trickle-cli 0.1.191 → 0.1.193
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/dist/commands/compliance.d.ts +17 -0
- package/dist/commands/compliance.js +251 -0
- package/dist/commands/init.js +14 -5
- package/dist/index.js +7 -0
- package/package.json +1 -1
- package/src/commands/compliance.ts +262 -0
- package/src/commands/init.ts +14 -5
- package/src/index.ts +7 -0
|
@@ -0,0 +1,17 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* trickle audit --compliance — Generate compliance audit report.
|
|
3
|
+
*
|
|
4
|
+
* Exports trickle's JSONL data as a structured compliance report for
|
|
5
|
+
* EU AI Act and Colorado AI Act requirements:
|
|
6
|
+
* - Decision lineage (LLM call → tool call → output)
|
|
7
|
+
* - Timestamped event log
|
|
8
|
+
* - Risk classification
|
|
9
|
+
* - Security scan results
|
|
10
|
+
* - Data processing summary
|
|
11
|
+
*
|
|
12
|
+
* Local-first: sensitive audit data never leaves the developer's machine.
|
|
13
|
+
*/
|
|
14
|
+
export declare function generateComplianceReport(opts: {
|
|
15
|
+
json?: boolean;
|
|
16
|
+
out?: string;
|
|
17
|
+
}): void;
|
|
@@ -0,0 +1,251 @@
|
|
|
1
|
+
"use strict";
|
|
2
|
+
/**
|
|
3
|
+
* trickle audit --compliance — Generate compliance audit report.
|
|
4
|
+
*
|
|
5
|
+
* Exports trickle's JSONL data as a structured compliance report for
|
|
6
|
+
* EU AI Act and Colorado AI Act requirements:
|
|
7
|
+
* - Decision lineage (LLM call → tool call → output)
|
|
8
|
+
* - Timestamped event log
|
|
9
|
+
* - Risk classification
|
|
10
|
+
* - Security scan results
|
|
11
|
+
* - Data processing summary
|
|
12
|
+
*
|
|
13
|
+
* Local-first: sensitive audit data never leaves the developer's machine.
|
|
14
|
+
*/
|
|
15
|
+
var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
|
|
16
|
+
if (k2 === undefined) k2 = k;
|
|
17
|
+
var desc = Object.getOwnPropertyDescriptor(m, k);
|
|
18
|
+
if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
|
|
19
|
+
desc = { enumerable: true, get: function() { return m[k]; } };
|
|
20
|
+
}
|
|
21
|
+
Object.defineProperty(o, k2, desc);
|
|
22
|
+
}) : (function(o, m, k, k2) {
|
|
23
|
+
if (k2 === undefined) k2 = k;
|
|
24
|
+
o[k2] = m[k];
|
|
25
|
+
}));
|
|
26
|
+
var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {
|
|
27
|
+
Object.defineProperty(o, "default", { enumerable: true, value: v });
|
|
28
|
+
}) : function(o, v) {
|
|
29
|
+
o["default"] = v;
|
|
30
|
+
});
|
|
31
|
+
var __importStar = (this && this.__importStar) || (function () {
|
|
32
|
+
var ownKeys = function(o) {
|
|
33
|
+
ownKeys = Object.getOwnPropertyNames || function (o) {
|
|
34
|
+
var ar = [];
|
|
35
|
+
for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;
|
|
36
|
+
return ar;
|
|
37
|
+
};
|
|
38
|
+
return ownKeys(o);
|
|
39
|
+
};
|
|
40
|
+
return function (mod) {
|
|
41
|
+
if (mod && mod.__esModule) return mod;
|
|
42
|
+
var result = {};
|
|
43
|
+
if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]);
|
|
44
|
+
__setModuleDefault(result, mod);
|
|
45
|
+
return result;
|
|
46
|
+
};
|
|
47
|
+
})();
|
|
48
|
+
var __importDefault = (this && this.__importDefault) || function (mod) {
|
|
49
|
+
return (mod && mod.__esModule) ? mod : { "default": mod };
|
|
50
|
+
};
|
|
51
|
+
Object.defineProperty(exports, "__esModule", { value: true });
|
|
52
|
+
exports.generateComplianceReport = generateComplianceReport;
|
|
53
|
+
const fs = __importStar(require("fs"));
|
|
54
|
+
const path = __importStar(require("path"));
|
|
55
|
+
const chalk_1 = __importDefault(require("chalk"));
|
|
56
|
+
function readJsonl(fp) {
|
|
57
|
+
if (!fs.existsSync(fp))
|
|
58
|
+
return [];
|
|
59
|
+
return fs.readFileSync(fp, 'utf-8').split('\n').filter(Boolean)
|
|
60
|
+
.map(l => { try {
|
|
61
|
+
return JSON.parse(l);
|
|
62
|
+
}
|
|
63
|
+
catch {
|
|
64
|
+
return null;
|
|
65
|
+
} }).filter(Boolean);
|
|
66
|
+
}
|
|
67
|
+
function generateComplianceReport(opts) {
|
|
68
|
+
const dir = process.env.TRICKLE_LOCAL_DIR || path.join(process.cwd(), '.trickle');
|
|
69
|
+
if (!fs.existsSync(dir)) {
|
|
70
|
+
console.log(chalk_1.default.yellow(' No .trickle/ data. Run trickle first.'));
|
|
71
|
+
return;
|
|
72
|
+
}
|
|
73
|
+
const llmCalls = readJsonl(path.join(dir, 'llm.jsonl'));
|
|
74
|
+
const agentEvents = readJsonl(path.join(dir, 'agents.jsonl'));
|
|
75
|
+
const mcpCalls = readJsonl(path.join(dir, 'mcp.jsonl'));
|
|
76
|
+
const errors = readJsonl(path.join(dir, 'errors.jsonl'));
|
|
77
|
+
const observations = readJsonl(path.join(dir, 'observations.jsonl'));
|
|
78
|
+
const calltrace = readJsonl(path.join(dir, 'calltrace.jsonl'));
|
|
79
|
+
// Build decision lineage — chronological event log
|
|
80
|
+
const lineage = [];
|
|
81
|
+
for (const e of agentEvents) {
|
|
82
|
+
lineage.push({
|
|
83
|
+
timestamp: e.timestamp || 0,
|
|
84
|
+
event: `agent:${e.event}`,
|
|
85
|
+
description: buildAgentDescription(e),
|
|
86
|
+
data: { framework: e.framework, tool: e.tool, chain: e.chain },
|
|
87
|
+
});
|
|
88
|
+
}
|
|
89
|
+
for (const c of llmCalls) {
|
|
90
|
+
lineage.push({
|
|
91
|
+
timestamp: c.timestamp || 0,
|
|
92
|
+
event: 'llm:call',
|
|
93
|
+
description: `${c.provider}/${c.model}: ${(c.inputPreview || '').substring(0, 80)} → ${(c.outputPreview || '').substring(0, 80)}`,
|
|
94
|
+
data: { model: c.model, tokens: c.totalTokens, cost: c.estimatedCostUsd, error: c.error },
|
|
95
|
+
});
|
|
96
|
+
}
|
|
97
|
+
for (const m of mcpCalls) {
|
|
98
|
+
if (m.tool === '__list_tools')
|
|
99
|
+
continue;
|
|
100
|
+
lineage.push({
|
|
101
|
+
timestamp: m.timestamp || 0,
|
|
102
|
+
event: `mcp:${m.direction}`,
|
|
103
|
+
description: `${m.tool}(${JSON.stringify(m.args || {}).substring(0, 60)}) → ${(m.resultPreview || '').substring(0, 60)}`,
|
|
104
|
+
data: { tool: m.tool, direction: m.direction, isError: m.isError },
|
|
105
|
+
});
|
|
106
|
+
}
|
|
107
|
+
lineage.sort((a, b) => a.timestamp - b.timestamp);
|
|
108
|
+
// Data processing summary
|
|
109
|
+
const providers = [...new Set(llmCalls.map(c => c.provider).filter(Boolean))];
|
|
110
|
+
const models = [...new Set(llmCalls.map(c => `${c.provider}/${c.model}`).filter(Boolean))];
|
|
111
|
+
const totalTokens = llmCalls.reduce((s, c) => s + (c.totalTokens || 0), 0);
|
|
112
|
+
const totalCost = llmCalls.reduce((s, c) => s + (c.estimatedCostUsd || 0), 0);
|
|
113
|
+
const agentTools = [...new Set(agentEvents.filter(e => e.tool).map(e => e.tool))];
|
|
114
|
+
const mcpTools = [...new Set(mcpCalls.filter(c => c.tool && c.tool !== '__list_tools').map(c => c.tool))];
|
|
115
|
+
const dataSources = [...new Set(observations.map(o => o.module).filter(Boolean))];
|
|
116
|
+
// Security scan
|
|
117
|
+
let securityFindings = [];
|
|
118
|
+
try {
|
|
119
|
+
const { runSecurityScan } = require('./security');
|
|
120
|
+
const origLog = console.log;
|
|
121
|
+
console.log = () => { };
|
|
122
|
+
const result = runSecurityScan({ dir });
|
|
123
|
+
console.log = origLog;
|
|
124
|
+
securityFindings = result.findings.map((f) => ({
|
|
125
|
+
severity: f.severity, category: f.category,
|
|
126
|
+
message: f.message, evidence: (f.evidence || '').substring(0, 100),
|
|
127
|
+
}));
|
|
128
|
+
}
|
|
129
|
+
catch { }
|
|
130
|
+
// Eval score
|
|
131
|
+
let evalScore = null;
|
|
132
|
+
try {
|
|
133
|
+
const { evalCommand } = require('./eval');
|
|
134
|
+
// We can't easily call evalCommand and get the result, so build a lightweight score
|
|
135
|
+
const completionRate = agentEvents.filter(e => e.event === 'crew_end').length /
|
|
136
|
+
Math.max(1, agentEvents.filter(e => e.event === 'crew_start').length);
|
|
137
|
+
const errorRate = errors.length / Math.max(1, lineage.length);
|
|
138
|
+
evalScore = {
|
|
139
|
+
overall: Math.round(Math.max(0, (1 - errorRate) * completionRate * 100)),
|
|
140
|
+
grade: completionRate >= 0.9 && errorRate < 0.1 ? 'A' : completionRate >= 0.7 ? 'B' : 'C',
|
|
141
|
+
dimensions: { completion: Math.round(completionRate * 100), errorRate: Math.round((1 - errorRate) * 100) },
|
|
142
|
+
};
|
|
143
|
+
}
|
|
144
|
+
catch { }
|
|
145
|
+
// Risk classification
|
|
146
|
+
const riskFactors = [];
|
|
147
|
+
if (llmCalls.length > 0)
|
|
148
|
+
riskFactors.push('Uses AI/LLM for decision-making');
|
|
149
|
+
if (agentEvents.length > 0)
|
|
150
|
+
riskFactors.push('Autonomous agent workflow');
|
|
151
|
+
if (agentTools.length > 0)
|
|
152
|
+
riskFactors.push(`Executes ${agentTools.length} tools autonomously`);
|
|
153
|
+
if (securityFindings.filter(f => f.severity === 'critical').length > 0)
|
|
154
|
+
riskFactors.push('Critical security findings');
|
|
155
|
+
if (errors.length > 0)
|
|
156
|
+
riskFactors.push(`${errors.length} runtime errors`);
|
|
157
|
+
const riskLevel = riskFactors.length >= 3 ? 'high' : riskFactors.length >= 1 ? 'medium' : 'low';
|
|
158
|
+
// Human oversight
|
|
159
|
+
const permissionEvents = agentEvents.filter(e => e.event === 'permission_request' || (e.tool || '').toLowerCase().includes('approval'));
|
|
160
|
+
const escalationEvents = agentEvents.filter(e => e.event?.includes('error') || e.event === 'crew_error');
|
|
161
|
+
const report = {
|
|
162
|
+
meta: {
|
|
163
|
+
generatedAt: new Date().toISOString(),
|
|
164
|
+
trickleVersion: 'CLI 0.1.191',
|
|
165
|
+
framework: [...new Set(agentEvents.map(e => e.framework).filter(Boolean))].join(', ') || 'N/A',
|
|
166
|
+
dataDir: dir,
|
|
167
|
+
},
|
|
168
|
+
riskClassification: { level: riskLevel, factors: riskFactors },
|
|
169
|
+
decisionLineage: lineage,
|
|
170
|
+
dataProcessing: {
|
|
171
|
+
llmProviders: providers, modelsUsed: models,
|
|
172
|
+
totalLlmCalls: llmCalls.length, totalTokens, estimatedCost: Math.round(totalCost * 10000) / 10000,
|
|
173
|
+
toolsUsed: agentTools, mcpToolsUsed: mcpTools,
|
|
174
|
+
dataSourcesAccessed: dataSources.slice(0, 20),
|
|
175
|
+
},
|
|
176
|
+
securityFindings,
|
|
177
|
+
evalScore,
|
|
178
|
+
humanOversight: {
|
|
179
|
+
hasHumanInLoop: permissionEvents.length > 0,
|
|
180
|
+
approvalCheckpoints: permissionEvents.length,
|
|
181
|
+
escalationEvents: escalationEvents.length,
|
|
182
|
+
},
|
|
183
|
+
};
|
|
184
|
+
// Output
|
|
185
|
+
if (opts.out) {
|
|
186
|
+
fs.writeFileSync(opts.out, JSON.stringify(report, null, 2), 'utf-8');
|
|
187
|
+
console.log(chalk_1.default.green(` Compliance report written to ${opts.out}`));
|
|
188
|
+
return;
|
|
189
|
+
}
|
|
190
|
+
if (opts.json) {
|
|
191
|
+
console.log(JSON.stringify(report, null, 2));
|
|
192
|
+
return;
|
|
193
|
+
}
|
|
194
|
+
// Pretty print
|
|
195
|
+
console.log('');
|
|
196
|
+
console.log(chalk_1.default.bold(' trickle audit --compliance'));
|
|
197
|
+
console.log(chalk_1.default.gray(' ' + '─'.repeat(60)));
|
|
198
|
+
const riskColor = riskLevel === 'high' ? chalk_1.default.red : riskLevel === 'medium' ? chalk_1.default.yellow : chalk_1.default.green;
|
|
199
|
+
console.log(` Risk: ${riskColor(riskLevel.toUpperCase())} (${riskFactors.length} factors)`);
|
|
200
|
+
for (const f of riskFactors)
|
|
201
|
+
console.log(chalk_1.default.gray(` • ${f}`));
|
|
202
|
+
console.log(chalk_1.default.gray('\n ' + '─'.repeat(60)));
|
|
203
|
+
console.log(chalk_1.default.bold(' Data Processing'));
|
|
204
|
+
console.log(` LLM providers: ${providers.join(', ') || 'none'}`);
|
|
205
|
+
console.log(` Models: ${models.join(', ') || 'none'}`);
|
|
206
|
+
console.log(` Calls: ${llmCalls.length} Tokens: ${totalTokens} Cost: $${totalCost.toFixed(4)}`);
|
|
207
|
+
console.log(` Tools: ${agentTools.join(', ') || 'none'}`);
|
|
208
|
+
if (mcpTools.length > 0)
|
|
209
|
+
console.log(` MCP tools: ${mcpTools.join(', ')}`);
|
|
210
|
+
console.log(chalk_1.default.gray('\n ' + '─'.repeat(60)));
|
|
211
|
+
console.log(chalk_1.default.bold(' Decision Lineage'));
|
|
212
|
+
console.log(` ${lineage.length} events recorded`);
|
|
213
|
+
for (const e of lineage.slice(0, 10)) {
|
|
214
|
+
const ts = new Date(e.timestamp).toISOString().substring(11, 23);
|
|
215
|
+
console.log(chalk_1.default.gray(` ${ts}`) + ` ${e.event.padEnd(20)} ${e.description.substring(0, 60)}`);
|
|
216
|
+
}
|
|
217
|
+
if (lineage.length > 10)
|
|
218
|
+
console.log(chalk_1.default.gray(` ... and ${lineage.length - 10} more`));
|
|
219
|
+
if (securityFindings.length > 0) {
|
|
220
|
+
console.log(chalk_1.default.gray('\n ' + '─'.repeat(60)));
|
|
221
|
+
console.log(chalk_1.default.bold(' Security'));
|
|
222
|
+
const crit = securityFindings.filter(f => f.severity === 'critical').length;
|
|
223
|
+
const warn = securityFindings.filter(f => f.severity === 'warning').length;
|
|
224
|
+
console.log(` ${chalk_1.default.red(String(crit))} critical, ${chalk_1.default.yellow(String(warn))} warnings`);
|
|
225
|
+
}
|
|
226
|
+
console.log(chalk_1.default.gray('\n ' + '─'.repeat(60)));
|
|
227
|
+
console.log(chalk_1.default.bold(' Human Oversight'));
|
|
228
|
+
console.log(` Human-in-the-loop: ${report.humanOversight.hasHumanInLoop ? chalk_1.default.green('Yes') : chalk_1.default.yellow('No')}`);
|
|
229
|
+
console.log(` Approval checkpoints: ${report.humanOversight.approvalCheckpoints}`);
|
|
230
|
+
console.log(` Escalation events: ${report.humanOversight.escalationEvents}`);
|
|
231
|
+
console.log(chalk_1.default.gray('\n ' + '─'.repeat(60)));
|
|
232
|
+
console.log(chalk_1.default.gray(' Export: trickle audit --compliance --json > audit-report.json'));
|
|
233
|
+
console.log(chalk_1.default.gray(' Export: trickle audit --compliance -o audit-report.json'));
|
|
234
|
+
console.log('');
|
|
235
|
+
}
|
|
236
|
+
function buildAgentDescription(e) {
|
|
237
|
+
const parts = [];
|
|
238
|
+
if (e.chain)
|
|
239
|
+
parts.push(e.chain);
|
|
240
|
+
if (e.tool)
|
|
241
|
+
parts.push(`tool:${e.tool}`);
|
|
242
|
+
if (e.toolInput)
|
|
243
|
+
parts.push(`input:${String(e.toolInput).substring(0, 40)}`);
|
|
244
|
+
if (e.output)
|
|
245
|
+
parts.push(`output:${String(e.output).substring(0, 40)}`);
|
|
246
|
+
if (e.error)
|
|
247
|
+
parts.push(`error:${String(e.error).substring(0, 40)}`);
|
|
248
|
+
if (e.thought)
|
|
249
|
+
parts.push(`thought:${String(e.thought).substring(0, 40)}`);
|
|
250
|
+
return parts.join(' | ') || e.event || '?';
|
|
251
|
+
}
|
package/dist/commands/init.js
CHANGED
|
@@ -515,8 +515,12 @@ trickle flamegraph
|
|
|
515
515
|
|
|
516
516
|
1. **Get overview**: \`trickle summary\` — status, errors, N+1 queries, root causes, fix recommendations
|
|
517
517
|
2. **Understand a file**: \`trickle explain ${exampleFile}\` — functions, call graph, data flow, queries, variables
|
|
518
|
-
3. **Debug errors**: \`trickle
|
|
518
|
+
3. **Debug errors**: \`trickle why\` — causal chain: error → call trace → variables → LLM reasoning
|
|
519
519
|
4. **Fix & verify**: \`trickle verify --baseline\` → fix code → \`${runCmd}\` → \`trickle verify\`
|
|
520
|
+
5. **Evaluate quality**: \`trickle eval\` — A-F reliability score (completion, errors, cost, tools, latency)
|
|
521
|
+
6. **Check security**: \`trickle security\` — prompt injection, privilege escalation, data exfiltration
|
|
522
|
+
7. **LLM costs**: \`trickle cost-report\` — cost breakdown by provider/model with budget checking
|
|
523
|
+
8. **Compliance**: \`trickle audit --compliance\` — risk classification, decision lineage, audit trail
|
|
520
524
|
|
|
521
525
|
### MCP Server (recommended for Claude Code)
|
|
522
526
|
|
|
@@ -533,19 +537,21 @@ Add to \`.claude/settings.json\` or \`claude_desktop_config.json\`:
|
|
|
533
537
|
}
|
|
534
538
|
\`\`\`
|
|
535
539
|
|
|
536
|
-
Key MCP tools (
|
|
540
|
+
Key MCP tools (38 total):
|
|
537
541
|
|
|
538
542
|
| Tool | Use for |
|
|
539
543
|
|------|---------|
|
|
540
544
|
| \`get_recommended_actions\` | Start here — tells you exactly what to do next |
|
|
541
545
|
| \`get_last_run_summary\` | Complete overview: errors, queries, alerts, root causes |
|
|
546
|
+
| \`why\` | Causal debugging: trace back from error to root cause |
|
|
542
547
|
| \`explain_file\` | Understand a file: functions, call graph, data flow, queries |
|
|
543
548
|
| \`run_tests\` | Run tests with structured pass/fail + runtime context |
|
|
544
549
|
| \`get_flamegraph\` | Performance hotspots sorted by execution time |
|
|
545
|
-
| \`
|
|
550
|
+
| \`get_llm_calls\` | LLM API calls with tokens, cost, latency |
|
|
551
|
+
| \`get_agent_trace\` | Agent execution tree with parent-child relationships |
|
|
552
|
+
| \`get_mcp_tool_calls\` | MCP tool invocations with latency and direction |
|
|
553
|
+
| \`get_cost_report\` | Cost breakdown by provider/model with budget check |
|
|
546
554
|
| \`save_baseline\` / \`compare_with_baseline\` | Before/after comparison for fix verification |
|
|
547
|
-
| \`get_new_alerts\` | Polling-based monitoring for new issues |
|
|
548
|
-
| \`refresh_runtime_data\` | Re-run the app to capture fresh data |
|
|
549
555
|
|
|
550
556
|
### What trickle Captures (automatically, zero config)
|
|
551
557
|
|
|
@@ -556,6 +562,9 @@ Key MCP tools (26 total):
|
|
|
556
562
|
- **Logs**: winston, pino, bunyan (JS); logging, loguru, structlog (Python)
|
|
557
563
|
- **HTTP requests**: fetch/requests calls with status + latency
|
|
558
564
|
- **Call traces**: execution flow with parent-child relationships
|
|
565
|
+
- **LLM calls**: OpenAI, Anthropic, Gemini — model, tokens, cost, latency (auto-detected)
|
|
566
|
+
- **Agent workflows**: LangChain, CrewAI, Claude Agent SDK, OpenAI Agents SDK (auto-detected)
|
|
567
|
+
- **MCP tool calls**: tool name, arguments, response, latency, direction
|
|
559
568
|
- **Memory**: RSS + heap snapshots
|
|
560
569
|
`;
|
|
561
570
|
fs.writeFileSync(claudePath, content, "utf-8");
|
package/dist/index.js
CHANGED
|
@@ -396,8 +396,15 @@ program
|
|
|
396
396
|
.option("--json", "Output raw JSON (for CI integration)")
|
|
397
397
|
.option("--fail-on-error", "Exit 1 if any errors are found")
|
|
398
398
|
.option("--fail-on-warning", "Exit 1 if any errors or warnings are found")
|
|
399
|
+
.option("--compliance", "Generate compliance audit report (EU AI Act / Colorado AI Act)")
|
|
400
|
+
.option("-o, --out <file>", "Write compliance report to file")
|
|
399
401
|
.option("--local", "Read from local .trickle/observations.jsonl instead of backend")
|
|
400
402
|
.action(async (opts) => {
|
|
403
|
+
if (opts.compliance) {
|
|
404
|
+
const { generateComplianceReport } = await Promise.resolve().then(() => __importStar(require("./commands/compliance")));
|
|
405
|
+
generateComplianceReport({ json: opts.json, out: opts.out });
|
|
406
|
+
return;
|
|
407
|
+
}
|
|
401
408
|
await (0, audit_1.auditCommand)(opts);
|
|
402
409
|
});
|
|
403
410
|
// trickle capture <method> <url>
|
package/package.json
CHANGED
|
@@ -0,0 +1,262 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* trickle audit --compliance — Generate compliance audit report.
|
|
3
|
+
*
|
|
4
|
+
* Exports trickle's JSONL data as a structured compliance report for
|
|
5
|
+
* EU AI Act and Colorado AI Act requirements:
|
|
6
|
+
* - Decision lineage (LLM call → tool call → output)
|
|
7
|
+
* - Timestamped event log
|
|
8
|
+
* - Risk classification
|
|
9
|
+
* - Security scan results
|
|
10
|
+
* - Data processing summary
|
|
11
|
+
*
|
|
12
|
+
* Local-first: sensitive audit data never leaves the developer's machine.
|
|
13
|
+
*/
|
|
14
|
+
|
|
15
|
+
import * as fs from 'fs';
|
|
16
|
+
import * as path from 'path';
|
|
17
|
+
import chalk from 'chalk';
|
|
18
|
+
|
|
19
|
+
function readJsonl(fp: string): any[] {
|
|
20
|
+
if (!fs.existsSync(fp)) return [];
|
|
21
|
+
return fs.readFileSync(fp, 'utf-8').split('\n').filter(Boolean)
|
|
22
|
+
.map(l => { try { return JSON.parse(l); } catch { return null; } }).filter(Boolean);
|
|
23
|
+
}
|
|
24
|
+
|
|
25
|
+
interface ComplianceReport {
|
|
26
|
+
meta: {
|
|
27
|
+
generatedAt: string;
|
|
28
|
+
trickleVersion: string;
|
|
29
|
+
framework: string;
|
|
30
|
+
dataDir: string;
|
|
31
|
+
};
|
|
32
|
+
riskClassification: {
|
|
33
|
+
level: 'high' | 'medium' | 'low';
|
|
34
|
+
factors: string[];
|
|
35
|
+
};
|
|
36
|
+
decisionLineage: Array<{
|
|
37
|
+
timestamp: number;
|
|
38
|
+
event: string;
|
|
39
|
+
description: string;
|
|
40
|
+
data?: Record<string, unknown>;
|
|
41
|
+
}>;
|
|
42
|
+
dataProcessing: {
|
|
43
|
+
llmProviders: string[];
|
|
44
|
+
modelsUsed: string[];
|
|
45
|
+
totalLlmCalls: number;
|
|
46
|
+
totalTokens: number;
|
|
47
|
+
estimatedCost: number;
|
|
48
|
+
toolsUsed: string[];
|
|
49
|
+
mcpToolsUsed: string[];
|
|
50
|
+
dataSourcesAccessed: string[];
|
|
51
|
+
};
|
|
52
|
+
securityFindings: Array<{
|
|
53
|
+
severity: string;
|
|
54
|
+
category: string;
|
|
55
|
+
message: string;
|
|
56
|
+
evidence: string;
|
|
57
|
+
}>;
|
|
58
|
+
evalScore: {
|
|
59
|
+
overall: number;
|
|
60
|
+
grade: string;
|
|
61
|
+
dimensions: Record<string, number>;
|
|
62
|
+
} | null;
|
|
63
|
+
humanOversight: {
|
|
64
|
+
hasHumanInLoop: boolean;
|
|
65
|
+
approvalCheckpoints: number;
|
|
66
|
+
escalationEvents: number;
|
|
67
|
+
};
|
|
68
|
+
}
|
|
69
|
+
|
|
70
|
+
export function generateComplianceReport(opts: { json?: boolean; out?: string }): void {
|
|
71
|
+
const dir = process.env.TRICKLE_LOCAL_DIR || path.join(process.cwd(), '.trickle');
|
|
72
|
+
|
|
73
|
+
if (!fs.existsSync(dir)) {
|
|
74
|
+
console.log(chalk.yellow(' No .trickle/ data. Run trickle first.'));
|
|
75
|
+
return;
|
|
76
|
+
}
|
|
77
|
+
|
|
78
|
+
const llmCalls = readJsonl(path.join(dir, 'llm.jsonl'));
|
|
79
|
+
const agentEvents = readJsonl(path.join(dir, 'agents.jsonl'));
|
|
80
|
+
const mcpCalls = readJsonl(path.join(dir, 'mcp.jsonl'));
|
|
81
|
+
const errors = readJsonl(path.join(dir, 'errors.jsonl'));
|
|
82
|
+
const observations = readJsonl(path.join(dir, 'observations.jsonl'));
|
|
83
|
+
const calltrace = readJsonl(path.join(dir, 'calltrace.jsonl'));
|
|
84
|
+
|
|
85
|
+
// Build decision lineage — chronological event log
|
|
86
|
+
const lineage: ComplianceReport['decisionLineage'] = [];
|
|
87
|
+
|
|
88
|
+
for (const e of agentEvents) {
|
|
89
|
+
lineage.push({
|
|
90
|
+
timestamp: e.timestamp || 0,
|
|
91
|
+
event: `agent:${e.event}`,
|
|
92
|
+
description: buildAgentDescription(e),
|
|
93
|
+
data: { framework: e.framework, tool: e.tool, chain: e.chain },
|
|
94
|
+
});
|
|
95
|
+
}
|
|
96
|
+
|
|
97
|
+
for (const c of llmCalls) {
|
|
98
|
+
lineage.push({
|
|
99
|
+
timestamp: c.timestamp || 0,
|
|
100
|
+
event: 'llm:call',
|
|
101
|
+
description: `${c.provider}/${c.model}: ${(c.inputPreview || '').substring(0, 80)} → ${(c.outputPreview || '').substring(0, 80)}`,
|
|
102
|
+
data: { model: c.model, tokens: c.totalTokens, cost: c.estimatedCostUsd, error: c.error },
|
|
103
|
+
});
|
|
104
|
+
}
|
|
105
|
+
|
|
106
|
+
for (const m of mcpCalls) {
|
|
107
|
+
if (m.tool === '__list_tools') continue;
|
|
108
|
+
lineage.push({
|
|
109
|
+
timestamp: m.timestamp || 0,
|
|
110
|
+
event: `mcp:${m.direction}`,
|
|
111
|
+
description: `${m.tool}(${JSON.stringify(m.args || {}).substring(0, 60)}) → ${(m.resultPreview || '').substring(0, 60)}`,
|
|
112
|
+
data: { tool: m.tool, direction: m.direction, isError: m.isError },
|
|
113
|
+
});
|
|
114
|
+
}
|
|
115
|
+
|
|
116
|
+
lineage.sort((a, b) => a.timestamp - b.timestamp);
|
|
117
|
+
|
|
118
|
+
// Data processing summary
|
|
119
|
+
const providers = [...new Set(llmCalls.map(c => c.provider).filter(Boolean))];
|
|
120
|
+
const models = [...new Set(llmCalls.map(c => `${c.provider}/${c.model}`).filter(Boolean))];
|
|
121
|
+
const totalTokens = llmCalls.reduce((s: number, c: any) => s + (c.totalTokens || 0), 0);
|
|
122
|
+
const totalCost = llmCalls.reduce((s: number, c: any) => s + (c.estimatedCostUsd || 0), 0);
|
|
123
|
+
const agentTools = [...new Set(agentEvents.filter(e => e.tool).map(e => e.tool))];
|
|
124
|
+
const mcpTools = [...new Set(mcpCalls.filter(c => c.tool && c.tool !== '__list_tools').map(c => c.tool))];
|
|
125
|
+
const dataSources = [...new Set(observations.map(o => o.module).filter(Boolean))];
|
|
126
|
+
|
|
127
|
+
// Security scan
|
|
128
|
+
let securityFindings: ComplianceReport['securityFindings'] = [];
|
|
129
|
+
try {
|
|
130
|
+
const { runSecurityScan } = require('./security');
|
|
131
|
+
const origLog = console.log;
|
|
132
|
+
console.log = () => {};
|
|
133
|
+
const result = runSecurityScan({ dir });
|
|
134
|
+
console.log = origLog;
|
|
135
|
+
securityFindings = result.findings.map((f: any) => ({
|
|
136
|
+
severity: f.severity, category: f.category,
|
|
137
|
+
message: f.message, evidence: (f.evidence || '').substring(0, 100),
|
|
138
|
+
}));
|
|
139
|
+
} catch {}
|
|
140
|
+
|
|
141
|
+
// Eval score
|
|
142
|
+
let evalScore: ComplianceReport['evalScore'] = null;
|
|
143
|
+
try {
|
|
144
|
+
const { evalCommand } = require('./eval');
|
|
145
|
+
// We can't easily call evalCommand and get the result, so build a lightweight score
|
|
146
|
+
const completionRate = agentEvents.filter(e => e.event === 'crew_end').length /
|
|
147
|
+
Math.max(1, agentEvents.filter(e => e.event === 'crew_start').length);
|
|
148
|
+
const errorRate = errors.length / Math.max(1, lineage.length);
|
|
149
|
+
evalScore = {
|
|
150
|
+
overall: Math.round(Math.max(0, (1 - errorRate) * completionRate * 100)),
|
|
151
|
+
grade: completionRate >= 0.9 && errorRate < 0.1 ? 'A' : completionRate >= 0.7 ? 'B' : 'C',
|
|
152
|
+
dimensions: { completion: Math.round(completionRate * 100), errorRate: Math.round((1 - errorRate) * 100) },
|
|
153
|
+
};
|
|
154
|
+
} catch {}
|
|
155
|
+
|
|
156
|
+
// Risk classification
|
|
157
|
+
const riskFactors: string[] = [];
|
|
158
|
+
if (llmCalls.length > 0) riskFactors.push('Uses AI/LLM for decision-making');
|
|
159
|
+
if (agentEvents.length > 0) riskFactors.push('Autonomous agent workflow');
|
|
160
|
+
if (agentTools.length > 0) riskFactors.push(`Executes ${agentTools.length} tools autonomously`);
|
|
161
|
+
if (securityFindings.filter(f => f.severity === 'critical').length > 0) riskFactors.push('Critical security findings');
|
|
162
|
+
if (errors.length > 0) riskFactors.push(`${errors.length} runtime errors`);
|
|
163
|
+
const riskLevel: 'high' | 'medium' | 'low' = riskFactors.length >= 3 ? 'high' : riskFactors.length >= 1 ? 'medium' : 'low';
|
|
164
|
+
|
|
165
|
+
// Human oversight
|
|
166
|
+
const permissionEvents = agentEvents.filter(e =>
|
|
167
|
+
e.event === 'permission_request' || (e.tool || '').toLowerCase().includes('approval'));
|
|
168
|
+
const escalationEvents = agentEvents.filter(e =>
|
|
169
|
+
e.event?.includes('error') || e.event === 'crew_error');
|
|
170
|
+
|
|
171
|
+
const report: ComplianceReport = {
|
|
172
|
+
meta: {
|
|
173
|
+
generatedAt: new Date().toISOString(),
|
|
174
|
+
trickleVersion: 'CLI 0.1.191',
|
|
175
|
+
framework: [...new Set(agentEvents.map(e => e.framework).filter(Boolean))].join(', ') || 'N/A',
|
|
176
|
+
dataDir: dir,
|
|
177
|
+
},
|
|
178
|
+
riskClassification: { level: riskLevel, factors: riskFactors },
|
|
179
|
+
decisionLineage: lineage,
|
|
180
|
+
dataProcessing: {
|
|
181
|
+
llmProviders: providers, modelsUsed: models,
|
|
182
|
+
totalLlmCalls: llmCalls.length, totalTokens, estimatedCost: Math.round(totalCost * 10000) / 10000,
|
|
183
|
+
toolsUsed: agentTools, mcpToolsUsed: mcpTools,
|
|
184
|
+
dataSourcesAccessed: dataSources.slice(0, 20),
|
|
185
|
+
},
|
|
186
|
+
securityFindings,
|
|
187
|
+
evalScore,
|
|
188
|
+
humanOversight: {
|
|
189
|
+
hasHumanInLoop: permissionEvents.length > 0,
|
|
190
|
+
approvalCheckpoints: permissionEvents.length,
|
|
191
|
+
escalationEvents: escalationEvents.length,
|
|
192
|
+
},
|
|
193
|
+
};
|
|
194
|
+
|
|
195
|
+
// Output
|
|
196
|
+
if (opts.out) {
|
|
197
|
+
fs.writeFileSync(opts.out, JSON.stringify(report, null, 2), 'utf-8');
|
|
198
|
+
console.log(chalk.green(` Compliance report written to ${opts.out}`));
|
|
199
|
+
return;
|
|
200
|
+
}
|
|
201
|
+
|
|
202
|
+
if (opts.json) {
|
|
203
|
+
console.log(JSON.stringify(report, null, 2));
|
|
204
|
+
return;
|
|
205
|
+
}
|
|
206
|
+
|
|
207
|
+
// Pretty print
|
|
208
|
+
console.log('');
|
|
209
|
+
console.log(chalk.bold(' trickle audit --compliance'));
|
|
210
|
+
console.log(chalk.gray(' ' + '─'.repeat(60)));
|
|
211
|
+
|
|
212
|
+
const riskColor = riskLevel === 'high' ? chalk.red : riskLevel === 'medium' ? chalk.yellow : chalk.green;
|
|
213
|
+
console.log(` Risk: ${riskColor(riskLevel.toUpperCase())} (${riskFactors.length} factors)`);
|
|
214
|
+
for (const f of riskFactors) console.log(chalk.gray(` • ${f}`));
|
|
215
|
+
|
|
216
|
+
console.log(chalk.gray('\n ' + '─'.repeat(60)));
|
|
217
|
+
console.log(chalk.bold(' Data Processing'));
|
|
218
|
+
console.log(` LLM providers: ${providers.join(', ') || 'none'}`);
|
|
219
|
+
console.log(` Models: ${models.join(', ') || 'none'}`);
|
|
220
|
+
console.log(` Calls: ${llmCalls.length} Tokens: ${totalTokens} Cost: $${totalCost.toFixed(4)}`);
|
|
221
|
+
console.log(` Tools: ${agentTools.join(', ') || 'none'}`);
|
|
222
|
+
if (mcpTools.length > 0) console.log(` MCP tools: ${mcpTools.join(', ')}`);
|
|
223
|
+
|
|
224
|
+
console.log(chalk.gray('\n ' + '─'.repeat(60)));
|
|
225
|
+
console.log(chalk.bold(' Decision Lineage'));
|
|
226
|
+
console.log(` ${lineage.length} events recorded`);
|
|
227
|
+
for (const e of lineage.slice(0, 10)) {
|
|
228
|
+
const ts = new Date(e.timestamp).toISOString().substring(11, 23);
|
|
229
|
+
console.log(chalk.gray(` ${ts}`) + ` ${e.event.padEnd(20)} ${e.description.substring(0, 60)}`);
|
|
230
|
+
}
|
|
231
|
+
if (lineage.length > 10) console.log(chalk.gray(` ... and ${lineage.length - 10} more`));
|
|
232
|
+
|
|
233
|
+
if (securityFindings.length > 0) {
|
|
234
|
+
console.log(chalk.gray('\n ' + '─'.repeat(60)));
|
|
235
|
+
console.log(chalk.bold(' Security'));
|
|
236
|
+
const crit = securityFindings.filter(f => f.severity === 'critical').length;
|
|
237
|
+
const warn = securityFindings.filter(f => f.severity === 'warning').length;
|
|
238
|
+
console.log(` ${chalk.red(String(crit))} critical, ${chalk.yellow(String(warn))} warnings`);
|
|
239
|
+
}
|
|
240
|
+
|
|
241
|
+
console.log(chalk.gray('\n ' + '─'.repeat(60)));
|
|
242
|
+
console.log(chalk.bold(' Human Oversight'));
|
|
243
|
+
console.log(` Human-in-the-loop: ${report.humanOversight.hasHumanInLoop ? chalk.green('Yes') : chalk.yellow('No')}`);
|
|
244
|
+
console.log(` Approval checkpoints: ${report.humanOversight.approvalCheckpoints}`);
|
|
245
|
+
console.log(` Escalation events: ${report.humanOversight.escalationEvents}`);
|
|
246
|
+
|
|
247
|
+
console.log(chalk.gray('\n ' + '─'.repeat(60)));
|
|
248
|
+
console.log(chalk.gray(' Export: trickle audit --compliance --json > audit-report.json'));
|
|
249
|
+
console.log(chalk.gray(' Export: trickle audit --compliance -o audit-report.json'));
|
|
250
|
+
console.log('');
|
|
251
|
+
}
|
|
252
|
+
|
|
253
|
+
function buildAgentDescription(e: any): string {
|
|
254
|
+
const parts: string[] = [];
|
|
255
|
+
if (e.chain) parts.push(e.chain);
|
|
256
|
+
if (e.tool) parts.push(`tool:${e.tool}`);
|
|
257
|
+
if (e.toolInput) parts.push(`input:${String(e.toolInput).substring(0, 40)}`);
|
|
258
|
+
if (e.output) parts.push(`output:${String(e.output).substring(0, 40)}`);
|
|
259
|
+
if (e.error) parts.push(`error:${String(e.error).substring(0, 40)}`);
|
|
260
|
+
if (e.thought) parts.push(`thought:${String(e.thought).substring(0, 40)}`);
|
|
261
|
+
return parts.join(' | ') || e.event || '?';
|
|
262
|
+
}
|
package/src/commands/init.ts
CHANGED
|
@@ -546,8 +546,12 @@ trickle flamegraph
|
|
|
546
546
|
|
|
547
547
|
1. **Get overview**: \`trickle summary\` — status, errors, N+1 queries, root causes, fix recommendations
|
|
548
548
|
2. **Understand a file**: \`trickle explain ${exampleFile}\` — functions, call graph, data flow, queries, variables
|
|
549
|
-
3. **Debug errors**: \`trickle
|
|
549
|
+
3. **Debug errors**: \`trickle why\` — causal chain: error → call trace → variables → LLM reasoning
|
|
550
550
|
4. **Fix & verify**: \`trickle verify --baseline\` → fix code → \`${runCmd}\` → \`trickle verify\`
|
|
551
|
+
5. **Evaluate quality**: \`trickle eval\` — A-F reliability score (completion, errors, cost, tools, latency)
|
|
552
|
+
6. **Check security**: \`trickle security\` — prompt injection, privilege escalation, data exfiltration
|
|
553
|
+
7. **LLM costs**: \`trickle cost-report\` — cost breakdown by provider/model with budget checking
|
|
554
|
+
8. **Compliance**: \`trickle audit --compliance\` — risk classification, decision lineage, audit trail
|
|
551
555
|
|
|
552
556
|
### MCP Server (recommended for Claude Code)
|
|
553
557
|
|
|
@@ -564,19 +568,21 @@ Add to \`.claude/settings.json\` or \`claude_desktop_config.json\`:
|
|
|
564
568
|
}
|
|
565
569
|
\`\`\`
|
|
566
570
|
|
|
567
|
-
Key MCP tools (
|
|
571
|
+
Key MCP tools (38 total):
|
|
568
572
|
|
|
569
573
|
| Tool | Use for |
|
|
570
574
|
|------|---------|
|
|
571
575
|
| \`get_recommended_actions\` | Start here — tells you exactly what to do next |
|
|
572
576
|
| \`get_last_run_summary\` | Complete overview: errors, queries, alerts, root causes |
|
|
577
|
+
| \`why\` | Causal debugging: trace back from error to root cause |
|
|
573
578
|
| \`explain_file\` | Understand a file: functions, call graph, data flow, queries |
|
|
574
579
|
| \`run_tests\` | Run tests with structured pass/fail + runtime context |
|
|
575
580
|
| \`get_flamegraph\` | Performance hotspots sorted by execution time |
|
|
576
|
-
| \`
|
|
581
|
+
| \`get_llm_calls\` | LLM API calls with tokens, cost, latency |
|
|
582
|
+
| \`get_agent_trace\` | Agent execution tree with parent-child relationships |
|
|
583
|
+
| \`get_mcp_tool_calls\` | MCP tool invocations with latency and direction |
|
|
584
|
+
| \`get_cost_report\` | Cost breakdown by provider/model with budget check |
|
|
577
585
|
| \`save_baseline\` / \`compare_with_baseline\` | Before/after comparison for fix verification |
|
|
578
|
-
| \`get_new_alerts\` | Polling-based monitoring for new issues |
|
|
579
|
-
| \`refresh_runtime_data\` | Re-run the app to capture fresh data |
|
|
580
586
|
|
|
581
587
|
### What trickle Captures (automatically, zero config)
|
|
582
588
|
|
|
@@ -587,6 +593,9 @@ Key MCP tools (26 total):
|
|
|
587
593
|
- **Logs**: winston, pino, bunyan (JS); logging, loguru, structlog (Python)
|
|
588
594
|
- **HTTP requests**: fetch/requests calls with status + latency
|
|
589
595
|
- **Call traces**: execution flow with parent-child relationships
|
|
596
|
+
- **LLM calls**: OpenAI, Anthropic, Gemini — model, tokens, cost, latency (auto-detected)
|
|
597
|
+
- **Agent workflows**: LangChain, CrewAI, Claude Agent SDK, OpenAI Agents SDK (auto-detected)
|
|
598
|
+
- **MCP tool calls**: tool name, arguments, response, latency, direction
|
|
590
599
|
- **Memory**: RSS + heap snapshots
|
|
591
600
|
`;
|
|
592
601
|
|
package/src/index.ts
CHANGED
|
@@ -384,8 +384,15 @@ program
|
|
|
384
384
|
.option("--json", "Output raw JSON (for CI integration)")
|
|
385
385
|
.option("--fail-on-error", "Exit 1 if any errors are found")
|
|
386
386
|
.option("--fail-on-warning", "Exit 1 if any errors or warnings are found")
|
|
387
|
+
.option("--compliance", "Generate compliance audit report (EU AI Act / Colorado AI Act)")
|
|
388
|
+
.option("-o, --out <file>", "Write compliance report to file")
|
|
387
389
|
.option("--local", "Read from local .trickle/observations.jsonl instead of backend")
|
|
388
390
|
.action(async (opts) => {
|
|
391
|
+
if (opts.compliance) {
|
|
392
|
+
const { generateComplianceReport } = await import("./commands/compliance");
|
|
393
|
+
generateComplianceReport({ json: opts.json, out: opts.out });
|
|
394
|
+
return;
|
|
395
|
+
}
|
|
389
396
|
await auditCommand(opts);
|
|
390
397
|
});
|
|
391
398
|
|