[1];
128
134
  const specific = available.filter((v) => family && v.startsWith(`claude-${family}-`));
129
135
  if (specific.length) return specific.some((v) => id === v || id.startsWith(`${v}-`));
130
- return available.some((v) => id === v || id.startsWith(`${v}-`) || (family && v === family));
136
+ return available.some(
137
+ (v) => id === v || (id.startsWith('claude-') && id.startsWith(`${v}-`)) || (family && v === family),
138
+ );
131
139
  }
132
140
 
133
141
  function economicConfig(config, loop) {
@@ -210,15 +218,18 @@ function routeTurn(config, loop, { facts, advice, pin, nativeModel, contextKnown
210
218
  (!loop.decision && Object.values(config.models).find((m) => isSameModel(m.id, nativeModel))) || specOf(last);
211
219
  const shrinksUnmeasured = (cell) =>
212
220
  (!contextKnown || !loop.historyMeasured) && specOf(cell).contextWindow < lastSpec.contextWindow;
213
- const fits = (cell) => {
221
+ const runs = (cell) => {
214
222
  const model = specOf(cell);
215
223
  return (
216
224
  !shrinksUnmeasured(cell) &&
225
+ (!facts.hasImages || model.imageInput !== false) &&
217
226
  isModelAllowed(model.id, availableModels, nativeModel) &&
218
- !loop.ineligible.includes(model.id) &&
219
- (!contextKnown || costs.nextContextTokens(facts) <= model.contextWindow * CONTEXT_FILL)
227
+ !loop.ineligible.includes(model.id)
220
228
  );
221
229
  };
230
+ const roomFor = (cell) =>
231
+ !contextKnown || costs.nextContextTokens(facts) <= specOf(cell).contextWindow * CONTEXT_FILL;
232
+ const fits = (cell) => runs(cell) && roomFor(cell);
222
233
  if (contextKnown) {
223
234
  const cell = result;
224
235
  const routeFor = (tier) => (tier === cell.tier ? routeOf(cell) : resolveRoute(config, tier, cell.activity));
@@ -234,7 +245,9 @@ function routeTurn(config, loop, { facts, advice, pin, nativeModel, contextKnown
234
245
  const fallback = cells.find(
235
246
  (cell) => cell && (cell.activity === null || cell.activity === activity) && fits(cell),
236
247
  ) ?? { tier: null, activity: null };
237
- result = { ...result, ...fallback, reason: 'model-unavailable', lateral: null };
248
+ // Only the window failed: no route is withheld, the conversation is just too full for every one of them.
249
+ const reason = runs(result) ? 'context-full' : 'model-unavailable';
250
+ result = { ...result, ...fallback, reason, lateral: null };
238
251
  }
239
252
  const route = result.tier ? routeOf(result) : null;
240
253
  const model = route ? routeSpec(config, route).id : nativeModel;
@@ -365,7 +378,11 @@ export function observedMove(config, loop, { observed, facts, contextKnown, now
365
378
 
366
379
  // A tool continuation keeps the turn's route and activity; only the context, the model list, or the turn's own tool
367
380
  // calls (observedMove) can move it.
368
- export function continueRoute(config, loop, { nativeModel, contextTokens, contextKnown, availableModels, effort }) {
381
+ export function continueRoute(
382
+ config,
383
+ loop,
384
+ { nativeModel, contextTokens, contextKnown, availableModels, effort, hasImages = false },
385
+ ) {
369
386
  if (!loop.decision) return loop;
370
387
  // A fit to another tier serves this turn only: `loop.route` stays the route of `lastRoute`, as the tier does.
371
388
  let decision = loop.decision;
@@ -388,6 +405,27 @@ export function continueRoute(config, loop, { nativeModel, contextTokens, contex
388
405
  };
389
406
  }
390
407
  }
408
+ if (hasImages && decision.tier && routeSpec(config, runningRoute(config, decision))?.imageInput === false) {
409
+ const runningWindow = routeSpec(config, runningRoute(config, decision)).contextWindow;
410
+ const tier = [...TIERS].reverse().find((t) => {
411
+ const model = routeSpec(config, resolveRoute(config, t, decision.activity));
412
+ return (
413
+ model.imageInput !== false &&
414
+ isModelAllowed(model.id, availableModels, nativeModel) &&
415
+ !loop.ineligible.includes(model.id) &&
416
+ (contextKnown ? contextTokens <= model.contextWindow * CONTEXT_FILL : model.contextWindow >= runningWindow)
417
+ );
418
+ });
419
+ const route = tier ? resolveRoute(config, tier, decision.activity) : null;
420
+ decision = {
421
+ ...decision,
422
+ tier: tier ?? null,
423
+ route,
424
+ model: route ? routeSpec(config, route).id : nativeModel,
425
+ effort: route ? costs.routeEffort(config, route, effort) : effort,
426
+ reason: 'model-unavailable',
427
+ };
428
+ }
391
429
  if (!isModelAllowed(decision.model, availableModels, nativeModel) || loop.ineligible.includes(decision.model)) {
392
430
  decision = {
393
431
  ...decision,
package/lib/savings.mjs CHANGED
@@ -4,6 +4,7 @@
4
4
  // records. The records under SAVINGS_PREFIX are undated history from before those, read and never written.
5
5
 
6
6
  import { effortSharesCache, ratesAt } from './cost.mjs';
7
+ import { canonicalModel } from './providers.mjs';
7
8
  import { afterResetOf, earlierSince, isRecord } from './records.mjs';
8
9
  import { HISTORY_SHRINK, isSameModel } from './route.mjs';
9
10
 
@@ -20,8 +21,9 @@ const TTLS = ['5m', '1h'];
20
21
  // suffix such as `[1m]`. Null when no entry prices it.
21
22
  export function modelSpec(config, id) {
22
23
  if (typeof id !== 'string') return null;
23
- const bare = id.replace(/\[[^\]]*\]$/, '');
24
- const dated = (model) => bare.startsWith(model.id) && /^-\d{8}$/.test(bare.slice(model.id.length));
24
+ const bare = canonicalModel(id);
25
+ const dated = (model) =>
26
+ model.id.startsWith('claude-') && bare.startsWith(model.id) && /^-\d{8}$/.test(bare.slice(model.id.length));
25
27
  return Object.values(config.models).find((model) => isSameModel(model.id, bare) || dated(model)) ?? null;
26
28
  }
27
29
 
@@ -38,11 +40,33 @@ export function modelAlias(config, id) {
38
40
 
39
41
  // One request's list price in USD: `counts` are its uncached input, cache read, cache write and output tokens, the
40
42
  // writes at `writeMultiplier` times input. Long-context rates follow the whole prompt.
41
- export function requestUsd(model, { input, cacheRead, cacheWrite, output }, writeMultiplier) {
43
+ export function requestUsd(model, { input, cacheRead, cacheWrite, output }, writeMultiplier, cacheMode = 'explicit') {
42
44
  const rates = ratesAt(model, input + cacheRead + cacheWrite);
43
- return (
44
- (rates.input * (input + cacheWrite * writeMultiplier) + rates.cacheRead * cacheRead + rates.output * output) / 1e6
45
- );
45
+ const inputUsd =
46
+ cacheMode === 'disabled'
47
+ ? rates.input * (input + cacheRead + cacheWrite)
48
+ : rates.input * (input + cacheWrite * writeMultiplier) + rates.cacheRead * cacheRead;
49
+ return (inputUsd + rates.output * output) / 1e6;
50
+ }
51
+
52
+ // Actual observed counters can be priced without inventing the cache on another route.
53
+ export function observedReplyUsd(config, usage, ttl) {
54
+ const counts = {
55
+ input: usage?.input_tokens,
56
+ cacheRead: usage?.cache_read_input_tokens,
57
+ cacheWrite: usage?.cache_creation_input_tokens,
58
+ output: usage?.output_tokens,
59
+ };
60
+ const model = modelSpec(config, usage?.model);
61
+ const write = config.cache.mode === 'explicit' ? config.cache.writeMultiplier[ttl] : 1;
62
+ if (
63
+ !model ||
64
+ !Number.isFinite(model.output) ||
65
+ !Number.isFinite(write) ||
66
+ !Object.values(counts).every((n) => Number.isFinite(n) && n >= 0)
67
+ )
68
+ return null;
69
+ return requestUsd(model, counts, write, config.cache.mode);
46
70
  }
47
71
 
48
72
  const truthy = (value) => /^(1|true|yes|on)$/i.test(String(value ?? '').trim());
@@ -100,7 +124,7 @@ export function compareReply(config, state, { usage, served, yours, ttl, now })
100
124
  const next = { total, at: now, onYours, ...identity };
101
125
  // A cache simulated for another model or effort of yours is none of this one's: the new one starts cold.
102
126
  const prior = state?.model === identity.model && (sharedEffort || state.effort === identity.effort) ? state : null;
103
- const write = config.cache.writeMultiplier[ttl];
127
+ const write = config.cache.mode === 'explicit' ? config.cache.writeMultiplier[ttl] : 1;
104
128
  if (
105
129
  !servedSpec ||
106
130
  !yoursSpec ||
@@ -109,16 +133,22 @@ export function compareReply(config, state, { usage, served, yours, ttl, now })
109
133
  !Number.isFinite(write)
110
134
  )
111
135
  return { state: next, reply: null };
112
- const routedUsd = requestUsd(servedSpec, counts, write);
136
+ // Implicit caches have no guaranteed TTL or cross-effort hit rate. Only identical observed routes compare exactly.
137
+ if (config.cache.mode === 'automatic' && !(onYours && (prior?.onYours ?? true))) return { state: next, reply: null };
138
+ const routedUsd = requestUsd(servedSpec, counts, write, config.cache.mode);
113
139
  let yoursUsd = routedUsd;
114
140
  let servedOnYoursUsd = routedUsd;
115
141
  if (!(onYours && (prior?.onYours ?? true))) {
116
142
  const warm =
117
- prior && now - prior.at < config.cache.ttlMs[ttl] && total >= prior.total * HISTORY_SHRINK && prior.total > 0;
143
+ config.cache.mode !== 'disabled' &&
144
+ prior &&
145
+ now - prior.at < config.cache.ttlMs[ttl] &&
146
+ total >= prior.total * HISTORY_SHRINK &&
147
+ prior.total > 0;
118
148
  const cacheRead = Math.max(counts.cacheRead, warm ? Math.min(prior.total, total - counts.input) : 0);
119
149
  const single = { ...counts, cacheRead, cacheWrite: total - counts.input - cacheRead };
120
- yoursUsd = requestUsd(yoursSpec, single, write);
121
- servedOnYoursUsd = requestUsd(servedSpec, single, write);
150
+ yoursUsd = requestUsd(yoursSpec, single, write, config.cache.mode);
151
+ servedOnYoursUsd = requestUsd(servedSpec, single, write, config.cache.mode);
122
152
  }
123
153
  const modelUsd = servedOnYoursUsd - yoursUsd;
124
154
  return {
@@ -149,7 +179,8 @@ export function unpricedCause(config, { usage, served, yours, ttl }) {
149
179
  const [servedSpec, yoursSpec] = [modelSpec(config, served.model), modelSpec(config, yours.model)];
150
180
  if (!servedSpec || !yoursSpec || !Number.isFinite(servedSpec.output) || !Number.isFinite(yoursSpec.output))
151
181
  return 'model';
152
- return Number.isFinite(config.cache.writeMultiplier[ttl]) ? null : 'ttl';
182
+ if (config.cache.mode === 'automatic') return 'ttl';
183
+ return config.cache.mode === 'disabled' || Number.isFinite(config.cache.writeMultiplier[ttl]) ? null : 'ttl';
153
184
  }
154
185
 
155
186
  // One main reply against every configured model as yours, at your effort, so the session readout can follow a
package/lib/subagents.mjs CHANGED
@@ -180,15 +180,16 @@ export function addAgentReply(config, agent, usage, ttl) {
180
180
  output: usage?.output_tokens,
181
181
  };
182
182
  const next = { ...agent, requests: agent.requests + 1 };
183
- const write = config.cache.writeMultiplier[ttl];
183
+ const write = config.cache.mode === 'explicit' ? config.cache.writeMultiplier[ttl] : 1;
184
184
  const actual = modelSpec(config, usage?.model);
185
185
  const other = agent.pass ? null : modelSpec(config, agent.coreModel);
186
186
  const priced = (spec) => spec && Number.isFinite(spec.output);
187
187
  if (!Object.values(counts).every(isCount) || !Number.isFinite(write) || !priced(actual)) return next;
188
188
  if (!agent.pass && !priced(other)) return next;
189
+ if (config.cache.mode === 'automatic' && other && actual !== other) return next;
189
190
  next.pricedRequests += 1;
190
- next.actualUsd += requestUsd(actual, counts, write);
191
- if (other) next.otherUsd += requestUsd(other, counts, write);
191
+ next.actualUsd += requestUsd(actual, counts, write, config.cache.mode);
192
+ if (other) next.otherUsd += requestUsd(other, counts, write, config.cache.mode);
192
193
  return next;
193
194
  }
194
195
 
package/package.json CHANGED
@@ -1,7 +1,7 @@
1
1
  {
2
2
  "name": "@alexeiled/claude-router",
3
- "version": "1.16.0",
4
- "description": "Native Claude Code Mod: model and effort routing for the main conversation, advised by Jev, Clef, Clef Flash, Clef Omni, OpenAI or Ollama.",
3
+ "version": "1.17.1",
4
+ "description": "Native Claude Code model and effort routing with Anthropic, DeepSeek and z.ai defaults, custom provider profiles, and classifier advice.",
5
5
  "license": "MIT",
6
6
  "author": "Alexei Ledenev",
7
7
  "repository": {
@@ -34,6 +34,7 @@
34
34
  "typecheck": "tools/node_modules/.bin/tsc -p tsconfig.json",
35
35
  "validate": "claude plugin validate .claude-plugin/plugin.json && claude plugin validate .claude-plugin/marketplace.json",
36
36
  "test:plugin": "bash scripts/plugin-test.sh",
37
+ "test:providers": "node experiments/mod-router/scripts/provider-acceptance.mjs",
37
38
  "format": "tools/node_modules/.bin/biome format --write .",
38
39
  "docs:images": "node scripts/render-doc-images.mjs",
39
40
  "docs:modes": "node scripts/mode-matrix.mjs",