imprint-mcp 0.6.0 → 0.6.2
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/README.md +2 -0
- package/package.json +4 -2
- package/prompts/audit-agent.md +4 -2
- package/prompts/auth-compile-agent.md +149 -158
- package/prompts/build-planning.md +14 -9
- package/prompts/compile-agent.md +46 -23
- package/prompts/live-verifier-agent.md +101 -0
- package/prompts/playbook-compilation.md +8 -1
- package/prompts/tool-candidate-detection.md +9 -29
- package/src/cli.ts +83 -1
- package/src/imprint/audit.ts +219 -20
- package/src/imprint/auth-compile-agent.ts +215 -155
- package/src/imprint/auth-compile-tools.ts +300 -118
- package/src/imprint/auth-verifier.ts +167 -180
- package/src/imprint/backend-cache.ts +18 -11
- package/src/imprint/backend-ladder.ts +225 -177
- package/src/imprint/build-plan.ts +3 -8
- package/src/imprint/cdp-browser-fetch.ts +930 -83
- package/src/imprint/claude-cli-compile.ts +46 -20
- package/src/imprint/codex-cli-compile.ts +436 -139
- package/src/imprint/compile-agent-types.ts +15 -10
- package/src/imprint/compile-agent.ts +88 -28
- package/src/imprint/compile-tools.ts +228 -39
- package/src/imprint/compile-verification.ts +187 -0
- package/src/imprint/compile.ts +71 -2
- package/src/imprint/credential-store.ts +57 -44
- package/src/imprint/emit.ts +3 -1
- package/src/imprint/freeform-redact.ts +27 -6
- package/src/imprint/live-verifier.ts +1665 -0
- package/src/imprint/llm.ts +22 -3
- package/src/imprint/login.ts +107 -25
- package/src/imprint/mcp-compile-server.ts +447 -114
- package/src/imprint/mcp-live-verifier-server.ts +359 -0
- package/src/imprint/mcp-server.ts +116 -35
- package/src/imprint/probe-backends.ts +88 -7
- package/src/imprint/recording-request.ts +36 -0
- package/src/imprint/request-capture.ts +40 -17
- package/src/imprint/runtime.ts +490 -337
- package/src/imprint/teach-state.ts +36 -2
- package/src/imprint/teach.ts +302 -146
- package/src/imprint/tool-candidates.ts +8 -10
- package/src/imprint/tool-loader.ts +12 -4
- package/src/imprint/types.ts +137 -68
|
@@ -50,22 +50,16 @@ export interface CompileAgentProgress extends AgentProgress {
|
|
|
50
50
|
/** 1-based verification cycle. Cycle 1 is the initial agent run. Subsequent cycles
|
|
51
51
|
* happen when the agent claims done() but external verification fails. */
|
|
52
52
|
verificationCycle: number;
|
|
53
|
-
/**
|
|
54
|
-
maxVerificationCycles
|
|
53
|
+
/** Verification cap for bounded data-tool compiles; auth is deadline-bounded. */
|
|
54
|
+
maxVerificationCycles?: number;
|
|
55
55
|
// ── Auth segments only (all optional; data-compile + codex paths leave unset) ──
|
|
56
56
|
/** 1-based current segment index in the resumable auth loop. */
|
|
57
57
|
segment?: number;
|
|
58
|
-
/** Total segment budget (MAX_AUTH_SEGMENTS). */
|
|
59
|
-
maxSegments?: number;
|
|
60
|
-
/** Live `initiate` attempts spent so far (AuthVerifier.attemptsUsed). */
|
|
61
|
-
attempt?: number;
|
|
62
|
-
/** Attempt cap (AuthVerifier.maxInitiateAttempts). */
|
|
63
|
-
maxAttempts?: number;
|
|
64
58
|
/** The most recent live verification result, so the orchestrator's progress
|
|
65
59
|
* line can surface a failure (e.g. a 403) the instant it happens instead of
|
|
66
60
|
* only feeding it to the agent. Grounded purely in AuthPhaseResult fields. */
|
|
67
61
|
lastVerification?: {
|
|
68
|
-
|
|
62
|
+
action: string;
|
|
69
63
|
ok: boolean;
|
|
70
64
|
status?: number;
|
|
71
65
|
error?: string;
|
|
@@ -82,7 +76,18 @@ export interface CompileAgentProgress extends AgentProgress {
|
|
|
82
76
|
* the cooldown — then resumes the agent (`claude --resume`) with the result as a
|
|
83
77
|
* follow-up user message. Site/channel-agnostic. */
|
|
84
78
|
export type AuthCheckpoint =
|
|
85
|
-
| {
|
|
79
|
+
| {
|
|
80
|
+
kind: 'run_verification';
|
|
81
|
+
action: string;
|
|
82
|
+
parameters?: Record<string, string | number | boolean>;
|
|
83
|
+
freshSession?: boolean;
|
|
84
|
+
cleanSession?: boolean;
|
|
85
|
+
}
|
|
86
|
+
| {
|
|
87
|
+
kind: 'inspect_verification_page';
|
|
88
|
+
maxChars?: number;
|
|
89
|
+
includeCookies?: boolean;
|
|
90
|
+
}
|
|
86
91
|
| { kind: 'prompt_user'; message: string; options?: string[] }
|
|
87
92
|
| { kind: 'wait_for_cooldown'; minutes: number; reason?: string };
|
|
88
93
|
|
|
@@ -28,6 +28,11 @@ import {
|
|
|
28
28
|
externalVerification,
|
|
29
29
|
} from './compile-tools.ts';
|
|
30
30
|
import { type Replacement, extractCredentials } from './credential-extract.ts';
|
|
31
|
+
import {
|
|
32
|
+
mergeSemanticParamVerification,
|
|
33
|
+
runLiveSemanticVerification,
|
|
34
|
+
semanticVerificationFailures,
|
|
35
|
+
} from './live-verifier.ts';
|
|
31
36
|
import {
|
|
32
37
|
type LLMOptions,
|
|
33
38
|
type ProviderName,
|
|
@@ -39,6 +44,7 @@ import {
|
|
|
39
44
|
import { loadJsonFile } from './load-json.ts';
|
|
40
45
|
import { createLog } from './log.ts';
|
|
41
46
|
import { localSiteDir } from './paths.ts';
|
|
47
|
+
import { rebindExistingBackendsCacheToWorkflow } from './probe-backends.ts';
|
|
42
48
|
import { detectPageMintedHeaders, redactSession } from './redact.ts';
|
|
43
49
|
import type { ClassifiedValue } from './session-diff.ts';
|
|
44
50
|
import type { SharedCompileContext, ToolCandidate } from './tool-candidates.ts';
|
|
@@ -51,6 +57,19 @@ const log = createLog('compile-agent');
|
|
|
51
57
|
const REPO_ROOT = pathJoin(import.meta.dir, '..', '..');
|
|
52
58
|
const PROMPTS_DIR = pathJoin(REPO_ROOT, 'prompts');
|
|
53
59
|
|
|
60
|
+
let claudeCliCompiler = compileViaClaudeCli;
|
|
61
|
+
let codexCliCompiler = compileViaCodexCli;
|
|
62
|
+
|
|
63
|
+
export function __setCompileAgentCliCompilersForTest(
|
|
64
|
+
compilers: {
|
|
65
|
+
claude?: typeof compileViaClaudeCli;
|
|
66
|
+
codex?: typeof compileViaCodexCli;
|
|
67
|
+
} | null,
|
|
68
|
+
): void {
|
|
69
|
+
claudeCliCompiler = compilers?.claude ?? compileViaClaudeCli;
|
|
70
|
+
codexCliCompiler = compilers?.codex ?? compileViaCodexCli;
|
|
71
|
+
}
|
|
72
|
+
|
|
54
73
|
/** Re-exported for callers (cli, teach) that need to display the selected
|
|
55
74
|
* model before kicking off the agent loop. */
|
|
56
75
|
export function resolveCompileAgentModel(provider: ProviderName): string {
|
|
@@ -101,6 +120,19 @@ interface CompileAgentOptions {
|
|
|
101
120
|
* response parsing, shared-module imports). Injected into the agent's initial
|
|
102
121
|
* message so the compile follows it. Generic — not tied to any site. */
|
|
103
122
|
toolPlan?: string;
|
|
123
|
+
/** Revise an existing generated artifact, using durable verification feedback
|
|
124
|
+
* as the starting point instead of rebuilding it from raw capture. */
|
|
125
|
+
revisionMode?: boolean;
|
|
126
|
+
}
|
|
127
|
+
|
|
128
|
+
function formatRevisionMode(enabled: boolean | undefined): string {
|
|
129
|
+
return enabled
|
|
130
|
+
? [
|
|
131
|
+
'REVISION MODE: this is a bounded resume of an existing generated tool, not a from-scratch compile.',
|
|
132
|
+
'Use read_session_summary.revisionContext, then read the listed current artifacts and durable verification feedback before inspecting raw response bodies.',
|
|
133
|
+
'Preserve proven behavior and make the smallest evidence-backed repair or contract reduction.',
|
|
134
|
+
].join(' ')
|
|
135
|
+
: '';
|
|
104
136
|
}
|
|
105
137
|
|
|
106
138
|
export async function compileAgent(opts: CompileAgentOptions): Promise<CompileAgentResult> {
|
|
@@ -203,6 +235,7 @@ export async function compileAgent(opts: CompileAgentOptions): Promise<CompileAg
|
|
|
203
235
|
teachCredentials: opts.teachCredentials,
|
|
204
236
|
buildPlanPath: opts.buildPlanPath,
|
|
205
237
|
sharedModules: opts.sharedModules,
|
|
238
|
+
revisionMode: opts.revisionMode,
|
|
206
239
|
}),
|
|
207
240
|
doneTool(),
|
|
208
241
|
giveUpTool(),
|
|
@@ -216,11 +249,12 @@ Tool directory: ${absoluteToolDir}
|
|
|
216
249
|
You will write artifacts into the tool directory.
|
|
217
250
|
${formatCandidateContext(opts.candidate, opts.sharedContext, assignedSharedModules)}
|
|
218
251
|
${formatToolPlan(opts.toolPlan)}
|
|
252
|
+
${formatRevisionMode(opts.revisionMode)}
|
|
219
253
|
|
|
220
254
|
Begin by calling read_session_summary to orient yourself, then proceed per the system prompt.`;
|
|
221
255
|
|
|
222
256
|
// 7. Compute deadline
|
|
223
|
-
|
|
257
|
+
let deadlineMs = Date.now() + (opts.maxDurationMs ?? 20 * 60 * 1000);
|
|
224
258
|
|
|
225
259
|
// 8. Instantiate provider (or use injected one for testing).
|
|
226
260
|
// CLI providers take a different path: they don't implement Anthropic
|
|
@@ -232,7 +266,7 @@ Begin by calling read_session_summary to orient yourself, then proceed per the s
|
|
|
232
266
|
} else {
|
|
233
267
|
const resolvedProvider = resolveProvider(opts.llmConfig);
|
|
234
268
|
if (resolvedProvider.name === 'claude-cli') {
|
|
235
|
-
return await
|
|
269
|
+
return await claudeCliCompiler({
|
|
236
270
|
session,
|
|
237
271
|
absoluteToolDir,
|
|
238
272
|
sessionPath: opts.sessionPath,
|
|
@@ -247,10 +281,12 @@ Begin by calling read_session_summary to orient yourself, then proceed per the s
|
|
|
247
281
|
buildPlanPath: opts.buildPlanPath,
|
|
248
282
|
sharedModules: opts.sharedModules,
|
|
249
283
|
toolPlan: opts.toolPlan,
|
|
284
|
+
revisionMode: opts.revisionMode,
|
|
285
|
+
model: opts.llmConfig?.model,
|
|
250
286
|
});
|
|
251
287
|
}
|
|
252
288
|
if (resolvedProvider.name === 'codex-cli') {
|
|
253
|
-
return await
|
|
289
|
+
return await codexCliCompiler({
|
|
254
290
|
session,
|
|
255
291
|
absoluteToolDir,
|
|
256
292
|
sessionPath: opts.sessionPath,
|
|
@@ -264,6 +300,8 @@ Begin by calling read_session_summary to orient yourself, then proceed per the s
|
|
|
264
300
|
buildPlanPath: opts.buildPlanPath,
|
|
265
301
|
sharedModules: opts.sharedModules,
|
|
266
302
|
toolPlan: opts.toolPlan,
|
|
303
|
+
revisionMode: opts.revisionMode,
|
|
304
|
+
model: opts.llmConfig?.model,
|
|
267
305
|
});
|
|
268
306
|
}
|
|
269
307
|
if (!isToolUseProvider(resolvedProvider)) {
|
|
@@ -336,11 +374,9 @@ Begin by calling read_session_summary to orient yourself, then proceed per the s
|
|
|
336
374
|
}
|
|
337
375
|
|
|
338
376
|
// Perform external verification
|
|
339
|
-
const
|
|
340
|
-
|
|
341
|
-
session,
|
|
342
|
-
sessionPathAbs,
|
|
343
|
-
{
|
|
377
|
+
const deterministicVerificationStartedAt = Date.now();
|
|
378
|
+
const { failures, warnings, paramVerification, integrationEvidence } =
|
|
379
|
+
await externalVerification(absoluteToolDir, session, sessionPathAbs, {
|
|
344
380
|
expectedToolName: opts.candidate?.toolName,
|
|
345
381
|
likelyParams: opts.candidate?.likelyParams,
|
|
346
382
|
candidateRequestSeqs: opts.candidate?.requestSeqs,
|
|
@@ -355,35 +391,58 @@ Begin by calling read_session_summary to orient yourself, then proceed per the s
|
|
|
355
391
|
requiredInputs,
|
|
356
392
|
credentialValues: opts.teachCredentials?.values,
|
|
357
393
|
credentialNames: opts.sharedContext?.credentialNames,
|
|
358
|
-
|
|
359
|
-
|
|
394
|
+
deferLiveIntegrationToSemanticAgent: true,
|
|
395
|
+
});
|
|
396
|
+
deadlineMs += Date.now() - deterministicVerificationStartedAt;
|
|
360
397
|
|
|
361
398
|
if (warnings.length > 0) {
|
|
362
399
|
log(`verification warnings (non-blocking):\n${warnings.join('\n')}`);
|
|
363
400
|
}
|
|
364
401
|
|
|
365
402
|
if (failures.length === 0) {
|
|
366
|
-
|
|
367
|
-
|
|
368
|
-
|
|
369
|
-
|
|
370
|
-
|
|
371
|
-
|
|
372
|
-
|
|
373
|
-
|
|
374
|
-
|
|
375
|
-
|
|
376
|
-
|
|
377
|
-
|
|
378
|
-
|
|
403
|
+
const semanticVerificationStartedAt = Date.now();
|
|
404
|
+
const semantic = await runLiveSemanticVerification({
|
|
405
|
+
provider: provider.name,
|
|
406
|
+
toolDir: absoluteToolDir,
|
|
407
|
+
evidence: integrationEvidence,
|
|
408
|
+
// Live semantic verification is a distinct phase. Do not inherit a
|
|
409
|
+
// nearly exhausted compile-agent deadline.
|
|
410
|
+
});
|
|
411
|
+
deadlineMs += Date.now() - semanticVerificationStartedAt;
|
|
412
|
+
failures.push(...semanticVerificationFailures(semantic.report));
|
|
413
|
+
if (semantic.report.status === 'approved_with_gaps') {
|
|
414
|
+
warnings.push(
|
|
415
|
+
`independent semantic verification approved with gaps: ${semantic.report.gaps.join('; ')}`,
|
|
416
|
+
);
|
|
379
417
|
}
|
|
380
|
-
if (
|
|
381
|
-
|
|
382
|
-
|
|
383
|
-
|
|
418
|
+
if (failures.length === 0) {
|
|
419
|
+
// Success (possibly with explicit semantic gaps). Persist per-parameter
|
|
420
|
+
// flags and mint liveVerified=true only after independent approval.
|
|
421
|
+
applyLiveVerification(absoluteToolDir, undefined);
|
|
422
|
+
const paramWarnings = applyParamVerification(
|
|
423
|
+
absoluteToolDir,
|
|
424
|
+
mergeSemanticParamVerification(paramVerification, semantic.report),
|
|
425
|
+
);
|
|
426
|
+
// liveVerified/parameter annotations change the hash-strict cache key,
|
|
427
|
+
// but not transport capabilities. Rebind the backend already exercised
|
|
428
|
+
// by the verifier instead of triggering another probe.
|
|
429
|
+
rebindExistingBackendsCacheToWorkflow(absoluteToolDir);
|
|
430
|
+
const allWarnings = [...warnings, ...paramWarnings];
|
|
431
|
+
if (paramWarnings.length > 0) {
|
|
432
|
+
log(`parameter verification:\n${paramWarnings.join('\n')}`);
|
|
384
433
|
}
|
|
434
|
+
message = result.doneSummary ?? 'Task completed';
|
|
435
|
+
if (allWarnings.length > 0) {
|
|
436
|
+
message += `\n\nWarnings:\n${allWarnings.join('\n')}`;
|
|
437
|
+
}
|
|
438
|
+
if (!opts.keepTest) {
|
|
439
|
+
for (const f of ['parser.test.ts', 'integration.test.ts']) {
|
|
440
|
+
const testPath = pathJoin(absoluteToolDir, f);
|
|
441
|
+
if (existsSync(testPath)) unlinkSync(testPath);
|
|
442
|
+
}
|
|
443
|
+
}
|
|
444
|
+
break;
|
|
385
445
|
}
|
|
386
|
-
break;
|
|
387
446
|
}
|
|
388
447
|
|
|
389
448
|
// Verification failed — re-enter the loop with a continuation message
|
|
@@ -399,6 +458,7 @@ Begin by calling read_session_summary to orient yourself, then proceed per the s
|
|
|
399
458
|
${failures.map((f) => `- ${f}`).join('\n')}
|
|
400
459
|
|
|
401
460
|
Resume your work. Read the files you wrote (workflow.json, parser.ts, parser.test.ts), fix the issues, re-run tests, and call done again when fixed.`;
|
|
461
|
+
currentInitialMessage += `\n\nUse the verifier's semantic evidence, not a requirement for 100% of the original candidate inputs. If the core intent works but a secondary parameter cannot be supported from the recording and live evidence without guessing, remove that parameter and add workflow.limitations with omittedParameters and a specific reason. Do not use a limitation to hide a broken core tool, and do not keep a known broken or ignored public input.`;
|
|
402
462
|
}
|
|
403
463
|
|
|
404
464
|
// 10. Final flush of the complete conversation log
|