@struktur/sdk 2.2.0 → 2.3.1
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/dist/artifacts/fileToArtifact.d.ts +8 -0
- package/dist/artifacts/fileToArtifact.d.ts.map +1 -0
- package/dist/artifacts/input.d.ts +60 -0
- package/dist/artifacts/input.d.ts.map +1 -0
- package/dist/artifacts/providers.d.ts +5 -0
- package/dist/artifacts/providers.d.ts.map +1 -0
- package/dist/artifacts/urlToArtifact.d.ts +3 -0
- package/dist/artifacts/urlToArtifact.d.ts.map +1 -0
- package/dist/auth/config.d.ts +34 -0
- package/dist/auth/config.d.ts.map +1 -0
- package/dist/auth/tokens.d.ts +18 -0
- package/dist/auth/tokens.d.ts.map +1 -0
- package/dist/chunking/ArtifactBatcher.d.ts +11 -0
- package/dist/chunking/ArtifactBatcher.d.ts.map +1 -0
- package/dist/chunking/ArtifactSplitter.d.ts +10 -0
- package/dist/chunking/ArtifactSplitter.d.ts.map +1 -0
- package/dist/debug/logger.d.ts +169 -0
- package/dist/debug/logger.d.ts.map +1 -0
- package/dist/extract.d.ts +3 -0
- package/dist/extract.d.ts.map +1 -0
- package/dist/fields.d.ts +75 -0
- package/dist/fields.d.ts.map +1 -0
- package/dist/index.d.ts +24 -0
- package/dist/index.d.ts.map +1 -0
- package/dist/index.js +1493 -1
- package/dist/index.js.map +1 -1
- package/dist/llm/LLMClient.d.ts +40 -0
- package/dist/llm/LLMClient.d.ts.map +1 -0
- package/dist/llm/RetryingRunner.d.ts +37 -0
- package/dist/llm/RetryingRunner.d.ts.map +1 -0
- package/dist/llm/message.d.ts +12 -0
- package/dist/llm/message.d.ts.map +1 -0
- package/dist/llm/models.d.ts +13 -0
- package/dist/llm/models.d.ts.map +1 -0
- package/dist/llm/resolveModel.d.ts +3 -0
- package/dist/llm/resolveModel.d.ts.map +1 -0
- package/dist/merge/Deduplicator.d.ts +4 -0
- package/dist/merge/Deduplicator.d.ts.map +1 -0
- package/dist/merge/SmartDataMerger.d.ts +7 -0
- package/dist/merge/SmartDataMerger.d.ts.map +1 -0
- package/dist/parsers/collect.d.ts +7 -0
- package/dist/parsers/collect.d.ts.map +1 -0
- package/dist/parsers/index.d.ts +7 -0
- package/dist/parsers/index.d.ts.map +1 -0
- package/dist/parsers/mime.d.ts +12 -0
- package/dist/parsers/mime.d.ts.map +1 -0
- package/dist/parsers/npm.d.ts +16 -0
- package/dist/parsers/npm.d.ts.map +1 -0
- package/dist/parsers/pdf.d.ts +36 -0
- package/dist/parsers/pdf.d.ts.map +1 -0
- package/dist/parsers/runner.d.ts +4 -0
- package/dist/parsers/runner.d.ts.map +1 -0
- package/dist/parsers/types.d.ts +27 -0
- package/dist/parsers/types.d.ts.map +1 -0
- package/dist/parsers.d.ts +1 -0
- package/dist/prompts/DeduplicationPrompt.d.ts +5 -0
- package/dist/prompts/DeduplicationPrompt.d.ts.map +1 -0
- package/dist/prompts/ExtractorPrompt.d.ts +6 -0
- package/dist/prompts/ExtractorPrompt.d.ts.map +1 -0
- package/dist/prompts/ParallelMergerPrompt.d.ts +5 -0
- package/dist/prompts/ParallelMergerPrompt.d.ts.map +1 -0
- package/dist/prompts/SequentialExtractorPrompt.d.ts +6 -0
- package/dist/prompts/SequentialExtractorPrompt.d.ts.map +1 -0
- package/dist/prompts/formatArtifacts.d.ts +3 -0
- package/dist/prompts/formatArtifacts.d.ts.map +1 -0
- package/dist/strategies/DoublePassAutoMergeStrategy.d.ts +23 -0
- package/dist/strategies/DoublePassAutoMergeStrategy.d.ts.map +1 -0
- package/dist/strategies/DoublePassStrategy.d.ts +22 -0
- package/dist/strategies/DoublePassStrategy.d.ts.map +1 -0
- package/dist/strategies/ParallelAutoMergeStrategy.d.ts +27 -0
- package/dist/strategies/ParallelAutoMergeStrategy.d.ts.map +1 -0
- package/dist/strategies/ParallelStrategy.d.ts +22 -0
- package/dist/strategies/ParallelStrategy.d.ts.map +1 -0
- package/dist/strategies/SequentialAutoMergeStrategy.d.ts +22 -0
- package/dist/strategies/SequentialAutoMergeStrategy.d.ts.map +1 -0
- package/dist/strategies/SequentialStrategy.d.ts +20 -0
- package/dist/strategies/SequentialStrategy.d.ts.map +1 -0
- package/dist/strategies/SimpleStrategy.d.ts +18 -0
- package/dist/strategies/SimpleStrategy.d.ts.map +1 -0
- package/dist/strategies/agent/AgentStrategy.d.ts +44 -0
- package/dist/strategies/agent/AgentStrategy.d.ts.map +1 -0
- package/dist/strategies/agent/AgentTools.d.ts +55 -0
- package/dist/strategies/agent/AgentTools.d.ts.map +1 -0
- package/dist/strategies/agent/ArtifactFilesystem.d.ts +51 -0
- package/dist/strategies/agent/ArtifactFilesystem.d.ts.map +1 -0
- package/dist/strategies/agent/index.d.ts +4 -0
- package/dist/strategies/agent/index.d.ts.map +1 -0
- package/dist/strategies/concurrency.d.ts +2 -0
- package/dist/strategies/concurrency.d.ts.map +1 -0
- package/dist/strategies/index.d.ts +9 -0
- package/dist/strategies/index.d.ts.map +1 -0
- package/dist/strategies/utils.d.ts +39 -0
- package/dist/strategies/utils.d.ts.map +1 -0
- package/dist/strategies.d.ts +1 -0
- package/dist/strategies.js +1495 -0
- package/dist/strategies.js.map +1 -1
- package/dist/tokenization.d.ts +11 -0
- package/dist/tokenization.d.ts.map +1 -0
- package/dist/types.d.ts +178 -0
- package/dist/types.d.ts.map +1 -0
- package/dist/validation/validator.d.ts +20 -0
- package/dist/validation/validator.d.ts.map +1 -0
- package/package.json +14 -10
package/dist/strategies.js
CHANGED
|
@@ -2416,7 +2416,1501 @@ var DoublePassAutoMergeStrategy = class {
|
|
|
2416
2416
|
var doublePassAutoMerge = (config) => {
|
|
2417
2417
|
return new DoublePassAutoMergeStrategy(config);
|
|
2418
2418
|
};
|
|
2419
|
+
|
|
2420
|
+
// src/strategies/agent/AgentStrategy.ts
|
|
2421
|
+
import {
|
|
2422
|
+
createAgentSession,
|
|
2423
|
+
AuthStorage,
|
|
2424
|
+
ModelRegistry,
|
|
2425
|
+
SessionManager,
|
|
2426
|
+
SettingsManager,
|
|
2427
|
+
DefaultResourceLoader
|
|
2428
|
+
} from "@mariozechner/pi-coding-agent";
|
|
2429
|
+
import { Bash as Bash2 } from "just-bash";
|
|
2430
|
+
|
|
2431
|
+
// src/strategies/agent/ArtifactFilesystem.ts
|
|
2432
|
+
var detectImageFormat = (base64) => {
|
|
2433
|
+
if (base64.startsWith("/9j/")) {
|
|
2434
|
+
return "jpg";
|
|
2435
|
+
}
|
|
2436
|
+
if (base64.startsWith("iVBOR")) {
|
|
2437
|
+
return "png";
|
|
2438
|
+
}
|
|
2439
|
+
if (base64.startsWith("R0lGOD")) {
|
|
2440
|
+
return "gif";
|
|
2441
|
+
}
|
|
2442
|
+
if (base64.startsWith("UklGR")) {
|
|
2443
|
+
return "webp";
|
|
2444
|
+
}
|
|
2445
|
+
if (base64.startsWith("Qk")) {
|
|
2446
|
+
return "bmp";
|
|
2447
|
+
}
|
|
2448
|
+
if (base64.startsWith("PHN2Zy") || base64.startsWith("data:image/svg")) {
|
|
2449
|
+
return "svg";
|
|
2450
|
+
}
|
|
2451
|
+
return "bin";
|
|
2452
|
+
};
|
|
2453
|
+
var sanitizeArtifactName = (name) => {
|
|
2454
|
+
return name.replace(/[^a-zA-Z0-9]/g, "-").replace(/-+/g, "-").replace(/^-|-$/g, "").toLowerCase();
|
|
2455
|
+
};
|
|
2456
|
+
var createVirtualFilesystem = (artifacts) => {
|
|
2457
|
+
const virtualFiles = /* @__PURE__ */ new Map();
|
|
2458
|
+
const transformedArtifacts = artifacts.map((artifact) => {
|
|
2459
|
+
const artifactName = sanitizeArtifactName(artifact.id);
|
|
2460
|
+
return {
|
|
2461
|
+
id: artifact.id,
|
|
2462
|
+
type: artifact.type,
|
|
2463
|
+
metadata: artifact.metadata,
|
|
2464
|
+
tokens: artifact.tokens,
|
|
2465
|
+
contents: artifact.contents.map((content, contentIndex) => {
|
|
2466
|
+
const pageNumber = content.page;
|
|
2467
|
+
return {
|
|
2468
|
+
page: content.page,
|
|
2469
|
+
text: content.text,
|
|
2470
|
+
media: content.media?.map((media, mediaIndex) => {
|
|
2471
|
+
if (media.base64 && media.base64.length > 0) {
|
|
2472
|
+
const extension = detectImageFormat(media.base64);
|
|
2473
|
+
let virtualPath;
|
|
2474
|
+
if (pageNumber !== void 0) {
|
|
2475
|
+
virtualPath = `/images/${artifactName}-page-${pageNumber}-image-${mediaIndex}.${extension}`;
|
|
2476
|
+
} else {
|
|
2477
|
+
virtualPath = `/images/${artifactName}-image-${mediaIndex}.${extension}`;
|
|
2478
|
+
}
|
|
2479
|
+
virtualFiles.set(virtualPath, media.base64);
|
|
2480
|
+
return {
|
|
2481
|
+
type: media.type,
|
|
2482
|
+
url: media.url,
|
|
2483
|
+
text: media.text,
|
|
2484
|
+
x: media.x,
|
|
2485
|
+
y: media.y,
|
|
2486
|
+
width: media.width,
|
|
2487
|
+
height: media.height,
|
|
2488
|
+
imageType: media.imageType,
|
|
2489
|
+
virtualPath,
|
|
2490
|
+
originalBase64: `[BASE64: ${media.base64.length} chars]`
|
|
2491
|
+
// For debugging
|
|
2492
|
+
};
|
|
2493
|
+
}
|
|
2494
|
+
return {
|
|
2495
|
+
type: media.type,
|
|
2496
|
+
url: media.url,
|
|
2497
|
+
text: media.text,
|
|
2498
|
+
x: media.x,
|
|
2499
|
+
y: media.y,
|
|
2500
|
+
width: media.width,
|
|
2501
|
+
height: media.height,
|
|
2502
|
+
imageType: media.imageType
|
|
2503
|
+
};
|
|
2504
|
+
})
|
|
2505
|
+
};
|
|
2506
|
+
})
|
|
2507
|
+
};
|
|
2508
|
+
});
|
|
2509
|
+
const totalImagesWithBase64 = Array.from(virtualFiles.keys()).length;
|
|
2510
|
+
const manifest = {
|
|
2511
|
+
count: artifacts.length,
|
|
2512
|
+
artifacts: transformedArtifacts,
|
|
2513
|
+
totalTokens: transformedArtifacts.reduce((sum, a) => sum + (a.tokens || 0), 0),
|
|
2514
|
+
summary: {
|
|
2515
|
+
textArtifacts: artifacts.filter((a) => a.type === "text").length,
|
|
2516
|
+
imageArtifacts: artifacts.filter((a) => a.type === "image").length,
|
|
2517
|
+
pdfArtifacts: artifacts.filter((a) => a.type === "pdf").length,
|
|
2518
|
+
fileArtifacts: artifacts.filter((a) => a.type === "file").length
|
|
2519
|
+
},
|
|
2520
|
+
virtualFiles: {
|
|
2521
|
+
count: totalImagesWithBase64,
|
|
2522
|
+
paths: Array.from(virtualFiles.keys())
|
|
2523
|
+
}
|
|
2524
|
+
};
|
|
2525
|
+
const getImageByPath = (path) => {
|
|
2526
|
+
return virtualFiles.get(path);
|
|
2527
|
+
};
|
|
2528
|
+
const filesystem = {
|
|
2529
|
+
"/artifact.json": JSON.stringify(transformedArtifacts, null, 2),
|
|
2530
|
+
"/manifest.json": JSON.stringify(manifest, null, 2),
|
|
2531
|
+
virtualFiles,
|
|
2532
|
+
getImageByPath
|
|
2533
|
+
};
|
|
2534
|
+
return filesystem;
|
|
2535
|
+
};
|
|
2536
|
+
|
|
2537
|
+
// src/strategies/agent/AgentTools.ts
|
|
2538
|
+
import "just-bash";
|
|
2539
|
+
import { Type } from "@sinclair/typebox";
|
|
2540
|
+
var BashParams = Type.Object({
|
|
2541
|
+
command: Type.String({
|
|
2542
|
+
description: "The bash command to execute"
|
|
2543
|
+
}),
|
|
2544
|
+
timeout: Type.Optional(
|
|
2545
|
+
Type.Number({
|
|
2546
|
+
description: "Timeout in milliseconds (default: 30000)"
|
|
2547
|
+
})
|
|
2548
|
+
)
|
|
2549
|
+
});
|
|
2550
|
+
var ReadParams = Type.Object({
|
|
2551
|
+
file_path: Type.String({
|
|
2552
|
+
description: "The absolute path to the file to read"
|
|
2553
|
+
}),
|
|
2554
|
+
offset: Type.Optional(
|
|
2555
|
+
Type.Number({
|
|
2556
|
+
description: "Line number to start reading from (1-indexed, default: 1)"
|
|
2557
|
+
})
|
|
2558
|
+
),
|
|
2559
|
+
limit: Type.Optional(
|
|
2560
|
+
Type.Number({
|
|
2561
|
+
description: "Maximum number of lines to read"
|
|
2562
|
+
})
|
|
2563
|
+
)
|
|
2564
|
+
});
|
|
2565
|
+
var GrepParams = Type.Object({
|
|
2566
|
+
pattern: Type.String({
|
|
2567
|
+
description: "The search pattern"
|
|
2568
|
+
}),
|
|
2569
|
+
path: Type.String({
|
|
2570
|
+
description: "The file or directory to search in"
|
|
2571
|
+
}),
|
|
2572
|
+
options: Type.Optional(
|
|
2573
|
+
Type.String({
|
|
2574
|
+
description: "Additional grep options (e.g., '-r' for recursive, '-i' for case-insensitive)"
|
|
2575
|
+
})
|
|
2576
|
+
)
|
|
2577
|
+
});
|
|
2578
|
+
var FindParams = Type.Object({
|
|
2579
|
+
path: Type.String({
|
|
2580
|
+
description: "The directory to search in"
|
|
2581
|
+
}),
|
|
2582
|
+
name: Type.Optional(
|
|
2583
|
+
Type.String({
|
|
2584
|
+
description: "Filename pattern to match (e.g., '*.json')"
|
|
2585
|
+
})
|
|
2586
|
+
)
|
|
2587
|
+
});
|
|
2588
|
+
var LsParams = Type.Object({
|
|
2589
|
+
path: Type.String({
|
|
2590
|
+
description: "The directory to list"
|
|
2591
|
+
}),
|
|
2592
|
+
recursive: Type.Optional(
|
|
2593
|
+
Type.Boolean({
|
|
2594
|
+
description: "List recursively"
|
|
2595
|
+
})
|
|
2596
|
+
)
|
|
2597
|
+
});
|
|
2598
|
+
var SetOutputDataParams = Type.Object({
|
|
2599
|
+
data: Type.Any({
|
|
2600
|
+
description: "The output data to set. Can be any shape - will be validated against the schema."
|
|
2601
|
+
})
|
|
2602
|
+
});
|
|
2603
|
+
var UpdateOutputDataParams = Type.Object({
|
|
2604
|
+
changes: Type.Record(Type.String(), Type.Any(), {
|
|
2605
|
+
description: "Changes to merge into the existing output data. Uses deep merge. Missing fields are preserved."
|
|
2606
|
+
})
|
|
2607
|
+
});
|
|
2608
|
+
var ViewImageParams = Type.Object({
|
|
2609
|
+
image_path: Type.String({
|
|
2610
|
+
description: "The absolute path to the image file to view (e.g., '/artifacts/images/artifact-name-page-1-image-0.png')"
|
|
2611
|
+
})
|
|
2612
|
+
});
|
|
2613
|
+
var FinishParams = Type.Object({});
|
|
2614
|
+
var FailParams = Type.Object({
|
|
2615
|
+
reason: Type.String({
|
|
2616
|
+
description: "Explanation of why extraction failed or what data could not be found."
|
|
2617
|
+
})
|
|
2618
|
+
});
|
|
2619
|
+
var createVirtualFilesystemTools = (bash, getImageByPath) => {
|
|
2620
|
+
const bashTool = {
|
|
2621
|
+
name: "bash",
|
|
2622
|
+
label: "Bash",
|
|
2623
|
+
description: "Execute bash commands in the virtual filesystem. Use this to explore artifacts with commands like cat, grep, head, tail, jq, etc.",
|
|
2624
|
+
parameters: BashParams,
|
|
2625
|
+
execute: async (toolCallId, params, signal, onUpdate, ctx) => {
|
|
2626
|
+
try {
|
|
2627
|
+
const result = await bash.exec(params.command);
|
|
2628
|
+
return {
|
|
2629
|
+
content: [
|
|
2630
|
+
{
|
|
2631
|
+
type: "text",
|
|
2632
|
+
text: result.exitCode === 0 ? result.stdout || "(no output)" : `Exit code ${result.exitCode}: ${result.stderr || result.stdout || "(no output)"}`
|
|
2633
|
+
}
|
|
2634
|
+
],
|
|
2635
|
+
details: {
|
|
2636
|
+
exitCode: result.exitCode,
|
|
2637
|
+
...result.stderr && { stderr: result.stderr }
|
|
2638
|
+
}
|
|
2639
|
+
};
|
|
2640
|
+
} catch (error) {
|
|
2641
|
+
const errorMsg = error.message;
|
|
2642
|
+
console.error(`[AgentTools] Bash command error: ${errorMsg}`);
|
|
2643
|
+
console.error(`[AgentTools] Command: ${params.command}`);
|
|
2644
|
+
return {
|
|
2645
|
+
content: [
|
|
2646
|
+
{
|
|
2647
|
+
type: "text",
|
|
2648
|
+
text: `Error: ${errorMsg}`
|
|
2649
|
+
}
|
|
2650
|
+
],
|
|
2651
|
+
details: {
|
|
2652
|
+
error: errorMsg
|
|
2653
|
+
},
|
|
2654
|
+
isError: true
|
|
2655
|
+
};
|
|
2656
|
+
}
|
|
2657
|
+
}
|
|
2658
|
+
};
|
|
2659
|
+
const readTool = {
|
|
2660
|
+
name: "read",
|
|
2661
|
+
label: "Read File",
|
|
2662
|
+
description: "Read the contents of a file from the virtual filesystem. Supports pagination with offset and limit parameters. Can read virtual image files (binary/base64 content).",
|
|
2663
|
+
parameters: ReadParams,
|
|
2664
|
+
execute: async (toolCallId, params, signal, onUpdate, ctx) => {
|
|
2665
|
+
try {
|
|
2666
|
+
if (getImageByPath && params.file_path.startsWith("/images/")) {
|
|
2667
|
+
const imageData = getImageByPath(params.file_path);
|
|
2668
|
+
if (imageData) {
|
|
2669
|
+
const displayData = imageData.length > 1e3 ? imageData.slice(0, 1e3) + "... [truncated]" : imageData;
|
|
2670
|
+
return {
|
|
2671
|
+
content: [
|
|
2672
|
+
{
|
|
2673
|
+
type: "text",
|
|
2674
|
+
text: `[IMAGE FILE: ${params.file_path}]
|
|
2675
|
+
Base64 content (${imageData.length} chars):
|
|
2676
|
+
${displayData}`
|
|
2677
|
+
}
|
|
2678
|
+
],
|
|
2679
|
+
details: {
|
|
2680
|
+
path: params.file_path,
|
|
2681
|
+
size: imageData.length,
|
|
2682
|
+
truncated: imageData.length > 1e3
|
|
2683
|
+
}
|
|
2684
|
+
};
|
|
2685
|
+
}
|
|
2686
|
+
}
|
|
2687
|
+
let command;
|
|
2688
|
+
const offset = params.offset || 1;
|
|
2689
|
+
const limit = params.limit;
|
|
2690
|
+
const endLine = limit ? offset + limit - 1 : void 0;
|
|
2691
|
+
if (limit && endLine) {
|
|
2692
|
+
command = `sed -n '${offset},${endLine}p' "${params.file_path}"`;
|
|
2693
|
+
} else if (offset > 1) {
|
|
2694
|
+
command = `sed -n '${offset},$p' "${params.file_path}"`;
|
|
2695
|
+
} else {
|
|
2696
|
+
command = `cat "${params.file_path}"`;
|
|
2697
|
+
}
|
|
2698
|
+
const result = await bash.exec(command);
|
|
2699
|
+
if (result.exitCode !== 0) {
|
|
2700
|
+
return {
|
|
2701
|
+
content: [
|
|
2702
|
+
{
|
|
2703
|
+
type: "text",
|
|
2704
|
+
text: `Error reading file: ${result.stderr || "File not found"}`
|
|
2705
|
+
}
|
|
2706
|
+
],
|
|
2707
|
+
details: {
|
|
2708
|
+
error: result.stderr || "File not found"
|
|
2709
|
+
},
|
|
2710
|
+
isError: true
|
|
2711
|
+
};
|
|
2712
|
+
}
|
|
2713
|
+
return {
|
|
2714
|
+
content: [
|
|
2715
|
+
{
|
|
2716
|
+
type: "text",
|
|
2717
|
+
text: result.stdout
|
|
2718
|
+
}
|
|
2719
|
+
],
|
|
2720
|
+
details: {
|
|
2721
|
+
lines: result.stdout.split("\n").length,
|
|
2722
|
+
characters: result.stdout.length
|
|
2723
|
+
}
|
|
2724
|
+
};
|
|
2725
|
+
} catch (error) {
|
|
2726
|
+
const errorMsg = error.message;
|
|
2727
|
+
console.error(`[AgentTools] Read error: ${errorMsg}`);
|
|
2728
|
+
console.error(`[AgentTools] File path: ${params.file_path}`);
|
|
2729
|
+
return {
|
|
2730
|
+
content: [
|
|
2731
|
+
{
|
|
2732
|
+
type: "text",
|
|
2733
|
+
text: `Error: ${errorMsg}`
|
|
2734
|
+
}
|
|
2735
|
+
],
|
|
2736
|
+
details: {
|
|
2737
|
+
error: errorMsg
|
|
2738
|
+
},
|
|
2739
|
+
isError: true
|
|
2740
|
+
};
|
|
2741
|
+
}
|
|
2742
|
+
}
|
|
2743
|
+
};
|
|
2744
|
+
const grepTool = {
|
|
2745
|
+
name: "grep",
|
|
2746
|
+
label: "Grep",
|
|
2747
|
+
description: "Search for patterns in files using grep",
|
|
2748
|
+
parameters: GrepParams,
|
|
2749
|
+
execute: async (toolCallId, params, signal, onUpdate, ctx) => {
|
|
2750
|
+
try {
|
|
2751
|
+
const options = params.options || "";
|
|
2752
|
+
const command = `grep ${options} "${params.pattern}" "${params.path}" 2>/dev/null || echo "(no matches found)"`;
|
|
2753
|
+
const result = await bash.exec(command);
|
|
2754
|
+
return {
|
|
2755
|
+
content: [
|
|
2756
|
+
{
|
|
2757
|
+
type: "text",
|
|
2758
|
+
text: result.stdout || "(no matches found)"
|
|
2759
|
+
}
|
|
2760
|
+
],
|
|
2761
|
+
details: {
|
|
2762
|
+
matches: result.stdout.split("\n").filter((line) => line.trim()).length
|
|
2763
|
+
}
|
|
2764
|
+
};
|
|
2765
|
+
} catch (error) {
|
|
2766
|
+
return {
|
|
2767
|
+
content: [
|
|
2768
|
+
{
|
|
2769
|
+
type: "text",
|
|
2770
|
+
text: `(no matches found)`
|
|
2771
|
+
}
|
|
2772
|
+
],
|
|
2773
|
+
details: {}
|
|
2774
|
+
};
|
|
2775
|
+
}
|
|
2776
|
+
}
|
|
2777
|
+
};
|
|
2778
|
+
const findTool = {
|
|
2779
|
+
name: "find",
|
|
2780
|
+
label: "Find",
|
|
2781
|
+
description: "Find files by name or pattern",
|
|
2782
|
+
parameters: FindParams,
|
|
2783
|
+
execute: async (toolCallId, params, signal, onUpdate, ctx) => {
|
|
2784
|
+
try {
|
|
2785
|
+
let command = `find "${params.path}" -type f`;
|
|
2786
|
+
if (params.name) {
|
|
2787
|
+
command = `find "${params.path}" -type f -name "${params.name}"`;
|
|
2788
|
+
}
|
|
2789
|
+
const result = await bash.exec(command);
|
|
2790
|
+
return {
|
|
2791
|
+
content: [
|
|
2792
|
+
{
|
|
2793
|
+
type: "text",
|
|
2794
|
+
text: result.stdout || "(no files found)"
|
|
2795
|
+
}
|
|
2796
|
+
],
|
|
2797
|
+
details: {
|
|
2798
|
+
files: result.stdout.split("\n").filter((line) => line.trim()).length
|
|
2799
|
+
}
|
|
2800
|
+
};
|
|
2801
|
+
} catch (error) {
|
|
2802
|
+
return {
|
|
2803
|
+
content: [
|
|
2804
|
+
{
|
|
2805
|
+
type: "text",
|
|
2806
|
+
text: `Error: ${error.message}`
|
|
2807
|
+
}
|
|
2808
|
+
],
|
|
2809
|
+
details: {
|
|
2810
|
+
error: error.message
|
|
2811
|
+
},
|
|
2812
|
+
isError: true
|
|
2813
|
+
};
|
|
2814
|
+
}
|
|
2815
|
+
}
|
|
2816
|
+
};
|
|
2817
|
+
const lsTool = {
|
|
2818
|
+
name: "ls",
|
|
2819
|
+
label: "List Directory",
|
|
2820
|
+
description: "List files and directories",
|
|
2821
|
+
parameters: LsParams,
|
|
2822
|
+
execute: async (toolCallId, params, signal, onUpdate, ctx) => {
|
|
2823
|
+
try {
|
|
2824
|
+
let command = `ls -la "${params.path}"`;
|
|
2825
|
+
if (params.recursive) {
|
|
2826
|
+
command = `ls -laR "${params.path}"`;
|
|
2827
|
+
}
|
|
2828
|
+
const result = await bash.exec(command);
|
|
2829
|
+
if (result.exitCode !== 0) {
|
|
2830
|
+
return {
|
|
2831
|
+
content: [
|
|
2832
|
+
{
|
|
2833
|
+
type: "text",
|
|
2834
|
+
text: `Error: ${result.stderr || "Directory not found"}`
|
|
2835
|
+
}
|
|
2836
|
+
],
|
|
2837
|
+
details: {
|
|
2838
|
+
error: result.stderr || "Directory not found"
|
|
2839
|
+
},
|
|
2840
|
+
isError: true
|
|
2841
|
+
};
|
|
2842
|
+
}
|
|
2843
|
+
return {
|
|
2844
|
+
content: [
|
|
2845
|
+
{
|
|
2846
|
+
type: "text",
|
|
2847
|
+
text: result.stdout
|
|
2848
|
+
}
|
|
2849
|
+
],
|
|
2850
|
+
details: {
|
|
2851
|
+
entries: result.stdout.split("\n").filter((line) => line.trim() && !line.startsWith("total ")).length
|
|
2852
|
+
}
|
|
2853
|
+
};
|
|
2854
|
+
} catch (error) {
|
|
2855
|
+
return {
|
|
2856
|
+
content: [
|
|
2857
|
+
{
|
|
2858
|
+
type: "text",
|
|
2859
|
+
text: `Error: ${error.message}`
|
|
2860
|
+
}
|
|
2861
|
+
],
|
|
2862
|
+
details: {
|
|
2863
|
+
error: error.message
|
|
2864
|
+
},
|
|
2865
|
+
isError: true
|
|
2866
|
+
};
|
|
2867
|
+
}
|
|
2868
|
+
}
|
|
2869
|
+
};
|
|
2870
|
+
const viewImageTool = {
|
|
2871
|
+
name: "view_image",
|
|
2872
|
+
label: "View Image",
|
|
2873
|
+
description: "View an image file from the virtual filesystem. This injects the image as a visual message that the AI can see and analyze. Use this to examine images, screenshots, diagrams, or any visual content in the artifacts.",
|
|
2874
|
+
parameters: ViewImageParams,
|
|
2875
|
+
execute: async (toolCallId, params, signal, onUpdate, ctx) => {
|
|
2876
|
+
try {
|
|
2877
|
+
if (!getImageByPath) {
|
|
2878
|
+
return {
|
|
2879
|
+
content: [
|
|
2880
|
+
{
|
|
2881
|
+
type: "text",
|
|
2882
|
+
text: "Error: Image viewing is not available"
|
|
2883
|
+
}
|
|
2884
|
+
],
|
|
2885
|
+
details: {
|
|
2886
|
+
error: "Image viewing is not available"
|
|
2887
|
+
},
|
|
2888
|
+
isError: true
|
|
2889
|
+
};
|
|
2890
|
+
}
|
|
2891
|
+
const imageData = getImageByPath(params.image_path);
|
|
2892
|
+
if (!imageData) {
|
|
2893
|
+
return {
|
|
2894
|
+
content: [
|
|
2895
|
+
{
|
|
2896
|
+
type: "text",
|
|
2897
|
+
text: `Error: Image not found at ${params.image_path}`
|
|
2898
|
+
}
|
|
2899
|
+
],
|
|
2900
|
+
details: {
|
|
2901
|
+
error: `Image not found at ${params.image_path}`
|
|
2902
|
+
},
|
|
2903
|
+
isError: true
|
|
2904
|
+
};
|
|
2905
|
+
}
|
|
2906
|
+
const getImageFormat = (path, data) => {
|
|
2907
|
+
if (path.endsWith(".png")) return "image/png";
|
|
2908
|
+
if (path.endsWith(".jpg") || path.endsWith(".jpeg")) return "image/jpeg";
|
|
2909
|
+
if (path.endsWith(".gif")) return "image/gif";
|
|
2910
|
+
if (path.endsWith(".webp")) return "image/webp";
|
|
2911
|
+
if (path.endsWith(".bmp")) return "image/bmp";
|
|
2912
|
+
if (path.endsWith(".svg")) return "image/svg+xml";
|
|
2913
|
+
if (data.startsWith("/9j/")) return "image/jpeg";
|
|
2914
|
+
if (data.startsWith("iVBORw0KGgo")) return "image/png";
|
|
2915
|
+
if (data.startsWith("R0lGOD")) return "image/gif";
|
|
2916
|
+
if (data.startsWith("Qk")) return "image/bmp";
|
|
2917
|
+
if (data.startsWith("PHN2Zy")) return "image/svg+xml";
|
|
2918
|
+
return "image/png";
|
|
2919
|
+
};
|
|
2920
|
+
const mimeType = getImageFormat(params.image_path, imageData);
|
|
2921
|
+
return {
|
|
2922
|
+
content: [
|
|
2923
|
+
{
|
|
2924
|
+
type: "text",
|
|
2925
|
+
text: `[Viewing image: ${params.image_path}]`
|
|
2926
|
+
},
|
|
2927
|
+
{
|
|
2928
|
+
type: "image",
|
|
2929
|
+
data: imageData,
|
|
2930
|
+
mimeType
|
|
2931
|
+
}
|
|
2932
|
+
],
|
|
2933
|
+
details: {
|
|
2934
|
+
path: params.image_path,
|
|
2935
|
+
format: mimeType,
|
|
2936
|
+
size: imageData.length
|
|
2937
|
+
}
|
|
2938
|
+
};
|
|
2939
|
+
} catch (error) {
|
|
2940
|
+
const errorMsg = error.message;
|
|
2941
|
+
console.error(`[AgentTools] View image error: ${errorMsg}`);
|
|
2942
|
+
console.error(`[AgentTools] Image path: ${params.image_path}`);
|
|
2943
|
+
return {
|
|
2944
|
+
content: [
|
|
2945
|
+
{
|
|
2946
|
+
type: "text",
|
|
2947
|
+
text: `Error viewing image: ${errorMsg}`
|
|
2948
|
+
}
|
|
2949
|
+
],
|
|
2950
|
+
details: {
|
|
2951
|
+
error: errorMsg
|
|
2952
|
+
},
|
|
2953
|
+
isError: true
|
|
2954
|
+
};
|
|
2955
|
+
}
|
|
2956
|
+
}
|
|
2957
|
+
};
|
|
2958
|
+
return {
|
|
2959
|
+
bash: bashTool,
|
|
2960
|
+
read: readTool,
|
|
2961
|
+
grep: grepTool,
|
|
2962
|
+
find: findTool,
|
|
2963
|
+
ls: lsTool,
|
|
2964
|
+
view_image: viewImageTool
|
|
2965
|
+
};
|
|
2966
|
+
};
|
|
2967
|
+
|
|
2968
|
+
// src/strategies/agent/AgentStrategy.ts
|
|
2969
|
+
var defaultSystemPrompt = (schema, outputInstructions) => {
|
|
2970
|
+
return `You are an autonomous data extraction agent. Your task is to explore the provided artifacts and extract structured data according to the given JSON schema.
|
|
2971
|
+
|
|
2972
|
+
## Your Environment
|
|
2973
|
+
|
|
2974
|
+
You have access to a virtual filesystem containing the artifacts to extract from:
|
|
2975
|
+
- "/artifact.json" - All artifacts in a structured JSON format (with embedded images replaced by virtual file paths)
|
|
2976
|
+
- "/manifest.json" - Summary and metadata about the artifacts
|
|
2977
|
+
- "/images/" - Virtual directory containing extracted image files (when artifacts have embedded images)
|
|
2978
|
+
|
|
2979
|
+
## Virtual Image Files
|
|
2980
|
+
|
|
2981
|
+
When artifacts contain embedded images (base64-encoded), they are extracted to separate files in "/images/" for easier access:
|
|
2982
|
+
- Image files are named: "/images/{artifact-name}-page-{n}-image-{i}.{ext}"
|
|
2983
|
+
- {artifact-name}: Sanitized artifact ID (lowercase, special chars become dashes)
|
|
2984
|
+
- page-{n}: Page number from the artifact (if available)
|
|
2985
|
+
- image-{i}: Image index within that page
|
|
2986
|
+
- {ext}: File extension determined from base64 (jpg, png, gif, webp, bmp, svg, or bin)
|
|
2987
|
+
- Examples: "/images/invoice-page-3-image-0.jpg" or "/images/report-image-1.png"
|
|
2988
|
+
- Use the "/images/" directory to access image data
|
|
2989
|
+
- The manifest shows which virtual files are available
|
|
2990
|
+
- Image format is shown in the file extension for easy identification
|
|
2991
|
+
|
|
2992
|
+
## IMPORTANT: Do NOT Install Tools
|
|
2993
|
+
|
|
2994
|
+
This is a **sandboxed environment** - you CANNOT install packages or tools:
|
|
2995
|
+
- \u274C DO NOT run: apt-get, pip install, npm install, brew install, etc.
|
|
2996
|
+
- \u274C DO NOT try to install tesseract, ocrmypdf, poppler, or any OCR tools
|
|
2997
|
+
- \u274C DO NOT check if tools exist with "which" or "command -v"
|
|
2998
|
+
- \u2705 ONLY use the provided tools listed below
|
|
2999
|
+
- \u2705 If a tool is missing, work with what you have or report it via fail()
|
|
3000
|
+
|
|
3001
|
+
## Available Tools
|
|
3002
|
+
|
|
3003
|
+
### Exploration Tools
|
|
3004
|
+
- **read** - Read file contents with pagination support (e.g., read {"file_path": "/manifest.json", "limit": 50})
|
|
3005
|
+
- **view_image** - View an image to see its contents visually (e.g., view_image {"image_path": "/images/doc-page-1-image-0.png"})
|
|
3006
|
+
- **bash** - Run shell commands (e.g., bash {"command": "head -20 /artifact.json"})
|
|
3007
|
+
- **grep** - Search for patterns in files
|
|
3008
|
+
- **find** - Find files by name or pattern
|
|
3009
|
+
- **ls** - List files and directories
|
|
3010
|
+
|
|
3011
|
+
### Output Management Tools (IMPORTANT - Use These!)
|
|
3012
|
+
- **set_output_data** - Set the initial extraction output. Call this as soon as you find the first piece of data.
|
|
3013
|
+
- Example: set_output_data({"data": {"company_name": "Acme Corp"}})
|
|
3014
|
+
- The data can be any shape - you'll update it incrementally
|
|
3015
|
+
|
|
3016
|
+
- **update_output_data** - Add or modify fields in the existing output data
|
|
3017
|
+
- Example: update_output_data({"changes": {"address": "123 Main St"}})
|
|
3018
|
+
- This merges new data with existing data (deep merge)
|
|
3019
|
+
- Call this frequently as you discover more information
|
|
3020
|
+
|
|
3021
|
+
- **finish** - Call this when extraction is complete and data validates against the schema
|
|
3022
|
+
- Only works if the data is valid according to the schema
|
|
3023
|
+
- This ends the extraction successfully
|
|
3024
|
+
|
|
3025
|
+
- **fail** - Call this if the schema cannot be satisfied with available data
|
|
3026
|
+
- Provide a reason explaining what data was missing or why extraction failed
|
|
3027
|
+
|
|
3028
|
+
## CRITICAL: Incremental Data Updates
|
|
3029
|
+
|
|
3030
|
+
**You MUST update the output data continuously as you explore!**
|
|
3031
|
+
|
|
3032
|
+
1. **Start immediately**: As soon as you find the first field, call set_output_data
|
|
3033
|
+
2. **Update frequently**: Every time you find new information, call update_output_data
|
|
3034
|
+
3. **Build incrementally**: Don't wait until the end - keep adding data as you go
|
|
3035
|
+
4. **Use all tools**: Combine exploration tools with output tools
|
|
3036
|
+
|
|
3037
|
+
### Example Workflow
|
|
3038
|
+
|
|
3039
|
+
1. Read manifest: read {"file_path": "/manifest.json", "limit": 20}
|
|
3040
|
+
2. Find first data point: grep "company_name" /artifact.json
|
|
3041
|
+
3. **Set initial data**: set_output_data({"data": {"company_name": "Acme Inc"}})
|
|
3042
|
+
4. Continue exploring: read {"file_path": "/artifact.json", "offset": 50, "limit": 30}
|
|
3043
|
+
5. **Update with more data**: update_output_data({"changes": {"address": "123 Main St", "city": "Berlin"}})
|
|
3044
|
+
6. Check images: view_image {"image_path": "/images/doc-page-1-image-0.png"}
|
|
3045
|
+
7. **Update again**: update_output_data({"changes": {"has_logo": true}})
|
|
3046
|
+
8. Verify complete: Check all schema fields are present
|
|
3047
|
+
9. **Finish**: finish()
|
|
3048
|
+
|
|
3049
|
+
## Efficient Exploration Strategy
|
|
3050
|
+
|
|
3051
|
+
**Don't read entire files at once.** Files may be large. Instead:
|
|
3052
|
+
|
|
3053
|
+
1. **Start small**: Read just the first 20-50 lines to understand the structure
|
|
3054
|
+
2. **Navigate selectively**: Use offset and limit to jump to relevant sections
|
|
3055
|
+
3. **Search first**: Use grep to find specific data before reading full content
|
|
3056
|
+
4. **Iterate**: Make multiple small reads rather than one giant read
|
|
3057
|
+
5. **Update as you go**: Call update_output_data immediately when you find data
|
|
3058
|
+
|
|
3059
|
+
### Pagination Examples
|
|
3060
|
+
|
|
3061
|
+
Read first 30 lines:
|
|
3062
|
+
read {"file_path": "/artifact.json", "limit": 30}
|
|
3063
|
+
|
|
3064
|
+
Read lines 31-60 (page 2):
|
|
3065
|
+
read {"file_path": "/artifact.json", "offset": 31, "limit": 30}
|
|
3066
|
+
|
|
3067
|
+
Read from line 100 to end:
|
|
3068
|
+
read {"file_path": "/artifact.json", "offset": 100}
|
|
3069
|
+
|
|
3070
|
+
## Output Rules
|
|
3071
|
+
|
|
3072
|
+
- **Update continuously**: Call update_output_data every time you find new information
|
|
3073
|
+
- **Start early**: Don't wait until the end - set initial data as soon as possible
|
|
3074
|
+
- **Use null for missing values**: If a field can't be found, set it to null
|
|
3075
|
+
- **Never guess**: Only extract information explicitly present in the artifacts
|
|
3076
|
+
- **Validate as you go**: The tools will tell you if your data has validation issues
|
|
3077
|
+
- **Finish properly**: You MUST call finish() to complete extraction successfully
|
|
3078
|
+
- **Fail if needed**: Use fail() if the schema truly cannot be satisfied
|
|
3079
|
+
|
|
3080
|
+
${outputInstructions ? `
|
|
3081
|
+
## Additional Instructions
|
|
3082
|
+
|
|
3083
|
+
${outputInstructions}
|
|
3084
|
+
` : ""}
|
|
3085
|
+
|
|
3086
|
+
## JSON Schema
|
|
3087
|
+
|
|
3088
|
+
${schema}
|
|
3089
|
+
|
|
3090
|
+
## CRITICAL: Tool Calling Format
|
|
3091
|
+
|
|
3092
|
+
When calling tools, you MUST provide the correct parameters:
|
|
3093
|
+
|
|
3094
|
+
**CORRECT - read with file_path:**
|
|
3095
|
+
read {"file_path": "/manifest.json"}
|
|
3096
|
+
|
|
3097
|
+
**CORRECT - read with pagination:**
|
|
3098
|
+
read {"file_path": "/artifact.json", "offset": 1, "limit": 50}
|
|
3099
|
+
|
|
3100
|
+
**CORRECT - view image:**
|
|
3101
|
+
view_image {"image_path": "/images/doc-page-1-image-0.png"}
|
|
3102
|
+
|
|
3103
|
+
**CORRECT - set output data:**
|
|
3104
|
+
set_output_data {"data": {"company_name": "Acme Corp"}}
|
|
3105
|
+
|
|
3106
|
+
**CORRECT - update output:**
|
|
3107
|
+
update_output_data {"changes": {"address": "123 Main St"}}
|
|
3108
|
+
|
|
3109
|
+
**CORRECT - finish:**
|
|
3110
|
+
finish {}
|
|
3111
|
+
|
|
3112
|
+
**CORRECT - fail:**
|
|
3113
|
+
fail {"reason": "Document is not an invoice"}
|
|
3114
|
+
|
|
3115
|
+
## Common Mistakes to AVOID
|
|
3116
|
+
|
|
3117
|
+
\u274C WRONG: read {} (missing file_path)
|
|
3118
|
+
\u274C WRONG: read {file_path: "/path"} (missing quotes around property names)
|
|
3119
|
+
\u274C WRONG: read /path (not using JSON format)
|
|
3120
|
+
\u274C WRONG: set_output_data {company: "Name"} (missing quotes and data wrapper)
|
|
3121
|
+
\u274C WRONG: Trying to install tools with apt-get, pip, npm, etc. (not allowed in sandbox)
|
|
3122
|
+
|
|
3123
|
+
## Remember
|
|
3124
|
+
|
|
3125
|
+
1. **ALWAYS** use set_output_data or update_output_data when you find information
|
|
3126
|
+
2. **ALWAYS** call finish() when done (or fail() if impossible)
|
|
3127
|
+
3. **ALWAYS** provide required parameters when calling tools (file_path for read, data for set_output_data, etc.)
|
|
3128
|
+
4. **NEVER** try to install packages or external tools - work with what you have
|
|
3129
|
+
5. The output tools will validate your data and report issues
|
|
3130
|
+
6. You can update data multiple times - keep refining as you explore
|
|
3131
|
+
7. The CLI shows your progress in real-time as you update the output`;
|
|
3132
|
+
};
|
|
3133
|
+
var AgentStrategy = class {
|
|
3134
|
+
name = "agent";
|
|
3135
|
+
config;
|
|
3136
|
+
constructor(config) {
|
|
3137
|
+
this.config = config;
|
|
3138
|
+
}
|
|
3139
|
+
getEstimatedSteps() {
|
|
3140
|
+
return this.config.maxSteps ?? 50;
|
|
3141
|
+
}
|
|
3142
|
+
async run(options) {
|
|
3143
|
+
const debug = options.debug ?? this.config.debug;
|
|
3144
|
+
const { telemetry } = options;
|
|
3145
|
+
const maxSteps = this.config.maxSteps ?? 50;
|
|
3146
|
+
const agentSpan = telemetry?.startSpan({
|
|
3147
|
+
name: "strategy.agent",
|
|
3148
|
+
kind: "AGENT",
|
|
3149
|
+
attributes: {
|
|
3150
|
+
"strategy.name": this.name,
|
|
3151
|
+
"agent.max_steps": maxSteps,
|
|
3152
|
+
"agent.model": this.config.model ? "custom" : `${this.config.provider}/${this.config.modelId}`,
|
|
3153
|
+
"agent.artifacts.count": options.artifacts.length
|
|
3154
|
+
}
|
|
3155
|
+
});
|
|
3156
|
+
const activeMessageSpans = /* @__PURE__ */ new Map();
|
|
3157
|
+
const activeToolSpans = /* @__PURE__ */ new Map();
|
|
3158
|
+
await options.events?.onStep?.({
|
|
3159
|
+
step: 1,
|
|
3160
|
+
total: this.getEstimatedSteps(),
|
|
3161
|
+
label: "agent_explore"
|
|
3162
|
+
});
|
|
3163
|
+
debug?.step({
|
|
3164
|
+
step: 1,
|
|
3165
|
+
total: this.getEstimatedSteps(),
|
|
3166
|
+
label: "agent_explore",
|
|
3167
|
+
strategy: this.name
|
|
3168
|
+
});
|
|
3169
|
+
const filesystem = createVirtualFilesystem(options.artifacts);
|
|
3170
|
+
const files = {
|
|
3171
|
+
"/artifact.json": filesystem["/artifact.json"],
|
|
3172
|
+
"/manifest.json": filesystem["/manifest.json"]
|
|
3173
|
+
};
|
|
3174
|
+
for (const [path, content] of filesystem.virtualFiles) {
|
|
3175
|
+
files[path] = content;
|
|
3176
|
+
}
|
|
3177
|
+
const bash = new Bash2({
|
|
3178
|
+
files,
|
|
3179
|
+
cwd: "/"
|
|
3180
|
+
});
|
|
3181
|
+
const virtualTools = createVirtualFilesystemTools(bash, filesystem.getImageByPath);
|
|
3182
|
+
const schema = JSON.stringify(options.schema, null, 2);
|
|
3183
|
+
const systemPrompt = this.config.systemPrompt ?? defaultSystemPrompt(schema, this.config.outputInstructions);
|
|
3184
|
+
const callId = `agent_${Date.now()}_${Math.random().toString(36).slice(2, 11)}`;
|
|
3185
|
+
debug?.llmCallStart({
|
|
3186
|
+
callId,
|
|
3187
|
+
model: this.config.model ? JSON.stringify(this.config.model) : "default",
|
|
3188
|
+
schemaName: "extract",
|
|
3189
|
+
systemLength: systemPrompt.length,
|
|
3190
|
+
userLength: 0,
|
|
3191
|
+
artifactCount: options.artifacts.length
|
|
3192
|
+
});
|
|
3193
|
+
debug?.promptSystem({ callId, system: systemPrompt });
|
|
3194
|
+
const startTime = Date.now();
|
|
3195
|
+
const agentDir = this.config.agentDir;
|
|
3196
|
+
const authStorage = agentDir ? AuthStorage.create(`${agentDir}/auth.json`) : AuthStorage.create();
|
|
3197
|
+
if (this.config.apiKey && this.config.provider) {
|
|
3198
|
+
authStorage.setRuntimeApiKey(this.config.provider, this.config.apiKey);
|
|
3199
|
+
}
|
|
3200
|
+
const modelRegistry = new ModelRegistry(authStorage);
|
|
3201
|
+
let model = this.config.model;
|
|
3202
|
+
if (!model && this.config.provider && this.config.modelId) {
|
|
3203
|
+
if (this.config.verbose) {
|
|
3204
|
+
console.error(`[AgentStrategy] Looking up model: ${this.config.provider}/${this.config.modelId}`);
|
|
3205
|
+
}
|
|
3206
|
+
model = modelRegistry.find(this.config.provider, this.config.modelId);
|
|
3207
|
+
if (this.config.verbose) {
|
|
3208
|
+
console.error(`[AgentStrategy] Model resolved: ${model ? "success" : "failed"}`);
|
|
3209
|
+
if (model) {
|
|
3210
|
+
console.error(`[AgentStrategy] Model info:`, JSON.stringify(model).slice(0, 200));
|
|
3211
|
+
}
|
|
3212
|
+
}
|
|
3213
|
+
} else if (model && this.config.verbose) {
|
|
3214
|
+
console.error(`[AgentStrategy] Using pre-configured model`);
|
|
3215
|
+
}
|
|
3216
|
+
const loader = new DefaultResourceLoader({
|
|
3217
|
+
cwd: "/artifacts",
|
|
3218
|
+
agentDir: agentDir || void 0,
|
|
3219
|
+
systemPromptOverride: () => systemPrompt
|
|
3220
|
+
});
|
|
3221
|
+
await loader.reload();
|
|
3222
|
+
const settingsManager = SettingsManager.inMemory({
|
|
3223
|
+
compaction: { enabled: false }
|
|
3224
|
+
// Disable compaction for extraction tasks
|
|
3225
|
+
});
|
|
3226
|
+
let currentOutput = null;
|
|
3227
|
+
let isFinished = false;
|
|
3228
|
+
let finishError = null;
|
|
3229
|
+
let extractionFailed = false;
|
|
3230
|
+
let failureReason = null;
|
|
3231
|
+
const validateData = (data) => {
|
|
3232
|
+
try {
|
|
3233
|
+
JSON.stringify(data);
|
|
3234
|
+
return { valid: true, errors: [] };
|
|
3235
|
+
} catch (e) {
|
|
3236
|
+
return { valid: false, errors: [e.message] };
|
|
3237
|
+
}
|
|
3238
|
+
};
|
|
3239
|
+
const deepMerge = (target, source) => {
|
|
3240
|
+
const output = Object.assign({}, target);
|
|
3241
|
+
if (isObject(target) && isObject(source)) {
|
|
3242
|
+
Object.keys(source).forEach((key) => {
|
|
3243
|
+
if (isObject(source[key])) {
|
|
3244
|
+
if (!(key in target)) {
|
|
3245
|
+
Object.assign(output, { [key]: source[key] });
|
|
3246
|
+
} else {
|
|
3247
|
+
output[key] = deepMerge(target[key], source[key]);
|
|
3248
|
+
}
|
|
3249
|
+
} else {
|
|
3250
|
+
Object.assign(output, { [key]: source[key] });
|
|
3251
|
+
}
|
|
3252
|
+
});
|
|
3253
|
+
}
|
|
3254
|
+
return output;
|
|
3255
|
+
};
|
|
3256
|
+
const isObject = (item) => {
|
|
3257
|
+
return item && typeof item === "object" && !Array.isArray(item);
|
|
3258
|
+
};
|
|
3259
|
+
await options.events?.onStep?.({
|
|
3260
|
+
step: 2,
|
|
3261
|
+
total: this.getEstimatedSteps(),
|
|
3262
|
+
label: "agent_init"
|
|
3263
|
+
});
|
|
3264
|
+
debug?.step({
|
|
3265
|
+
step: 2,
|
|
3266
|
+
total: this.getEstimatedSteps(),
|
|
3267
|
+
label: "agent_init",
|
|
3268
|
+
strategy: this.name
|
|
3269
|
+
});
|
|
3270
|
+
const { Type: Type2 } = await import("@sinclair/typebox");
|
|
3271
|
+
const setOutputDataTool = {
|
|
3272
|
+
name: "set_output_data",
|
|
3273
|
+
label: "Set Output Data",
|
|
3274
|
+
description: "Set the initial output data. You can use any structure - it will be validated against the schema.",
|
|
3275
|
+
parameters: Type2.Object({
|
|
3276
|
+
data: Type2.Any({ description: "The output data to set" })
|
|
3277
|
+
}),
|
|
3278
|
+
execute: async (toolCallId, params) => {
|
|
3279
|
+
currentOutput = params.data;
|
|
3280
|
+
const validation = validateData(params.data);
|
|
3281
|
+
const status = validation.valid ? "\u2713 Valid structure" : `\u2717 Validation issues: ${validation.errors.join(", ")}`;
|
|
3282
|
+
await options.events?.onStep?.({
|
|
3283
|
+
step: stepCount + 1,
|
|
3284
|
+
total: this.getEstimatedSteps(),
|
|
3285
|
+
label: `Output: ${JSON.stringify(params.data).slice(0, 50)}...`
|
|
3286
|
+
});
|
|
3287
|
+
return {
|
|
3288
|
+
content: [{ type: "text", text: `Output data set. ${status}` }],
|
|
3289
|
+
details: { validation }
|
|
3290
|
+
};
|
|
3291
|
+
}
|
|
3292
|
+
};
|
|
3293
|
+
const updateOutputDataTool = {
|
|
3294
|
+
name: "update_output_data",
|
|
3295
|
+
label: "Update Output Data",
|
|
3296
|
+
description: "Update the output data by merging changes. Existing fields are preserved, new fields are added.",
|
|
3297
|
+
parameters: Type2.Object({
|
|
3298
|
+
changes: Type2.Record(Type2.String(), Type2.Any(), {
|
|
3299
|
+
description: "Changes to merge into existing data"
|
|
3300
|
+
})
|
|
3301
|
+
}),
|
|
3302
|
+
execute: async (toolCallId, params) => {
|
|
3303
|
+
if (currentOutput === null) {
|
|
3304
|
+
return {
|
|
3305
|
+
content: [{ type: "text", text: "Error: No output data set yet. Use set_output_data first." }],
|
|
3306
|
+
isError: true
|
|
3307
|
+
};
|
|
3308
|
+
}
|
|
3309
|
+
currentOutput = deepMerge(currentOutput, params.changes);
|
|
3310
|
+
const validation = validateData(currentOutput);
|
|
3311
|
+
const status = validation.valid ? "\u2713 Valid structure" : `\u2717 Validation issues: ${validation.errors.join(", ")}`;
|
|
3312
|
+
await options.events?.onStep?.({
|
|
3313
|
+
step: stepCount + 1,
|
|
3314
|
+
total: this.getEstimatedSteps(),
|
|
3315
|
+
label: `Updated: ${JSON.stringify(params.changes).slice(0, 50)}...`
|
|
3316
|
+
});
|
|
3317
|
+
return {
|
|
3318
|
+
content: [{ type: "text", text: `Output data updated. ${status}` }],
|
|
3319
|
+
details: { validation, currentOutput }
|
|
3320
|
+
};
|
|
3321
|
+
}
|
|
3322
|
+
};
|
|
3323
|
+
const finishTool = {
|
|
3324
|
+
name: "finish",
|
|
3325
|
+
label: "Finish Extraction",
|
|
3326
|
+
description: "Complete the extraction. Can only be called when data validates against the schema.",
|
|
3327
|
+
parameters: Type2.Object({}),
|
|
3328
|
+
execute: async (toolCallId) => {
|
|
3329
|
+
if (extractionFailed) {
|
|
3330
|
+
return {
|
|
3331
|
+
content: [{ type: "text", text: "Cannot finish - extraction was marked as failed." }],
|
|
3332
|
+
isError: true
|
|
3333
|
+
};
|
|
3334
|
+
}
|
|
3335
|
+
if (currentOutput === null) {
|
|
3336
|
+
return {
|
|
3337
|
+
content: [{ type: "text", text: "Error: No output data set. Extract data first." }],
|
|
3338
|
+
isError: true
|
|
3339
|
+
};
|
|
3340
|
+
}
|
|
3341
|
+
const validation = validateData(currentOutput);
|
|
3342
|
+
if (!validation.valid) {
|
|
3343
|
+
finishError = `Schema validation failed: ${validation.errors.join(", ")}`;
|
|
3344
|
+
return {
|
|
3345
|
+
content: [{
|
|
3346
|
+
type: "text",
|
|
3347
|
+
text: `Cannot finish: ${finishError}
|
|
3348
|
+
|
|
3349
|
+
Fix the data and try again, or use fail() if extraction is impossible.`
|
|
3350
|
+
}],
|
|
3351
|
+
isError: true
|
|
3352
|
+
};
|
|
3353
|
+
}
|
|
3354
|
+
isFinished = true;
|
|
3355
|
+
return {
|
|
3356
|
+
content: [{ type: "text", text: "\u2713 Extraction complete! Data validated successfully." }]
|
|
3357
|
+
};
|
|
3358
|
+
}
|
|
3359
|
+
};
|
|
3360
|
+
const failTool = {
|
|
3361
|
+
name: "fail",
|
|
3362
|
+
label: "Fail Extraction",
|
|
3363
|
+
description: "Mark extraction as failed when the schema cannot be satisfied with the available data.",
|
|
3364
|
+
parameters: Type2.Object({
|
|
3365
|
+
reason: Type2.String({ description: "Why extraction failed or what data was missing" })
|
|
3366
|
+
}),
|
|
3367
|
+
execute: async (toolCallId, params) => {
|
|
3368
|
+
extractionFailed = true;
|
|
3369
|
+
failureReason = params.reason;
|
|
3370
|
+
return {
|
|
3371
|
+
content: [{ type: "text", text: `Extraction marked as failed: ${params.reason}` }]
|
|
3372
|
+
};
|
|
3373
|
+
}
|
|
3374
|
+
};
|
|
3375
|
+
const allTools = [
|
|
3376
|
+
virtualTools.read,
|
|
3377
|
+
virtualTools.bash,
|
|
3378
|
+
virtualTools.grep,
|
|
3379
|
+
virtualTools.find,
|
|
3380
|
+
virtualTools.ls,
|
|
3381
|
+
setOutputDataTool,
|
|
3382
|
+
updateOutputDataTool,
|
|
3383
|
+
finishTool,
|
|
3384
|
+
failTool
|
|
3385
|
+
];
|
|
3386
|
+
if (this.config.verbose) {
|
|
3387
|
+
console.error(`[AgentStrategy] Creating session with ${allTools.length} tools`);
|
|
3388
|
+
console.error(`[AgentStrategy] Tool names: ${allTools.map((t) => t.name).join(", ")}`);
|
|
3389
|
+
allTools.forEach((tool) => {
|
|
3390
|
+
console.error(`[AgentStrategy] Tool "${tool.name}" details:`);
|
|
3391
|
+
console.error(` - label: ${tool.label}`);
|
|
3392
|
+
console.error(` - description: ${tool.description?.slice(0, 100)}...`);
|
|
3393
|
+
console.error(` - has parameters: ${!!tool.parameters}`);
|
|
3394
|
+
if (tool.parameters) {
|
|
3395
|
+
console.error(` - parameters type: ${tool.parameters.type}`);
|
|
3396
|
+
console.error(` - required fields: ${JSON.stringify(tool.parameters.required)}`);
|
|
3397
|
+
}
|
|
3398
|
+
});
|
|
3399
|
+
}
|
|
3400
|
+
const { session } = await createAgentSession({
|
|
3401
|
+
model,
|
|
3402
|
+
authStorage,
|
|
3403
|
+
modelRegistry,
|
|
3404
|
+
resourceLoader: loader,
|
|
3405
|
+
sessionManager: SessionManager.inMemory(),
|
|
3406
|
+
settingsManager,
|
|
3407
|
+
tools: [],
|
|
3408
|
+
// No default tools
|
|
3409
|
+
customTools: allTools
|
|
3410
|
+
});
|
|
3411
|
+
if (this.config.verbose) {
|
|
3412
|
+
console.error(`[AgentStrategy] Session created successfully`);
|
|
3413
|
+
}
|
|
3414
|
+
await options.events?.onStep?.({
|
|
3415
|
+
step: 3,
|
|
3416
|
+
total: this.getEstimatedSteps(),
|
|
3417
|
+
label: "agent_session_ready"
|
|
3418
|
+
});
|
|
3419
|
+
debug?.step({
|
|
3420
|
+
step: 3,
|
|
3421
|
+
total: this.getEstimatedSteps(),
|
|
3422
|
+
label: "agent_session_ready",
|
|
3423
|
+
strategy: this.name
|
|
3424
|
+
});
|
|
3425
|
+
let usage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
|
|
3426
|
+
let stepCount = 0;
|
|
3427
|
+
let finalResponse = "";
|
|
3428
|
+
const maxToolCalls = maxSteps;
|
|
3429
|
+
let textBuffer = "";
|
|
3430
|
+
try {
|
|
3431
|
+
const unsubscribe = session.subscribe((event) => {
|
|
3432
|
+
try {
|
|
3433
|
+
switch (event.type) {
|
|
3434
|
+
case "message_update": {
|
|
3435
|
+
if (event.assistantMessageEvent.type === "text_delta") {
|
|
3436
|
+
const delta = event.assistantMessageEvent.delta;
|
|
3437
|
+
finalResponse += delta;
|
|
3438
|
+
options.events?.onAgentMessage?.({
|
|
3439
|
+
content: delta,
|
|
3440
|
+
role: "assistant"
|
|
3441
|
+
});
|
|
3442
|
+
textBuffer += delta;
|
|
3443
|
+
let newlineIndex;
|
|
3444
|
+
while ((newlineIndex = textBuffer.indexOf("\n")) !== -1) {
|
|
3445
|
+
const line = textBuffer.slice(0, newlineIndex).trim();
|
|
3446
|
+
textBuffer = textBuffer.slice(newlineIndex + 1);
|
|
3447
|
+
if (line.length > 0) {
|
|
3448
|
+
options.events?.onStep?.({
|
|
3449
|
+
step: stepCount,
|
|
3450
|
+
total: this.getEstimatedSteps(),
|
|
3451
|
+
label: `\u2192 ${line.slice(0, 120)}`
|
|
3452
|
+
});
|
|
3453
|
+
}
|
|
3454
|
+
}
|
|
3455
|
+
if (textBuffer.length > 100) {
|
|
3456
|
+
const line = textBuffer.trim();
|
|
3457
|
+
if (line.length > 0) {
|
|
3458
|
+
options.events?.onStep?.({
|
|
3459
|
+
step: stepCount,
|
|
3460
|
+
total: this.getEstimatedSteps(),
|
|
3461
|
+
label: `\u2192 ${line.slice(0, 120)}`
|
|
3462
|
+
});
|
|
3463
|
+
}
|
|
3464
|
+
textBuffer = "";
|
|
3465
|
+
}
|
|
3466
|
+
}
|
|
3467
|
+
break;
|
|
3468
|
+
}
|
|
3469
|
+
case "tool_execution_start": {
|
|
3470
|
+
stepCount++;
|
|
3471
|
+
if (telemetry && agentSpan) {
|
|
3472
|
+
const toolSpan = telemetry.startSpan({
|
|
3473
|
+
name: `agent.tool.${event.toolName}`,
|
|
3474
|
+
kind: "TOOL",
|
|
3475
|
+
parentSpan: agentSpan,
|
|
3476
|
+
attributes: {
|
|
3477
|
+
"tool.name": event.toolName,
|
|
3478
|
+
"tool.call_id": event.toolCallId,
|
|
3479
|
+
"tool.args": JSON.stringify(event.args || {})
|
|
3480
|
+
}
|
|
3481
|
+
});
|
|
3482
|
+
activeToolSpans.set(event.toolCallId, toolSpan);
|
|
3483
|
+
}
|
|
3484
|
+
let label;
|
|
3485
|
+
let detail;
|
|
3486
|
+
const args = event.args;
|
|
3487
|
+
const toolName = event.toolName;
|
|
3488
|
+
const toolStartEvent = {
|
|
3489
|
+
toolName,
|
|
3490
|
+
toolCallId: event.toolCallId,
|
|
3491
|
+
args: args || {}
|
|
3492
|
+
};
|
|
3493
|
+
options.events?.onAgentToolStart?.(toolStartEvent);
|
|
3494
|
+
if (toolName === "read" && args?.file_path) {
|
|
3495
|
+
const fileName = args.file_path.split("/").pop() || args.file_path;
|
|
3496
|
+
const pagination = [];
|
|
3497
|
+
if (args.offset && args.offset > 1) {
|
|
3498
|
+
pagination.push(`offset ${args.offset}`);
|
|
3499
|
+
}
|
|
3500
|
+
if (args.limit) {
|
|
3501
|
+
pagination.push(`limit ${args.limit}`);
|
|
3502
|
+
}
|
|
3503
|
+
const paginationStr = pagination.length > 0 ? ` (${pagination.join(", ")})` : "";
|
|
3504
|
+
label = `Read ${fileName}${paginationStr}`;
|
|
3505
|
+
detail = "";
|
|
3506
|
+
} else if (toolName === "bash" && args?.command) {
|
|
3507
|
+
const cmd = args.command.length > 40 ? args.command.slice(0, 37) + "..." : args.command;
|
|
3508
|
+
label = `Bash: ${cmd}`;
|
|
3509
|
+
detail = "";
|
|
3510
|
+
} else if (toolName === "grep" && args?.pattern) {
|
|
3511
|
+
label = `Grep "${args.pattern}"`;
|
|
3512
|
+
detail = args.path ? `in ${args.path.split("/").pop()}` : "";
|
|
3513
|
+
} else if (toolName === "find" && args?.path) {
|
|
3514
|
+
label = `Find`;
|
|
3515
|
+
detail = args.name ? `"${args.name}" in ${args.path}` : `in ${args.path}`;
|
|
3516
|
+
} else if (toolName === "ls" && args?.path) {
|
|
3517
|
+
label = `List ${args.path}`;
|
|
3518
|
+
detail = args.recursive ? "recursive" : "";
|
|
3519
|
+
} else if (toolName === "set_output_data") {
|
|
3520
|
+
label = "Set Output";
|
|
3521
|
+
detail = args?.data ? JSON.stringify(args.data).slice(0, 80) : "";
|
|
3522
|
+
} else if (toolName === "update_output_data") {
|
|
3523
|
+
label = "Update Output";
|
|
3524
|
+
detail = args?.changes ? JSON.stringify(args.changes).slice(0, 80) : "";
|
|
3525
|
+
} else if (toolName === "finish") {
|
|
3526
|
+
label = "Finish";
|
|
3527
|
+
detail = "";
|
|
3528
|
+
} else if (toolName === "fail") {
|
|
3529
|
+
label = "Fail";
|
|
3530
|
+
detail = args?.reason || "";
|
|
3531
|
+
} else {
|
|
3532
|
+
label = toolName;
|
|
3533
|
+
detail = args ? JSON.stringify(args).slice(0, 100) : "";
|
|
3534
|
+
}
|
|
3535
|
+
options.events?.onStep?.({
|
|
3536
|
+
step: stepCount + 1,
|
|
3537
|
+
total: this.getEstimatedSteps(),
|
|
3538
|
+
label,
|
|
3539
|
+
detail
|
|
3540
|
+
});
|
|
3541
|
+
debug?.step({
|
|
3542
|
+
step: stepCount + 1,
|
|
3543
|
+
total: this.getEstimatedSteps(),
|
|
3544
|
+
label,
|
|
3545
|
+
strategy: this.name
|
|
3546
|
+
});
|
|
3547
|
+
break;
|
|
3548
|
+
}
|
|
3549
|
+
case "tool_execution_end": {
|
|
3550
|
+
const toolEndEvent = event;
|
|
3551
|
+
const toolSpan = activeToolSpans.get(toolEndEvent.toolCallId);
|
|
3552
|
+
if (toolSpan && telemetry) {
|
|
3553
|
+
const hasError = toolEndEvent.isError || toolEndEvent.error;
|
|
3554
|
+
telemetry.endSpan(toolSpan, {
|
|
3555
|
+
status: hasError ? "error" : "ok",
|
|
3556
|
+
error: hasError ? new Error(toolEndEvent.error || "Tool execution failed") : void 0,
|
|
3557
|
+
output: toolEndEvent.result
|
|
3558
|
+
});
|
|
3559
|
+
activeToolSpans.delete(toolEndEvent.toolCallId);
|
|
3560
|
+
}
|
|
3561
|
+
options.events?.onAgentToolEnd?.({
|
|
3562
|
+
toolCallId: toolEndEvent.toolCallId,
|
|
3563
|
+
result: toolEndEvent.result,
|
|
3564
|
+
error: toolEndEvent.error || toolEndEvent.isError ? toolEndEvent.error || "Tool execution failed" : void 0
|
|
3565
|
+
});
|
|
3566
|
+
if (toolEndEvent.isError || toolEndEvent.error) {
|
|
3567
|
+
const errorMsg = toolEndEvent.error || "Unknown tool error";
|
|
3568
|
+
const toolName = toolEndEvent.toolName || "unknown";
|
|
3569
|
+
const toolCallId = toolEndEvent.toolCallId || "unknown";
|
|
3570
|
+
if (this.config.verbose) {
|
|
3571
|
+
console.error(`[AgentStrategy] Tool execution failed: ${errorMsg}`);
|
|
3572
|
+
console.error(`[AgentStrategy] Tool: ${toolName}, Call ID: ${toolCallId}`);
|
|
3573
|
+
if (toolEndEvent.result) {
|
|
3574
|
+
console.error(`[AgentStrategy] Result:`, JSON.stringify(toolEndEvent.result));
|
|
3575
|
+
}
|
|
3576
|
+
}
|
|
3577
|
+
}
|
|
3578
|
+
break;
|
|
3579
|
+
}
|
|
3580
|
+
case "agent_end": {
|
|
3581
|
+
if (textBuffer.trim().length > 0) {
|
|
3582
|
+
options.events?.onStep?.({
|
|
3583
|
+
step: stepCount,
|
|
3584
|
+
total: this.getEstimatedSteps(),
|
|
3585
|
+
label: `\u2192 ${textBuffer.trim().slice(0, 120)}`
|
|
3586
|
+
});
|
|
3587
|
+
textBuffer = "";
|
|
3588
|
+
}
|
|
3589
|
+
if (event.messages && event.messages.length > 0) {
|
|
3590
|
+
const inputTokens = event.messages.reduce((sum, msg) => {
|
|
3591
|
+
if (msg.role === "user") {
|
|
3592
|
+
return sum + Math.ceil(JSON.stringify(msg.content).length / 4);
|
|
3593
|
+
}
|
|
3594
|
+
return sum;
|
|
3595
|
+
}, 0);
|
|
3596
|
+
const outputTokens = event.messages.reduce((sum, msg) => {
|
|
3597
|
+
if (msg.role === "assistant") {
|
|
3598
|
+
return sum + Math.ceil(JSON.stringify(msg.content).length / 4);
|
|
3599
|
+
}
|
|
3600
|
+
return sum;
|
|
3601
|
+
}, 0);
|
|
3602
|
+
usage = {
|
|
3603
|
+
inputTokens,
|
|
3604
|
+
outputTokens,
|
|
3605
|
+
totalTokens: inputTokens + outputTokens
|
|
3606
|
+
};
|
|
3607
|
+
}
|
|
3608
|
+
options.events?.onStep?.({
|
|
3609
|
+
step: stepCount + 1,
|
|
3610
|
+
total: this.getEstimatedSteps(),
|
|
3611
|
+
label: "agent_complete"
|
|
3612
|
+
});
|
|
3613
|
+
debug?.step({
|
|
3614
|
+
step: stepCount + 1,
|
|
3615
|
+
total: this.getEstimatedSteps(),
|
|
3616
|
+
label: "agent_complete",
|
|
3617
|
+
strategy: this.name
|
|
3618
|
+
});
|
|
3619
|
+
break;
|
|
3620
|
+
}
|
|
3621
|
+
case "message_start": {
|
|
3622
|
+
const messageKey = `msg_${Date.now()}_${Math.random().toString(36).slice(2, 11)}`;
|
|
3623
|
+
if (telemetry && agentSpan) {
|
|
3624
|
+
const llmSpan = telemetry.startSpan({
|
|
3625
|
+
name: "agent.llm.generate",
|
|
3626
|
+
kind: "LLM",
|
|
3627
|
+
parentSpan: agentSpan,
|
|
3628
|
+
attributes: {
|
|
3629
|
+
"llm.message_type": event.message?.role || "unknown",
|
|
3630
|
+
"llm.type": "agent_message"
|
|
3631
|
+
}
|
|
3632
|
+
});
|
|
3633
|
+
activeMessageSpans.set(messageKey, llmSpan);
|
|
3634
|
+
event._telemetryKey = messageKey;
|
|
3635
|
+
}
|
|
3636
|
+
break;
|
|
3637
|
+
}
|
|
3638
|
+
case "message_end": {
|
|
3639
|
+
const messageKey = event._telemetryKey;
|
|
3640
|
+
if (messageKey) {
|
|
3641
|
+
const llmSpan = activeMessageSpans.get(messageKey);
|
|
3642
|
+
if (llmSpan && telemetry) {
|
|
3643
|
+
telemetry.endSpan(llmSpan, {
|
|
3644
|
+
status: "ok",
|
|
3645
|
+
output: finalResponse.slice(-200)
|
|
3646
|
+
// Last 200 chars as output preview
|
|
3647
|
+
});
|
|
3648
|
+
activeMessageSpans.delete(messageKey);
|
|
3649
|
+
}
|
|
3650
|
+
}
|
|
3651
|
+
break;
|
|
3652
|
+
}
|
|
3653
|
+
case "agent_start": {
|
|
3654
|
+
if (this.config.verbose) {
|
|
3655
|
+
console.error("[AgentStrategy] Agent started processing");
|
|
3656
|
+
}
|
|
3657
|
+
break;
|
|
3658
|
+
}
|
|
3659
|
+
case "turn_start":
|
|
3660
|
+
case "turn_end": {
|
|
3661
|
+
break;
|
|
3662
|
+
}
|
|
3663
|
+
case "auto_compaction_start":
|
|
3664
|
+
case "auto_compaction_end":
|
|
3665
|
+
case "auto_retry_start":
|
|
3666
|
+
case "auto_retry_end": {
|
|
3667
|
+
break;
|
|
3668
|
+
}
|
|
3669
|
+
default: {
|
|
3670
|
+
const unhandledEvent = event;
|
|
3671
|
+
if (unhandledEvent.type && !unhandledEvent.type.includes("_")) {
|
|
3672
|
+
if (this.config.verbose) {
|
|
3673
|
+
console.error(`[AgentStrategy] Unexpected event type: ${unhandledEvent.type}`);
|
|
3674
|
+
}
|
|
3675
|
+
}
|
|
3676
|
+
break;
|
|
3677
|
+
}
|
|
3678
|
+
}
|
|
3679
|
+
} catch (eventHandlerError) {
|
|
3680
|
+
if (this.config.verbose) {
|
|
3681
|
+
console.error(`[AgentStrategy] Error in event handler: ${eventHandlerError.message}`);
|
|
3682
|
+
console.error("AgentStrategy event handler error:", eventHandlerError);
|
|
3683
|
+
}
|
|
3684
|
+
throw eventHandlerError;
|
|
3685
|
+
}
|
|
3686
|
+
});
|
|
3687
|
+
await session.prompt(
|
|
3688
|
+
"Begin exploring the artifacts and extract the required data according to the schema. Start by reading the manifest file.",
|
|
3689
|
+
{
|
|
3690
|
+
// Ensure the agent keeps running until it calls finish() or fail()
|
|
3691
|
+
}
|
|
3692
|
+
);
|
|
3693
|
+
if (currentOutput === null && !extractionFailed && !isFinished) {
|
|
3694
|
+
if (this.config.verbose) {
|
|
3695
|
+
console.error("[AgentStrategy] No output after first run. Sending retry prompt...");
|
|
3696
|
+
}
|
|
3697
|
+
await options.events?.onStep?.({
|
|
3698
|
+
step: stepCount + 1,
|
|
3699
|
+
total: this.getEstimatedSteps(),
|
|
3700
|
+
label: "Retry: forcing output extraction"
|
|
3701
|
+
});
|
|
3702
|
+
const retryUnsubscribe = session.subscribe((event) => {
|
|
3703
|
+
try {
|
|
3704
|
+
switch (event.type) {
|
|
3705
|
+
case "message_update": {
|
|
3706
|
+
if (event.assistantMessageEvent.type === "text_delta") {
|
|
3707
|
+
const delta = event.assistantMessageEvent.delta;
|
|
3708
|
+
finalResponse += delta;
|
|
3709
|
+
options.events?.onAgentMessage?.({
|
|
3710
|
+
content: delta,
|
|
3711
|
+
role: "assistant"
|
|
3712
|
+
});
|
|
3713
|
+
textBuffer += delta;
|
|
3714
|
+
let newlineIndex;
|
|
3715
|
+
while ((newlineIndex = textBuffer.indexOf("\n")) !== -1) {
|
|
3716
|
+
const line = textBuffer.slice(0, newlineIndex).trim();
|
|
3717
|
+
textBuffer = textBuffer.slice(newlineIndex + 1);
|
|
3718
|
+
if (line.length > 0) {
|
|
3719
|
+
options.events?.onStep?.({
|
|
3720
|
+
step: stepCount,
|
|
3721
|
+
total: this.getEstimatedSteps(),
|
|
3722
|
+
label: `\u2192 ${line.slice(0, 120)}`
|
|
3723
|
+
});
|
|
3724
|
+
}
|
|
3725
|
+
}
|
|
3726
|
+
if (textBuffer.length > 100) {
|
|
3727
|
+
const line = textBuffer.trim();
|
|
3728
|
+
if (line.length > 0) {
|
|
3729
|
+
options.events?.onStep?.({
|
|
3730
|
+
step: stepCount,
|
|
3731
|
+
total: this.getEstimatedSteps(),
|
|
3732
|
+
label: `\u2192 ${line.slice(0, 120)}`
|
|
3733
|
+
});
|
|
3734
|
+
}
|
|
3735
|
+
textBuffer = "";
|
|
3736
|
+
}
|
|
3737
|
+
}
|
|
3738
|
+
break;
|
|
3739
|
+
}
|
|
3740
|
+
case "tool_execution_start": {
|
|
3741
|
+
stepCount++;
|
|
3742
|
+
const toolName = event.toolName;
|
|
3743
|
+
const args = event.args;
|
|
3744
|
+
let label = toolName;
|
|
3745
|
+
if (toolName === "set_output_data") {
|
|
3746
|
+
label = "Set Output (retry)";
|
|
3747
|
+
} else if (toolName === "update_output_data") {
|
|
3748
|
+
label = "Update Output (retry)";
|
|
3749
|
+
} else if (toolName === "finish") {
|
|
3750
|
+
label = "Finish (retry)";
|
|
3751
|
+
} else if (toolName === "fail") {
|
|
3752
|
+
label = "Fail (retry)";
|
|
3753
|
+
}
|
|
3754
|
+
options.events?.onStep?.({
|
|
3755
|
+
step: stepCount + 1,
|
|
3756
|
+
total: this.getEstimatedSteps(),
|
|
3757
|
+
label
|
|
3758
|
+
});
|
|
3759
|
+
break;
|
|
3760
|
+
}
|
|
3761
|
+
}
|
|
3762
|
+
} catch (eventHandlerError) {
|
|
3763
|
+
if (this.config.verbose) {
|
|
3764
|
+
console.error(`[AgentStrategy] Error in retry event handler: ${eventHandlerError.message}`);
|
|
3765
|
+
}
|
|
3766
|
+
}
|
|
3767
|
+
});
|
|
3768
|
+
await session.prompt(
|
|
3769
|
+
`You have explored the artifacts but haven't called any output tools yet. You MUST now extract data and call either:
|
|
3770
|
+
1. set_output_data with the extracted data, then finish()
|
|
3771
|
+
2. fail() if the document doesn't contain the required information
|
|
3772
|
+
|
|
3773
|
+
The schema requires: ${JSON.stringify(options.schema).slice(0, 200)}...
|
|
3774
|
+
|
|
3775
|
+
Extract what you can from the artifacts and set the output data NOW.`,
|
|
3776
|
+
{}
|
|
3777
|
+
);
|
|
3778
|
+
retryUnsubscribe();
|
|
3779
|
+
}
|
|
3780
|
+
unsubscribe();
|
|
3781
|
+
const durationMs = Date.now() - startTime;
|
|
3782
|
+
let extractedData;
|
|
3783
|
+
if (extractionFailed) {
|
|
3784
|
+
throw new Error(`Extraction failed: ${failureReason}`);
|
|
3785
|
+
}
|
|
3786
|
+
if (!isFinished) {
|
|
3787
|
+
if (currentOutput !== null) {
|
|
3788
|
+
if (this.config.verbose) {
|
|
3789
|
+
console.error("[AgentStrategy] Warning: Agent did not call finish(). Using collected data.");
|
|
3790
|
+
}
|
|
3791
|
+
const validation = validateData(currentOutput);
|
|
3792
|
+
if (!validation.valid && this.config.verbose) {
|
|
3793
|
+
console.error(`[AgentStrategy] Data validation issues: ${validation.errors.join(", ")}`);
|
|
3794
|
+
}
|
|
3795
|
+
extractedData = currentOutput;
|
|
3796
|
+
} else {
|
|
3797
|
+
const toolCallsMade = stepCount > 0;
|
|
3798
|
+
const toolsFailed = toolCallsMade && currentOutput === null;
|
|
3799
|
+
if (toolsFailed) {
|
|
3800
|
+
const errorMsg = `Agent did not produce any output data. The model may not support tool calling properly.
|
|
3801
|
+
|
|
3802
|
+
Troubleshooting:
|
|
3803
|
+
1. Check if your model supports function calling/tool use
|
|
3804
|
+
2. Try a different model like anthropic/claude-sonnet-4 or openai/gpt-4o
|
|
3805
|
+
3. See MODEL_COMPATIBILITY.md for supported models
|
|
3806
|
+
|
|
3807
|
+
If you continue to see "Tool execution failed" errors with empty tool names,
|
|
3808
|
+
the model is not compatible with the agent strategy. Use --strategy simple instead.`;
|
|
3809
|
+
throw new Error(errorMsg);
|
|
3810
|
+
} else {
|
|
3811
|
+
const errorMsg = `Agent did not produce any output data. No data was extracted.
|
|
3812
|
+
|
|
3813
|
+
This can happen when:
|
|
3814
|
+
- The model doesn't support tool calling properly
|
|
3815
|
+
- The agent got confused and didn't use the output tools
|
|
3816
|
+
- The document doesn't contain extractable data
|
|
3817
|
+
|
|
3818
|
+
Suggestions:
|
|
3819
|
+
1. Try a different model with better tool support (anthropic/claude-sonnet-4)
|
|
3820
|
+
2. Use --strategy simple for models without tool calling
|
|
3821
|
+
3. Check if the document actually contains the data specified in your schema
|
|
3822
|
+
|
|
3823
|
+
Retry was attempted but the agent still didn't produce output.`;
|
|
3824
|
+
throw new Error(errorMsg);
|
|
3825
|
+
}
|
|
3826
|
+
}
|
|
3827
|
+
} else {
|
|
3828
|
+
if (currentOutput === null) {
|
|
3829
|
+
throw new Error("Agent called finish() but no output data was set.");
|
|
3830
|
+
}
|
|
3831
|
+
extractedData = currentOutput;
|
|
3832
|
+
}
|
|
3833
|
+
debug?.rawResponse({ callId, response: extractedData });
|
|
3834
|
+
debug?.llmCallComplete({
|
|
3835
|
+
callId,
|
|
3836
|
+
success: true,
|
|
3837
|
+
inputTokens: usage.inputTokens,
|
|
3838
|
+
outputTokens: usage.outputTokens,
|
|
3839
|
+
totalTokens: usage.totalTokens,
|
|
3840
|
+
durationMs
|
|
3841
|
+
});
|
|
3842
|
+
await options.events?.onStep?.({
|
|
3843
|
+
step: this.getEstimatedSteps(),
|
|
3844
|
+
total: this.getEstimatedSteps(),
|
|
3845
|
+
label: "extract"
|
|
3846
|
+
});
|
|
3847
|
+
debug?.step({
|
|
3848
|
+
step: this.getEstimatedSteps(),
|
|
3849
|
+
total: this.getEstimatedSteps(),
|
|
3850
|
+
label: "extract",
|
|
3851
|
+
strategy: this.name
|
|
3852
|
+
});
|
|
3853
|
+
if (telemetry) {
|
|
3854
|
+
for (const [key, span] of activeMessageSpans.entries()) {
|
|
3855
|
+
telemetry.endSpan(span, { status: "ok" });
|
|
3856
|
+
}
|
|
3857
|
+
activeMessageSpans.clear();
|
|
3858
|
+
for (const [key, span] of activeToolSpans.entries()) {
|
|
3859
|
+
telemetry.endSpan(span, { status: "ok" });
|
|
3860
|
+
}
|
|
3861
|
+
activeToolSpans.clear();
|
|
3862
|
+
if (agentSpan) {
|
|
3863
|
+
telemetry.endSpan(agentSpan, {
|
|
3864
|
+
status: "ok",
|
|
3865
|
+
output: extractedData
|
|
3866
|
+
});
|
|
3867
|
+
}
|
|
3868
|
+
}
|
|
3869
|
+
return { data: extractedData, usage };
|
|
3870
|
+
} catch (error) {
|
|
3871
|
+
const durationMs = Date.now() - startTime;
|
|
3872
|
+
debug?.llmCallComplete({
|
|
3873
|
+
callId,
|
|
3874
|
+
success: false,
|
|
3875
|
+
inputTokens: usage.inputTokens,
|
|
3876
|
+
outputTokens: usage.outputTokens,
|
|
3877
|
+
totalTokens: usage.totalTokens,
|
|
3878
|
+
durationMs,
|
|
3879
|
+
error: error.message
|
|
3880
|
+
});
|
|
3881
|
+
if (telemetry) {
|
|
3882
|
+
for (const [key, span] of activeMessageSpans.entries()) {
|
|
3883
|
+
telemetry.endSpan(span, {
|
|
3884
|
+
status: "error",
|
|
3885
|
+
error: error instanceof Error ? error : new Error(String(error))
|
|
3886
|
+
});
|
|
3887
|
+
}
|
|
3888
|
+
activeMessageSpans.clear();
|
|
3889
|
+
for (const [key, span] of activeToolSpans.entries()) {
|
|
3890
|
+
telemetry.endSpan(span, {
|
|
3891
|
+
status: "error",
|
|
3892
|
+
error: error instanceof Error ? error : new Error(String(error))
|
|
3893
|
+
});
|
|
3894
|
+
}
|
|
3895
|
+
activeToolSpans.clear();
|
|
3896
|
+
if (agentSpan) {
|
|
3897
|
+
telemetry.endSpan(agentSpan, {
|
|
3898
|
+
status: "error",
|
|
3899
|
+
error: error instanceof Error ? error : new Error(String(error))
|
|
3900
|
+
});
|
|
3901
|
+
}
|
|
3902
|
+
}
|
|
3903
|
+
throw error;
|
|
3904
|
+
} finally {
|
|
3905
|
+
session.dispose();
|
|
3906
|
+
}
|
|
3907
|
+
}
|
|
3908
|
+
};
|
|
3909
|
+
var agent = (config) => {
|
|
3910
|
+
return new AgentStrategy(config);
|
|
3911
|
+
};
|
|
2419
3912
|
export {
|
|
3913
|
+
AgentStrategy,
|
|
2420
3914
|
DoublePassAutoMergeStrategy,
|
|
2421
3915
|
DoublePassStrategy,
|
|
2422
3916
|
ParallelAutoMergeStrategy,
|
|
@@ -2424,6 +3918,7 @@ export {
|
|
|
2424
3918
|
SequentialAutoMergeStrategy,
|
|
2425
3919
|
SequentialStrategy,
|
|
2426
3920
|
SimpleStrategy,
|
|
3921
|
+
agent,
|
|
2427
3922
|
doublePass,
|
|
2428
3923
|
doublePassAutoMerge,
|
|
2429
3924
|
parallel,
|