From 64a83ff4c143c57a5064b49e28b02465e13547cb Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=AE=9D=E5=84=BF?= <274762368+xcjy8bao@users.noreply.github.com> Date: Mon, 12 Oct 2026 18:57:07 +0800 Subject: [PATCH] feat: add BYO Jev and multimodal providers --- CHANGELOG.md | 7 + README.md | 30 +- README.zh-CN.md | 30 +- package.json | 2 +- plugins/sift-light/.claude-plugin/plugin.json | 2 +- plugins/sift-light/.codex-plugin/plugin.json | 2 +- plugins/sift-light/.mcp.json | 2 +- plugins/sift-light/kimi.plugin.json | 4 +- plugins/sift-light/omp-extension.mjs | 740 ++++++++++++--- plugins/sift-light/package.json | 2 +- src/config-reader.ts | 191 +++- src/config.ts | 1 + src/mcp-semantic-judge.ts | 3 + src/mcp-server.mjs | 881 +++++++++++++----- src/mcp-server.ts | 9 +- src/mcp.ts | 5 +- src/media-config.ts | 21 + src/media-search.ts | 74 +- src/native-text-config.ts | 2 +- src/remote-media-client.ts | 363 ++++++++ src/semantic-judge.ts | 323 ++++--- src/service.ts | 7 +- 22 files changed, 2143 insertions(+), 558 deletions(-) create mode 100644 src/media-config.ts create mode 100644 src/remote-media-client.ts diff --git a/CHANGELOG.md b/CHANGELOG.md index c66e50c2..29ccc373 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,12 @@ # Changelog +## 1.1.0 — 2026-10-12 + +- Add opt-in BYO HTTP JSON providers for Jev-compatible semantic judging and multimodal retrieval. +- Keep local native text and visual workers as the default paths, with local source hashes, PDF page mapping, inspection and validation retained after remote ranking. +- Bound remote request and response sizes, timeouts, retries, file identifiers and credentials; API keys remain environment references rather than configuration values. +- Add configuration, protocol, MCP boundary and end-to-end local provider regression coverage. + ## 1.1.0-1 — 2026-10-12 - Add an independent native text retrieval lane and bounded rank fusion while retaining exact source inspection, validation and both text result lanes. diff --git a/README.md b/README.md index 3ed75d2d..9f29bb8b 100644 --- a/README.md +++ b/README.md @@ -235,6 +235,34 @@ The additional `sift-light-media` tool accepts `mode: "search"`, an exact local Results retain up to ten distinct files, content hashes, original PDF page numbers, and exact `inspect`/`validate` requests for this tool. `inspect` returns actual PNG pixels at a maximum dimension of 1024; scan results render the corresponding original PDF page. Changed sources reject inspection and validation reports stale evidence. Scanned pages in mixed PDFs retain their original page numbers. Visual similarity remains candidate evidence; small print recognition and region-level text location are not guaranteed. -Cold indexing returns `nextRequest` with `mode: "await"`; copy it to the same tool. `mode: "cancel"` and its `operationId` stop that work. The service reuses one resident visual model for consecutive requests, releases it after 60 seconds idle, and owns shutdown and cancellation. Each request admits at most 512 media files, 20 MiB per file and 256 MiB total; corpus limits and enumeration limits are explicit partial coverage. Request deadline is five minutes. Results expire after ten minutes and are scoped to the originating workspace. Office documents, audio and video are outside this tool's current contract. +### Bring your own provider + +The local model path remains available through `nativeMedia`. To use a provider you supply, add `media` to the same `sift-light.json`. The endpoint must implement the HTTP JSON contract below; it may be a third-party service or a local gateway that translates another vendor's API. A local service can omit `apiKeyEnv`; remote services should use HTTPS. + +```json +{ + "media": { + "provider": "http-json", + "endpoint": "https://your-provider.example/v1/media-search", + "model": "your-vision-model", + "apiKeyEnv": "YOUR_MEDIA_KEY", + "timeoutMs": 120000, + "maxRetries": 2, + "maxRequestBytes": 33554432, + "renderer": { + "executable": "/absolute/native-media", + "modelDirectory": "/absolute/model-directory" + } + } +} +``` + +The plugin sends admitted files as bounded Base64 content with opaque `fileId` values; it never sends local absolute paths. A response must contain `version: 1`, `indexedFiles`, and `hits`. Each hit contains `fileId`, `score`, `kind`, `chunkCount`, `width`, and `height`; scanned PDF hits also contain a one-based `page`. The plugin maps IDs back to local sources and then performs hash, page, `inspect`, and `validate` checks. The remote service does not decide whether source evidence is trustworthy. + +Jev also accepts `semanticJudge.provider: "http-json"`. The request is `{ "version": 1, "model", "query", "candidates" }`; the response is `{ "version": 1, "model", "judgments", "usage" }`. Each judgment contains a candidate `id`, a `choice`, and a `probability` from 0 to 1. When `apiKeyEnv` is configured, the key is read only from that environment variable and is never written to configuration, logs, or result content. A keyless local provider is valid. Remote failures retain local candidates and mark coverage as partial. + +The remote path uses the same `sift-light-media` tool and source verification as the local path. PNG hits can be inspected directly; PDF or other image previews require a local renderer in `media.renderer`. Validation does not require a renderer. + +Cold indexing returns `nextRequest` with `mode: "await"`; copy it to the same tool. `mode: "cancel"` and its `operationId` stop that work. The local provider reuses one resident visual model for consecutive requests and releases it after 60 seconds idle; remote provider timeouts, retries, and request limits come from its configuration. The service owns shutdown and cancellation. Each request admits at most 512 media files, 20 MiB per file and 256 MiB total; corpus limits and enumeration limits are explicit partial coverage. Request deadline is five minutes. Results expire after ten minutes and are scoped to the originating workspace. Office documents, audio and video are outside this tool's current contract. Native text and visual workers reserve 1 GB of inference headroom above the loaded model and its working space, and return unused compute cache after each query. This is not a hard limit on total process memory. The model still uses several GB, with additional allocations during indexing and inference. Each MCP service session owns its models; multiple connections and simultaneous text/visual workers add their memory use. Ordinary exact search does not load these models. Measure memory and latency for the actual model, machine and file scope. diff --git a/README.zh-CN.md b/README.zh-CN.md index c2a03424..77f108be 100644 --- a/README.zh-CN.md +++ b/README.zh-CN.md @@ -230,6 +230,34 @@ Kimi Code 的 web 模式可能从安装目录启动插件 MCP 服务。如果相 结果最多保留十个不同文件，包含原文件内容哈希、原 PDF 页码和可直接执行的 `inspect`、`validate` 请求。把这些请求继续交给本工具：`inspect` 返回最长边不超过 1024 像素的真实 PNG；扫描页返回原 PDF 中对应页面。原文件发生变化时，查看会明确拒绝，校验会报告过期。混合 PDF 中的扫描页也保留原页码。相似度只是候选依据，不保证识别所有小字，也不提供文字区域坐标。 -首次索引较慢时，工具返回 `nextRequest` 和 `mode: "await"`；照原请求继续等待即可。使用 `mode: "cancel"` 和对应 `operationId` 可以取消。连续请求复用一个视觉模型，空闲 60 秒后释放，取消和关闭由服务回收进程。每次最多准入 512 个媒体文件，每个文件不超过 20 MiB，总量不超过 256 MiB；数量、总量和枚举限制会明确标为部分覆盖。单次任务期限为五分钟。结果保留十分钟，且只在原工作区中可用。当前入口不包含 Office、音频或视频。 +### 第三方模型（BYO） + +本地模型仍然使用 `nativeMedia`。如果要接入自己提供的多模态服务，在同一份 `sift-light.json` 中增加 `media`。服务必须接受下面约定的 HTTP JSON 请求；它可以是你的第三方服务，也可以是你在本机部署的转换网关。没有 API Key 的本机服务可以省略 `apiKeyEnv`；远程服务应使用 HTTPS。 + +```json +{ + "media": { + "provider": "http-json", + "endpoint": "https://your-provider.example/v1/media-search", + "model": "your-vision-model", + "apiKeyEnv": "YOUR_MEDIA_KEY", + "timeoutMs": 120000, + "maxRetries": 2, + "maxRequestBytes": 33554432, + "renderer": { + "executable": "/absolute/native-media", + "modelDirectory": "/absolute/model-directory" + } + } +} +``` + +插件会把准入的文件以受限大小的 Base64 内容发送，并使用不含本地绝对路径的 `fileId`；返回结果必须包含 `version: 1`、`indexedFiles` 和 `hits`，每个命中使用 `fileId`、`score`、`kind`、`chunkCount`、`width`、`height`，扫描 PDF 还要返回从 1 开始的 `page`。插件收到结果后才把 ID 映射回本地原文件，并继续做哈希、页码、`inspect` 和 `validate` 校验。远程服务不会决定源码证据是否可信。 + +Jev 也支持 `semanticJudge.provider: "http-json"`。它接收 `{ "version": 1, "model", "query", "candidates" }`，返回 `{ "version": 1, "model", "judgments", "usage" }`；每个 judgment 包含候选 `id`、分类 `choice` 和 0 到 1 的 `probability`。如果配置了 `apiKeyEnv`，插件只从该环境变量读取密钥，不把密钥写入配置、日志或请求内容以外的地方。没有密钥的本机 provider 可以直接使用；远程 provider 失败时，本地候选会保留并标记为 partial。 + +远程多模态结果仍沿用同一个 `sift-light-media` 工具和同一套源文件核验。PNG 命中可以直接 inspect；PDF 或其他图片格式要返回预览时，请在 `media.renderer` 中配置本地渲染 worker。只做 `validate` 不需要 renderer。 + +首次索引较慢时，工具返回 `nextRequest` 和 `mode: "await"`；照原请求继续等待即可。使用 `mode: "cancel"` 和对应 `operationId` 可以取消。本地 provider 的连续请求复用一个视觉模型，空闲 60 秒后释放；远程 provider 的超时、重试和请求大小由配置控制。取消和关闭由服务回收自有资源。每次最多准入 512 个媒体文件，每个文件不超过 20 MiB，总量不超过 256 MiB；数量、总量和枚举限制会明确标为部分覆盖。单次任务期限为五分钟。结果保留十分钟，且只在原工作区中可用。当前入口不包含 Office、音频或视频。 原生文本和视觉程序在已加载模型之外设置 1 GB 推理余量，使用计算所需的工作空间，并在每次查询后归还空闲计算缓存；这不是进程总内存的硬上限。模型本身仍会占用数 GB，索引和计算期间还有额外内存。每个 MCP 服务会话分别持有自己的模型；多个连接或文本与视觉同时工作时，占用会相加。普通精确搜索不加载这些模型。运行内存和速度必须按实际模型、机器与文件范围测量。 diff --git a/package.json b/package.json index dddecc76..28949cc5 100644 --- a/package.json +++ b/package.json @@ -1,6 +1,6 @@ { "name": "sift-light", - "version": "1.1.0-1", + "version": "1.1.0", "description": "Context-efficient local search for files, documents, notes and logs across Pi, OMP and MCP clients", "keywords": [ "ai-agent", diff --git a/plugins/sift-light/.claude-plugin/plugin.json b/plugins/sift-light/.claude-plugin/plugin.json index e3654d61..ef24ff83 100644 --- a/plugins/sift-light/.claude-plugin/plugin.json +++ b/plugins/sift-light/.claude-plugin/plugin.json @@ -1,6 +1,6 @@ { "name": "sift-light", - "version": "1.1.0-1", + "version": "1.1.0", "description": "Require sift-light for conventional local searches while keeping development tools available.", "author": { "name": "baoer" diff --git a/plugins/sift-light/.codex-plugin/plugin.json b/plugins/sift-light/.codex-plugin/plugin.json index 928bed1c..481c1bb5 100644 --- a/plugins/sift-light/.codex-plugin/plugin.json +++ b/plugins/sift-light/.codex-plugin/plugin.json @@ -1,6 +1,6 @@ { "name": "sift-light", - "version": "1.1.0-1", + "version": "1.1.0", "description": "Require sift-light for conventional local searches while keeping development tools available.", "author": { "name": "baoer" diff --git a/plugins/sift-light/.mcp.json b/plugins/sift-light/.mcp.json index dd5d8c78..14299a62 100644 --- a/plugins/sift-light/.mcp.json +++ b/plugins/sift-light/.mcp.json @@ -5,7 +5,7 @@ "args": [ "--yes", "--package", - "sift-light-runtime@npm:sift-light@1.1.0-1", + "sift-light-runtime@npm:sift-light@1.1.0", "sift-light-mcp", "--stdio" ], diff --git a/plugins/sift-light/kimi.plugin.json b/plugins/sift-light/kimi.plugin.json index 5d65fbcd..3b571f32 100644 --- a/plugins/sift-light/kimi.plugin.json +++ b/plugins/sift-light/kimi.plugin.json @@ -1,6 +1,6 @@ { "name": "sift-light", - "version": "1.1.0-1", + "version": "1.1.0", "description": "Require sift-light for conventional local searches while keeping development tools available.", "author": { "name": "baoer" @@ -13,7 +13,7 @@ "args": [ "--yes", "--package", - "sift-light-runtime@npm:sift-light@1.1.0-1", + "sift-light-runtime@npm:sift-light@1.1.0", "sift-light-mcp", "--stdio" ], diff --git a/plugins/sift-light/omp-extension.mjs b/plugins/sift-light/omp-extension.mjs index 034c3bd5..65c40dd9 100644 --- a/plugins/sift-light/omp-extension.mjs +++ b/plugins/sift-light/omp-extension.mjs @@ -176,6 +176,42 @@ function isRawSiftLightConfig(value) { function isRawSemanticJudgeConfig(value) { return typeof value === "object" && value !== null && !Array.isArray(value); } +function isRawMediaConfig(value) { + return typeof value === "object" && value !== null && !Array.isArray(value); +} +function parseEndpoint(value, path, field) { + if (typeof value !== "string" || value.length === 0 || value.length > 2048) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be a nonempty URL`); + } + let parsedEndpoint; + try { + parsedEndpoint = new URL(value); + } catch (error) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be a URL`, { + cause: error + }); + } + const loopbackHosts = new Set(["localhost", "127.0.0.1", "[::1]"]); + const secureEndpoint = parsedEndpoint.protocol === "https:" || parsedEndpoint.protocol === "http:" && loopbackHosts.has(parsedEndpoint.hostname); + if (!secureEndpoint || parsedEndpoint.username || parsedEndpoint.password) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must use HTTPS, except that HTTP is allowed for localhost loopback development; URL credentials are not allowed`); + } + return value; +} +function parseApiKeyEnv(value, path, field) { + if (value === undefined) + return; + if (typeof value !== "string" || !/^[A-Z][A-Z0-9_]{0,127}$/u.test(value)) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be an uppercase environment variable name`); + } + return value; +} +function parseBoundedModel(value, path, field) { + if (typeof value !== "string" || value.length === 0 || value.length > 128 || /[\r\n\0]/u.test(value)) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be bounded single-line text`); + } + return value; +} function boundedInteger(value, fallback, minimum, maximum, field) { const candidate = value ?? fallback; if (typeof candidate !== "number" || !Number.isSafeInteger(candidate) || candidate < minimum || candidate > maximum) { @@ -207,45 +243,61 @@ function parseSemanticJudge(value, path) { throw new Error(`Invalid sift-light config at ${path}: semanticJudge.enabled must be a boolean`); } const provider = value.provider ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.provider; - if (provider !== "jev") { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.provider must be "jev"`); - } - const endpoint = value.endpoint ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.endpoint; - if (typeof endpoint !== "string" || endpoint.length === 0 || endpoint.length > 2048) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.endpoint must be a nonempty URL`); - } - let parsedEndpoint; - try { - parsedEndpoint = new URL(endpoint); - } catch (error) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.endpoint must be a URL`, { - cause: error - }); - } - const loopbackHosts = new Set(["localhost", "127.0.0.1", "[::1]"]); - const secureEndpoint = parsedEndpoint.protocol === "https:" || parsedEndpoint.protocol === "http:" && loopbackHosts.has(parsedEndpoint.hostname); - if (!secureEndpoint || parsedEndpoint.username || parsedEndpoint.password) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.endpoint must use HTTPS, except that HTTP is allowed for localhost loopback development; URL credentials are not allowed`); - } - const apiKeyEnv = value.apiKeyEnv ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.apiKeyEnv; - if (typeof apiKeyEnv !== "string" || !/^[A-Z][A-Z0-9_]{0,127}$/u.test(apiKeyEnv)) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.apiKeyEnv must be an uppercase environment variable name`); - } - const model = value.model ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.model; - if (typeof model !== "string" || model.length === 0 || model.length > 128 || /[\r\n\0]/u.test(model)) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.model must be bounded single-line text`); + if (provider !== "jev" && provider !== "http-json") { + throw new Error(`Invalid sift-light config at ${path}: semanticJudge.provider must be "jev" or "http-json"`); } + const endpoint = parseEndpoint(value.endpoint ?? (provider === "jev" ? DEFAULT_SEMANTIC_JUDGE_CONFIG.endpoint : undefined), path, "semanticJudge.endpoint"); + const apiKeyEnv = parseApiKeyEnv(value.apiKeyEnv, path, "semanticJudge.apiKeyEnv") ?? (provider === "jev" ? DEFAULT_SEMANTIC_JUDGE_CONFIG.apiKeyEnv : undefined); + const model = parseBoundedModel(value.model ?? (provider === "jev" ? DEFAULT_SEMANTIC_JUDGE_CONFIG.model : undefined), path, "semanticJudge.model"); return { enabled, provider, endpoint, - apiKeyEnv, + ...apiKeyEnv === undefined ? {} : { apiKeyEnv }, model, timeoutMs: boundedInteger(value.timeoutMs, DEFAULT_SEMANTIC_JUDGE_CONFIG.timeoutMs, 1000, 1200000, `config at ${path}: semanticJudge.timeoutMs`), maxCandidates: boundedInteger(value.maxCandidates, DEFAULT_SEMANTIC_JUDGE_CONFIG.maxCandidates, 1, 20, `config at ${path}: semanticJudge.maxCandidates`), maxRetries: boundedInteger(value.maxRetries, DEFAULT_SEMANTIC_JUDGE_CONFIG.maxRetries, 0, 5, `config at ${path}: semanticJudge.maxRetries`) }; } +function parseMedia(value, path) { + if (value === undefined) + return; + if (!isRawMediaConfig(value)) { + throw new Error(`Invalid sift-light config at ${path}: media must be an object`); + } + const unknown = Object.keys(value).filter((key) => ![ + "provider", + "endpoint", + "apiKeyEnv", + "model", + "timeoutMs", + "maxRetries", + "maxRequestBytes", + "renderer" + ].includes(key)); + if (unknown.length > 0) { + throw new Error(`Invalid sift-light config at ${path}: unsupported media fields; accepted fields are provider, endpoint, apiKeyEnv, model, timeoutMs, maxRetries, maxRequestBytes and renderer`); + } + const provider = value.provider ?? "http-json"; + if (provider !== "http-json") { + throw new Error(`Invalid sift-light config at ${path}: media.provider must be "http-json"`); + } + const endpoint = parseEndpoint(value.endpoint, path, "media.endpoint"); + const apiKeyEnv = parseApiKeyEnv(value.apiKeyEnv, path, "media.apiKeyEnv"); + const model = parseBoundedModel(value.model, path, "media.model"); + const renderer = parseNativeTextConfig(value.renderer, "media.renderer"); + return { + provider, + endpoint, + ...apiKeyEnv === undefined ? {} : { apiKeyEnv }, + model, + timeoutMs: boundedInteger(value.timeoutMs, 120000, 1000, 1200000, `config at ${path}: media.timeoutMs`), + maxRetries: boundedInteger(value.maxRetries, 2, 0, 5, `config at ${path}: media.maxRetries`), + maxRequestBytes: boundedInteger(value.maxRequestBytes, 32 * 1024 * 1024, 1 * 1024 * 1024, 64 * 1024 * 1024, `config at ${path}: media.maxRequestBytes`), + ...renderer === undefined ? {} : { renderer } + }; +} function parseConfig(value, path) { if (!isRawSiftLightConfig(value)) { throw new Error(`Invalid sift-light config at ${path}: expected a JSON object`); @@ -257,10 +309,11 @@ function parseConfig(value, path) { "semanticJudge", "dataSources", "nativeText", - "nativeMedia" + "nativeMedia", + "media" ].includes(key)); if (unknown.length > 0) { - throw new Error(`Invalid sift-light config at ${path}: unsupported configuration fields; only locale, enforceSearch, vectorSearchEnabled, semanticJudge, dataSources, nativeText and nativeMedia are accepted`); + throw new Error(`Invalid sift-light config at ${path}: unsupported configuration fields; only locale, enforceSearch, vectorSearchEnabled, semanticJudge, dataSources, nativeText, nativeMedia and media are accepted`); } const { locale, enforceSearch, vectorSearchEnabled } = value; if (locale !== undefined && locale !== "en" && locale !== "zh-CN") { @@ -272,11 +325,15 @@ function parseConfig(value, path) { const enforcement = normalizeSearchEnforcement(enforceSearch, `config at ${path}`); const nativeText = parseNativeTextConfig(value.nativeText); const nativeMedia = parseNativeTextConfig(value.nativeMedia, "nativeMedia"); + const media = parseMedia(value.media, `${path}: media`); + if (nativeMedia && media) + throw new Error(`Invalid sift-light config at ${path}: choose nativeMedia or media, not both`); if (nativeText && vectorSearchEnabled !== true) throw new Error("nativeText requires vectorSearchEnabled:true"); return { ...nativeText ? { nativeText } : {}, ...nativeMedia ? { nativeMedia } : {}, + ...media ? { media } : {}, locale: locale ?? DEFAULT_SIFT_LIGHT_CONFIG.locale, enforceSearch: enforcement, vectorSearchEnabled: vectorSearchEnabled ?? DEFAULT_SIFT_LIGHT_CONFIG.vectorSearchEnabled ?? false, @@ -2060,7 +2117,7 @@ function createCtagsStructureProvider(options = {}) { // package.json var package_default = { name: "sift-light", - version: "1.1.0-1", + version: "1.1.0", description: "Context-efficient local search for files, documents, notes and logs across Pi, OMP and MCP clients", keywords: [ "ai-agent", @@ -2676,10 +2733,10 @@ class OperationLifecycle { } // src/media-search.ts -import { createHash, randomUUID as randomUUID3 } from "crypto"; -import { mkdtemp, mkdir, readFile as readFile3, realpath as realpath3, rm, stat as stat4, writeFile } from "fs/promises"; +import { createHash, randomUUID as randomUUID4 } from "crypto"; +import { mkdtemp, mkdir, readFile as readFile4, realpath as realpath3, rm, stat as stat5, writeFile } from "fs/promises"; import { tmpdir } from "os"; -import { basename as basename2, dirname as dirname2, extname, join as join3, resolve as resolve8 } from "path"; +import { basename as basename2, dirname as dirname2, extname as extname2, join as join3, resolve as resolve8 } from "path"; // node_modules/zod/v4/core/util.js function getEnumValues(entries) { @@ -8295,6 +8352,278 @@ class ResidentWorkerClient { } } +// src/remote-media-client.ts +import { randomUUID as randomUUID3 } from "crypto"; +import { readFile as readFile3, stat as stat4 } from "fs/promises"; +import { extname } from "path"; +var MAX_RESPONSE_BYTES = 4 * 1024 * 1024; +var MAX_FILE_ID_LENGTH = 128; + +class RemoteMediaRequestError extends SiftLightError { + retryable; + constructor(message, retryable, options) { + super(message, options); + this.name = "RemoteMediaRequestError"; + this.retryable = retryable; + } +} +var remoteResponseSchema = object({ + version: literal(1), + indexedFiles: number2().int().nonnegative(), + hits: array(object({ + fileId: string2().min(1).max(MAX_FILE_ID_LENGTH), + score: number2().finite(), + kind: _enum(["image", "scan"]), + page: number2().int().positive().nullable().optional(), + chunkCount: number2().int().positive(), + width: number2().int().nonnegative(), + height: number2().int().nonnegative() + }).strict()).max(10) +}).strict(); +function mimeType(path) { + switch (extname(path).toLowerCase()) { + case ".png": + return "image/png"; + case ".jpg": + case ".jpeg": + return "image/jpeg"; + case ".gif": + return "image/gif"; + case ".bmp": + return "image/bmp"; + case ".tif": + case ".tiff": + return "image/tiff"; + case ".heic": + return "image/heic"; + case ".heif": + return "image/heif"; + case ".webp": + return "image/webp"; + case ".pdf": + return "application/pdf"; + default: + return "application/octet-stream"; + } +} +async function boundedResponseText(response) { + const contentLength = response.headers.get("content-length"); + if (contentLength !== null && Number(contentLength) > MAX_RESPONSE_BYTES) + throw new SiftLightError("Remote media response exceeded the 4 MiB limit"); + if (!response.body) { + const text = await response.text(); + if (new TextEncoder().encode(text).byteLength > MAX_RESPONSE_BYTES) + throw new SiftLightError("Remote media response exceeded the 4 MiB limit"); + return text; + } + const reader = response.body.getReader(); + const chunks = []; + let bytes = 0; + try { + while (true) { + const result = await reader.read(); + if (result.done) + break; + bytes += result.value.byteLength; + if (bytes > MAX_RESPONSE_BYTES) { + await reader.cancel(); + throw new SiftLightError("Remote media response exceeded the 4 MiB limit"); + } + chunks.push(result.value); + } + } finally { + reader.releaseLock(); + } + const all = new Uint8Array(bytes); + let offset = 0; + for (const chunk of chunks) { + all.set(chunk, offset); + offset += chunk.byteLength; + } + return new TextDecoder("utf-8", { fatal: true }).decode(all); +} +function retryableStatus(status) { + return status === 408 || status === 429 || status === 529 || status >= 500; +} +function retryDelay(attempt) { + return Math.min(20000, 400 * 2 ** attempt); +} +async function waitForRetry(delayMs, signal) { + if (signal?.aborted) + throw signal.reason instanceof Error ? signal.reason : new Error("Operation aborted"); + await new Promise((resolve, reject) => { + const timer = setTimeout(() => { + signal?.removeEventListener("abort", onAbort); + resolve(); + }, delayMs); + const onAbort = () => { + clearTimeout(timer); + signal?.removeEventListener("abort", onAbort); + reject(signal?.reason instanceof Error ? signal.reason : new Error("Operation aborted")); + }; + signal?.addEventListener("abort", onAbort, { once: true }); + }); +} +function base642(bytes) { + return bytes.toString("base64"); +} +function base64Length(bytes) { + return Math.ceil(bytes / 3) * 4; +} + +class RemoteMediaClient { + #config; + #key; + #fetcher; + constructor(config, environment = process.env, fetcher = fetch) { + this.#config = config; + this.#key = config.apiKeyEnv ? environment[config.apiKeyEnv]?.trim() : undefined; + if (config.apiKeyEnv && !this.#key) { + throw new SiftLightError(`Media provider is configured with ${config.apiKeyEnv}, but that environment variable is missing or empty`); + } + this.#fetcher = fetcher; + } + async search(request) { + const id = randomUUID3(); + const paths = new Map; + const files = []; + const fileSizes = []; + for (const file of request.files) { + if (file.id.length === 0 || file.id.length > MAX_FILE_ID_LENGTH || /[\r\n\0]/u.test(file.id) || paths.has(file.id)) + throw new SiftLightError("Remote media file ids must be unique bounded text"); + paths.set(file.id, file.path); + const metadata2 = await stat4(file.path); + if (!metadata2.isFile()) + throw new SiftLightError("Remote media files must be regular files"); + fileSizes.push(metadata2.size); + } + let imageSize; + if (request.imagePath !== undefined) { + const metadata2 = await stat4(request.imagePath); + if (!metadata2.isFile()) + throw new SiftLightError("Remote media query image must be a regular file"); + imageSize = metadata2.size; + } + const requestSkeleton = JSON.stringify({ + version: 1, + model: this.#config.model, + ...request.query === undefined ? {} : { query: request.query }, + ...request.imagePath === undefined ? {} : { image: { mimeType: mimeType(request.imagePath), data: "" } }, + kind: request.kind, + expectedFiles: request.expectedFiles, + files: [] + }); + let estimatedBytes = new TextEncoder().encode(requestSkeleton).byteLength; + if (imageSize !== undefined) + estimatedBytes += base64Length(imageSize); + for (let index = 0;index < request.files.length; index += 1) { + const file = request.files[index]; + const fileSkeleton = JSON.stringify({ + id: file?.id, + mimeType: file ? mimeType(file.path) : "application/octet-stream", + data: "" + }); + estimatedBytes += new TextEncoder().encode(fileSkeleton).byteLength; + estimatedBytes += base64Length(fileSizes[index] ?? 0); + } + if (request.files.length > 1) + estimatedBytes += request.files.length - 1; + if (estimatedBytes > this.#config.maxRequestBytes) + throw new SiftLightError(`Remote media request exceeded the configured ${String(this.#config.maxRequestBytes)} byte limit`); + for (const file of request.files) { + files.push({ + id: file.id, + mimeType: mimeType(file.path), + data: base642(await readFile3(file.path, { signal: request.signal })) + }); + } + const image = request.imagePath === undefined ? undefined : { + mimeType: mimeType(request.imagePath), + data: base642(await readFile3(request.imagePath, { signal: request.signal })) + }; + const body2 = JSON.stringify({ + version: 1, + model: this.#config.model, + ...request.query === undefined ? {} : { query: request.query }, + ...image === undefined ? {} : { image }, + kind: request.kind, + expectedFiles: request.expectedFiles, + files + }); + if (new TextEncoder().encode(body2).byteLength > this.#config.maxRequestBytes) + throw new SiftLightError(`Remote media request exceeded the configured ${String(this.#config.maxRequestBytes)} byte limit`); + for (let attempt = 0;attempt <= this.#config.maxRetries; attempt += 1) { + const controller = new AbortController; + const onParentAbort = () => controller.abort(request.signal?.reason); + request.signal?.addEventListener("abort", onParentAbort, { once: true }); + const timeout = setTimeout(() => controller.abort(new Error("Remote media request timed out")), this.#config.timeoutMs); + try { + const headers = { "content-type": "application/json" }; + if (this.#key) + headers.authorization = `Bearer ${this.#key}`; + const response = await this.#fetcher(this.#config.endpoint, { + method: "POST", + headers, + body: body2, + signal: controller.signal + }); + if (!response.ok) { + throw new RemoteMediaRequestError(`Remote media provider returned HTTP ${String(response.status)}`, retryableStatus(response.status)); + } + let parsed; + try { + const raw = JSON.parse(await boundedResponseText(response)); + parsed = remoteResponseSchema.parse(raw); + } catch (error) { + throw new SiftLightError("Remote media provider returned an invalid response", { + cause: error + }); + } + if (parsed.indexedFiles !== request.expectedFiles) + throw new SiftLightError("Remote media provider indexed an unexpected file count"); + return { + version: 1, + id, + indexedFiles: parsed.indexedFiles, + modelLoads: 0, + modelReused: true, + processId: process.pid, + hits: parsed.hits.map((hit) => { + const path = paths.get(hit.fileId); + if (!path) + throw new SiftLightError("Remote media provider returned an unknown file id"); + return { + path, + score: hit.score, + kind: hit.kind, + page: hit.page ?? null, + chunkCount: hit.chunkCount, + width: hit.width, + height: hit.height + }; + }) + }; + } catch (error) { + if (request.signal?.aborted) + throw error; + const retryable = error instanceof RemoteMediaRequestError ? error.retryable : !(error instanceof SiftLightError); + if (!retryable || attempt >= this.#config.maxRetries) + throw error; + await waitForRetry(retryDelay(attempt), request.signal); + } finally { + clearTimeout(timeout); + request.signal?.removeEventListener("abort", onParentAbort); + } + } + throw new SiftLightError("Remote media request failed"); + } + async close() {} +} +function mediaMimeType(path) { + const type = mimeType(path); + return type === "application/octet-stream" ? undefined : type; +} + // src/workspace-files.ts import { relative as relative3, resolve as resolve7, sep as sep2 } from "path"; class EnumerationLimit extends Error { @@ -8417,7 +8746,7 @@ var responseSchema = object({ version: literal(1), id: string2(), indexedFiles: number2().int().nonnegative(), - modelLoads: literal(1), + modelLoads: number2().int().nonnegative(), modelReused: boolean2(), processId: number2().int().positive(), hits: array(object({ @@ -8453,11 +8782,11 @@ async function sourceBytes(cwd, path, signal) { const canonical = await policy.resolveExistingPath(path); if (!canonical) throw new SiftLightError(`Media source unavailable: ${path}`); - const before = await stat4(canonical); + const before = await stat5(canonical); if (!before.isFile() || before.size > maxFileBytes) throw new SiftLightError(`Media source must be a regular file at most ${String(maxFileBytes)} bytes: ${path}`); - const bytes = await readFile3(canonical, { signal }); - const after = await stat4(canonical); + const bytes = await readFile4(canonical, { signal }); + const after = await stat5(canonical); if (bytes.length > maxFileBytes || before.ino !== after.ino || before.size !== after.size || before.mtimeMs !== after.mtimeMs || before.ctimeMs !== after.ctimeMs || await policy.resolveExistingPath(path) !== canonical) throw new SiftLightError(`Media source changed during read: ${path}`); return { canonical, bytes }; @@ -8465,6 +8794,7 @@ async function sourceBytes(cwd, path, signal) { class MediaSearch { #config; + #remoteConfig; #operations = new OperationLifecycle({ deadlineMs: 300000 }); #queue = new OwnedTaskQueue; #lifecycle = new AbortController; @@ -8472,9 +8802,13 @@ class MediaSearch { #known = new Map; #directory; #worker; + #remote; #closed = false; constructor(config) { - this.#config = config; + if ("provider" in config) + this.#remoteConfig = config; + else + this.#config = config; } async search(value, cwd, signal) { const request = mediaRequestSchema.parse(value); @@ -8530,7 +8864,7 @@ class MediaSearch { const reasons = [...inventory.reasons]; let total = 0; for (const path of inventory.paths) { - if (!extensions.has(extname(path).toLowerCase())) + if (!extensions.has(extname2(path).toLowerCase())) continue; if (signal.aborted) throw abortError(); @@ -8568,20 +8902,37 @@ class MediaSearch { let imageSource; let queryImage; if (r.image) { - if (!extensions.has(extname(r.image).toLowerCase()) || extname(r.image).toLowerCase() === ".pdf") + if (!extensions.has(extname2(r.image).toLowerCase()) || extname2(r.image).toLowerCase() === ".pdf") throw new SiftLightError("image must be a supported local image file"); imageSource = await sourceBytes(cwd, r.image, signal); - queryImage = join3(root, "query" + extname(r.image).toLowerCase()); + queryImage = join3(root, "query" + extname2(r.image).toLowerCase()); await writeFile(queryImage, imageSource.bytes, { mode: 384 }); } let response; if (admitted.size) { - this.#worker ??= new ResidentWorkerClient(this.#config, root, corpus, join3(root, "index.sqlite"), { parse: (value) => responseSchema.parse(value), maxResponseBytes: 128 * 1024 }); - response = await this.#worker.search({ - ...queryImage ? { image: queryImage } : { query: r.query }, - kind: r.kind, - expectedFiles: admitted.size - }, signal); + if (this.#config) { + this.#worker ??= new ResidentWorkerClient(this.#config, root, corpus, join3(root, "index.sqlite"), { parse: (value) => responseSchema.parse(value), maxResponseBytes: 128 * 1024 }); + response = await this.#worker.search({ + ...queryImage ? { image: queryImage } : { query: r.query }, + kind: r.kind, + expectedFiles: admitted.size + }, signal); + } else if (this.#remoteConfig) { + this.#remote ??= new RemoteMediaClient(this.#remoteConfig); + const remoteFiles = [...admitted.values()].map((entry) => ({ + id: createHash("sha256").update(entry.staged).digest("hex").slice(0, 32), + path: entry.staged + })); + response = await this.#remote.search({ + ...queryImage ? { imagePath: queryImage } : { query: r.query }, + kind: r.kind, + expectedFiles: admitted.size, + files: remoteFiles, + signal + }); + } + if (!response) + throw new SiftLightError("Media provider is not configured"); if (response.indexedFiles !== admitted.size) throw new SiftLightError("Media index inventory mismatch"); } @@ -8589,7 +8940,7 @@ class MediaSearch { const seen = new Set; for (const hit of response?.hits ?? []) { const source = admitted.get(resolve8(hit.path)); - if (!source || hit.kind === "scan" !== (extname(source.path).toLowerCase() === ".pdf") || seen.has(hit.path) || hit.kind === "scan" !== (hit.page !== null) || hit.page !== null && hit.page > hit.chunkCount) + if (!source || hit.kind === "scan" !== (extname2(source.path).toLowerCase() === ".pdf") || seen.has(hit.path) || hit.kind === "scan" !== (hit.page !== null) || hit.page !== null && hit.page > hit.chunkCount) throw new SiftLightError("Media result source or page is inconsistent"); seen.add(hit.path); saved.push({ hit, source }); @@ -8609,7 +8960,7 @@ class MediaSearch { } if (signal.aborted) throw abortError(); - const cursor = randomUUID3(); + const cursor = randomUUID4(); for (const [id, snapshot] of this.#snapshots) if (Date.now() - snapshot.createdAt > 600000) this.#snapshots.delete(id); @@ -8672,11 +9023,23 @@ ${JSON.stringify(details)}`, throw new SiftLightError("Media source changed; search again before inspection"); const folder = await mkdtemp(join3(await this.#root(), "inspect-")); try { - const input = join3(folder, "source" + extname(item.source.path)); + const input = join3(folder, "source" + extname2(item.source.path)); const output = join3(folder, "preview.png"); await writeFile(input, current.bytes, { mode: 384 }); + const renderer = this.#config ?? this.#remoteConfig?.renderer; + if (!renderer) { + const mimeType = mediaMimeType(item.source.path); + if (mimeType?.startsWith("image/") && mimeType === "image/png") { + return { + text: JSON.stringify(details), + details, + image: { data: current.bytes.toString("base64"), mimeType } + }; + } + throw new SiftLightError("Media preview requires nativeMedia or media.renderer; validate still works without a local renderer"); + } const result = await runOwnedProcess({ - executable: this.#config.executable, + executable: renderer.executable, args: ["--render", input, String(item.hit.page ?? 0), output], cwd: folder, signal @@ -8687,7 +9050,7 @@ ${JSON.stringify(details)}`, }); if (result.code !== 0) throw new SiftLightError(`Media preview failed: ${result.stderr}`); - const preview = await readFile3(output); + const preview = await readFile4(output); if (preview.length > 4 * 1024 * 1024) throw new SiftLightError("Media preview exceeds 4 MiB"); return { @@ -8712,6 +9075,7 @@ ${JSON.stringify(details)}`, await this.#queue.run(async () => { try { await this.#worker?.close(); + await this.#remote?.close(); } finally { this.#snapshots.clear(); if (this.#directory) @@ -8730,7 +9094,7 @@ class DataSourceError extends Error { this.code = code; } } -var MAX_RESPONSE_BYTES = 1024 * 1024; +var MAX_RESPONSE_BYTES2 = 1024 * 1024; async function credential(config) { let value; if (config.tokenEnv) @@ -8785,7 +9149,7 @@ async function sourcePost(config, path, payload, signal) { if (part.done) break; size += part.value.byteLength; - if (size > MAX_RESPONSE_BYTES) { + if (size > MAX_RESPONSE_BYTES2) { await reader.cancel(); throw new DataSourceError("protocol"); } @@ -9035,14 +9399,14 @@ class DataSources { import { createHash as createHash7 } from "crypto"; // src/concept-worker-client.ts -import { randomUUID as randomUUID4 } from "crypto"; +import { randomUUID as randomUUID5 } from "crypto"; import { mkdir as mkdir2, rm as rm2 } from "fs/promises"; import { dirname as dirname3, join as join5 } from "path"; import { fileURLToPath } from "url"; import { StringDecoder as StringDecoder2 } from "string_decoder"; // src/concept-model.ts -import { stat as stat5 } from "fs/promises"; +import { stat as stat6 } from "fs/promises"; import { homedir as homedir2 } from "os"; import { join as join4, resolve as resolve9 } from "path"; var CONCEPT_MODEL = "Xenova/multilingual-e5-small"; @@ -9085,7 +9449,7 @@ function conceptModelDirectory() { async function inspectConceptModel(directory = conceptModelDirectory()) { for (const asset of CONCEPT_ASSETS) { try { - const information = await stat5(join4(directory, asset.path)); + const information = await stat6(join4(directory, asset.path)); if (!information.isFile() || information.size !== asset.bytes) return { installed: false, @@ -9181,7 +9545,7 @@ class ConceptWorkerClient { throw new SiftLightError("Concept worker idle timeout must be 1\u2013300000 ms"); } async#start() { - const stagingRoot = join5(conceptCacheDirectory(), ".staging", randomUUID4()); + const stagingRoot = join5(conceptCacheDirectory(), ".staging", randomUUID5()); await mkdir2(stagingRoot, { recursive: true }); const ready = Promise.withResolvers(); ready.promise.catch(() => { @@ -9315,7 +9679,7 @@ class ConceptWorkerClient { const input = await session.ready; if (combined.aborted) throw abortError(); - const id = randomUUID4(); + const id = randomUUID5(); const body2 = JSON.stringify({ id, query, @@ -10109,7 +10473,7 @@ function deriveSyntaxFacts(tree, language, text) { } // src/language-capability-definitions.ts -import { extname as extname2 } from "path"; +import { extname as extname3 } from "path"; var TYPESCRIPT_LANGUAGE_CAPABILITIES = [ { id: "ast-grep.outline", @@ -10258,14 +10622,14 @@ var DEFAULT_NEUTRAL_CAPABILITIES = [ } ]; function descriptorForPath(descriptors, path) { - const extension = extname2(path).toLowerCase(); + const extension = extname3(path).toLowerCase(); return descriptors.find((item) => item.extensions.includes(extension)); } function languageForPath(descriptors, path) { return descriptorForPath(descriptors, path)?.language; } function outlineExtension(path) { - const extension = extname2(path).toLowerCase(); + const extension = extname3(path).toLowerCase(); return extension || undefined; } @@ -10441,7 +10805,7 @@ async function parseSyntax(path, text, signal, pattern) { import { createHash as createHash4 } from "crypto"; // src/source-access.ts -import { extname as extname3, resolve as resolve13 } from "path"; +import { extname as extname4, resolve as resolve13 } from "path"; // src/historical-paths.ts import { lstat, mkdir as mkdir3, mkdtemp as mkdtemp2, open as open3, rm as rm3, writeFile as writeFile2 } from "fs/promises"; @@ -11574,7 +11938,7 @@ class SyntaxQueue { throw new SourceDocumentError("encoding", "Syntax requires lossless UTF-8 source"); const origin = document2.reference.origin; const revision = origin.kind === "worktree" ? origin.contentHash : origin.blob; - const key = `${extname3(document2.path).toLowerCase()}\x00${revision}\x00${pattern ?? ""}`; + const key = `${extname4(document2.path).toLowerCase()}\x00${revision}\x00${pattern ?? ""}`; const cached = this.#cache.get(key); if (cached) { this.#cache.delete(key); @@ -12353,7 +12717,7 @@ async function structuralSearch(input, access) { // src/native-text-search.ts import { createHash as createHash6 } from "crypto"; -import { mkdtemp as mkdtemp3, mkdir as mkdir4, readFile as readFile4, realpath as realpath5, rm as rm4, writeFile as writeFile3 } from "fs/promises"; +import { mkdtemp as mkdtemp3, mkdir as mkdir4, readFile as readFile5, realpath as realpath5, rm as rm4, writeFile as writeFile3 } from "fs/promises"; import { tmpdir as tmpdir3 } from "os"; import { basename as basename3, dirname as dirname6, isAbsolute as isAbsolute8, join as join7, resolve as resolve15 } from "path"; @@ -12477,7 +12841,7 @@ function createNativeTextSearch(config, search) { if (!entry || seen.has(entry.staged)) throw new SiftLightError("Native text returned an unknown or duplicate source"); seen.add(entry.staged); - const content = await readFile4(entry.staged); + const content = await readFile5(entry.staged); if (createHash6("sha256").update(content).digest("hex") !== entry.hash) throw new SiftLightError("Native staged source changed"); const document2 = new SourceDocument(entry.source, content); @@ -12558,10 +12922,10 @@ function rankRetrievalFiles(current, native) { } // src/evidence-service.ts -import { dirname as dirname7, extname as extname4, resolve as resolve24 } from "path"; +import { dirname as dirname7, extname as extname5, resolve as resolve24 } from "path"; // src/analysis-store.ts -import { randomUUID as randomUUID5 } from "crypto"; +import { randomUUID as randomUUID6 } from "crypto"; // src/analysis-types.ts var SEMANTIC_JUDGE_CLASSIFICATIONS = [ @@ -12946,7 +13310,7 @@ class AnalysisStore { } create(result, summarize) { this.#expire(); - const id = randomUUID5(); + const id = randomUUID6(); const bounded = { ...result, ...result.retrievalLanes ? { @@ -15191,7 +15555,7 @@ async function discoverFiles(input, cwd, signal) { } // src/source-continuations.ts -import { randomUUID as randomUUID6 } from "crypto"; +import { randomUUID as randomUUID7 } from "crypto"; // src/source-pages.ts function mergeByteRanges(ranges) { @@ -15295,7 +15659,7 @@ class SourceContinuations { create(source, target, gaps, boundary) { this.#sweep(); const item = { - id: randomUUID6(), + id: randomUUID7(), source: structuredClone(source), target: mergeByteRanges(target), gaps: mergeByteRanges(gaps), @@ -15891,7 +16255,7 @@ async function continueSource(cursor, access, continuations, expectedPath) { } // src/evidence-validation.ts -import { realpath as realpath7, stat as stat6 } from "fs/promises"; +import { realpath as realpath7, stat as stat7 } from "fs/promises"; import { resolve as resolve22 } from "path"; // src/evidence-validity.ts @@ -15929,9 +16293,9 @@ async function confirmWorktreeState(path, cwd, signal) { const policy = new SearchPathPolicy(cwd); try { policy.assertPath(absolute); - const before = await stat6(absolute); + const before = await stat7(absolute); const beforeCanonical = await realpath7(absolute); - const after = await stat6(absolute); + const after = await stat7(absolute); const afterCanonical = await realpath7(absolute); if (!before.isFile() || !after.isFile()) { return { status: "stale", reason: "Source is no longer a regular file" }; @@ -16073,9 +16437,9 @@ async function validateSnapshotTarget(target, cwd, policy, signal) { const result = await confirmWorktreeState(target.path, cwd, signal); return { path: target.path, role: "source", ...result }; } - const before = await stat6(absolute); + const before = await stat7(absolute); const beforeCanonical = canonical; - const after = await stat6(absolute); + const after = await stat7(absolute); const afterCanonical = await realpath7(absolute); if (!before.isFile() || !after.isFile()) { return { @@ -17029,7 +17393,7 @@ async function judgeSemanticCandidateBatches(runner, query, batches, signal) { // src/semantic-judge.ts var MAX_CANDIDATE_CHARS = 4000; -var MAX_RESPONSE_BYTES2 = 4 * 1024 * 1024; +var MAX_RESPONSE_BYTES3 = 4 * 1024 * 1024; var MAX_ERROR_CHARS = 512; var CLASSIFICATION_PRIORITY = { "implementation-candidate": 0, @@ -17119,14 +17483,22 @@ function requestBody(query, candidates, model) { questions }; } +function httpJsonRequestBody(query, candidates, model) { + return { + version: 1, + model, + query, + candidates: candidates.map(({ id, excerpt }) => ({ id, excerpt })) + }; +} async function responseText(response) { if (!response.body) { const contentLength = response.headers.get("content-length"); - if (contentLength !== null && Number(contentLength) > MAX_RESPONSE_BYTES2) { + if (contentLength !== null && Number(contentLength) > MAX_RESPONSE_BYTES3) { throw new SemanticJudgeRequestError("Semantic judge response exceeded the 4 MiB limit", false); } const text = await response.text(); - if (new TextEncoder().encode(text).byteLength > MAX_RESPONSE_BYTES2) { + if (new TextEncoder().encode(text).byteLength > MAX_RESPONSE_BYTES3) { throw new SemanticJudgeRequestError("Semantic judge response exceeded the 4 MiB limit", false); } return text; @@ -17141,7 +17513,7 @@ async function responseText(response) { break; const chunk = result.value; bytes += chunk.byteLength; - if (bytes > MAX_RESPONSE_BYTES2) { + if (bytes > MAX_RESPONSE_BYTES3) { await reader.cancel(); throw new SemanticJudgeRequestError("Semantic judge response exceeded the 4 MiB limit", false); } @@ -17158,10 +17530,10 @@ async function responseText(response) { } return new TextDecoder("utf-8", { fatal: true }).decode(all); } -function retryDelay(attempt) { +function retryDelay2(attempt) { return Math.min(20000, 400 * 2 ** attempt); } -async function waitForRetry(delayMs, signal) { +async function waitForRetry2(delayMs, signal) { if (signal?.aborted) throw signal.reason instanceof Error ? signal.reason : new Error("Operation aborted"); await new Promise((resolve, reject) => { @@ -17179,68 +17551,127 @@ async function waitForRetry(delayMs, signal) { signal?.addEventListener("abort", onAbort, { once: true }); }); } +async function fetchSemanticResponse(config, key, fetcher, body2, parentSignal) { + if (new TextEncoder().encode(body2).byteLength > MAX_SEMANTIC_JUDGE_REQUEST_BYTES) { + throw new SemanticJudgePayloadTooLargeError("Semantic judge request exceeded the 64 KiB request limit"); + } + for (let attempt = 0;attempt <= config.maxRetries; attempt += 1) { + const controller = new AbortController; + const onParentAbort = () => controller.abort(parentSignal?.reason); + parentSignal?.addEventListener("abort", onParentAbort, { once: true }); + const timeout = setTimeout(() => controller.abort(new Error("Semantic judge request timed out")), config.timeoutMs); + try { + const headers = { "content-type": "application/json" }; + if (key) + headers.authorization = `Bearer ${key}`; + const response = await fetcher(config.endpoint, { + method: "POST", + headers, + body: body2, + signal: controller.signal + }); + if (!response.ok) { + if (response.status === 413) { + throw new SemanticJudgePayloadTooLargeError("Semantic judge rejected an oversized request with HTTP 413"); + } + const retryable = response.status === 408 || response.status === 429 || response.status === 529 || response.status >= 500; + throw new SemanticJudgeRequestError(`Semantic judge returned HTTP ${String(response.status)}`, retryable); + } + const raw = JSON.parse(await responseText(response)); + if (!isRecord3(raw)) { + throw new SemanticJudgeRequestError("Semantic judge response must be an object", false); + } + return raw; + } catch (error) { + if (parentSignal?.aborted) + throw error; + const retryable = error instanceof SemanticJudgePayloadTooLargeError ? false : error instanceof SemanticJudgeRequestError ? error.retryable : true; + if (!retryable || attempt >= config.maxRetries) + throw error; + await waitForRetry2(retryDelay2(attempt), parentSignal); + } finally { + clearTimeout(timeout); + parentSignal?.removeEventListener("abort", onParentAbort); + } + } + throw new SemanticJudgeRequestError("Semantic judge request failed", false); +} +function responseModel(raw, config) { + const model = raw.model ?? config.model; + if (typeof model !== "string" || model.length === 0 || model.length > 128 || /[\r\n\0]/u.test(model)) { + throw new SemanticJudgeRequestError("Semantic judge response model must be bounded single-line text", false); + } + return model; +} +function responseUsage(raw) { + const usage = isRecord3(raw.usage) ? raw.usage : undefined; + const inputTokens = typeof usage?.input_tokens === "number" && Number.isSafeInteger(usage.input_tokens) && usage.input_tokens >= 0 ? usage.input_tokens : undefined; + const outputTokens = typeof usage?.output_tokens === "number" && Number.isSafeInteger(usage.output_tokens) && usage.output_tokens >= 0 ? usage.output_tokens : undefined; + return { + ...inputTokens === undefined ? {} : { inputTokens }, + ...outputTokens === undefined ? {} : { outputTokens } + }; +} +function parseHttpJsonResult(raw, candidates, config, elapsedMs) { + if (raw.version !== 1 || !Array.isArray(raw.judgments)) { + throw new SemanticJudgeRequestError("HTTP JSON semantic judge response must contain version 1 judgments", false); + } + const candidateIndexes = new Map(candidates.map((candidate, index) => [candidate.id, index])); + const judgments = new Map; + for (const value of raw.judgments) { + if (!isRecord3(value) || typeof value.id !== "string") { + throw new SemanticJudgeRequestError("HTTP JSON semantic judge judgment has an invalid id", false); + } + const candidateIndex = candidateIndexes.get(value.id); + if (candidateIndex === undefined || judgments.has(candidateIndex)) { + throw new SemanticJudgeRequestError("HTTP JSON semantic judge response contains an unknown or duplicate candidate", false); + } + if (!isClassification(value.choice)) { + throw new SemanticJudgeRequestError(`HTTP JSON semantic judge response has an invalid choice for ${value.id}`, false); + } + const confidence = value.confidence === undefined ? undefined : probability(value.confidence, "confidence"); + judgments.set(candidateIndex, { + candidateIndex, + classification: value.choice, + probability: probability(value.probability, "probability"), + ...confidence === undefined ? {} : { confidence } + }); + } + if (judgments.size !== candidates.length) { + throw new SemanticJudgeRequestError(`HTTP JSON semantic judge response judged ${String(judgments.size)} of ${String(candidates.length)} candidates`, false); + } + const usage = responseUsage(raw); + return { + model: responseModel(raw, config), + judgments: candidates.map((_candidate, index) => judgments.get(index)), + ...usage, + elapsedMs + }; +} function createJevRunner(config, key, fetcher) { return async (query, candidates, parentSignal) => { const startedAt = performance.now(); const body2 = JSON.stringify(requestBody(query, candidates, config.model)); - if (new TextEncoder().encode(body2).byteLength > MAX_SEMANTIC_JUDGE_REQUEST_BYTES) { - throw new SemanticJudgePayloadTooLargeError("Semantic judge request exceeded the 64 KiB request limit"); - } - for (let attempt = 0;attempt <= config.maxRetries; attempt += 1) { - const controller = new AbortController; - const onParentAbort = () => controller.abort(parentSignal?.reason); - parentSignal?.addEventListener("abort", onParentAbort, { once: true }); - const timeout = setTimeout(() => controller.abort(new Error("Semantic judge request timed out")), config.timeoutMs); - try { - const response = await fetcher(config.endpoint, { - method: "POST", - headers: { - "content-type": "application/json", - authorization: `Bearer ${key}` - }, - body: body2, - signal: controller.signal - }); - if (!response.ok) { - if (response.status === 413) { - throw new SemanticJudgePayloadTooLargeError("Semantic judge rejected an oversized request with HTTP 413"); - } - const retryable = response.status === 408 || response.status === 429 || response.status === 529 || response.status >= 500; - throw new SemanticJudgeRequestError(`Semantic judge returned HTTP ${String(response.status)}`, retryable); - } - const raw = JSON.parse(await responseText(response)); - if (!isRecord3(raw) || !isRecord3(raw.answers)) { - throw new SemanticJudgeRequestError("Semantic judge response did not contain answers", false); - } - const answers = raw.answers; - const judgments = candidates.map((candidate, index) => parseJudgment(answers[candidate.id], index)); - const model = raw.model ?? config.model; - if (typeof model !== "string" || model.length === 0 || model.length > 128 || /[\r\n\0]/u.test(model)) { - throw new SemanticJudgeRequestError("Semantic judge response model must be bounded single-line text", false); - } - const usage = isRecord3(raw.usage) ? raw.usage : undefined; - const inputTokens = typeof usage?.input_tokens === "number" && Number.isSafeInteger(usage.input_tokens) && usage.input_tokens >= 0 ? usage.input_tokens : undefined; - const outputTokens = typeof usage?.output_tokens === "number" && Number.isSafeInteger(usage.output_tokens) && usage.output_tokens >= 0 ? usage.output_tokens : undefined; - return { - model, - judgments, - ...inputTokens === undefined ? {} : { inputTokens }, - ...outputTokens === undefined ? {} : { outputTokens }, - elapsedMs: Math.round(performance.now() - startedAt) - }; - } catch (error) { - if (parentSignal?.aborted) - throw error; - const retryable = error instanceof SemanticJudgePayloadTooLargeError ? false : error instanceof SemanticJudgeRequestError ? error.retryable : true; - if (!retryable || attempt >= config.maxRetries) - throw error; - await waitForRetry(retryDelay(attempt), parentSignal); - } finally { - clearTimeout(timeout); - parentSignal?.removeEventListener("abort", onParentAbort); - } + const raw = await fetchSemanticResponse(config, key, fetcher, body2, parentSignal); + if (!isRecord3(raw.answers)) { + throw new SemanticJudgeRequestError("Semantic judge response did not contain answers", false); } - throw new SemanticJudgeRequestError("Semantic judge request failed", false); + const answers = raw.answers; + const judgments = candidates.map((candidate, index) => parseJudgment(answers[candidate.id], index)); + const usage = responseUsage(raw); + return { + model: responseModel(raw, config), + judgments, + ...usage, + elapsedMs: Math.round(performance.now() - startedAt) + }; + }; +} +function createHttpJsonRunner(config, key, fetcher) { + return async (query, candidates, parentSignal) => { + const startedAt = performance.now(); + const raw = await fetchSemanticResponse(config, key, fetcher, JSON.stringify(httpJsonRequestBody(query, candidates, config.model)), parentSignal); + return parseHttpJsonResult(raw, candidates, config, Math.round(performance.now() - startedAt)); }; } function createDisabledSemanticJudgeIntegration(config) { @@ -17249,14 +17680,16 @@ function createDisabledSemanticJudgeIntegration(config) { function createSemanticJudgeIntegration(config, environment = process.env, fetcher = fetch) { if (!config.enabled) return createDisabledSemanticJudgeIntegration(config); - const key = environment[config.apiKeyEnv]?.trim(); - if (!key) { - throw new SemanticJudgeConfigurationError(`Semantic judge is enabled, but environment variable ${config.apiKeyEnv} is missing or empty`); + const key = config.apiKeyEnv ? environment[config.apiKeyEnv]?.trim() : undefined; + if (config.provider === "jev" && !key) { + throw new SemanticJudgeConfigurationError(`Semantic judge is enabled, but environment variable ${config.apiKeyEnv ?? "TYPESAFE_API_KEY"} is missing or empty`); } - if (config.provider !== "jev") { - throw new SemanticJudgeConfigurationError(`Unsupported semantic judge provider: ${String(config.provider)}`); + if (config.provider === "http-json" && config.apiKeyEnv && !key) { + throw new SemanticJudgeConfigurationError(`Semantic judge is enabled, but environment variable ${config.apiKeyEnv} is missing or empty`); } - return { config, runner: createJevRunner(config, key, fetcher) }; + if (config.provider === "jev") + return { config, runner: createJevRunner(config, key, fetcher) }; + return { config, runner: createHttpJsonRunner(config, key, fetcher) }; } function createConfiguredSemanticJudgeIntegration(config, environment = process.env) { const judge = config.semanticJudge ?? DEFAULT_SEMANTIC_JUDGE_CONFIG; @@ -19078,7 +19511,7 @@ class EvidenceService { throw new SiftLightError(`No outline provider is registered for ${document2.path}; use mode=capabilities to inspect available language modes`); } if (!language && !isPython || language === "go") { - const extension = extname4(document2.path) || "extensionless source"; + const extension = extname5(document2.path) || "extensionless source"; throw new SiftLightError(`${input.mode} is unavailable for ${extension}; choose a language capability from mode=capabilities or use ordinary content search`); } if (input.mode === "outline") { @@ -19271,7 +19704,7 @@ function normalizeRequestAliases(request) { } // src/format.ts -import { readFile as readFile5 } from "fs/promises"; +import { readFile as readFile6 } from "fs/promises"; var RESULT_METADATA_RESERVE_BYTES = 1024; var RESULT_METADATA_RESERVE_CHARACTERS = 512; @@ -19405,7 +19838,7 @@ async function loadContextLines(match, expectedRevision, cache, signal) { cache.set(match.absolutePath, changed); return changed; } - const content = await readFile5(match.absolutePath, { encoding: "utf8", signal }); + const content = await readFile6(match.absolutePath, { encoding: "utf8", signal }); const afterRevision = await getSourceRevision(match.absolutePath); if (!afterRevision || !sameSourceRevision(expectedRevision, afterRevision)) { const changed = { status: "changed" }; @@ -19730,7 +20163,7 @@ ${lineRows.join(` } // src/snapshot-store.ts -import { randomUUID as randomUUID7 } from "crypto"; +import { randomUUID as randomUUID8 } from "crypto"; class SnapshotStore { #snapshots = new Map; #expired = new Set; @@ -19755,7 +20188,7 @@ class SnapshotStore { const now = this.#now(); const snapshot = { ...scan, - id: randomUUID7(), + id: randomUUID8(), createdAt: now, lastAccessedAt: now }; @@ -20525,7 +20958,7 @@ class SiftLightService { } mediaSearch(value, cwd, signal) { if (!this.#media) - throw new SiftLightError("Visual search is disabled; configure nativeMedia and restart the host"); + throw new SiftLightError("Visual search is disabled; configure nativeMedia or media and restart the host"); return this.#media.search(value, cwd, signal); } #dataSources; @@ -20541,7 +20974,8 @@ class SiftLightService { #operations; #reusableSummarySnapshots = new WeakSet; constructor(options) { - this.#media = options.nativeMedia ? new MediaSearch(options.nativeMedia) : undefined; + const mediaConfig = options.nativeMedia ?? options.media; + this.#media = mediaConfig ? new MediaSearch(mediaConfig) : undefined; this.#dataSources = new DataSources(options.dataSources); this.#runRipgrep = options.runRipgrep; this.#snapshots = options.snapshots ?? new SnapshotStore; diff --git a/plugins/sift-light/package.json b/plugins/sift-light/package.json index b0b78e97..4146ff54 100644 --- a/plugins/sift-light/package.json +++ b/plugins/sift-light/package.json @@ -1,6 +1,6 @@ { "name": "sift-light", - "version": "1.1.0-1", + "version": "1.1.0", "private": true, "type": "module", "omp": { diff --git a/src/config-reader.ts b/src/config-reader.ts index 661e0817..643cb639 100644 --- a/src/config-reader.ts +++ b/src/config-reader.ts @@ -2,6 +2,9 @@ import { parseDataSources, type DataSourceConfig } from "./data-source-config.js import { readFile } from "node:fs/promises"; import { join } from "node:path"; import { parseNativeTextConfig, type NativeTextConfig } from "./native-text-config.js"; +import type { MediaConfig } from "./media-config.js"; + +export type { MediaConfig } from "./media-config.js"; export const SIFT_LIGHT_CONFIG_FILE = "sift-light.json"; export const SIFT_LIGHT_CONFIG_ENV = "SIFT_LIGHT_CONFIG"; @@ -14,6 +17,7 @@ export const SIFT_LIGHT_ENFORCEMENT_ENV = "SIFT_LIGHT_ENFORCE_SEARCH"; export interface SiftLightConfig { nativeText?: NativeTextConfig; nativeMedia?: NativeTextConfig; + media?: MediaConfig; dataSources?: DataSourceConfig[]; locale: SiftLightLocale; enforceSearch?: SearchEnforcementMode; @@ -21,7 +25,7 @@ export interface SiftLightConfig { semanticJudge?: SemanticJudgeConfig; } -export type SemanticJudgeProvider = "jev"; +export type SemanticJudgeProvider = "jev" | "http-json"; export const SEMANTIC_JUDGE_API_KEY_ENVS = ["TYPESAFE_API_KEY", "SIFT_LIGHT_JEV_API_KEY"] as const; @@ -29,7 +33,7 @@ export interface SemanticJudgeConfig { enabled: boolean; provider: SemanticJudgeProvider; endpoint: string; - apiKeyEnv: string; + apiKeyEnv?: string; model: string; timeoutMs: number; maxCandidates: number; @@ -57,6 +61,7 @@ export const DEFAULT_SIFT_LIGHT_CONFIG: Readonly = { interface RawSiftLightConfig { nativeText?: unknown; nativeMedia?: unknown; + media?: unknown; dataSources?: unknown; locale?: unknown; enforceSearch?: unknown; @@ -87,6 +92,17 @@ interface RawSemanticJudgeConfig { maxRetries?: unknown; } +interface RawMediaConfig { + provider?: unknown; + endpoint?: unknown; + apiKeyEnv?: unknown; + model?: unknown; + timeoutMs?: unknown; + maxRetries?: unknown; + maxRequestBytes?: unknown; + renderer?: unknown; +} + function hasErrorCode(error: unknown, codes: string[]): boolean { return error instanceof Error && "code" in error && codes.includes(String(error.code)); } @@ -103,6 +119,58 @@ function isRawSemanticJudgeConfig(value: unknown): value is RawSemanticJudgeConf return typeof value === "object" && value !== null && !Array.isArray(value); } +function isRawMediaConfig(value: unknown): value is RawMediaConfig { + return typeof value === "object" && value !== null && !Array.isArray(value); +} + +function parseEndpoint(value: unknown, path: string, field: string): string { + if (typeof value !== "string" || value.length === 0 || value.length > 2_048) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be a nonempty URL`); + } + let parsedEndpoint: URL; + try { + parsedEndpoint = new URL(value); + } catch (error) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be a URL`, { + cause: error, + }); + } + const loopbackHosts = new Set(["localhost", "127.0.0.1", "[::1]"]); + const secureEndpoint = + parsedEndpoint.protocol === "https:" || + (parsedEndpoint.protocol === "http:" && loopbackHosts.has(parsedEndpoint.hostname)); + if (!secureEndpoint || parsedEndpoint.username || parsedEndpoint.password) { + throw new Error( + `Invalid sift-light config at ${path}: ${field} must use HTTPS, except that HTTP is allowed for localhost loopback development; URL credentials are not allowed`, + ); + } + return value; +} + +function parseApiKeyEnv(value: unknown, path: string, field: string): string | undefined { + if (value === undefined) return undefined; + if (typeof value !== "string" || !/^[A-Z][A-Z0-9_]{0,127}$/u.test(value)) { + throw new Error( + `Invalid sift-light config at ${path}: ${field} must be an uppercase environment variable name`, + ); + } + return value; +} + +function parseBoundedModel(value: unknown, path: string, field: string): string { + if ( + typeof value !== "string" || + value.length === 0 || + value.length > 128 || + /[\r\n\0]/u.test(value) + ) { + throw new Error( + `Invalid sift-light config at ${path}: ${field} must be bounded single-line text`, + ); + } + return value; +} + function boundedInteger( value: unknown, fallback: number, @@ -154,54 +222,29 @@ function parseSemanticJudge(value: unknown, path: string): SemanticJudgeConfig { ); } const provider = value.provider ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.provider; - if (provider !== "jev") { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.provider must be "jev"`); - } - const endpoint = value.endpoint ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.endpoint; - if (typeof endpoint !== "string" || endpoint.length === 0 || endpoint.length > 2_048) { - throw new Error( - `Invalid sift-light config at ${path}: semanticJudge.endpoint must be a nonempty URL`, - ); - } - let parsedEndpoint: URL; - try { - parsedEndpoint = new URL(endpoint); - } catch (error) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.endpoint must be a URL`, { - cause: error, - }); - } - const loopbackHosts = new Set(["localhost", "127.0.0.1", "[::1]"]); - const secureEndpoint = - parsedEndpoint.protocol === "https:" || - (parsedEndpoint.protocol === "http:" && loopbackHosts.has(parsedEndpoint.hostname)); - if (!secureEndpoint || parsedEndpoint.username || parsedEndpoint.password) { - throw new Error( - `Invalid sift-light config at ${path}: semanticJudge.endpoint must use HTTPS, except that HTTP is allowed for localhost loopback development; URL credentials are not allowed`, - ); - } - const apiKeyEnv = value.apiKeyEnv ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.apiKeyEnv; - if (typeof apiKeyEnv !== "string" || !/^[A-Z][A-Z0-9_]{0,127}$/u.test(apiKeyEnv)) { - throw new Error( - `Invalid sift-light config at ${path}: semanticJudge.apiKeyEnv must be an uppercase environment variable name`, - ); - } - const model = value.model ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.model; - if ( - typeof model !== "string" || - model.length === 0 || - model.length > 128 || - /[\r\n\0]/u.test(model) - ) { + if (provider !== "jev" && provider !== "http-json") { throw new Error( - `Invalid sift-light config at ${path}: semanticJudge.model must be bounded single-line text`, + `Invalid sift-light config at ${path}: semanticJudge.provider must be "jev" or "http-json"`, ); } + const endpoint = parseEndpoint( + value.endpoint ?? (provider === "jev" ? DEFAULT_SEMANTIC_JUDGE_CONFIG.endpoint : undefined), + path, + "semanticJudge.endpoint", + ); + const apiKeyEnv = + parseApiKeyEnv(value.apiKeyEnv, path, "semanticJudge.apiKeyEnv") ?? + (provider === "jev" ? DEFAULT_SEMANTIC_JUDGE_CONFIG.apiKeyEnv : undefined); + const model = parseBoundedModel( + value.model ?? (provider === "jev" ? DEFAULT_SEMANTIC_JUDGE_CONFIG.model : undefined), + path, + "semanticJudge.model", + ); return { enabled, provider, endpoint, - apiKeyEnv, + ...(apiKeyEnv === undefined ? {} : { apiKeyEnv }), model, timeoutMs: boundedInteger( value.timeoutMs, @@ -227,6 +270,61 @@ function parseSemanticJudge(value: unknown, path: string): SemanticJudgeConfig { }; } +function parseMedia(value: unknown, path: string): MediaConfig | undefined { + if (value === undefined) return undefined; + if (!isRawMediaConfig(value)) { + throw new Error(`Invalid sift-light config at ${path}: media must be an object`); + } + const unknown = Object.keys(value).filter( + (key) => + ![ + "provider", + "endpoint", + "apiKeyEnv", + "model", + "timeoutMs", + "maxRetries", + "maxRequestBytes", + "renderer", + ].includes(key), + ); + if (unknown.length > 0) { + throw new Error( + `Invalid sift-light config at ${path}: unsupported media fields; accepted fields are provider, endpoint, apiKeyEnv, model, timeoutMs, maxRetries, maxRequestBytes and renderer`, + ); + } + const provider = value.provider ?? "http-json"; + if (provider !== "http-json") { + throw new Error(`Invalid sift-light config at ${path}: media.provider must be "http-json"`); + } + const endpoint = parseEndpoint(value.endpoint, path, "media.endpoint"); + const apiKeyEnv = parseApiKeyEnv(value.apiKeyEnv, path, "media.apiKeyEnv"); + const model = parseBoundedModel(value.model, path, "media.model"); + const renderer = parseNativeTextConfig(value.renderer, "media.renderer"); + return { + provider, + endpoint, + ...(apiKeyEnv === undefined ? {} : { apiKeyEnv }), + model, + timeoutMs: boundedInteger( + value.timeoutMs, + 120_000, + 1_000, + 1_200_000, + `config at ${path}: media.timeoutMs`, + ), + maxRetries: boundedInteger(value.maxRetries, 2, 0, 5, `config at ${path}: media.maxRetries`), + maxRequestBytes: boundedInteger( + value.maxRequestBytes, + 32 * 1024 * 1024, + 1 * 1024 * 1024, + 64 * 1024 * 1024, + `config at ${path}: media.maxRequestBytes`, + ), + ...(renderer === undefined ? {} : { renderer }), + }; +} + function parseConfig(value: unknown, path: string): SiftLightConfig { if (!isRawSiftLightConfig(value)) { throw new Error(`Invalid sift-light config at ${path}: expected a JSON object`); @@ -241,11 +339,12 @@ function parseConfig(value: unknown, path: string): SiftLightConfig { "dataSources", "nativeText", "nativeMedia", + "media", ].includes(key), ); if (unknown.length > 0) { throw new Error( - `Invalid sift-light config at ${path}: unsupported configuration fields; only locale, enforceSearch, vectorSearchEnabled, semanticJudge, dataSources, nativeText and nativeMedia are accepted`, + `Invalid sift-light config at ${path}: unsupported configuration fields; only locale, enforceSearch, vectorSearchEnabled, semanticJudge, dataSources, nativeText, nativeMedia and media are accepted`, ); } const { locale, enforceSearch, vectorSearchEnabled } = value; @@ -258,11 +357,15 @@ function parseConfig(value: unknown, path: string): SiftLightConfig { const enforcement = normalizeSearchEnforcement(enforceSearch, `config at ${path}`); const nativeText = parseNativeTextConfig(value.nativeText); const nativeMedia = parseNativeTextConfig(value.nativeMedia, "nativeMedia"); + const media = parseMedia(value.media, `${path}: media`); + if (nativeMedia && media) + throw new Error(`Invalid sift-light config at ${path}: choose nativeMedia or media, not both`); if (nativeText && vectorSearchEnabled !== true) throw new Error("nativeText requires vectorSearchEnabled:true"); return { ...(nativeText ? { nativeText } : {}), ...(nativeMedia ? { nativeMedia } : {}), + ...(media ? { media } : {}), locale: locale ?? DEFAULT_SIFT_LIGHT_CONFIG.locale, enforceSearch: enforcement, vectorSearchEnabled: diff --git a/src/config.ts b/src/config.ts index f96fbcc2..6c16a518 100644 --- a/src/config.ts +++ b/src/config.ts @@ -18,6 +18,7 @@ export { type SearchEnforcementMode, type SemanticJudgeConfig, type SemanticJudgeProvider, + type MediaConfig, } from "./config-reader.js"; export function siftLightConfigPath( diff --git a/src/mcp-semantic-judge.ts b/src/mcp-semantic-judge.ts index 21c36744..1d23977c 100644 --- a/src/mcp-semantic-judge.ts +++ b/src/mcp-semantic-judge.ts @@ -1,5 +1,6 @@ import type { DataSourceConfig } from "./data-source-config.js"; import type { NativeTextConfig } from "./native-text-config.js"; +import type { MediaConfig } from "./media-config.js"; import { DEFAULT_SIFT_LIGHT_CONFIG, DEFAULT_SEMANTIC_JUDGE_CONFIG, @@ -30,6 +31,7 @@ export async function createMcpSearchFeatures( dataSources: DataSourceConfig[]; nativeText?: NativeTextConfig; nativeMedia?: NativeTextConfig; + media?: MediaConfig; }> { const path = configuredPath(environment); if (!path) @@ -46,6 +48,7 @@ export async function createMcpSearchFeatures( dataSources: config.dataSources ?? [], ...(config.nativeText ? { nativeText: config.nativeText } : {}), ...(config.nativeMedia ? { nativeMedia: config.nativeMedia } : {}), + ...(config.media ? { media: config.media } : {}), }; } diff --git a/src/mcp-server.mjs b/src/mcp-server.mjs index 8e2946af..94441c2f 100755 --- a/src/mcp-server.mjs +++ b/src/mcp-server.mjs @@ -1,7 +1,7 @@ #!/usr/bin/env node // src/mcp.ts -import { z as z3 } from "zod"; +import { z as z4 } from "zod"; // src/operation-lifecycle.ts import { randomUUID } from "node:crypto"; @@ -334,11 +334,11 @@ class OperationLifecycle { } // src/media-search.ts -import { createHash, randomUUID as randomUUID3 } from "node:crypto"; -import { mkdtemp, mkdir, readFile as readFile2, realpath as realpath3, rm, stat as stat4, writeFile } from "node:fs/promises"; +import { createHash, randomUUID as randomUUID4 } from "node:crypto"; +import { mkdtemp, mkdir, readFile as readFile3, realpath as realpath3, rm, stat as stat5, writeFile } from "node:fs/promises"; import { tmpdir } from "node:os"; -import { basename as basename2, dirname as dirname2, extname, join as join2, resolve as resolve7 } from "node:path"; -import { z } from "zod"; +import { basename as basename2, dirname as dirname2, extname as extname2, join as join2, resolve as resolve7 } from "node:path"; +import { z as z2 } from "zod"; // src/resident-worker-client.ts import { randomUUID as randomUUID2 } from "node:crypto"; @@ -715,11 +715,284 @@ class ResidentWorkerClient { } } +// src/remote-media-client.ts +import { randomUUID as randomUUID3 } from "node:crypto"; +import { readFile, stat } from "node:fs/promises"; +import { extname } from "node:path"; +import { z } from "zod"; +var MAX_RESPONSE_BYTES = 4 * 1024 * 1024; +var MAX_FILE_ID_LENGTH = 128; + +class RemoteMediaRequestError extends SiftLightError { + retryable; + constructor(message, retryable, options) { + super(message, options); + this.name = "RemoteMediaRequestError"; + this.retryable = retryable; + } +} +var remoteResponseSchema = z.object({ + version: z.literal(1), + indexedFiles: z.number().int().nonnegative(), + hits: z.array(z.object({ + fileId: z.string().min(1).max(MAX_FILE_ID_LENGTH), + score: z.number().finite(), + kind: z.enum(["image", "scan"]), + page: z.number().int().positive().nullable().optional(), + chunkCount: z.number().int().positive(), + width: z.number().int().nonnegative(), + height: z.number().int().nonnegative() + }).strict()).max(10) +}).strict(); +function mimeType(path) { + switch (extname(path).toLowerCase()) { + case ".png": + return "image/png"; + case ".jpg": + case ".jpeg": + return "image/jpeg"; + case ".gif": + return "image/gif"; + case ".bmp": + return "image/bmp"; + case ".tif": + case ".tiff": + return "image/tiff"; + case ".heic": + return "image/heic"; + case ".heif": + return "image/heif"; + case ".webp": + return "image/webp"; + case ".pdf": + return "application/pdf"; + default: + return "application/octet-stream"; + } +} +async function boundedResponseText(response) { + const contentLength = response.headers.get("content-length"); + if (contentLength !== null && Number(contentLength) > MAX_RESPONSE_BYTES) + throw new SiftLightError("Remote media response exceeded the 4 MiB limit"); + if (!response.body) { + const text = await response.text(); + if (new TextEncoder().encode(text).byteLength > MAX_RESPONSE_BYTES) + throw new SiftLightError("Remote media response exceeded the 4 MiB limit"); + return text; + } + const reader = response.body.getReader(); + const chunks = []; + let bytes = 0; + try { + while (true) { + const result = await reader.read(); + if (result.done) + break; + bytes += result.value.byteLength; + if (bytes > MAX_RESPONSE_BYTES) { + await reader.cancel(); + throw new SiftLightError("Remote media response exceeded the 4 MiB limit"); + } + chunks.push(result.value); + } + } finally { + reader.releaseLock(); + } + const all = new Uint8Array(bytes); + let offset = 0; + for (const chunk of chunks) { + all.set(chunk, offset); + offset += chunk.byteLength; + } + return new TextDecoder("utf-8", { fatal: true }).decode(all); +} +function retryableStatus(status) { + return status === 408 || status === 429 || status === 529 || status >= 500; +} +function retryDelay(attempt) { + return Math.min(20000, 400 * 2 ** attempt); +} +async function waitForRetry(delayMs, signal) { + if (signal?.aborted) + throw signal.reason instanceof Error ? signal.reason : new Error("Operation aborted"); + await new Promise((resolve, reject) => { + const timer = setTimeout(() => { + signal?.removeEventListener("abort", onAbort); + resolve(); + }, delayMs); + const onAbort = () => { + clearTimeout(timer); + signal?.removeEventListener("abort", onAbort); + reject(signal?.reason instanceof Error ? signal.reason : new Error("Operation aborted")); + }; + signal?.addEventListener("abort", onAbort, { once: true }); + }); +} +function base64(bytes) { + return bytes.toString("base64"); +} +function base64Length(bytes) { + return Math.ceil(bytes / 3) * 4; +} + +class RemoteMediaClient { + #config; + #key; + #fetcher; + constructor(config, environment = process.env, fetcher = fetch) { + this.#config = config; + this.#key = config.apiKeyEnv ? environment[config.apiKeyEnv]?.trim() : undefined; + if (config.apiKeyEnv && !this.#key) { + throw new SiftLightError(`Media provider is configured with ${config.apiKeyEnv}, but that environment variable is missing or empty`); + } + this.#fetcher = fetcher; + } + async search(request) { + const id = randomUUID3(); + const paths = new Map; + const files = []; + const fileSizes = []; + for (const file of request.files) { + if (file.id.length === 0 || file.id.length > MAX_FILE_ID_LENGTH || /[\r\n\0]/u.test(file.id) || paths.has(file.id)) + throw new SiftLightError("Remote media file ids must be unique bounded text"); + paths.set(file.id, file.path); + const metadata = await stat(file.path); + if (!metadata.isFile()) + throw new SiftLightError("Remote media files must be regular files"); + fileSizes.push(metadata.size); + } + let imageSize; + if (request.imagePath !== undefined) { + const metadata = await stat(request.imagePath); + if (!metadata.isFile()) + throw new SiftLightError("Remote media query image must be a regular file"); + imageSize = metadata.size; + } + const requestSkeleton = JSON.stringify({ + version: 1, + model: this.#config.model, + ...request.query === undefined ? {} : { query: request.query }, + ...request.imagePath === undefined ? {} : { image: { mimeType: mimeType(request.imagePath), data: "" } }, + kind: request.kind, + expectedFiles: request.expectedFiles, + files: [] + }); + let estimatedBytes = new TextEncoder().encode(requestSkeleton).byteLength; + if (imageSize !== undefined) + estimatedBytes += base64Length(imageSize); + for (let index = 0;index < request.files.length; index += 1) { + const file = request.files[index]; + const fileSkeleton = JSON.stringify({ + id: file?.id, + mimeType: file ? mimeType(file.path) : "application/octet-stream", + data: "" + }); + estimatedBytes += new TextEncoder().encode(fileSkeleton).byteLength; + estimatedBytes += base64Length(fileSizes[index] ?? 0); + } + if (request.files.length > 1) + estimatedBytes += request.files.length - 1; + if (estimatedBytes > this.#config.maxRequestBytes) + throw new SiftLightError(`Remote media request exceeded the configured ${String(this.#config.maxRequestBytes)} byte limit`); + for (const file of request.files) { + files.push({ + id: file.id, + mimeType: mimeType(file.path), + data: base64(await readFile(file.path, { signal: request.signal })) + }); + } + const image = request.imagePath === undefined ? undefined : { + mimeType: mimeType(request.imagePath), + data: base64(await readFile(request.imagePath, { signal: request.signal })) + }; + const body = JSON.stringify({ + version: 1, + model: this.#config.model, + ...request.query === undefined ? {} : { query: request.query }, + ...image === undefined ? {} : { image }, + kind: request.kind, + expectedFiles: request.expectedFiles, + files + }); + if (new TextEncoder().encode(body).byteLength > this.#config.maxRequestBytes) + throw new SiftLightError(`Remote media request exceeded the configured ${String(this.#config.maxRequestBytes)} byte limit`); + for (let attempt = 0;attempt <= this.#config.maxRetries; attempt += 1) { + const controller = new AbortController; + const onParentAbort = () => controller.abort(request.signal?.reason); + request.signal?.addEventListener("abort", onParentAbort, { once: true }); + const timeout = setTimeout(() => controller.abort(new Error("Remote media request timed out")), this.#config.timeoutMs); + try { + const headers = { "content-type": "application/json" }; + if (this.#key) + headers.authorization = `Bearer ${this.#key}`; + const response = await this.#fetcher(this.#config.endpoint, { + method: "POST", + headers, + body, + signal: controller.signal + }); + if (!response.ok) { + throw new RemoteMediaRequestError(`Remote media provider returned HTTP ${String(response.status)}`, retryableStatus(response.status)); + } + let parsed; + try { + const raw = JSON.parse(await boundedResponseText(response)); + parsed = remoteResponseSchema.parse(raw); + } catch (error) { + throw new SiftLightError("Remote media provider returned an invalid response", { + cause: error + }); + } + if (parsed.indexedFiles !== request.expectedFiles) + throw new SiftLightError("Remote media provider indexed an unexpected file count"); + return { + version: 1, + id, + indexedFiles: parsed.indexedFiles, + modelLoads: 0, + modelReused: true, + processId: process.pid, + hits: parsed.hits.map((hit) => { + const path = paths.get(hit.fileId); + if (!path) + throw new SiftLightError("Remote media provider returned an unknown file id"); + return { + path, + score: hit.score, + kind: hit.kind, + page: hit.page ?? null, + chunkCount: hit.chunkCount, + width: hit.width, + height: hit.height + }; + }) + }; + } catch (error) { + if (request.signal?.aborted) + throw error; + const retryable = error instanceof RemoteMediaRequestError ? error.retryable : !(error instanceof SiftLightError); + if (!retryable || attempt >= this.#config.maxRetries) + throw error; + await waitForRetry(retryDelay(attempt), request.signal); + } finally { + clearTimeout(timeout); + request.signal?.removeEventListener("abort", onParentAbort); + } + } + throw new SiftLightError("Remote media request failed"); + } + async close() {} +} +function mediaMimeType(path) { + const type = mimeType(path); + return type === "application/octet-stream" ? undefined : type; +} + // src/workspace-files.ts import { relative as relative3, resolve as resolve6, sep as sep2 } from "node:path"; // src/path-policy.ts -import { realpath, stat } from "node:fs/promises"; +import { realpath, stat as stat2 } from "node:fs/promises"; import { homedir } from "node:os"; import { dirname, isAbsolute, join, relative, resolve, sep } from "node:path"; var POSIX_SPECIAL_ROOTS = ["/dev", "/proc", "/sys"]; @@ -870,7 +1143,7 @@ class SearchPathPolicy { return this.cwd; let isDirectory; try { - isDirectory = (await stat(absolute)).isDirectory(); + isDirectory = (await stat2(absolute)).isDirectory(); } catch (error) { if (error instanceof Error && "code" in error && error.code === "ENOENT") return this.cwd; @@ -1118,7 +1391,7 @@ async function consumeCappedLines(stream, onLine, options = {}) { // src/ripgrep-executable.ts import { constants } from "node:fs"; -import { access, stat as stat2 } from "node:fs/promises"; +import { access, stat as stat3 } from "node:fs/promises"; import { delimiter, isAbsolute as isAbsolute2, resolve as resolve2 } from "node:path"; var OVERRIDE_ENV = "SIFT_LIGHT_RG_PATH"; var PATH_RIPGREP_NAME = process.platform === "win32" ? "rg.exe" : "rg"; @@ -1130,7 +1403,7 @@ async function findExecutableOnPath() { const candidates = pathValue.split(delimiter).map((directory) => resolve2(directory || ".", PATH_RIPGREP_NAME)); const available = await Promise.all(candidates.map(async (candidate) => { try { - if (!(await stat2(candidate)).isFile()) + if (!(await stat3(candidate)).isFile()) return; await access(candidate, constants.X_OK); return candidate; @@ -1158,7 +1431,7 @@ async function resolveRipgrepExecutable() { } } try { - if (!(await stat2(executable)).isFile()) + if (!(await stat3(executable)).isFile()) throw new Error("Expected an executable file"); await access(executable, constants.X_OK); } catch (cause) { @@ -1413,12 +1686,12 @@ function describeUnreadableDiagnostics(diagnostics) { } // src/source.ts -import { readFile, realpath as realpath2, stat as stat3 } from "node:fs/promises"; +import { readFile as readFile2, realpath as realpath2, stat as stat4 } from "node:fs/promises"; var SOURCE_RANGE_METADATA_RESERVE_BYTES = 1024; var MAX_SOURCE_RANGE_BYTES = MAX_RESULT_BYTES - SOURCE_RANGE_METADATA_RESERVE_BYTES; async function getSourceRevision(path, onError) { try { - const metadata = await stat3(path); + const metadata = await stat4(path); return sourceRevisionFromStats(metadata); } catch (error) { onError?.(error); @@ -2195,40 +2468,40 @@ async function listWorkspaceFiles(cwd, signal, options = {}) { } // src/media-search.ts -var mediaRequestSchema = z.discriminatedUnion("mode", [ - z.object({ mode: z.enum(["await", "cancel"]), operationId: z.string().uuid() }).strict(), - z.object({ - mode: z.literal("search"), - path: z.string().min(1).max(4096), - query: z.string().min(1).max(256).optional(), - image: z.string().min(1).max(4096).optional(), - kind: z.enum(["all", "image", "scan"]).default("all"), - glob: z.array(z.string().max(4096)).max(64).optional(), - exclude: z.array(z.string().max(4096)).max(64).optional(), - hidden: z.boolean().default(true), - ignorePolicy: z.enum(["respect", "include"]).default("respect") +var mediaRequestSchema = z2.discriminatedUnion("mode", [ + z2.object({ mode: z2.enum(["await", "cancel"]), operationId: z2.string().uuid() }).strict(), + z2.object({ + mode: z2.literal("search"), + path: z2.string().min(1).max(4096), + query: z2.string().min(1).max(256).optional(), + image: z2.string().min(1).max(4096).optional(), + kind: z2.enum(["all", "image", "scan"]).default("all"), + glob: z2.array(z2.string().max(4096)).max(64).optional(), + exclude: z2.array(z2.string().max(4096)).max(64).optional(), + hidden: z2.boolean().default(true), + ignorePolicy: z2.enum(["respect", "include"]).default("respect") }).strict().refine((r) => r.query === undefined !== (r.image === undefined), "Supply exactly one query or image"), - z.object({ - mode: z.enum(["inspect", "validate"]), - cursor: z.string().uuid(), - matchIndex: z.number().int().min(1).max(10) + z2.object({ + mode: z2.enum(["inspect", "validate"]), + cursor: z2.string().uuid(), + matchIndex: z2.number().int().min(1).max(10) }).strict() ]); -var responseSchema = z.object({ - version: z.literal(1), - id: z.string(), - indexedFiles: z.number().int().nonnegative(), - modelLoads: z.literal(1), - modelReused: z.boolean(), - processId: z.number().int().positive(), - hits: z.array(z.object({ - path: z.string(), - score: z.number().finite(), - kind: z.enum(["image", "scan"]), - page: z.number().int().positive().nullish().transform((value) => value ?? null), - chunkCount: z.number().int().positive(), - width: z.number().int().nonnegative(), - height: z.number().int().nonnegative() +var responseSchema = z2.object({ + version: z2.literal(1), + id: z2.string(), + indexedFiles: z2.number().int().nonnegative(), + modelLoads: z2.number().int().nonnegative(), + modelReused: z2.boolean(), + processId: z2.number().int().positive(), + hits: z2.array(z2.object({ + path: z2.string(), + score: z2.number().finite(), + kind: z2.enum(["image", "scan"]), + page: z2.number().int().positive().nullish().transform((value) => value ?? null), + chunkCount: z2.number().int().positive(), + width: z2.number().int().nonnegative(), + height: z2.number().int().nonnegative() })).max(10) }); var extensions = new Set([ @@ -2254,11 +2527,11 @@ async function sourceBytes(cwd, path, signal) { const canonical = await policy.resolveExistingPath(path); if (!canonical) throw new SiftLightError(`Media source unavailable: ${path}`); - const before = await stat4(canonical); + const before = await stat5(canonical); if (!before.isFile() || before.size > maxFileBytes) throw new SiftLightError(`Media source must be a regular file at most ${String(maxFileBytes)} bytes: ${path}`); - const bytes = await readFile2(canonical, { signal }); - const after = await stat4(canonical); + const bytes = await readFile3(canonical, { signal }); + const after = await stat5(canonical); if (bytes.length > maxFileBytes || before.ino !== after.ino || before.size !== after.size || before.mtimeMs !== after.mtimeMs || before.ctimeMs !== after.ctimeMs || await policy.resolveExistingPath(path) !== canonical) throw new SiftLightError(`Media source changed during read: ${path}`); return { canonical, bytes }; @@ -2266,6 +2539,7 @@ async function sourceBytes(cwd, path, signal) { class MediaSearch { #config; + #remoteConfig; #operations = new OperationLifecycle({ deadlineMs: 300000 }); #queue = new OwnedTaskQueue; #lifecycle = new AbortController; @@ -2273,9 +2547,13 @@ class MediaSearch { #known = new Map; #directory; #worker; + #remote; #closed = false; constructor(config) { - this.#config = config; + if ("provider" in config) + this.#remoteConfig = config; + else + this.#config = config; } async search(value, cwd, signal) { const request = mediaRequestSchema.parse(value); @@ -2331,7 +2609,7 @@ class MediaSearch { const reasons = [...inventory.reasons]; let total = 0; for (const path of inventory.paths) { - if (!extensions.has(extname(path).toLowerCase())) + if (!extensions.has(extname2(path).toLowerCase())) continue; if (signal.aborted) throw abortError(); @@ -2369,20 +2647,37 @@ class MediaSearch { let imageSource; let queryImage; if (r.image) { - if (!extensions.has(extname(r.image).toLowerCase()) || extname(r.image).toLowerCase() === ".pdf") + if (!extensions.has(extname2(r.image).toLowerCase()) || extname2(r.image).toLowerCase() === ".pdf") throw new SiftLightError("image must be a supported local image file"); imageSource = await sourceBytes(cwd, r.image, signal); - queryImage = join2(root, "query" + extname(r.image).toLowerCase()); + queryImage = join2(root, "query" + extname2(r.image).toLowerCase()); await writeFile(queryImage, imageSource.bytes, { mode: 384 }); } let response; if (admitted.size) { - this.#worker ??= new ResidentWorkerClient(this.#config, root, corpus, join2(root, "index.sqlite"), { parse: (value) => responseSchema.parse(value), maxResponseBytes: 128 * 1024 }); - response = await this.#worker.search({ - ...queryImage ? { image: queryImage } : { query: r.query }, - kind: r.kind, - expectedFiles: admitted.size - }, signal); + if (this.#config) { + this.#worker ??= new ResidentWorkerClient(this.#config, root, corpus, join2(root, "index.sqlite"), { parse: (value) => responseSchema.parse(value), maxResponseBytes: 128 * 1024 }); + response = await this.#worker.search({ + ...queryImage ? { image: queryImage } : { query: r.query }, + kind: r.kind, + expectedFiles: admitted.size + }, signal); + } else if (this.#remoteConfig) { + this.#remote ??= new RemoteMediaClient(this.#remoteConfig); + const remoteFiles = [...admitted.values()].map((entry) => ({ + id: createHash("sha256").update(entry.staged).digest("hex").slice(0, 32), + path: entry.staged + })); + response = await this.#remote.search({ + ...queryImage ? { imagePath: queryImage } : { query: r.query }, + kind: r.kind, + expectedFiles: admitted.size, + files: remoteFiles, + signal + }); + } + if (!response) + throw new SiftLightError("Media provider is not configured"); if (response.indexedFiles !== admitted.size) throw new SiftLightError("Media index inventory mismatch"); } @@ -2390,7 +2685,7 @@ class MediaSearch { const seen = new Set; for (const hit of response?.hits ?? []) { const source = admitted.get(resolve7(hit.path)); - if (!source || hit.kind === "scan" !== (extname(source.path).toLowerCase() === ".pdf") || seen.has(hit.path) || hit.kind === "scan" !== (hit.page !== null) || hit.page !== null && hit.page > hit.chunkCount) + if (!source || hit.kind === "scan" !== (extname2(source.path).toLowerCase() === ".pdf") || seen.has(hit.path) || hit.kind === "scan" !== (hit.page !== null) || hit.page !== null && hit.page > hit.chunkCount) throw new SiftLightError("Media result source or page is inconsistent"); seen.add(hit.path); saved.push({ hit, source }); @@ -2410,7 +2705,7 @@ class MediaSearch { } if (signal.aborted) throw abortError(); - const cursor = randomUUID3(); + const cursor = randomUUID4(); for (const [id, snapshot] of this.#snapshots) if (Date.now() - snapshot.createdAt > 600000) this.#snapshots.delete(id); @@ -2473,11 +2768,23 @@ ${JSON.stringify(details)}`, throw new SiftLightError("Media source changed; search again before inspection"); const folder = await mkdtemp(join2(await this.#root(), "inspect-")); try { - const input = join2(folder, "source" + extname(item.source.path)); + const input = join2(folder, "source" + extname2(item.source.path)); const output = join2(folder, "preview.png"); await writeFile(input, current.bytes, { mode: 384 }); + const renderer = this.#config ?? this.#remoteConfig?.renderer; + if (!renderer) { + const mimeType = mediaMimeType(item.source.path); + if (mimeType?.startsWith("image/") && mimeType === "image/png") { + return { + text: JSON.stringify(details), + details, + image: { data: current.bytes.toString("base64"), mimeType } + }; + } + throw new SiftLightError("Media preview requires nativeMedia or media.renderer; validate still works without a local renderer"); + } const result = await runOwnedProcess({ - executable: this.#config.executable, + executable: renderer.executable, args: ["--render", input, String(item.hit.page ?? 0), output], cwd: folder, signal @@ -2488,7 +2795,7 @@ ${JSON.stringify(details)}`, }); if (result.code !== 0) throw new SiftLightError(`Media preview failed: ${result.stderr}`); - const preview = await readFile2(output); + const preview = await readFile3(output); if (preview.length > 4 * 1024 * 1024) throw new SiftLightError("Media preview exceeds 4 MiB"); return { @@ -2513,6 +2820,7 @@ ${JSON.stringify(details)}`, await this.#queue.run(async () => { try { await this.#worker?.close(); + await this.#remote?.close(); } finally { this.#snapshots.clear(); if (this.#directory) @@ -2523,13 +2831,13 @@ ${JSON.stringify(details)}`, } // src/mcp.ts -import { randomUUID as randomUUID8 } from "node:crypto"; +import { randomUUID as randomUUID9 } from "node:crypto"; import { createServer } from "node:http"; import { URL as URL2 } from "node:url"; // package.json var package_default = { name: "sift-light", - version: "1.1.0-1", + version: "1.1.0", description: "Context-efficient local search for files, documents, notes and logs across Pi, OMP and MCP clients", keywords: [ "ai-agent", @@ -3035,7 +3343,7 @@ function boundedRequestContractDetails(details) { } // src/language-capability-definitions.ts -import { extname as extname2 } from "node:path"; +import { extname as extname3 } from "node:path"; var TYPESCRIPT_LANGUAGE_CAPABILITIES = [ { id: "ast-grep.outline", @@ -3184,14 +3492,14 @@ var DEFAULT_NEUTRAL_CAPABILITIES = [ } ]; function descriptorForPath(descriptors, path) { - const extension = extname2(path).toLowerCase(); + const extension = extname3(path).toLowerCase(); return descriptors.find((item) => item.extensions.includes(extension)); } function languageForPath(descriptors, path) { return descriptorForPath(descriptors, path)?.language; } function outlineExtension(path) { - const extension = extname2(path).toLowerCase(); + const extension = extname3(path).toLowerCase(); return extension || undefined; } @@ -4134,7 +4442,7 @@ class DataSourceError extends Error { this.code = code; } } -var MAX_RESPONSE_BYTES = 1024 * 1024; +var MAX_RESPONSE_BYTES2 = 1024 * 1024; async function credential(config) { let value; if (config.tokenEnv) @@ -4189,7 +4497,7 @@ async function sourcePost(config, path, payload, signal) { if (part.done) break; size += part.value.byteLength; - if (size > MAX_RESPONSE_BYTES) { + if (size > MAX_RESPONSE_BYTES2) { await reader.cancel(); throw new DataSourceError("protocol"); } @@ -4439,14 +4747,14 @@ class DataSources { import { createHash as createHash7 } from "node:crypto"; // src/concept-worker-client.ts -import { randomUUID as randomUUID4 } from "node:crypto"; +import { randomUUID as randomUUID5 } from "node:crypto"; import { mkdir as mkdir2, rm as rm2 } from "node:fs/promises"; import { dirname as dirname3, join as join4 } from "node:path"; import { fileURLToPath } from "node:url"; import { StringDecoder as StringDecoder2 } from "node:string_decoder"; // src/concept-model.ts -import { stat as stat5 } from "node:fs/promises"; +import { stat as stat6 } from "node:fs/promises"; import { homedir as homedir2 } from "node:os"; import { join as join3, resolve as resolve9 } from "node:path"; var CONCEPT_MODEL = "Xenova/multilingual-e5-small"; @@ -4489,7 +4797,7 @@ function conceptModelDirectory() { async function inspectConceptModel(directory = conceptModelDirectory()) { for (const asset of CONCEPT_ASSETS) { try { - const information = await stat5(join3(directory, asset.path)); + const information = await stat6(join3(directory, asset.path)); if (!information.isFile() || information.size !== asset.bytes) return { installed: false, @@ -4585,7 +4893,7 @@ class ConceptWorkerClient { throw new SiftLightError("Concept worker idle timeout must be 1–300000 ms"); } async#start() { - const stagingRoot = join4(conceptCacheDirectory(), ".staging", randomUUID4()); + const stagingRoot = join4(conceptCacheDirectory(), ".staging", randomUUID5()); await mkdir2(stagingRoot, { recursive: true }); const ready = Promise.withResolvers(); ready.promise.catch(() => { @@ -4719,7 +5027,7 @@ class ConceptWorkerClient { const input = await session.ready; if (combined.aborted) throw abortError(); - const id = randomUUID4(); + const id = randomUUID5(); const body = JSON.stringify({ id, query, @@ -5657,7 +5965,7 @@ async function parseSyntax(path, text, signal, pattern) { import { createHash as createHash4 } from "node:crypto"; // src/source-access.ts -import { extname as extname3, resolve as resolve13 } from "node:path"; +import { extname as extname4, resolve as resolve13 } from "node:path"; // src/historical-paths.ts import { lstat, mkdir as mkdir3, mkdtemp as mkdtemp2, open as open2, rm as rm3, writeFile as writeFile2 } from "node:fs/promises"; @@ -6790,7 +7098,7 @@ class SyntaxQueue { throw new SourceDocumentError("encoding", "Syntax requires lossless UTF-8 source"); const origin = document.reference.origin; const revision = origin.kind === "worktree" ? origin.contentHash : origin.blob; - const key = `${extname3(document.path).toLowerCase()}\x00${revision}\x00${pattern ?? ""}`; + const key = `${extname4(document.path).toLowerCase()}\x00${revision}\x00${pattern ?? ""}`; const cached = this.#cache.get(key); if (cached) { this.#cache.delete(key); @@ -7569,25 +7877,25 @@ async function structuralSearch(input, access) { // src/native-text-search.ts import { createHash as createHash6 } from "node:crypto"; -import { mkdtemp as mkdtemp3, mkdir as mkdir4, readFile as readFile3, realpath as realpath5, rm as rm4, writeFile as writeFile3 } from "node:fs/promises"; +import { mkdtemp as mkdtemp3, mkdir as mkdir4, readFile as readFile4, realpath as realpath5, rm as rm4, writeFile as writeFile3 } from "node:fs/promises"; import { tmpdir as tmpdir3 } from "node:os"; import { basename as basename3, dirname as dirname6, isAbsolute as isAbsolute7, join as join6, resolve as resolve15 } from "node:path"; // src/native-text-worker-client.ts -import { z as z2 } from "zod"; -var responseSchema2 = z2.object({ - version: z2.literal(2), - id: z2.string(), - indexedFiles: z2.number().int().nonnegative(), - modelLoads: z2.literal(1), - modelReused: z2.boolean(), - processId: z2.number().int().positive(), - hits: z2.array(z2.object({ - path: z2.string(), - score: z2.number().finite(), - line: z2.number().int().positive(), - chunkIndex: z2.number().int().nonnegative(), - locator: z2.string() +import { z as z3 } from "zod"; +var responseSchema2 = z3.object({ + version: z3.literal(2), + id: z3.string(), + indexedFiles: z3.number().int().nonnegative(), + modelLoads: z3.literal(1), + modelReused: z3.boolean(), + processId: z3.number().int().positive(), + hits: z3.array(z3.object({ + path: z3.string(), + score: z3.number().finite(), + line: z3.number().int().positive(), + chunkIndex: z3.number().int().nonnegative(), + locator: z3.string() })).max(10) }); @@ -7694,7 +8002,7 @@ function createNativeTextSearch(config, search) { if (!entry || seen.has(entry.staged)) throw new SiftLightError("Native text returned an unknown or duplicate source"); seen.add(entry.staged); - const content = await readFile3(entry.staged); + const content = await readFile4(entry.staged); if (createHash6("sha256").update(content).digest("hex") !== entry.hash) throw new SiftLightError("Native staged source changed"); const document = new SourceDocument(entry.source, content); @@ -7775,10 +8083,10 @@ function rankRetrievalFiles(current, native) { } // src/evidence-service.ts -import { dirname as dirname7, extname as extname4, resolve as resolve24 } from "node:path"; +import { dirname as dirname7, extname as extname5, resolve as resolve24 } from "node:path"; // src/analysis-store.ts -import { randomUUID as randomUUID5 } from "node:crypto"; +import { randomUUID as randomUUID6 } from "node:crypto"; // src/analysis-types.ts var SEMANTIC_JUDGE_CLASSIFICATIONS = [ @@ -8060,7 +8368,7 @@ class AnalysisStore { } create(result, summarize) { this.#expire(); - const id = randomUUID5(); + const id = randomUUID6(); const bounded = { ...result, ...result.retrievalLanes ? { @@ -10305,7 +10613,7 @@ async function discoverFiles(input, cwd, signal) { } // src/source-continuations.ts -import { randomUUID as randomUUID6 } from "node:crypto"; +import { randomUUID as randomUUID7 } from "node:crypto"; // src/source-pages.ts function mergeByteRanges(ranges) { @@ -10409,7 +10717,7 @@ class SourceContinuations { create(source, target, gaps, boundary) { this.#sweep(); const item = { - id: randomUUID6(), + id: randomUUID7(), source: structuredClone(source), target: mergeByteRanges(target), gaps: mergeByteRanges(gaps), @@ -11005,7 +11313,7 @@ async function continueSource(cursor, access, continuations, expectedPath) { } // src/evidence-validation.ts -import { realpath as realpath7, stat as stat6 } from "node:fs/promises"; +import { realpath as realpath7, stat as stat7 } from "node:fs/promises"; import { resolve as resolve22 } from "node:path"; // src/evidence-validity.ts @@ -11043,9 +11351,9 @@ async function confirmWorktreeState(path, cwd, signal) { const policy = new SearchPathPolicy(cwd); try { policy.assertPath(absolute); - const before = await stat6(absolute); + const before = await stat7(absolute); const beforeCanonical = await realpath7(absolute); - const after = await stat6(absolute); + const after = await stat7(absolute); const afterCanonical = await realpath7(absolute); if (!before.isFile() || !after.isFile()) { return { status: "stale", reason: "Source is no longer a regular file" }; @@ -11187,9 +11495,9 @@ async function validateSnapshotTarget(target, cwd, policy, signal) { const result = await confirmWorktreeState(target.path, cwd, signal); return { path: target.path, role: "source", ...result }; } - const before = await stat6(absolute); + const before = await stat7(absolute); const beforeCanonical = canonical; - const after = await stat6(absolute); + const after = await stat7(absolute); const afterCanonical = await realpath7(absolute); if (!before.isFile() || !after.isFile()) { return { @@ -12063,7 +12371,7 @@ function pythonRoleAnalysis(document) { import { resolve as resolve23 } from "node:path"; // src/config-reader.ts -import { readFile as readFile4 } from "node:fs/promises"; +import { readFile as readFile5 } from "node:fs/promises"; // src/native-text-config.ts import { isAbsolute as isAbsolute8 } from "node:path"; @@ -12111,6 +12419,42 @@ function isRawSiftLightConfig(value) { function isRawSemanticJudgeConfig(value) { return typeof value === "object" && value !== null && !Array.isArray(value); } +function isRawMediaConfig(value) { + return typeof value === "object" && value !== null && !Array.isArray(value); +} +function parseEndpoint(value, path, field) { + if (typeof value !== "string" || value.length === 0 || value.length > 2048) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be a nonempty URL`); + } + let parsedEndpoint; + try { + parsedEndpoint = new URL(value); + } catch (error) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be a URL`, { + cause: error + }); + } + const loopbackHosts = new Set(["localhost", "127.0.0.1", "[::1]"]); + const secureEndpoint = parsedEndpoint.protocol === "https:" || parsedEndpoint.protocol === "http:" && loopbackHosts.has(parsedEndpoint.hostname); + if (!secureEndpoint || parsedEndpoint.username || parsedEndpoint.password) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must use HTTPS, except that HTTP is allowed for localhost loopback development; URL credentials are not allowed`); + } + return value; +} +function parseApiKeyEnv(value, path, field) { + if (value === undefined) + return; + if (typeof value !== "string" || !/^[A-Z][A-Z0-9_]{0,127}$/u.test(value)) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be an uppercase environment variable name`); + } + return value; +} +function parseBoundedModel(value, path, field) { + if (typeof value !== "string" || value.length === 0 || value.length > 128 || /[\r\n\0]/u.test(value)) { + throw new Error(`Invalid sift-light config at ${path}: ${field} must be bounded single-line text`); + } + return value; +} function boundedInteger2(value, fallback, minimum, maximum, field) { const candidate = value ?? fallback; if (typeof candidate !== "number" || !Number.isSafeInteger(candidate) || candidate < minimum || candidate > maximum) { @@ -12142,45 +12486,61 @@ function parseSemanticJudge(value, path) { throw new Error(`Invalid sift-light config at ${path}: semanticJudge.enabled must be a boolean`); } const provider = value.provider ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.provider; - if (provider !== "jev") { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.provider must be "jev"`); - } - const endpoint = value.endpoint ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.endpoint; - if (typeof endpoint !== "string" || endpoint.length === 0 || endpoint.length > 2048) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.endpoint must be a nonempty URL`); - } - let parsedEndpoint; - try { - parsedEndpoint = new URL(endpoint); - } catch (error) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.endpoint must be a URL`, { - cause: error - }); - } - const loopbackHosts = new Set(["localhost", "127.0.0.1", "[::1]"]); - const secureEndpoint = parsedEndpoint.protocol === "https:" || parsedEndpoint.protocol === "http:" && loopbackHosts.has(parsedEndpoint.hostname); - if (!secureEndpoint || parsedEndpoint.username || parsedEndpoint.password) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.endpoint must use HTTPS, except that HTTP is allowed for localhost loopback development; URL credentials are not allowed`); - } - const apiKeyEnv = value.apiKeyEnv ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.apiKeyEnv; - if (typeof apiKeyEnv !== "string" || !/^[A-Z][A-Z0-9_]{0,127}$/u.test(apiKeyEnv)) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.apiKeyEnv must be an uppercase environment variable name`); - } - const model = value.model ?? DEFAULT_SEMANTIC_JUDGE_CONFIG.model; - if (typeof model !== "string" || model.length === 0 || model.length > 128 || /[\r\n\0]/u.test(model)) { - throw new Error(`Invalid sift-light config at ${path}: semanticJudge.model must be bounded single-line text`); + if (provider !== "jev" && provider !== "http-json") { + throw new Error(`Invalid sift-light config at ${path}: semanticJudge.provider must be "jev" or "http-json"`); } + const endpoint = parseEndpoint(value.endpoint ?? (provider === "jev" ? DEFAULT_SEMANTIC_JUDGE_CONFIG.endpoint : undefined), path, "semanticJudge.endpoint"); + const apiKeyEnv = parseApiKeyEnv(value.apiKeyEnv, path, "semanticJudge.apiKeyEnv") ?? (provider === "jev" ? DEFAULT_SEMANTIC_JUDGE_CONFIG.apiKeyEnv : undefined); + const model = parseBoundedModel(value.model ?? (provider === "jev" ? DEFAULT_SEMANTIC_JUDGE_CONFIG.model : undefined), path, "semanticJudge.model"); return { enabled, provider, endpoint, - apiKeyEnv, + ...apiKeyEnv === undefined ? {} : { apiKeyEnv }, model, timeoutMs: boundedInteger2(value.timeoutMs, DEFAULT_SEMANTIC_JUDGE_CONFIG.timeoutMs, 1000, 1200000, `config at ${path}: semanticJudge.timeoutMs`), maxCandidates: boundedInteger2(value.maxCandidates, DEFAULT_SEMANTIC_JUDGE_CONFIG.maxCandidates, 1, 20, `config at ${path}: semanticJudge.maxCandidates`), maxRetries: boundedInteger2(value.maxRetries, DEFAULT_SEMANTIC_JUDGE_CONFIG.maxRetries, 0, 5, `config at ${path}: semanticJudge.maxRetries`) }; } +function parseMedia(value, path) { + if (value === undefined) + return; + if (!isRawMediaConfig(value)) { + throw new Error(`Invalid sift-light config at ${path}: media must be an object`); + } + const unknown = Object.keys(value).filter((key) => ![ + "provider", + "endpoint", + "apiKeyEnv", + "model", + "timeoutMs", + "maxRetries", + "maxRequestBytes", + "renderer" + ].includes(key)); + if (unknown.length > 0) { + throw new Error(`Invalid sift-light config at ${path}: unsupported media fields; accepted fields are provider, endpoint, apiKeyEnv, model, timeoutMs, maxRetries, maxRequestBytes and renderer`); + } + const provider = value.provider ?? "http-json"; + if (provider !== "http-json") { + throw new Error(`Invalid sift-light config at ${path}: media.provider must be "http-json"`); + } + const endpoint = parseEndpoint(value.endpoint, path, "media.endpoint"); + const apiKeyEnv = parseApiKeyEnv(value.apiKeyEnv, path, "media.apiKeyEnv"); + const model = parseBoundedModel(value.model, path, "media.model"); + const renderer = parseNativeTextConfig(value.renderer, "media.renderer"); + return { + provider, + endpoint, + ...apiKeyEnv === undefined ? {} : { apiKeyEnv }, + model, + timeoutMs: boundedInteger2(value.timeoutMs, 120000, 1000, 1200000, `config at ${path}: media.timeoutMs`), + maxRetries: boundedInteger2(value.maxRetries, 2, 0, 5, `config at ${path}: media.maxRetries`), + maxRequestBytes: boundedInteger2(value.maxRequestBytes, 32 * 1024 * 1024, 1 * 1024 * 1024, 64 * 1024 * 1024, `config at ${path}: media.maxRequestBytes`), + ...renderer === undefined ? {} : { renderer } + }; +} function parseConfig(value, path) { if (!isRawSiftLightConfig(value)) { throw new Error(`Invalid sift-light config at ${path}: expected a JSON object`); @@ -12192,10 +12552,11 @@ function parseConfig(value, path) { "semanticJudge", "dataSources", "nativeText", - "nativeMedia" + "nativeMedia", + "media" ].includes(key)); if (unknown.length > 0) { - throw new Error(`Invalid sift-light config at ${path}: unsupported configuration fields; only locale, enforceSearch, vectorSearchEnabled, semanticJudge, dataSources, nativeText and nativeMedia are accepted`); + throw new Error(`Invalid sift-light config at ${path}: unsupported configuration fields; only locale, enforceSearch, vectorSearchEnabled, semanticJudge, dataSources, nativeText, nativeMedia and media are accepted`); } const { locale, enforceSearch, vectorSearchEnabled } = value; if (locale !== undefined && locale !== "en" && locale !== "zh-CN") { @@ -12207,11 +12568,15 @@ function parseConfig(value, path) { const enforcement = normalizeSearchEnforcement(enforceSearch, `config at ${path}`); const nativeText = parseNativeTextConfig(value.nativeText); const nativeMedia = parseNativeTextConfig(value.nativeMedia, "nativeMedia"); + const media = parseMedia(value.media, `${path}: media`); + if (nativeMedia && media) + throw new Error(`Invalid sift-light config at ${path}: choose nativeMedia or media, not both`); if (nativeText && vectorSearchEnabled !== true) throw new Error("nativeText requires vectorSearchEnabled:true"); return { ...nativeText ? { nativeText } : {}, ...nativeMedia ? { nativeMedia } : {}, + ...media ? { media } : {}, locale: locale ?? DEFAULT_SIFT_LIGHT_CONFIG.locale, enforceSearch: enforcement, vectorSearchEnabled: vectorSearchEnabled ?? DEFAULT_SIFT_LIGHT_CONFIG.vectorSearchEnabled ?? false, @@ -12230,7 +12595,7 @@ function normalizeSearchEnforcement(value, source) { } async function readSiftLightConfigFile(path, options = {}) { try { - const content = await readFile4(path, "utf8"); + const content = await readFile5(path, "utf8"); return parseConfig(JSON.parse(content), path); } catch (error) { if (isMissingFile(error)) { @@ -12329,7 +12694,7 @@ async function judgeSemanticCandidateBatches(runner, query, batches, signal) { // src/semantic-judge.ts var MAX_CANDIDATE_CHARS = 4000; -var MAX_RESPONSE_BYTES2 = 4 * 1024 * 1024; +var MAX_RESPONSE_BYTES3 = 4 * 1024 * 1024; var MAX_ERROR_CHARS = 512; var CLASSIFICATION_PRIORITY = { "implementation-candidate": 0, @@ -12419,14 +12784,22 @@ function requestBody(query, candidates, model) { questions }; } +function httpJsonRequestBody(query, candidates, model) { + return { + version: 1, + model, + query, + candidates: candidates.map(({ id, excerpt }) => ({ id, excerpt })) + }; +} async function responseText(response) { if (!response.body) { const contentLength = response.headers.get("content-length"); - if (contentLength !== null && Number(contentLength) > MAX_RESPONSE_BYTES2) { + if (contentLength !== null && Number(contentLength) > MAX_RESPONSE_BYTES3) { throw new SemanticJudgeRequestError("Semantic judge response exceeded the 4 MiB limit", false); } const text = await response.text(); - if (new TextEncoder().encode(text).byteLength > MAX_RESPONSE_BYTES2) { + if (new TextEncoder().encode(text).byteLength > MAX_RESPONSE_BYTES3) { throw new SemanticJudgeRequestError("Semantic judge response exceeded the 4 MiB limit", false); } return text; @@ -12441,7 +12814,7 @@ async function responseText(response) { break; const chunk = result.value; bytes += chunk.byteLength; - if (bytes > MAX_RESPONSE_BYTES2) { + if (bytes > MAX_RESPONSE_BYTES3) { await reader.cancel(); throw new SemanticJudgeRequestError("Semantic judge response exceeded the 4 MiB limit", false); } @@ -12458,10 +12831,10 @@ async function responseText(response) { } return new TextDecoder("utf-8", { fatal: true }).decode(all); } -function retryDelay(attempt) { +function retryDelay2(attempt) { return Math.min(20000, 400 * 2 ** attempt); } -async function waitForRetry(delayMs, signal) { +async function waitForRetry2(delayMs, signal) { if (signal?.aborted) throw signal.reason instanceof Error ? signal.reason : new Error("Operation aborted"); await new Promise((resolve, reject) => { @@ -12479,68 +12852,127 @@ async function waitForRetry(delayMs, signal) { signal?.addEventListener("abort", onAbort, { once: true }); }); } +async function fetchSemanticResponse(config, key, fetcher, body, parentSignal) { + if (new TextEncoder().encode(body).byteLength > MAX_SEMANTIC_JUDGE_REQUEST_BYTES) { + throw new SemanticJudgePayloadTooLargeError("Semantic judge request exceeded the 64 KiB request limit"); + } + for (let attempt = 0;attempt <= config.maxRetries; attempt += 1) { + const controller = new AbortController; + const onParentAbort = () => controller.abort(parentSignal?.reason); + parentSignal?.addEventListener("abort", onParentAbort, { once: true }); + const timeout = setTimeout(() => controller.abort(new Error("Semantic judge request timed out")), config.timeoutMs); + try { + const headers = { "content-type": "application/json" }; + if (key) + headers.authorization = `Bearer ${key}`; + const response = await fetcher(config.endpoint, { + method: "POST", + headers, + body, + signal: controller.signal + }); + if (!response.ok) { + if (response.status === 413) { + throw new SemanticJudgePayloadTooLargeError("Semantic judge rejected an oversized request with HTTP 413"); + } + const retryable = response.status === 408 || response.status === 429 || response.status === 529 || response.status >= 500; + throw new SemanticJudgeRequestError(`Semantic judge returned HTTP ${String(response.status)}`, retryable); + } + const raw = JSON.parse(await responseText(response)); + if (!isRecord4(raw)) { + throw new SemanticJudgeRequestError("Semantic judge response must be an object", false); + } + return raw; + } catch (error) { + if (parentSignal?.aborted) + throw error; + const retryable = error instanceof SemanticJudgePayloadTooLargeError ? false : error instanceof SemanticJudgeRequestError ? error.retryable : true; + if (!retryable || attempt >= config.maxRetries) + throw error; + await waitForRetry2(retryDelay2(attempt), parentSignal); + } finally { + clearTimeout(timeout); + parentSignal?.removeEventListener("abort", onParentAbort); + } + } + throw new SemanticJudgeRequestError("Semantic judge request failed", false); +} +function responseModel(raw, config) { + const model = raw.model ?? config.model; + if (typeof model !== "string" || model.length === 0 || model.length > 128 || /[\r\n\0]/u.test(model)) { + throw new SemanticJudgeRequestError("Semantic judge response model must be bounded single-line text", false); + } + return model; +} +function responseUsage(raw) { + const usage = isRecord4(raw.usage) ? raw.usage : undefined; + const inputTokens = typeof usage?.input_tokens === "number" && Number.isSafeInteger(usage.input_tokens) && usage.input_tokens >= 0 ? usage.input_tokens : undefined; + const outputTokens = typeof usage?.output_tokens === "number" && Number.isSafeInteger(usage.output_tokens) && usage.output_tokens >= 0 ? usage.output_tokens : undefined; + return { + ...inputTokens === undefined ? {} : { inputTokens }, + ...outputTokens === undefined ? {} : { outputTokens } + }; +} +function parseHttpJsonResult(raw, candidates, config, elapsedMs) { + if (raw.version !== 1 || !Array.isArray(raw.judgments)) { + throw new SemanticJudgeRequestError("HTTP JSON semantic judge response must contain version 1 judgments", false); + } + const candidateIndexes = new Map(candidates.map((candidate, index) => [candidate.id, index])); + const judgments = new Map; + for (const value of raw.judgments) { + if (!isRecord4(value) || typeof value.id !== "string") { + throw new SemanticJudgeRequestError("HTTP JSON semantic judge judgment has an invalid id", false); + } + const candidateIndex = candidateIndexes.get(value.id); + if (candidateIndex === undefined || judgments.has(candidateIndex)) { + throw new SemanticJudgeRequestError("HTTP JSON semantic judge response contains an unknown or duplicate candidate", false); + } + if (!isClassification(value.choice)) { + throw new SemanticJudgeRequestError(`HTTP JSON semantic judge response has an invalid choice for ${value.id}`, false); + } + const confidence = value.confidence === undefined ? undefined : probability(value.confidence, "confidence"); + judgments.set(candidateIndex, { + candidateIndex, + classification: value.choice, + probability: probability(value.probability, "probability"), + ...confidence === undefined ? {} : { confidence } + }); + } + if (judgments.size !== candidates.length) { + throw new SemanticJudgeRequestError(`HTTP JSON semantic judge response judged ${String(judgments.size)} of ${String(candidates.length)} candidates`, false); + } + const usage = responseUsage(raw); + return { + model: responseModel(raw, config), + judgments: candidates.map((_candidate, index) => judgments.get(index)), + ...usage, + elapsedMs + }; +} function createJevRunner(config, key, fetcher) { return async (query, candidates, parentSignal) => { const startedAt = performance.now(); const body = JSON.stringify(requestBody(query, candidates, config.model)); - if (new TextEncoder().encode(body).byteLength > MAX_SEMANTIC_JUDGE_REQUEST_BYTES) { - throw new SemanticJudgePayloadTooLargeError("Semantic judge request exceeded the 64 KiB request limit"); + const raw = await fetchSemanticResponse(config, key, fetcher, body, parentSignal); + if (!isRecord4(raw.answers)) { + throw new SemanticJudgeRequestError("Semantic judge response did not contain answers", false); } - for (let attempt = 0;attempt <= config.maxRetries; attempt += 1) { - const controller = new AbortController; - const onParentAbort = () => controller.abort(parentSignal?.reason); - parentSignal?.addEventListener("abort", onParentAbort, { once: true }); - const timeout = setTimeout(() => controller.abort(new Error("Semantic judge request timed out")), config.timeoutMs); - try { - const response = await fetcher(config.endpoint, { - method: "POST", - headers: { - "content-type": "application/json", - authorization: `Bearer ${key}` - }, - body, - signal: controller.signal - }); - if (!response.ok) { - if (response.status === 413) { - throw new SemanticJudgePayloadTooLargeError("Semantic judge rejected an oversized request with HTTP 413"); - } - const retryable = response.status === 408 || response.status === 429 || response.status === 529 || response.status >= 500; - throw new SemanticJudgeRequestError(`Semantic judge returned HTTP ${String(response.status)}`, retryable); - } - const raw = JSON.parse(await responseText(response)); - if (!isRecord4(raw) || !isRecord4(raw.answers)) { - throw new SemanticJudgeRequestError("Semantic judge response did not contain answers", false); - } - const answers = raw.answers; - const judgments = candidates.map((candidate, index) => parseJudgment(answers[candidate.id], index)); - const model = raw.model ?? config.model; - if (typeof model !== "string" || model.length === 0 || model.length > 128 || /[\r\n\0]/u.test(model)) { - throw new SemanticJudgeRequestError("Semantic judge response model must be bounded single-line text", false); - } - const usage = isRecord4(raw.usage) ? raw.usage : undefined; - const inputTokens = typeof usage?.input_tokens === "number" && Number.isSafeInteger(usage.input_tokens) && usage.input_tokens >= 0 ? usage.input_tokens : undefined; - const outputTokens = typeof usage?.output_tokens === "number" && Number.isSafeInteger(usage.output_tokens) && usage.output_tokens >= 0 ? usage.output_tokens : undefined; - return { - model, - judgments, - ...inputTokens === undefined ? {} : { inputTokens }, - ...outputTokens === undefined ? {} : { outputTokens }, - elapsedMs: Math.round(performance.now() - startedAt) - }; - } catch (error) { - if (parentSignal?.aborted) - throw error; - const retryable = error instanceof SemanticJudgePayloadTooLargeError ? false : error instanceof SemanticJudgeRequestError ? error.retryable : true; - if (!retryable || attempt >= config.maxRetries) - throw error; - await waitForRetry(retryDelay(attempt), parentSignal); - } finally { - clearTimeout(timeout); - parentSignal?.removeEventListener("abort", onParentAbort); - } - } - throw new SemanticJudgeRequestError("Semantic judge request failed", false); + const answers = raw.answers; + const judgments = candidates.map((candidate, index) => parseJudgment(answers[candidate.id], index)); + const usage = responseUsage(raw); + return { + model: responseModel(raw, config), + judgments, + ...usage, + elapsedMs: Math.round(performance.now() - startedAt) + }; + }; +} +function createHttpJsonRunner(config, key, fetcher) { + return async (query, candidates, parentSignal) => { + const startedAt = performance.now(); + const raw = await fetchSemanticResponse(config, key, fetcher, JSON.stringify(httpJsonRequestBody(query, candidates, config.model)), parentSignal); + return parseHttpJsonResult(raw, candidates, config, Math.round(performance.now() - startedAt)); }; } function createDisabledSemanticJudgeIntegration(config) { @@ -12549,14 +12981,16 @@ function createDisabledSemanticJudgeIntegration(config) { function createSemanticJudgeIntegration(config, environment = process.env, fetcher = fetch) { if (!config.enabled) return createDisabledSemanticJudgeIntegration(config); - const key = environment[config.apiKeyEnv]?.trim(); - if (!key) { - throw new SemanticJudgeConfigurationError(`Semantic judge is enabled, but environment variable ${config.apiKeyEnv} is missing or empty`); + const key = config.apiKeyEnv ? environment[config.apiKeyEnv]?.trim() : undefined; + if (config.provider === "jev" && !key) { + throw new SemanticJudgeConfigurationError(`Semantic judge is enabled, but environment variable ${config.apiKeyEnv ?? "TYPESAFE_API_KEY"} is missing or empty`); } - if (config.provider !== "jev") { - throw new SemanticJudgeConfigurationError(`Unsupported semantic judge provider: ${String(config.provider)}`); + if (config.provider === "http-json" && config.apiKeyEnv && !key) { + throw new SemanticJudgeConfigurationError(`Semantic judge is enabled, but environment variable ${config.apiKeyEnv} is missing or empty`); } - return { config, runner: createJevRunner(config, key, fetcher) }; + if (config.provider === "jev") + return { config, runner: createJevRunner(config, key, fetcher) }; + return { config, runner: createHttpJsonRunner(config, key, fetcher) }; } function createConfiguredSemanticJudgeIntegration(config, environment = process.env) { const judge = config.semanticJudge ?? DEFAULT_SEMANTIC_JUDGE_CONFIG; @@ -13733,7 +14167,7 @@ class EvidenceService { throw new SiftLightError(`No outline provider is registered for ${document.path}; use mode=capabilities to inspect available language modes`); } if (!language && !isPython || language === "go") { - const extension = extname4(document.path) || "extensionless source"; + const extension = extname5(document.path) || "extensionless source"; throw new SiftLightError(`${input.mode} is unavailable for ${extension}; choose a language capability from mode=capabilities or use ordinary content search`); } if (input.mode === "outline") { @@ -13926,7 +14360,7 @@ function normalizeRequestAliases(request) { } // src/format.ts -import { readFile as readFile5 } from "node:fs/promises"; +import { readFile as readFile6 } from "node:fs/promises"; var RESULT_METADATA_RESERVE_BYTES = 1024; var RESULT_METADATA_RESERVE_CHARACTERS = 512; @@ -14060,7 +14494,7 @@ async function loadContextLines(match, expectedRevision, cache, signal) { cache.set(match.absolutePath, changed); return changed; } - const content = await readFile5(match.absolutePath, { encoding: "utf8", signal }); + const content = await readFile6(match.absolutePath, { encoding: "utf8", signal }); const afterRevision = await getSourceRevision(match.absolutePath); if (!afterRevision || !sameSourceRevision(expectedRevision, afterRevision)) { const changed = { status: "changed" }; @@ -14385,7 +14819,7 @@ ${lineRows.join(` } // src/snapshot-store.ts -import { randomUUID as randomUUID7 } from "node:crypto"; +import { randomUUID as randomUUID8 } from "node:crypto"; class SnapshotStore { #snapshots = new Map; #expired = new Set; @@ -14410,7 +14844,7 @@ class SnapshotStore { const now = this.#now(); const snapshot = { ...scan, - id: randomUUID7(), + id: randomUUID8(), createdAt: now, lastAccessedAt: now }; @@ -15180,7 +15614,7 @@ class SiftLightService { } mediaSearch(value, cwd, signal) { if (!this.#media) - throw new SiftLightError("Visual search is disabled; configure nativeMedia and restart the host"); + throw new SiftLightError("Visual search is disabled; configure nativeMedia or media and restart the host"); return this.#media.search(value, cwd, signal); } #dataSources; @@ -15196,7 +15630,8 @@ class SiftLightService { #operations; #reusableSummarySnapshots = new WeakSet; constructor(options) { - this.#media = options.nativeMedia ? new MediaSearch(options.nativeMedia) : undefined; + const mediaConfig = options.nativeMedia ?? options.media; + this.#media = mediaConfig ? new MediaSearch(mediaConfig) : undefined; this.#dataSources = new DataSources(options.dataSources); this.#runRipgrep = options.runRipgrep; this.#snapshots = options.snapshots ?? new SnapshotStore; @@ -15908,7 +16343,7 @@ function siftLightTool(outputMode) { tool.outputSchema = SIFT_LIGHT_OUTPUT_SCHEMA; return tool; } -function createDefaultSiftLightMcpService(semanticJudge, vectorSearchEnabled = false, dataSources = [], nativeText, nativeMedia) { +function createDefaultSiftLightMcpService(semanticJudge, vectorSearchEnabled = false, dataSources = [], nativeText, nativeMedia, media) { const resolvedSemanticJudge = semanticJudge ?? createDisabledSemanticJudgeIntegration(DEFAULT_SEMANTIC_JUDGE_CONFIG); return new SiftLightService({ runRipgrep: createRipgrepRunner(), @@ -15917,7 +16352,8 @@ function createDefaultSiftLightMcpService(semanticJudge, vectorSearchEnabled = f dataSources, semanticJudge: resolvedSemanticJudge, ...nativeText ? { nativeText } : {}, - ...nativeMedia ? { nativeMedia } : {} + ...nativeMedia ? { nativeMedia } : {}, + ...media ? { media } : {} }); } function errorMessage2(error) { @@ -15974,9 +16410,9 @@ function createSiftLightMcpServer(service, cwd, outputMode = DEFAULT_MCP_OUTPUT_ { name: "sift-light-media", title: "Visual search", - description: "Search local image content or scanned PDF pages. mode=search needs path and exactly one query (text) or image (local image path). kind=all/image/scan. Honors scope, glob, exclude, hidden and ignorePolicy. Returns up to ten candidates with source hashes and PDF page numbers. Copy inspect or validate requests to this tool; inspect returns actual image pixels. Local resident vision model; slow search returns nextRequest mode=await to this tool. mode=cancel stops an operation. Text/code search stays in sift-light.", + description: "Search image content or scanned PDF pages through the configured local or BYO provider. mode=search needs path and exactly one query (text) or image (local image path). kind=all/image/scan. Honors scope, glob, exclude, hidden and ignorePolicy. Returns up to ten candidates with source hashes and PDF page numbers. Copy inspect or validate requests to this tool; inspect returns actual image pixels when a renderer is configured. Slow search returns nextRequest mode=await to this tool. mode=cancel stops an operation. Text/code search stays in sift-light.", inputSchema: { - ...z3.toJSONSchema(mediaRequestSchema, { unrepresentable: "any" }), + ...z4.toJSONSchema(mediaRequestSchema, { unrepresentable: "any" }), type: "object" }, annotations: { readOnlyHint: true, destructiveHint: false, openWorldHint: false } @@ -16194,7 +16630,7 @@ async function handleMcpRequest(request, response, state, createService, cwd) { try { const service = createService(); const transport = new StreamableHTTPServerTransport({ - sessionIdGenerator: () => randomUUID8(), + sessionIdGenerator: () => randomUUID9(), onsessioninitialized: (initializedSessionId) => { if (!pendingSession) throw new Error("MCP session initialized before registration"); @@ -16381,7 +16817,8 @@ async function createMcpSearchFeatures(environment = process.env) { vectorSearchEnabled: config.vectorSearchEnabled === true, dataSources: config.dataSources ?? [], ...config.nativeText ? { nativeText: config.nativeText } : {}, - ...config.nativeMedia ? { nativeMedia: config.nativeMedia } : {} + ...config.nativeMedia ? { nativeMedia: config.nativeMedia } : {}, + ...config.media ? { media: config.media } : {} }; } function mcpSemanticJudgeConfigSource(environment = process.env) { @@ -16501,12 +16938,12 @@ function logSemanticJudgeStatus(integration) { process.stderr.write(`sift-light MCP semantic judge: ${status}; source=${mcpSemanticJudgeConfigSource()} `); } -async function runHttpServer(outputMode, semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia) { +async function runHttpServer(outputMode, semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia, media) { const running = await startSiftLightMcpServer({ cwd: process.env.SIFT_LIGHT_MCP_CWD ?? process.cwd(), host: process.env.SIFT_LIGHT_MCP_HOST ?? DEFAULT_MCP_HOST, port: environmentInteger("SIFT_LIGHT_MCP_PORT", DEFAULT_MCP_PORT, 0, 65535), - createService: () => createDefaultSiftLightMcpService(semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia), + createService: () => createDefaultSiftLightMcpService(semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia, media), maxSessions: environmentInteger("SIFT_LIGHT_MCP_MAX_SESSIONS", DEFAULT_MCP_MAX_SESSIONS, 1, Number.MAX_SAFE_INTEGER), sessionIdleTimeoutMs: environmentInteger("SIFT_LIGHT_MCP_SESSION_IDLE_MS", DEFAULT_MCP_SESSION_IDLE_TIMEOUT_MS, 1, Number.MAX_SAFE_INTEGER), allowedOrigins: allowedOrigins(), @@ -16540,11 +16977,11 @@ async function runHttpServer(outputMode, semanticJudge, vectorSearchEnabled, dat process.once("SIGINT", shutdown); process.once("SIGTERM", shutdown); } -async function runStdioServer(outputMode, semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia) { +async function runStdioServer(outputMode, semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia, media) { const running = await startSiftLightMcpStdioServer({ cwd: process.env.SIFT_LIGHT_MCP_CWD ?? process.env.CLAUDE_PROJECT_DIR ?? process.cwd(), outputMode, - createService: () => createDefaultSiftLightMcpService(semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia) + createService: () => createDefaultSiftLightMcpService(semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia, media) }); process.stderr.write(`sift-light MCP serving one local client over stdio `); @@ -16573,15 +17010,15 @@ async function main() { return; } const outputMode = parseSiftLightMcpOutputMode(process.env.SIFT_LIGHT_MCP_OUTPUT_MODE); - const { semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia } = await createMcpSearchFeatures(); + const { semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia, media } = await createMcpSearchFeatures(); logSemanticJudgeStatus(semanticJudge); process.stderr.write(`sift-light MCP vector search: ${vectorSearchEnabled ? "enabled" : "disabled"} `); if (transport === "stdio") { - await runStdioServer(outputMode, semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia); + await runStdioServer(outputMode, semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia, media); return; } - await runHttpServer(outputMode, semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia); + await runHttpServer(outputMode, semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia, media); } try { await main(); diff --git a/src/mcp-server.ts b/src/mcp-server.ts index 2143ae7f..d0a68250 100644 --- a/src/mcp-server.ts +++ b/src/mcp-server.ts @@ -1,6 +1,7 @@ #!/usr/bin/env node import type { DataSourceConfig } from "./data-source-config.js"; import type { NativeTextConfig } from "./native-text-config.js"; +import type { MediaConfig } from "./media-config.js"; import { DEFAULT_MCP_HOST, DEFAULT_MCP_MAX_SESSIONS, @@ -54,6 +55,7 @@ async function runHttpServer( dataSources: DataSourceConfig[], nativeText?: NativeTextConfig, nativeMedia?: NativeTextConfig, + media?: MediaConfig, ): Promise { const running = await startSiftLightMcpServer({ cwd: process.env.SIFT_LIGHT_MCP_CWD ?? process.cwd(), @@ -66,6 +68,7 @@ async function runHttpServer( dataSources, nativeText, nativeMedia, + media, ), maxSessions: environmentInteger( "SIFT_LIGHT_MCP_MAX_SESSIONS", @@ -118,6 +121,7 @@ async function runStdioServer( dataSources: DataSourceConfig[], nativeText?: NativeTextConfig, nativeMedia?: NativeTextConfig, + media?: MediaConfig, ): Promise { const running = await startSiftLightMcpStdioServer({ cwd: process.env.SIFT_LIGHT_MCP_CWD ?? process.env.CLAUDE_PROJECT_DIR ?? process.cwd(), @@ -129,6 +133,7 @@ async function runStdioServer( dataSources, nativeText, nativeMedia, + media, ), }); process.stderr.write("sift-light MCP serving one local client over stdio\n"); @@ -157,7 +162,7 @@ async function main(): Promise { return; } const outputMode = parseSiftLightMcpOutputMode(process.env.SIFT_LIGHT_MCP_OUTPUT_MODE); - const { semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia } = + const { semanticJudge, vectorSearchEnabled, dataSources, nativeText, nativeMedia, media } = await createMcpSearchFeatures(); logSemanticJudgeStatus(semanticJudge); process.stderr.write( @@ -171,6 +176,7 @@ async function main(): Promise { dataSources, nativeText, nativeMedia, + media, ); return; } @@ -181,6 +187,7 @@ async function main(): Promise { dataSources, nativeText, nativeMedia, + media, ); } diff --git a/src/mcp.ts b/src/mcp.ts index 9937a65c..cb1147e2 100644 --- a/src/mcp.ts +++ b/src/mcp.ts @@ -2,6 +2,7 @@ import { z } from "zod"; import { mediaRequestSchema, type MediaResult } from "./media-search.js"; import type { DataSourceConfig } from "./data-source-config.js"; import type { NativeTextConfig } from "./native-text-config.js"; +import type { MediaConfig } from "./media-config.js"; import { randomUUID } from "node:crypto"; import { createServer, type IncomingMessage, type Server, type ServerResponse } from "node:http"; import { URL } from "node:url"; @@ -103,6 +104,7 @@ export function createDefaultSiftLightMcpService( dataSources: readonly DataSourceConfig[] = [], nativeText?: NativeTextConfig, nativeMedia?: NativeTextConfig, + media?: MediaConfig, ): SiftLightMcpService { const resolvedSemanticJudge = semanticJudge ?? createDisabledSemanticJudgeIntegration(DEFAULT_SEMANTIC_JUDGE_CONFIG); @@ -114,6 +116,7 @@ export function createDefaultSiftLightMcpService( semanticJudge: resolvedSemanticJudge, ...(nativeText ? { nativeText } : {}), ...(nativeMedia ? { nativeMedia } : {}), + ...(media ? { media } : {}), }); } @@ -193,7 +196,7 @@ export function createSiftLightMcpServer( name: "sift-light-media", title: "Visual search", description: - "Search local image content or scanned PDF pages. mode=search needs path and exactly one query (text) or image (local image path). kind=all/image/scan. Honors scope, glob, exclude, hidden and ignorePolicy. Returns up to ten candidates with source hashes and PDF page numbers. Copy inspect or validate requests to this tool; inspect returns actual image pixels. Local resident vision model; slow search returns nextRequest mode=await to this tool. mode=cancel stops an operation. Text/code search stays in sift-light.", + "Search image content or scanned PDF pages through the configured local or BYO provider. mode=search needs path and exactly one query (text) or image (local image path). kind=all/image/scan. Honors scope, glob, exclude, hidden and ignorePolicy. Returns up to ten candidates with source hashes and PDF page numbers. Copy inspect or validate requests to this tool; inspect returns actual image pixels when a renderer is configured. Slow search returns nextRequest mode=await to this tool. mode=cancel stops an operation. Text/code search stays in sift-light.", inputSchema: { ...z.toJSONSchema(mediaRequestSchema, { unrepresentable: "any" }), type: "object" as const, diff --git a/src/media-config.ts b/src/media-config.ts new file mode 100644 index 00000000..a165957c --- /dev/null +++ b/src/media-config.ts @@ -0,0 +1,21 @@ +import type { NativeTextConfig } from "./native-text-config.js"; + +/** + * A provider-owned multimodal retrieval endpoint. + * + * The endpoint receives bounded file bytes and a text or image query and + * returns normalized media hits. Source paths and page verification remain + * local to sift-light; the remote service only sees opaque file ids. + */ +export interface HttpMediaConfig { + provider: "http-json"; + endpoint: string; + apiKeyEnv?: string; + model: string; + timeoutMs: number; + maxRetries: number; + maxRequestBytes: number; + renderer?: NativeTextConfig; +} + +export type MediaConfig = HttpMediaConfig; diff --git a/src/media-search.ts b/src/media-search.ts index e86d99f5..25d2a147 100644 --- a/src/media-search.ts +++ b/src/media-search.ts @@ -5,7 +5,9 @@ import { tmpdir } from "node:os"; import { basename, dirname, extname, join, resolve } from "node:path"; import { z } from "zod"; import type { NativeTextConfig } from "./native-text-config.js"; +import type { HttpMediaConfig } from "./media-config.js"; import { ResidentWorkerClient } from "./resident-worker-client.js"; +import { RemoteMediaClient, mediaMimeType } from "./remote-media-client.js"; import { OwnedTaskQueue } from "./owned-task-queue.js"; import { listWorkspaceFiles } from "./workspace-files.js"; import { SearchPathPolicy } from "./path-policy.js"; @@ -44,7 +46,7 @@ const responseSchema = z.object({ version: z.literal(1), id: z.string(), indexedFiles: z.number().int().nonnegative(), - modelLoads: z.literal(1), + modelLoads: z.number().int().nonnegative(), modelReused: z.boolean(), processId: z.number().int().positive(), hits: z @@ -86,7 +88,7 @@ interface Snapshot { export interface MediaResult { text: string; details: Record; - image?: { data: string; mimeType: "image/png" }; + image?: { data: string; mimeType: string }; } const extensions = new Set([ ".png", @@ -134,7 +136,8 @@ async function sourceBytes( } /** Visual evidence owns its index and shares the local admission policy with text search. */ export class MediaSearch { - readonly #config: NativeTextConfig; + readonly #config: NativeTextConfig | undefined; + readonly #remoteConfig: HttpMediaConfig | undefined; readonly #operations = new OperationLifecycle({ deadlineMs: 300_000 }); readonly #queue = new OwnedTaskQueue(); #lifecycle = new AbortController(); @@ -142,9 +145,11 @@ export class MediaSearch { readonly #known = new Map(); #directory: Promise | undefined; #worker: ResidentWorkerClient> | undefined; + #remote: RemoteMediaClient | undefined; #closed = false; - constructor(config: NativeTextConfig) { - this.#config = config; + constructor(config: NativeTextConfig | HttpMediaConfig) { + if ("provider" in config) this.#remoteConfig = config; + else this.#config = config; } async search(value: unknown, cwd: string, signal?: AbortSignal): Promise { const request = mediaRequestSchema.parse(value); @@ -268,21 +273,37 @@ export class MediaSearch { } let response: z.infer | undefined; if (admitted.size) { - this.#worker ??= new ResidentWorkerClient( - this.#config, - root, - corpus, - join(root, "index.sqlite"), - { parse: (value) => responseSchema.parse(value), maxResponseBytes: 128 * 1024 }, - ); - response = await this.#worker.search( - { - ...(queryImage ? { image: queryImage } : { query: r.query }), + if (this.#config) { + this.#worker ??= new ResidentWorkerClient( + this.#config, + root, + corpus, + join(root, "index.sqlite"), + { parse: (value) => responseSchema.parse(value), maxResponseBytes: 128 * 1024 }, + ); + response = await this.#worker.search( + { + ...(queryImage ? { image: queryImage } : { query: r.query }), + kind: r.kind, + expectedFiles: admitted.size, + }, + signal, + ); + } else if (this.#remoteConfig) { + this.#remote ??= new RemoteMediaClient(this.#remoteConfig); + const remoteFiles = [...admitted.values()].map((entry) => ({ + id: createHash("sha256").update(entry.staged).digest("hex").slice(0, 32), + path: entry.staged, + })); + response = await this.#remote.search({ + ...(queryImage ? { imagePath: queryImage } : { query: r.query! }), kind: r.kind, expectedFiles: admitted.size, - }, - signal, - ); + files: remoteFiles, + signal, + }); + } + if (!response) throw new SiftLightError("Media provider is not configured"); if (response.indexedFiles !== admitted.size) throw new SiftLightError("Media index inventory mismatch"); } @@ -384,9 +405,23 @@ export class MediaSearch { const input = join(folder, "source" + extname(item.source.path)); const output = join(folder, "preview.png"); await writeFile(input, current.bytes, { mode: 0o600 }); + const renderer = this.#config ?? this.#remoteConfig?.renderer; + if (!renderer) { + const mimeType = mediaMimeType(item.source.path); + if (mimeType?.startsWith("image/") && mimeType === "image/png") { + return { + text: JSON.stringify(details), + details, + image: { data: current.bytes.toString("base64"), mimeType }, + }; + } + throw new SiftLightError( + "Media preview requires nativeMedia or media.renderer; validate still works without a local renderer", + ); + } const result = await runOwnedProcess( { - executable: this.#config.executable, + executable: renderer.executable, args: ["--render", input, String(item.hit.page ?? 0), output], cwd: folder, signal, @@ -421,6 +456,7 @@ export class MediaSearch { await this.#queue.run(async () => { try { await this.#worker?.close(); + await this.#remote?.close(); } finally { this.#snapshots.clear(); if (this.#directory) await rm(await this.#directory, { recursive: true, force: true }); diff --git a/src/native-text-config.ts b/src/native-text-config.ts index 012f9bfe..7f0be327 100644 --- a/src/native-text-config.ts +++ b/src/native-text-config.ts @@ -8,7 +8,7 @@ export interface NativeTextConfig { export function parseNativeTextConfig( value: unknown, - field: "nativeText" | "nativeMedia" = "nativeText", + field: string = "nativeText", ): NativeTextConfig | undefined { if (value === undefined) return undefined; if ( diff --git a/src/remote-media-client.ts b/src/remote-media-client.ts new file mode 100644 index 00000000..64952914 --- /dev/null +++ b/src/remote-media-client.ts @@ -0,0 +1,363 @@ +import { randomUUID } from "node:crypto"; +import { readFile, stat } from "node:fs/promises"; +import { extname } from "node:path"; +import { z } from "zod"; +import type { HttpMediaConfig } from "./media-config.js"; +import { SiftLightError } from "./errors.js"; + +const MAX_RESPONSE_BYTES = 4 * 1024 * 1024; +const MAX_FILE_ID_LENGTH = 128; + +class RemoteMediaRequestError extends SiftLightError { + readonly retryable: boolean; + + constructor(message: string, retryable: boolean, options?: ErrorOptions) { + super(message, options); + this.name = "RemoteMediaRequestError"; + this.retryable = retryable; + } +} + +export interface RemoteMediaFile { + id: string; + path: string; +} + +export interface RemoteMediaSearchRequest { + query?: string; + imagePath?: string; + kind: "all" | "image" | "scan"; + expectedFiles: number; + files: readonly RemoteMediaFile[]; + signal?: AbortSignal; +} + +export interface RemoteMediaSearchResponse { + version: 1; + id: string; + indexedFiles: number; + modelLoads: number; + modelReused: boolean; + processId: number; + hits: Array<{ + path: string; + score: number; + kind: "image" | "scan"; + page: number | null; + chunkCount: number; + width: number; + height: number; + }>; +} + +export type RemoteMediaFetcher = ( + input: string | URL | Request, + init?: RequestInit, +) => Promise; + +const remoteResponseSchema = z + .object({ + version: z.literal(1), + indexedFiles: z.number().int().nonnegative(), + hits: z + .array( + z + .object({ + fileId: z.string().min(1).max(MAX_FILE_ID_LENGTH), + score: z.number().finite(), + kind: z.enum(["image", "scan"]), + page: z.number().int().positive().nullable().optional(), + chunkCount: z.number().int().positive(), + width: z.number().int().nonnegative(), + height: z.number().int().nonnegative(), + }) + .strict(), + ) + .max(10), + }) + .strict(); + +function mimeType(path: string): string { + switch (extname(path).toLowerCase()) { + case ".png": + return "image/png"; + case ".jpg": + case ".jpeg": + return "image/jpeg"; + case ".gif": + return "image/gif"; + case ".bmp": + return "image/bmp"; + case ".tif": + case ".tiff": + return "image/tiff"; + case ".heic": + return "image/heic"; + case ".heif": + return "image/heif"; + case ".webp": + return "image/webp"; + case ".pdf": + return "application/pdf"; + default: + return "application/octet-stream"; + } +} + +async function boundedResponseText(response: Response): Promise { + const contentLength = response.headers.get("content-length"); + if (contentLength !== null && Number(contentLength) > MAX_RESPONSE_BYTES) + throw new SiftLightError("Remote media response exceeded the 4 MiB limit"); + if (!response.body) { + const text = await response.text(); + if (new TextEncoder().encode(text).byteLength > MAX_RESPONSE_BYTES) + throw new SiftLightError("Remote media response exceeded the 4 MiB limit"); + return text; + } + const reader = response.body.getReader(); + const chunks: Uint8Array[] = []; + let bytes = 0; + try { + while (true) { + // oxlint-disable-next-line no-await-in-loop -- response chunks must be consumed in order. + const result = await reader.read(); + if (result.done) break; + bytes += result.value.byteLength; + if (bytes > MAX_RESPONSE_BYTES) { + // oxlint-disable-next-line no-await-in-loop -- cancel the owned stream before reporting the bounded failure. + await reader.cancel(); + throw new SiftLightError("Remote media response exceeded the 4 MiB limit"); + } + chunks.push(result.value); + } + } finally { + reader.releaseLock(); + } + const all = new Uint8Array(bytes); + let offset = 0; + for (const chunk of chunks) { + all.set(chunk, offset); + offset += chunk.byteLength; + } + return new TextDecoder("utf-8", { fatal: true }).decode(all); +} + +function retryableStatus(status: number): boolean { + return status === 408 || status === 429 || status === 529 || status >= 500; +} + +function retryDelay(attempt: number): number { + return Math.min(20_000, 400 * 2 ** attempt); +} + +async function waitForRetry(delayMs: number, signal?: AbortSignal): Promise { + if (signal?.aborted) + throw signal.reason instanceof Error ? signal.reason : new Error("Operation aborted"); + await new Promise((resolve, reject) => { + const timer = setTimeout(() => { + signal?.removeEventListener("abort", onAbort); + resolve(); + }, delayMs); + const onAbort = (): void => { + clearTimeout(timer); + signal?.removeEventListener("abort", onAbort); + reject(signal?.reason instanceof Error ? signal.reason : new Error("Operation aborted")); + }; + signal?.addEventListener("abort", onAbort, { once: true }); + }); +} + +function base64(bytes: Buffer): string { + return bytes.toString("base64"); +} + +function base64Length(bytes: number): number { + return Math.ceil(bytes / 3) * 4; +} + +export class RemoteMediaClient { + readonly #config: HttpMediaConfig; + readonly #key: string | undefined; + readonly #fetcher: RemoteMediaFetcher; + + constructor( + config: HttpMediaConfig, + environment: NodeJS.ProcessEnv = process.env, + fetcher: RemoteMediaFetcher = fetch, + ) { + this.#config = config; + this.#key = config.apiKeyEnv ? environment[config.apiKeyEnv]?.trim() : undefined; + if (config.apiKeyEnv && !this.#key) { + throw new SiftLightError( + `Media provider is configured with ${config.apiKeyEnv}, but that environment variable is missing or empty`, + ); + } + this.#fetcher = fetcher; + } + + async search(request: RemoteMediaSearchRequest): Promise { + const id = randomUUID(); + const paths = new Map(); + const files: Array<{ id: string; mimeType: string; data: string }> = []; + const fileSizes: number[] = []; + for (const file of request.files) { + if ( + file.id.length === 0 || + file.id.length > MAX_FILE_ID_LENGTH || + /[\r\n\0]/u.test(file.id) || + paths.has(file.id) + ) + throw new SiftLightError("Remote media file ids must be unique bounded text"); + paths.set(file.id, file.path); + // oxlint-disable-next-line no-await-in-loop -- preflight file sizes before allocating request payloads. + const metadata = await stat(file.path); + if (!metadata.isFile()) throw new SiftLightError("Remote media files must be regular files"); + fileSizes.push(metadata.size); + } + let imageSize: number | undefined; + if (request.imagePath !== undefined) { + const metadata = await stat(request.imagePath); + if (!metadata.isFile()) + throw new SiftLightError("Remote media query image must be a regular file"); + imageSize = metadata.size; + } + const requestSkeleton = JSON.stringify({ + version: 1, + model: this.#config.model, + ...(request.query === undefined ? {} : { query: request.query }), + ...(request.imagePath === undefined + ? {} + : { image: { mimeType: mimeType(request.imagePath), data: "" } }), + kind: request.kind, + expectedFiles: request.expectedFiles, + files: [], + }); + let estimatedBytes = new TextEncoder().encode(requestSkeleton).byteLength; + if (imageSize !== undefined) estimatedBytes += base64Length(imageSize); + for (let index = 0; index < request.files.length; index += 1) { + const file = request.files[index]; + const fileSkeleton = JSON.stringify({ + id: file?.id, + mimeType: file ? mimeType(file.path) : "application/octet-stream", + data: "", + }); + estimatedBytes += new TextEncoder().encode(fileSkeleton).byteLength; + estimatedBytes += base64Length(fileSizes[index] ?? 0); + } + if (request.files.length > 1) estimatedBytes += request.files.length - 1; + // The estimate reserves the exact JSON metadata and Base64 payload sizes so a + // rejected request never needs to materialize a corpus larger than the limit. + if (estimatedBytes > this.#config.maxRequestBytes) + throw new SiftLightError( + `Remote media request exceeded the configured ${String(this.#config.maxRequestBytes)} byte limit`, + ); + for (const file of request.files) { + files.push({ + id: file.id, + mimeType: mimeType(file.path), + // oxlint-disable-next-line no-await-in-loop -- read and bound one admitted file at a time. + data: base64(await readFile(file.path, { signal: request.signal })), + }); + } + const image = + request.imagePath === undefined + ? undefined + : { + mimeType: mimeType(request.imagePath), + data: base64(await readFile(request.imagePath, { signal: request.signal })), + }; + const body = JSON.stringify({ + version: 1, + model: this.#config.model, + ...(request.query === undefined ? {} : { query: request.query }), + ...(image === undefined ? {} : { image }), + kind: request.kind, + expectedFiles: request.expectedFiles, + files, + }); + if (new TextEncoder().encode(body).byteLength > this.#config.maxRequestBytes) + throw new SiftLightError( + `Remote media request exceeded the configured ${String(this.#config.maxRequestBytes)} byte limit`, + ); + for (let attempt = 0; attempt <= this.#config.maxRetries; attempt += 1) { + const controller = new AbortController(); + const onParentAbort = (): void => controller.abort(request.signal?.reason); + request.signal?.addEventListener("abort", onParentAbort, { once: true }); + const timeout = setTimeout( + () => controller.abort(new Error("Remote media request timed out")), + this.#config.timeoutMs, + ); + try { + const headers: Record = { "content-type": "application/json" }; + if (this.#key) headers.authorization = `Bearer ${this.#key}`; + // oxlint-disable-next-line no-await-in-loop -- retry attempts share one bounded provider budget. + const response = await this.#fetcher(this.#config.endpoint, { + method: "POST", + headers, + body, + signal: controller.signal, + }); + if (!response.ok) { + throw new RemoteMediaRequestError( + `Remote media provider returned HTTP ${String(response.status)}`, + retryableStatus(response.status), + ); + } + let parsed: z.infer; + try { + // oxlint-disable-next-line no-await-in-loop -- consume the bounded response before deciding whether to retry. + const raw: unknown = JSON.parse(await boundedResponseText(response)); + parsed = remoteResponseSchema.parse(raw); + } catch (error) { + throw new SiftLightError("Remote media provider returned an invalid response", { + cause: error, + }); + } + if (parsed.indexedFiles !== request.expectedFiles) + throw new SiftLightError("Remote media provider indexed an unexpected file count"); + return { + version: 1, + id, + indexedFiles: parsed.indexedFiles, + modelLoads: 0, + modelReused: true, + processId: process.pid, + hits: parsed.hits.map((hit) => { + const path = paths.get(hit.fileId); + if (!path) + throw new SiftLightError("Remote media provider returned an unknown file id"); + return { + path, + score: hit.score, + kind: hit.kind, + page: hit.page ?? null, + chunkCount: hit.chunkCount, + width: hit.width, + height: hit.height, + }; + }), + }; + } catch (error) { + if (request.signal?.aborted) throw error; + const retryable = + error instanceof RemoteMediaRequestError + ? error.retryable + : !(error instanceof SiftLightError); + if (!retryable || attempt >= this.#config.maxRetries) throw error; + // oxlint-disable-next-line no-await-in-loop -- retries intentionally serialize attempts. + await waitForRetry(retryDelay(attempt), request.signal); + } finally { + clearTimeout(timeout); + request.signal?.removeEventListener("abort", onParentAbort); + } + } + throw new SiftLightError("Remote media request failed"); + } + + async close(): Promise {} +} + +export function mediaMimeType(path: string): string | undefined { + const type = mimeType(path); + return type === "application/octet-stream" ? undefined : type; +} diff --git a/src/semantic-judge.ts b/src/semantic-judge.ts index f94ffb6d..3bf11245 100644 --- a/src/semantic-judge.ts +++ b/src/semantic-judge.ts @@ -167,6 +167,19 @@ function requestBody( }; } +function httpJsonRequestBody( + query: string, + candidates: readonly SemanticJudgeCandidate[], + model: string, +): RawRecord { + return { + version: 1, + model, + query, + candidates: candidates.map(({ id, excerpt }) => ({ id, excerpt })), + }; +} + async function responseText(response: Response): Promise { if (!response.body) { const contentLength = response.headers.get("content-length"); @@ -240,6 +253,171 @@ async function waitForRetry(delayMs: number, signal?: AbortSignal): Promise { + if (new TextEncoder().encode(body).byteLength > MAX_SEMANTIC_JUDGE_REQUEST_BYTES) { + throw new SemanticJudgePayloadTooLargeError( + "Semantic judge request exceeded the 64 KiB request limit", + ); + } + for (let attempt = 0; attempt <= config.maxRetries; attempt += 1) { + const controller = new AbortController(); + const onParentAbort = (): void => controller.abort(parentSignal?.reason); + parentSignal?.addEventListener("abort", onParentAbort, { once: true }); + const timeout = setTimeout( + () => controller.abort(new Error("Semantic judge request timed out")), + config.timeoutMs, + ); + try { + const headers: Record = { "content-type": "application/json" }; + if (key) headers.authorization = `Bearer ${key}`; + // oxlint-disable-next-line no-await-in-loop -- retry attempts share one bounded provider budget. + const response = await fetcher(config.endpoint, { + method: "POST", + headers, + body, + signal: controller.signal, + }); + if (!response.ok) { + if (response.status === 413) { + throw new SemanticJudgePayloadTooLargeError( + "Semantic judge rejected an oversized request with HTTP 413", + ); + } + const retryable = + response.status === 408 || + response.status === 429 || + response.status === 529 || + response.status >= 500; + throw new SemanticJudgeRequestError( + `Semantic judge returned HTTP ${String(response.status)}`, + retryable, + ); + } + // oxlint-disable-next-line no-await-in-loop -- consume the bounded response before deciding whether to retry. + const raw: unknown = JSON.parse(await responseText(response)); + if (!isRecord(raw)) { + throw new SemanticJudgeRequestError("Semantic judge response must be an object", false); + } + return raw; + } catch (error) { + if (parentSignal?.aborted) throw error; + const retryable = + error instanceof SemanticJudgePayloadTooLargeError + ? false + : error instanceof SemanticJudgeRequestError + ? error.retryable + : true; + if (!retryable || attempt >= config.maxRetries) throw error; + // oxlint-disable-next-line no-await-in-loop -- retries intentionally serialize attempts. + await waitForRetry(retryDelay(attempt), parentSignal); + } finally { + clearTimeout(timeout); + parentSignal?.removeEventListener("abort", onParentAbort); + } + } + throw new SemanticJudgeRequestError("Semantic judge request failed", false); +} + +function responseModel(raw: RawRecord, config: SemanticJudgeConfig): string { + const model = raw.model ?? config.model; + if ( + typeof model !== "string" || + model.length === 0 || + model.length > 128 || + /[\r\n\0]/u.test(model) + ) { + throw new SemanticJudgeRequestError( + "Semantic judge response model must be bounded single-line text", + false, + ); + } + return model; +} + +function responseUsage(raw: RawRecord): { inputTokens?: number; outputTokens?: number } { + const usage = isRecord(raw.usage) ? raw.usage : undefined; + const inputTokens = + typeof usage?.input_tokens === "number" && + Number.isSafeInteger(usage.input_tokens) && + usage.input_tokens >= 0 + ? usage.input_tokens + : undefined; + const outputTokens = + typeof usage?.output_tokens === "number" && + Number.isSafeInteger(usage.output_tokens) && + usage.output_tokens >= 0 + ? usage.output_tokens + : undefined; + return { + ...(inputTokens === undefined ? {} : { inputTokens }), + ...(outputTokens === undefined ? {} : { outputTokens }), + }; +} + +function parseHttpJsonResult( + raw: RawRecord, + candidates: readonly SemanticJudgeCandidate[], + config: SemanticJudgeConfig, + elapsedMs: number, +): SemanticJudgeResult { + if (raw.version !== 1 || !Array.isArray(raw.judgments)) { + throw new SemanticJudgeRequestError( + "HTTP JSON semantic judge response must contain version 1 judgments", + false, + ); + } + const candidateIndexes = new Map(candidates.map((candidate, index) => [candidate.id, index])); + const judgments = new Map(); + for (const value of raw.judgments) { + if (!isRecord(value) || typeof value.id !== "string") { + throw new SemanticJudgeRequestError( + "HTTP JSON semantic judge judgment has an invalid id", + false, + ); + } + const candidateIndex = candidateIndexes.get(value.id); + if (candidateIndex === undefined || judgments.has(candidateIndex)) { + throw new SemanticJudgeRequestError( + "HTTP JSON semantic judge response contains an unknown or duplicate candidate", + false, + ); + } + if (!isClassification(value.choice)) { + throw new SemanticJudgeRequestError( + `HTTP JSON semantic judge response has an invalid choice for ${value.id}`, + false, + ); + } + const confidence = + value.confidence === undefined ? undefined : probability(value.confidence, "confidence"); + judgments.set(candidateIndex, { + candidateIndex, + classification: value.choice, + probability: probability(value.probability, "probability"), + ...(confidence === undefined ? {} : { confidence }), + }); + } + if (judgments.size !== candidates.length) { + throw new SemanticJudgeRequestError( + `HTTP JSON semantic judge response judged ${String(judgments.size)} of ${String(candidates.length)} candidates`, + false, + ); + } + const usage = responseUsage(raw); + return { + model: responseModel(raw, config), + judgments: candidates.map((_candidate, index) => judgments.get(index)!), + ...usage, + elapsedMs, + }; +} + function createJevRunner( config: SemanticJudgeConfig, key: string, @@ -248,107 +426,39 @@ function createJevRunner( return async (query, candidates, parentSignal) => { const startedAt = performance.now(); const body = JSON.stringify(requestBody(query, candidates, config.model)); - if (new TextEncoder().encode(body).byteLength > MAX_SEMANTIC_JUDGE_REQUEST_BYTES) { - throw new SemanticJudgePayloadTooLargeError( - "Semantic judge request exceeded the 64 KiB request limit", - ); - } - for (let attempt = 0; attempt <= config.maxRetries; attempt += 1) { - const controller = new AbortController(); - const onParentAbort = (): void => controller.abort(parentSignal?.reason); - parentSignal?.addEventListener("abort", onParentAbort, { once: true }); - const timeout = setTimeout( - () => controller.abort(new Error("Semantic judge request timed out")), - config.timeoutMs, - ); - try { - // oxlint-disable-next-line no-await-in-loop -- retries intentionally serialize attempts under one request budget. - const response = await fetcher(config.endpoint, { - method: "POST", - headers: { - "content-type": "application/json", - authorization: `Bearer ${key}`, - }, - body, - signal: controller.signal, - }); - if (!response.ok) { - if (response.status === 413) { - throw new SemanticJudgePayloadTooLargeError( - "Semantic judge rejected an oversized request with HTTP 413", - ); - } - const retryable = - response.status === 408 || - response.status === 429 || - response.status === 529 || - response.status >= 500; - throw new SemanticJudgeRequestError( - `Semantic judge returned HTTP ${String(response.status)}`, - retryable, - ); - } - // oxlint-disable-next-line no-await-in-loop -- parse the current bounded response before deciding whether to retry. - const raw: unknown = JSON.parse(await responseText(response)); - if (!isRecord(raw) || !isRecord(raw.answers)) { - throw new SemanticJudgeRequestError( - "Semantic judge response did not contain answers", - false, - ); - } - const answers = raw.answers; - const judgments = candidates.map((candidate, index) => - parseJudgment(answers[candidate.id], index), - ); - const model = raw.model ?? config.model; - if ( - typeof model !== "string" || - model.length === 0 || - model.length > 128 || - /[\r\n\0]/u.test(model) - ) { - throw new SemanticJudgeRequestError( - "Semantic judge response model must be bounded single-line text", - false, - ); - } - const usage = isRecord(raw.usage) ? raw.usage : undefined; - const inputTokens = - typeof usage?.input_tokens === "number" && - Number.isSafeInteger(usage.input_tokens) && - usage.input_tokens >= 0 - ? usage.input_tokens - : undefined; - const outputTokens = - typeof usage?.output_tokens === "number" && - Number.isSafeInteger(usage.output_tokens) && - usage.output_tokens >= 0 - ? usage.output_tokens - : undefined; - return { - model, - judgments, - ...(inputTokens === undefined ? {} : { inputTokens }), - ...(outputTokens === undefined ? {} : { outputTokens }), - elapsedMs: Math.round(performance.now() - startedAt), - }; - } catch (error) { - if (parentSignal?.aborted) throw error; - const retryable = - error instanceof SemanticJudgePayloadTooLargeError - ? false - : error instanceof SemanticJudgeRequestError - ? error.retryable - : true; - if (!retryable || attempt >= config.maxRetries) throw error; - // oxlint-disable-next-line no-await-in-loop -- provider retries share one bounded request budget. - await waitForRetry(retryDelay(attempt), parentSignal); - } finally { - clearTimeout(timeout); - parentSignal?.removeEventListener("abort", onParentAbort); - } + const raw = await fetchSemanticResponse(config, key, fetcher, body, parentSignal); + if (!isRecord(raw.answers)) { + throw new SemanticJudgeRequestError("Semantic judge response did not contain answers", false); } - throw new SemanticJudgeRequestError("Semantic judge request failed", false); + const answers = raw.answers; + const judgments = candidates.map((candidate, index) => + parseJudgment(answers[candidate.id], index), + ); + const usage = responseUsage(raw); + return { + model: responseModel(raw, config), + judgments, + ...usage, + elapsedMs: Math.round(performance.now() - startedAt), + }; + }; +} + +function createHttpJsonRunner( + config: SemanticJudgeConfig, + key: string | undefined, + fetcher: SemanticJudgeFetcher, +): SemanticJudgeRunner { + return async (query, candidates, parentSignal) => { + const startedAt = performance.now(); + const raw = await fetchSemanticResponse( + config, + key, + fetcher, + JSON.stringify(httpJsonRequestBody(query, candidates, config.model)), + parentSignal, + ); + return parseHttpJsonResult(raw, candidates, config, Math.round(performance.now() - startedAt)); }; } @@ -364,18 +474,19 @@ export function createSemanticJudgeIntegration( fetcher: SemanticJudgeFetcher = fetch, ): SemanticJudgeIntegration { if (!config.enabled) return createDisabledSemanticJudgeIntegration(config); - const key = environment[config.apiKeyEnv]?.trim(); - if (!key) { + const key = config.apiKeyEnv ? environment[config.apiKeyEnv]?.trim() : undefined; + if (config.provider === "jev" && !key) { throw new SemanticJudgeConfigurationError( - `Semantic judge is enabled, but environment variable ${config.apiKeyEnv} is missing or empty`, + `Semantic judge is enabled, but environment variable ${config.apiKeyEnv ?? "TYPESAFE_API_KEY"} is missing or empty`, ); } - if (config.provider !== "jev") { + if (config.provider === "http-json" && config.apiKeyEnv && !key) { throw new SemanticJudgeConfigurationError( - `Unsupported semantic judge provider: ${String(config.provider)}`, + `Semantic judge is enabled, but environment variable ${config.apiKeyEnv} is missing or empty`, ); } - return { config, runner: createJevRunner(config, key, fetcher) }; + if (config.provider === "jev") return { config, runner: createJevRunner(config, key!, fetcher) }; + return { config, runner: createHttpJsonRunner(config, key, fetcher) }; } export function createConfiguredSemanticJudgeIntegration( diff --git a/src/service.ts b/src/service.ts index f17fb240..e44b6232 100644 --- a/src/service.ts +++ b/src/service.ts @@ -2,6 +2,7 @@ import { MediaSearch, type MediaResult } from "./media-search.js"; import { DataSources } from "./data-sources.js"; import type { DataSourceConfig } from "./data-source-config.js"; import type { NativeTextConfig } from "./native-text-config.js"; +import type { MediaConfig } from "./media-config.js"; import { createHash } from "node:crypto"; import { EvidenceService, isEvidenceRequest } from "./evidence-service.js"; import { ConceptSourceChangedError } from "./concept-source-generation.js"; @@ -107,6 +108,7 @@ export interface SiftLightInput extends RawSearchInput { export interface SiftLightServiceOptions { nativeText?: NativeTextConfig; nativeMedia?: NativeTextConfig; + media?: MediaConfig; dataSources?: readonly DataSourceConfig[]; runRipgrep: RipgrepRunner; snapshots?: SnapshotStore; @@ -402,7 +404,7 @@ export class SiftLightService { mediaSearch(value: unknown, cwd: string, signal?: AbortSignal): Promise { if (!this.#media) throw new SiftLightError( - "Visual search is disabled; configure nativeMedia and restart the host", + "Visual search is disabled; configure nativeMedia or media and restart the host", ); return this.#media.search(value, cwd, signal); } @@ -420,7 +422,8 @@ export class SiftLightService { readonly #reusableSummarySnapshots = new WeakSet(); constructor(options: SiftLightServiceOptions) { - this.#media = options.nativeMedia ? new MediaSearch(options.nativeMedia) : undefined; + const mediaConfig = options.nativeMedia ?? options.media; + this.#media = mediaConfig ? new MediaSearch(mediaConfig) : undefined; this.#dataSources = new DataSources(options.dataSources); this.#runRipgrep = options.runRipgrep; this.#snapshots = options.snapshots ?? new SnapshotStore();