@nathapp/nax 0.46.0 → 0.46.1
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 +14 -0
- package/dist/nax.js +163 -82
- package/package.json +1 -1
- package/src/acceptance/generator.ts +1 -1
- package/src/acceptance/types.ts +2 -0
- package/src/agents/acp/cost.ts +5 -75
- package/src/agents/acp/spawn-client.ts +11 -1
- package/src/agents/claude/cost.ts +12 -264
- package/src/agents/claude/execution.ts +12 -1
- package/src/agents/cost/calculate.ts +154 -0
- package/src/agents/cost/index.ts +10 -0
- package/src/agents/cost/parse.ts +97 -0
- package/src/agents/cost/pricing.ts +59 -0
- package/src/agents/cost/types.ts +45 -0
- package/src/agents/index.ts +4 -2
- package/src/agents/types.ts +3 -0
- package/src/cli/init.ts +15 -1
- package/src/pipeline/stages/acceptance-setup.ts +1 -0
- package/src/precheck/checks-git.ts +28 -2
- package/src/precheck/checks-warnings.ts +30 -2
- package/src/precheck/checks.ts +1 -0
- package/src/precheck/index.ts +2 -0
|
@@ -1,268 +1,16 @@
|
|
|
1
1
|
/**
|
|
2
|
-
* Cost Tracking
|
|
2
|
+
* Cost Tracking — re-exports from the shared src/agents/cost/ module.
|
|
3
3
|
*
|
|
4
|
-
*
|
|
5
|
-
*
|
|
4
|
+
* Kept for zero-breakage backward compatibility.
|
|
5
|
+
* Import directly from src/agents/cost for new code.
|
|
6
6
|
*/
|
|
7
7
|
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
|
|
11
|
-
|
|
12
|
-
|
|
13
|
-
|
|
14
|
-
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
export interface TokenUsage {
|
|
18
|
-
inputTokens: number;
|
|
19
|
-
outputTokens: number;
|
|
20
|
-
}
|
|
21
|
-
|
|
22
|
-
/** Cost estimate with confidence indicator */
|
|
23
|
-
export interface CostEstimate {
|
|
24
|
-
cost: number;
|
|
25
|
-
confidence: "exact" | "estimated" | "fallback";
|
|
26
|
-
}
|
|
27
|
-
|
|
28
|
-
/** Model tier cost rates (as of 2025-01) */
|
|
29
|
-
export const COST_RATES: Record<ModelTier, ModelCostRates> = {
|
|
30
|
-
fast: {
|
|
31
|
-
// Haiku 4.5
|
|
32
|
-
inputPer1M: 0.8,
|
|
33
|
-
outputPer1M: 4.0,
|
|
34
|
-
},
|
|
35
|
-
balanced: {
|
|
36
|
-
// Sonnet 4.5
|
|
37
|
-
inputPer1M: 3.0,
|
|
38
|
-
outputPer1M: 15.0,
|
|
39
|
-
},
|
|
40
|
-
powerful: {
|
|
41
|
-
// Opus 4
|
|
42
|
-
inputPer1M: 15.0,
|
|
43
|
-
outputPer1M: 75.0,
|
|
44
|
-
},
|
|
45
|
-
};
|
|
46
|
-
|
|
47
|
-
/**
|
|
48
|
-
* Token usage with confidence indicator.
|
|
49
|
-
*/
|
|
50
|
-
export interface TokenUsageWithConfidence {
|
|
51
|
-
inputTokens: number;
|
|
52
|
-
outputTokens: number;
|
|
53
|
-
confidence: "exact" | "estimated";
|
|
54
|
-
}
|
|
55
|
-
|
|
56
|
-
/**
|
|
57
|
-
* Parse Claude Code output for token usage.
|
|
58
|
-
*
|
|
59
|
-
* Supports multiple formats with varying confidence levels:
|
|
60
|
-
* - JSON structured output → "exact" confidence
|
|
61
|
-
* - Markdown/plain text patterns → "estimated" confidence
|
|
62
|
-
*
|
|
63
|
-
* Uses specific regex patterns to reduce false positives.
|
|
64
|
-
*
|
|
65
|
-
* @param output - Agent stdout + stderr combined
|
|
66
|
-
* @returns Token usage with confidence indicator, or null if tokens cannot be parsed
|
|
67
|
-
*
|
|
68
|
-
* @example
|
|
69
|
-
* ```ts
|
|
70
|
-
* // JSON format (exact)
|
|
71
|
-
* const usage1 = parseTokenUsage('{"usage": {"input_tokens": 1234, "output_tokens": 5678}}');
|
|
72
|
-
* // { inputTokens: 1234, outputTokens: 5678, confidence: 'exact' }
|
|
73
|
-
*
|
|
74
|
-
* // Markdown format (estimated)
|
|
75
|
-
* const usage2 = parseTokenUsage('Input tokens: 1234\nOutput tokens: 5678');
|
|
76
|
-
* // { inputTokens: 1234, outputTokens: 5678, confidence: 'estimated' }
|
|
77
|
-
*
|
|
78
|
-
* // Unparseable
|
|
79
|
-
* const usage3 = parseTokenUsage('No token data here');
|
|
80
|
-
* // null
|
|
81
|
-
* ```
|
|
82
|
-
*/
|
|
83
|
-
export function parseTokenUsage(output: string): TokenUsageWithConfidence | null {
|
|
84
|
-
// Try JSON format first (most reliable) - confidence: exact
|
|
85
|
-
try {
|
|
86
|
-
const jsonMatch = output.match(
|
|
87
|
-
/\{[^}]*"usage"\s*:\s*\{[^}]*"input_tokens"\s*:\s*(\d+)[^}]*"output_tokens"\s*:\s*(\d+)[^}]*\}[^}]*\}/,
|
|
88
|
-
);
|
|
89
|
-
if (jsonMatch) {
|
|
90
|
-
return {
|
|
91
|
-
inputTokens: Number.parseInt(jsonMatch[1], 10),
|
|
92
|
-
outputTokens: Number.parseInt(jsonMatch[2], 10),
|
|
93
|
-
confidence: "exact",
|
|
94
|
-
};
|
|
95
|
-
}
|
|
96
|
-
|
|
97
|
-
// Try parsing as full JSON object
|
|
98
|
-
const lines = output.split("\n");
|
|
99
|
-
for (const line of lines) {
|
|
100
|
-
if (line.trim().startsWith("{")) {
|
|
101
|
-
try {
|
|
102
|
-
const parsed = JSON.parse(line);
|
|
103
|
-
if (parsed.usage?.input_tokens && parsed.usage?.output_tokens) {
|
|
104
|
-
return {
|
|
105
|
-
inputTokens: parsed.usage.input_tokens,
|
|
106
|
-
outputTokens: parsed.usage.output_tokens,
|
|
107
|
-
confidence: "exact",
|
|
108
|
-
};
|
|
109
|
-
}
|
|
110
|
-
} catch {
|
|
111
|
-
// Not valid JSON, continue
|
|
112
|
-
}
|
|
113
|
-
}
|
|
114
|
-
}
|
|
115
|
-
} catch {
|
|
116
|
-
// JSON parsing failed, try regex patterns
|
|
117
|
-
}
|
|
118
|
-
|
|
119
|
-
// Try specific markdown-style patterns (more specific to reduce false positives)
|
|
120
|
-
// Match "Input tokens: 1234" or "input_tokens: 1234" or "INPUT TOKENS: 1234"
|
|
121
|
-
// Use word boundary at start, require colon or space after keyword, then digits
|
|
122
|
-
// confidence: estimated (regex-based)
|
|
123
|
-
const inputMatch = output.match(/\b(?:input|input_tokens)\s*:\s*(\d{2,})|(?:input)\s+(?:tokens?)\s*:\s*(\d{2,})/i);
|
|
124
|
-
const outputMatch = output.match(
|
|
125
|
-
/\b(?:output|output_tokens)\s*:\s*(\d{2,})|(?:output)\s+(?:tokens?)\s*:\s*(\d{2,})/i,
|
|
126
|
-
);
|
|
127
|
-
|
|
128
|
-
if (inputMatch && outputMatch) {
|
|
129
|
-
// Extract token counts (may be in capture group 1 or 2)
|
|
130
|
-
const inputTokens = Number.parseInt(inputMatch[1] || inputMatch[2], 10);
|
|
131
|
-
const outputTokens = Number.parseInt(outputMatch[1] || outputMatch[2], 10);
|
|
132
|
-
|
|
133
|
-
// Sanity check: reject if tokens seem unreasonably large (> 1M each)
|
|
134
|
-
if (inputTokens > 1_000_000 || outputTokens > 1_000_000) {
|
|
135
|
-
return null;
|
|
136
|
-
}
|
|
137
|
-
|
|
138
|
-
return {
|
|
139
|
-
inputTokens,
|
|
140
|
-
outputTokens,
|
|
141
|
-
confidence: "estimated",
|
|
142
|
-
};
|
|
143
|
-
}
|
|
144
|
-
|
|
145
|
-
return null;
|
|
146
|
-
}
|
|
147
|
-
|
|
148
|
-
/**
|
|
149
|
-
* Estimate cost in USD based on token usage.
|
|
150
|
-
*
|
|
151
|
-
* Calculates total cost using tier-specific rates per 1M tokens.
|
|
152
|
-
*
|
|
153
|
-
* @param modelTier - Model tier (fast/balanced/powerful)
|
|
154
|
-
* @param inputTokens - Number of input tokens consumed
|
|
155
|
-
* @param outputTokens - Number of output tokens generated
|
|
156
|
-
* @returns Total cost in USD
|
|
157
|
-
*
|
|
158
|
-
* @example
|
|
159
|
-
* ```ts
|
|
160
|
-
* const cost = estimateCost("balanced", 10000, 5000);
|
|
161
|
-
* // Sonnet 4.5: (10000/1M * $3.00) + (5000/1M * $15.00) = $0.105
|
|
162
|
-
* ```
|
|
163
|
-
*/
|
|
164
|
-
export function estimateCost(
|
|
165
|
-
modelTier: ModelTier,
|
|
166
|
-
inputTokens: number,
|
|
167
|
-
outputTokens: number,
|
|
168
|
-
customRates?: ModelCostRates,
|
|
169
|
-
): number {
|
|
170
|
-
const rates = customRates ?? COST_RATES[modelTier];
|
|
171
|
-
const inputCost = (inputTokens / 1_000_000) * rates.inputPer1M;
|
|
172
|
-
const outputCost = (outputTokens / 1_000_000) * rates.outputPer1M;
|
|
173
|
-
return inputCost + outputCost;
|
|
174
|
-
}
|
|
175
|
-
|
|
176
|
-
/**
|
|
177
|
-
* Estimate cost from agent output by parsing token usage.
|
|
178
|
-
*
|
|
179
|
-
* Attempts to extract token counts from stdout/stderr, then calculates cost.
|
|
180
|
-
* Returns null if tokens cannot be parsed (caller should use fallback estimation).
|
|
181
|
-
*
|
|
182
|
-
* @param modelTier - Model tier for cost calculation
|
|
183
|
-
* @param output - Agent stdout + stderr combined
|
|
184
|
-
* @returns Cost estimate with confidence indicator, or null if unparseable
|
|
185
|
-
*
|
|
186
|
-
* @example
|
|
187
|
-
* ```ts
|
|
188
|
-
* const estimate = estimateCostFromOutput("balanced", agentOutput);
|
|
189
|
-
* if (estimate) {
|
|
190
|
-
* console.log(`Cost: $${estimate.cost.toFixed(4)} (${estimate.confidence})`);
|
|
191
|
-
* } else {
|
|
192
|
-
* // Fall back to duration-based estimation
|
|
193
|
-
* }
|
|
194
|
-
* ```
|
|
195
|
-
*/
|
|
196
|
-
export function estimateCostFromOutput(modelTier: ModelTier, output: string): CostEstimate | null {
|
|
197
|
-
const usage = parseTokenUsage(output);
|
|
198
|
-
if (!usage) {
|
|
199
|
-
return null;
|
|
200
|
-
}
|
|
201
|
-
const cost = estimateCost(modelTier, usage.inputTokens, usage.outputTokens);
|
|
202
|
-
return {
|
|
203
|
-
cost,
|
|
204
|
-
confidence: usage.confidence,
|
|
205
|
-
};
|
|
206
|
-
}
|
|
207
|
-
|
|
208
|
-
/**
|
|
209
|
-
* Fallback cost estimation based on runtime duration.
|
|
210
|
-
*
|
|
211
|
-
* Used when token usage cannot be parsed from agent output.
|
|
212
|
-
* Provides conservative estimates using per-minute rates.
|
|
213
|
-
*
|
|
214
|
-
* @param modelTier - Model tier for cost calculation
|
|
215
|
-
* @param durationMs - Agent runtime in milliseconds
|
|
216
|
-
* @returns Cost estimate with 'fallback' confidence
|
|
217
|
-
*
|
|
218
|
-
* @example
|
|
219
|
-
* ```ts
|
|
220
|
-
* const estimate = estimateCostByDuration("balanced", 120000); // 2 minutes
|
|
221
|
-
* // { cost: 0.10, confidence: 'fallback' }
|
|
222
|
-
* // Sonnet: 2 min * $0.05/min = $0.10
|
|
223
|
-
* ```
|
|
224
|
-
*/
|
|
225
|
-
export function estimateCostByDuration(modelTier: ModelTier, durationMs: number): CostEstimate {
|
|
226
|
-
const costPerMinute: Record<ModelTier, number> = {
|
|
227
|
-
fast: 0.01,
|
|
228
|
-
balanced: 0.05,
|
|
229
|
-
powerful: 0.15,
|
|
230
|
-
};
|
|
231
|
-
const minutes = durationMs / 60000;
|
|
232
|
-
const cost = minutes * costPerMinute[modelTier];
|
|
233
|
-
return {
|
|
234
|
-
cost,
|
|
235
|
-
confidence: "fallback",
|
|
236
|
-
};
|
|
237
|
-
}
|
|
238
|
-
|
|
239
|
-
/**
|
|
240
|
-
* Format cost estimate with confidence indicator for display.
|
|
241
|
-
*
|
|
242
|
-
* @param estimate - Cost estimate with confidence level
|
|
243
|
-
* @returns Formatted cost string with confidence indicator
|
|
244
|
-
*
|
|
245
|
-
* @example
|
|
246
|
-
* ```ts
|
|
247
|
-
* formatCostWithConfidence({ cost: 0.12, confidence: 'exact' });
|
|
248
|
-
* // "$0.12"
|
|
249
|
-
*
|
|
250
|
-
* formatCostWithConfidence({ cost: 0.15, confidence: 'estimated' });
|
|
251
|
-
* // "~$0.15"
|
|
252
|
-
*
|
|
253
|
-
* formatCostWithConfidence({ cost: 0.05, confidence: 'fallback' });
|
|
254
|
-
* // "~$0.05 (duration-based)"
|
|
255
|
-
* ```
|
|
256
|
-
*/
|
|
257
|
-
export function formatCostWithConfidence(estimate: CostEstimate): string {
|
|
258
|
-
const formattedCost = `$${estimate.cost.toFixed(2)}`;
|
|
259
|
-
|
|
260
|
-
switch (estimate.confidence) {
|
|
261
|
-
case "exact":
|
|
262
|
-
return formattedCost;
|
|
263
|
-
case "estimated":
|
|
264
|
-
return `~${formattedCost}`;
|
|
265
|
-
case "fallback":
|
|
266
|
-
return `~${formattedCost} (duration-based)`;
|
|
267
|
-
}
|
|
268
|
-
}
|
|
8
|
+
export type { ModelCostRates, TokenUsage, CostEstimate, TokenUsageWithConfidence } from "../cost";
|
|
9
|
+
export {
|
|
10
|
+
COST_RATES,
|
|
11
|
+
parseTokenUsage,
|
|
12
|
+
estimateCost,
|
|
13
|
+
estimateCostFromOutput,
|
|
14
|
+
estimateCostByDuration,
|
|
15
|
+
formatCostWithConfidence,
|
|
16
|
+
} from "../cost";
|
|
@@ -4,6 +4,8 @@
|
|
|
4
4
|
* Handles building commands, preparing environment, and process execution.
|
|
5
5
|
*/
|
|
6
6
|
|
|
7
|
+
import { homedir } from "node:os";
|
|
8
|
+
import { isAbsolute } from "node:path";
|
|
7
9
|
import { resolvePermissions } from "../../config/permissions";
|
|
8
10
|
import type { PidRegistry } from "../../execution/pid-registry";
|
|
9
11
|
import { withProcessTimeout } from "../../execution/timeout-handler";
|
|
@@ -65,13 +67,22 @@ export function buildCommand(binary: string, options: AgentRunOptions): string[]
|
|
|
65
67
|
export function buildAllowedEnv(options: AgentRunOptions): Record<string, string | undefined> {
|
|
66
68
|
const allowed: Record<string, string | undefined> = {};
|
|
67
69
|
|
|
68
|
-
const essentialVars = ["PATH", "
|
|
70
|
+
const essentialVars = ["PATH", "TMPDIR", "NODE_ENV", "USER", "LOGNAME"];
|
|
69
71
|
for (const varName of essentialVars) {
|
|
70
72
|
if (process.env[varName]) {
|
|
71
73
|
allowed[varName] = process.env[varName];
|
|
72
74
|
}
|
|
73
75
|
}
|
|
74
76
|
|
|
77
|
+
// Sanitize HOME — must be absolute path. Unexpanded "~" causes literal ~/dir in cwd.
|
|
78
|
+
const rawHome = process.env.HOME ?? "";
|
|
79
|
+
const safeHome = rawHome && isAbsolute(rawHome) ? rawHome : homedir();
|
|
80
|
+
if (rawHome !== safeHome) {
|
|
81
|
+
const logger = getLogger();
|
|
82
|
+
logger.warn("env", `HOME env is not absolute ("${rawHome}"), falling back to os.homedir(): ${safeHome}`);
|
|
83
|
+
}
|
|
84
|
+
allowed.HOME = safeHome;
|
|
85
|
+
|
|
75
86
|
const apiKeyVars = ["ANTHROPIC_API_KEY", "OPENAI_API_KEY"];
|
|
76
87
|
for (const varName of apiKeyVars) {
|
|
77
88
|
if (process.env[varName]) {
|
|
@@ -0,0 +1,154 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Cost calculation functions for all agent adapters.
|
|
3
|
+
*/
|
|
4
|
+
|
|
5
|
+
import type { ModelTier } from "../../config/schema";
|
|
6
|
+
import { parseTokenUsage } from "./parse";
|
|
7
|
+
import { COST_RATES, MODEL_PRICING } from "./pricing";
|
|
8
|
+
import type { CostEstimate, ModelCostRates, SessionTokenUsage } from "./types";
|
|
9
|
+
|
|
10
|
+
/**
|
|
11
|
+
* Estimate cost in USD based on token usage and model tier.
|
|
12
|
+
*
|
|
13
|
+
* @param modelTier - Model tier (fast/balanced/powerful)
|
|
14
|
+
* @param inputTokens - Number of input tokens consumed
|
|
15
|
+
* @param outputTokens - Number of output tokens generated
|
|
16
|
+
* @param customRates - Optional custom rates (overrides tier defaults)
|
|
17
|
+
* @returns Total cost in USD
|
|
18
|
+
*
|
|
19
|
+
* @example
|
|
20
|
+
* ```ts
|
|
21
|
+
* const cost = estimateCost("balanced", 10000, 5000);
|
|
22
|
+
* // Sonnet 4.5: (10000/1M * $3.00) + (5000/1M * $15.00) = $0.105
|
|
23
|
+
* ```
|
|
24
|
+
*/
|
|
25
|
+
export function estimateCost(
|
|
26
|
+
modelTier: ModelTier,
|
|
27
|
+
inputTokens: number,
|
|
28
|
+
outputTokens: number,
|
|
29
|
+
customRates?: ModelCostRates,
|
|
30
|
+
): number {
|
|
31
|
+
const rates = customRates ?? COST_RATES[modelTier];
|
|
32
|
+
const inputCost = (inputTokens / 1_000_000) * rates.inputPer1M;
|
|
33
|
+
const outputCost = (outputTokens / 1_000_000) * rates.outputPer1M;
|
|
34
|
+
return inputCost + outputCost;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
/**
|
|
38
|
+
* Estimate cost from agent output by parsing token usage.
|
|
39
|
+
*
|
|
40
|
+
* Attempts to extract token counts from stdout/stderr, then calculates cost.
|
|
41
|
+
* Returns null if tokens cannot be parsed.
|
|
42
|
+
*
|
|
43
|
+
* @param modelTier - Model tier for cost calculation
|
|
44
|
+
* @param output - Agent stdout + stderr combined
|
|
45
|
+
* @returns Cost estimate with confidence indicator, or null if unparseable
|
|
46
|
+
*/
|
|
47
|
+
export function estimateCostFromOutput(modelTier: ModelTier, output: string): CostEstimate | null {
|
|
48
|
+
const usage = parseTokenUsage(output);
|
|
49
|
+
if (!usage) {
|
|
50
|
+
return null;
|
|
51
|
+
}
|
|
52
|
+
const cost = estimateCost(modelTier, usage.inputTokens, usage.outputTokens);
|
|
53
|
+
return {
|
|
54
|
+
cost,
|
|
55
|
+
confidence: usage.confidence,
|
|
56
|
+
};
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
/**
|
|
60
|
+
* Fallback cost estimation based on runtime duration.
|
|
61
|
+
*
|
|
62
|
+
* Used when token usage cannot be parsed from agent output.
|
|
63
|
+
* Provides conservative estimates using per-minute rates.
|
|
64
|
+
*
|
|
65
|
+
* @param modelTier - Model tier for cost calculation
|
|
66
|
+
* @param durationMs - Agent runtime in milliseconds
|
|
67
|
+
* @returns Cost estimate with 'fallback' confidence
|
|
68
|
+
*
|
|
69
|
+
* @example
|
|
70
|
+
* ```ts
|
|
71
|
+
* const estimate = estimateCostByDuration("balanced", 120000); // 2 minutes
|
|
72
|
+
* // { cost: 0.10, confidence: 'fallback' }
|
|
73
|
+
* // Sonnet: 2 min * $0.05/min = $0.10
|
|
74
|
+
* ```
|
|
75
|
+
*/
|
|
76
|
+
export function estimateCostByDuration(modelTier: ModelTier, durationMs: number): CostEstimate {
|
|
77
|
+
const costPerMinute: Record<ModelTier, number> = {
|
|
78
|
+
fast: 0.01,
|
|
79
|
+
balanced: 0.05,
|
|
80
|
+
powerful: 0.15,
|
|
81
|
+
};
|
|
82
|
+
const minutes = durationMs / 60000;
|
|
83
|
+
const cost = minutes * costPerMinute[modelTier];
|
|
84
|
+
return {
|
|
85
|
+
cost,
|
|
86
|
+
confidence: "fallback",
|
|
87
|
+
};
|
|
88
|
+
}
|
|
89
|
+
|
|
90
|
+
/**
|
|
91
|
+
* Format cost estimate with confidence indicator for display.
|
|
92
|
+
*
|
|
93
|
+
* @param estimate - Cost estimate with confidence level
|
|
94
|
+
* @returns Formatted cost string with confidence indicator
|
|
95
|
+
*
|
|
96
|
+
* @example
|
|
97
|
+
* ```ts
|
|
98
|
+
* formatCostWithConfidence({ cost: 0.12, confidence: 'exact' });
|
|
99
|
+
* // "$0.12"
|
|
100
|
+
*
|
|
101
|
+
* formatCostWithConfidence({ cost: 0.15, confidence: 'estimated' });
|
|
102
|
+
* // "~$0.15"
|
|
103
|
+
*
|
|
104
|
+
* formatCostWithConfidence({ cost: 0.05, confidence: 'fallback' });
|
|
105
|
+
* // "~$0.05 (duration-based)"
|
|
106
|
+
* ```
|
|
107
|
+
*/
|
|
108
|
+
export function formatCostWithConfidence(estimate: CostEstimate): string {
|
|
109
|
+
const formattedCost = `$${estimate.cost.toFixed(2)}`;
|
|
110
|
+
|
|
111
|
+
switch (estimate.confidence) {
|
|
112
|
+
case "exact":
|
|
113
|
+
return formattedCost;
|
|
114
|
+
case "estimated":
|
|
115
|
+
return `~${formattedCost}`;
|
|
116
|
+
case "fallback":
|
|
117
|
+
return `~${formattedCost} (duration-based)`;
|
|
118
|
+
}
|
|
119
|
+
}
|
|
120
|
+
|
|
121
|
+
/**
|
|
122
|
+
* Calculate USD cost from ACP session token counts using per-model pricing.
|
|
123
|
+
*
|
|
124
|
+
* @param usage - Token counts from cumulative_token_usage
|
|
125
|
+
* @param model - Model identifier (e.g., 'claude-sonnet-4', 'claude-haiku-4-5')
|
|
126
|
+
* @returns Estimated cost in USD
|
|
127
|
+
*/
|
|
128
|
+
export function estimateCostFromTokenUsage(usage: SessionTokenUsage, model: string): number {
|
|
129
|
+
const pricing = MODEL_PRICING[model];
|
|
130
|
+
|
|
131
|
+
if (!pricing) {
|
|
132
|
+
// Fallback: use average rate for unknown models
|
|
133
|
+
const fallbackInputRate = 3 / 1_000_000;
|
|
134
|
+
const fallbackOutputRate = 15 / 1_000_000;
|
|
135
|
+
const inputCost = (usage.input_tokens ?? 0) * fallbackInputRate;
|
|
136
|
+
const outputCost = (usage.output_tokens ?? 0) * fallbackOutputRate;
|
|
137
|
+
const cacheReadCost = (usage.cache_read_input_tokens ?? 0) * (0.5 / 1_000_000);
|
|
138
|
+
const cacheCreationCost = (usage.cache_creation_input_tokens ?? 0) * (2 / 1_000_000);
|
|
139
|
+
return inputCost + outputCost + cacheReadCost + cacheCreationCost;
|
|
140
|
+
}
|
|
141
|
+
|
|
142
|
+
// Convert $/1M rates to $/token
|
|
143
|
+
const inputRate = pricing.input / 1_000_000;
|
|
144
|
+
const outputRate = pricing.output / 1_000_000;
|
|
145
|
+
const cacheReadRate = (pricing.cacheRead ?? pricing.input * 0.1) / 1_000_000;
|
|
146
|
+
const cacheCreationRate = (pricing.cacheCreation ?? pricing.input * 0.33) / 1_000_000;
|
|
147
|
+
|
|
148
|
+
const inputCost = (usage.input_tokens ?? 0) * inputRate;
|
|
149
|
+
const outputCost = (usage.output_tokens ?? 0) * outputRate;
|
|
150
|
+
const cacheReadCost = (usage.cache_read_input_tokens ?? 0) * cacheReadRate;
|
|
151
|
+
const cacheCreationCost = (usage.cache_creation_input_tokens ?? 0) * cacheCreationRate;
|
|
152
|
+
|
|
153
|
+
return inputCost + outputCost + cacheReadCost + cacheCreationCost;
|
|
154
|
+
}
|
|
@@ -0,0 +1,10 @@
|
|
|
1
|
+
export type { ModelCostRates, TokenUsage, CostEstimate, TokenUsageWithConfidence, SessionTokenUsage } from "./types";
|
|
2
|
+
export { COST_RATES, MODEL_PRICING } from "./pricing";
|
|
3
|
+
export { parseTokenUsage } from "./parse";
|
|
4
|
+
export {
|
|
5
|
+
estimateCost,
|
|
6
|
+
estimateCostFromOutput,
|
|
7
|
+
estimateCostByDuration,
|
|
8
|
+
formatCostWithConfidence,
|
|
9
|
+
estimateCostFromTokenUsage,
|
|
10
|
+
} from "./calculate";
|
|
@@ -0,0 +1,97 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Token usage parsing from raw agent output strings.
|
|
3
|
+
*/
|
|
4
|
+
|
|
5
|
+
import type { TokenUsageWithConfidence } from "./types";
|
|
6
|
+
|
|
7
|
+
/**
|
|
8
|
+
* Parse Claude Code output for token usage.
|
|
9
|
+
*
|
|
10
|
+
* Supports multiple formats with varying confidence levels:
|
|
11
|
+
* - JSON structured output → "exact" confidence
|
|
12
|
+
* - Markdown/plain text patterns → "estimated" confidence
|
|
13
|
+
*
|
|
14
|
+
* Uses specific regex patterns to reduce false positives.
|
|
15
|
+
*
|
|
16
|
+
* @param output - Agent stdout + stderr combined
|
|
17
|
+
* @returns Token usage with confidence indicator, or null if tokens cannot be parsed
|
|
18
|
+
*
|
|
19
|
+
* @example
|
|
20
|
+
* ```ts
|
|
21
|
+
* // JSON format (exact)
|
|
22
|
+
* const usage1 = parseTokenUsage('{"usage": {"input_tokens": 1234, "output_tokens": 5678}}');
|
|
23
|
+
* // { inputTokens: 1234, outputTokens: 5678, confidence: 'exact' }
|
|
24
|
+
*
|
|
25
|
+
* // Markdown format (estimated)
|
|
26
|
+
* const usage2 = parseTokenUsage('Input tokens: 1234\nOutput tokens: 5678');
|
|
27
|
+
* // { inputTokens: 1234, outputTokens: 5678, confidence: 'estimated' }
|
|
28
|
+
*
|
|
29
|
+
* // Unparseable
|
|
30
|
+
* const usage3 = parseTokenUsage('No token data here');
|
|
31
|
+
* // null
|
|
32
|
+
* ```
|
|
33
|
+
*/
|
|
34
|
+
export function parseTokenUsage(output: string): TokenUsageWithConfidence | null {
|
|
35
|
+
// Try JSON format first (most reliable) - confidence: exact
|
|
36
|
+
try {
|
|
37
|
+
const jsonMatch = output.match(
|
|
38
|
+
/\{[^}]*"usage"\s*:\s*\{[^}]*"input_tokens"\s*:\s*(\d+)[^}]*"output_tokens"\s*:\s*(\d+)[^}]*\}[^}]*\}/,
|
|
39
|
+
);
|
|
40
|
+
if (jsonMatch) {
|
|
41
|
+
return {
|
|
42
|
+
inputTokens: Number.parseInt(jsonMatch[1], 10),
|
|
43
|
+
outputTokens: Number.parseInt(jsonMatch[2], 10),
|
|
44
|
+
confidence: "exact",
|
|
45
|
+
};
|
|
46
|
+
}
|
|
47
|
+
|
|
48
|
+
// Try parsing as full JSON object
|
|
49
|
+
const lines = output.split("\n");
|
|
50
|
+
for (const line of lines) {
|
|
51
|
+
if (line.trim().startsWith("{")) {
|
|
52
|
+
try {
|
|
53
|
+
const parsed = JSON.parse(line);
|
|
54
|
+
if (parsed.usage?.input_tokens && parsed.usage?.output_tokens) {
|
|
55
|
+
return {
|
|
56
|
+
inputTokens: parsed.usage.input_tokens,
|
|
57
|
+
outputTokens: parsed.usage.output_tokens,
|
|
58
|
+
confidence: "exact",
|
|
59
|
+
};
|
|
60
|
+
}
|
|
61
|
+
} catch {
|
|
62
|
+
// Not valid JSON, continue
|
|
63
|
+
}
|
|
64
|
+
}
|
|
65
|
+
}
|
|
66
|
+
} catch {
|
|
67
|
+
// JSON parsing failed, try regex patterns
|
|
68
|
+
}
|
|
69
|
+
|
|
70
|
+
// Try specific markdown-style patterns (more specific to reduce false positives)
|
|
71
|
+
// Match "Input tokens: 1234" or "input_tokens: 1234" or "INPUT TOKENS: 1234"
|
|
72
|
+
// Use word boundary at start, require colon or space after keyword, then digits
|
|
73
|
+
// confidence: estimated (regex-based)
|
|
74
|
+
const inputMatch = output.match(/\b(?:input|input_tokens)\s*:\s*(\d{2,})|(?:input)\s+(?:tokens?)\s*:\s*(\d{2,})/i);
|
|
75
|
+
const outputMatch = output.match(
|
|
76
|
+
/\b(?:output|output_tokens)\s*:\s*(\d{2,})|(?:output)\s+(?:tokens?)\s*:\s*(\d{2,})/i,
|
|
77
|
+
);
|
|
78
|
+
|
|
79
|
+
if (inputMatch && outputMatch) {
|
|
80
|
+
// Extract token counts (may be in capture group 1 or 2)
|
|
81
|
+
const inputTokens = Number.parseInt(inputMatch[1] || inputMatch[2], 10);
|
|
82
|
+
const outputTokens = Number.parseInt(outputMatch[1] || outputMatch[2], 10);
|
|
83
|
+
|
|
84
|
+
// Sanity check: reject if tokens seem unreasonably large (> 1M each)
|
|
85
|
+
if (inputTokens > 1_000_000 || outputTokens > 1_000_000) {
|
|
86
|
+
return null;
|
|
87
|
+
}
|
|
88
|
+
|
|
89
|
+
return {
|
|
90
|
+
inputTokens,
|
|
91
|
+
outputTokens,
|
|
92
|
+
confidence: "estimated",
|
|
93
|
+
};
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
return null;
|
|
97
|
+
}
|
|
@@ -0,0 +1,59 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Cost rate tables for all supported model tiers and specific models.
|
|
3
|
+
*/
|
|
4
|
+
|
|
5
|
+
import type { ModelTier } from "../../config/schema";
|
|
6
|
+
import type { ModelCostRates } from "./types";
|
|
7
|
+
|
|
8
|
+
/** Model tier cost rates (as of 2025-01) */
|
|
9
|
+
export const COST_RATES: Record<ModelTier, ModelCostRates> = {
|
|
10
|
+
fast: {
|
|
11
|
+
// Haiku 4.5
|
|
12
|
+
inputPer1M: 0.8,
|
|
13
|
+
outputPer1M: 4.0,
|
|
14
|
+
},
|
|
15
|
+
balanced: {
|
|
16
|
+
// Sonnet 4.5
|
|
17
|
+
inputPer1M: 3.0,
|
|
18
|
+
outputPer1M: 15.0,
|
|
19
|
+
},
|
|
20
|
+
powerful: {
|
|
21
|
+
// Opus 4
|
|
22
|
+
inputPer1M: 15.0,
|
|
23
|
+
outputPer1M: 75.0,
|
|
24
|
+
},
|
|
25
|
+
};
|
|
26
|
+
|
|
27
|
+
/** Per-model pricing in $/1M tokens: { input, output } */
|
|
28
|
+
export const MODEL_PRICING: Record<
|
|
29
|
+
string,
|
|
30
|
+
{ input: number; output: number; cacheRead?: number; cacheCreation?: number }
|
|
31
|
+
> = {
|
|
32
|
+
// Anthropic Claude models (short aliases)
|
|
33
|
+
sonnet: { input: 3, output: 15 },
|
|
34
|
+
haiku: { input: 0.8, output: 4.0, cacheRead: 0.1, cacheCreation: 1.0 },
|
|
35
|
+
opus: { input: 15, output: 75 },
|
|
36
|
+
|
|
37
|
+
// Anthropic Claude models (full names)
|
|
38
|
+
"claude-sonnet-4": { input: 3, output: 15 },
|
|
39
|
+
"claude-sonnet-4-5": { input: 3, output: 15 },
|
|
40
|
+
"claude-sonnet-4-6": { input: 3, output: 15 },
|
|
41
|
+
"claude-haiku": { input: 0.8, output: 4.0, cacheRead: 0.1, cacheCreation: 1.0 },
|
|
42
|
+
"claude-haiku-4-5": { input: 0.8, output: 4.0, cacheRead: 0.1, cacheCreation: 1.0 },
|
|
43
|
+
"claude-opus": { input: 15, output: 75 },
|
|
44
|
+
"claude-opus-4": { input: 15, output: 75 },
|
|
45
|
+
"claude-opus-4-6": { input: 15, output: 75 },
|
|
46
|
+
|
|
47
|
+
// OpenAI models
|
|
48
|
+
"gpt-4.1": { input: 10, output: 30 },
|
|
49
|
+
"gpt-4": { input: 30, output: 60 },
|
|
50
|
+
"gpt-3.5-turbo": { input: 0.5, output: 1.5 },
|
|
51
|
+
|
|
52
|
+
// Google Gemini
|
|
53
|
+
"gemini-2.5-pro": { input: 0.075, output: 0.3 },
|
|
54
|
+
"gemini-2-pro": { input: 0.075, output: 0.3 },
|
|
55
|
+
|
|
56
|
+
// OpenAI Codex
|
|
57
|
+
codex: { input: 0.02, output: 0.06 },
|
|
58
|
+
"code-davinci-002": { input: 0.02, output: 0.06 },
|
|
59
|
+
};
|
|
@@ -0,0 +1,45 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Cost tracking types — shared across all agent adapters.
|
|
3
|
+
*/
|
|
4
|
+
|
|
5
|
+
import type { ModelTier } from "../../config/schema";
|
|
6
|
+
|
|
7
|
+
export type { ModelTier };
|
|
8
|
+
|
|
9
|
+
/** Cost rates per 1M tokens (USD) */
|
|
10
|
+
export interface ModelCostRates {
|
|
11
|
+
inputPer1M: number;
|
|
12
|
+
outputPer1M: number;
|
|
13
|
+
}
|
|
14
|
+
|
|
15
|
+
/** Token usage data (camelCase — nax-internal representation) */
|
|
16
|
+
export interface TokenUsage {
|
|
17
|
+
inputTokens: number;
|
|
18
|
+
outputTokens: number;
|
|
19
|
+
}
|
|
20
|
+
|
|
21
|
+
/** Cost estimate with confidence indicator */
|
|
22
|
+
export interface CostEstimate {
|
|
23
|
+
cost: number;
|
|
24
|
+
confidence: "exact" | "estimated" | "fallback";
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
/** Token usage with confidence indicator */
|
|
28
|
+
export interface TokenUsageWithConfidence {
|
|
29
|
+
inputTokens: number;
|
|
30
|
+
outputTokens: number;
|
|
31
|
+
confidence: "exact" | "estimated";
|
|
32
|
+
}
|
|
33
|
+
|
|
34
|
+
/**
|
|
35
|
+
* Token usage from an ACP session's cumulative_token_usage field.
|
|
36
|
+
* Uses snake_case to match the ACP wire format.
|
|
37
|
+
*/
|
|
38
|
+
export interface SessionTokenUsage {
|
|
39
|
+
input_tokens: number;
|
|
40
|
+
output_tokens: number;
|
|
41
|
+
/** Cache read tokens — billed at a reduced rate */
|
|
42
|
+
cache_read_input_tokens?: number;
|
|
43
|
+
/** Cache creation tokens — billed at a higher creation rate */
|
|
44
|
+
cache_creation_input_tokens?: number;
|
|
45
|
+
}
|