Unified normalized attachments, request versions, and provider files
统一规范化附件、请求版本与提供方文件
Durable image history, provider resolution, inline request size, and remote file reuse have different limits. Treating an admitted image as the bytes sent on every later request forced one byte cap and one raster to serve all four concerns. Large but ordinary input was refused, clean 16-bit PNG could pass into history and fail at DeepSeek, repeated base64 expanded long requests, and a provider rejection repeated beca
English
Problem
Durable image history, provider resolution, inline request size, and remote file reuse have different limits. Treating an admitted image as the bytes sent on every later request forced one byte cap and one raster to serve all four concerns. Large but ordinary input was refused, clean 16-bit PNG could pass into history and fail at DeepSeek, repeated base64 expanded long requests, and a provider rejection repeated because the same durable image stayed in every future request.
Decision
The image path has two explicit versions. The attachment backend owns a provider-independent durable normalized attachment. Each image-capable model route owns a deterministic request policy, and the attachment backend derives and caches the exact request version from that attachment. Session history contains only the normalized attachment reference; inline bytes and provider file ids remain transient request projections.
Provider-independent normalized attachment
Admission accepts at most 20 images and 200MiB of encoded source bytes per message. Each source is fully decoded under configurable 20MiB, 64,000,000-pixel, and 8192px-per-side limits. Normalization applies EXIF orientation, removes metadata and color profiles, converts to 8-bit sRGB/sRGBA, and preserves aspect ratio while scaling into the normalizedImageMaxPixels total-pixel budget (2048x2048 by default) under a normalizedImageMaxDimension long-edge cap (8192px by default). When scaling reduces the raster, originalDimensions records its orientation-applied width and height before normalization.
The normalized attachment has an independent normalizedImageMaxBytes encoded-byte target, 4MiB by default. Alpha is never flattened. Codec routing is by the decoded alpha fact alone — alpha input encodes as WebP (effort 0) and opaque input as JPEG, each down the shared 85/75/60 quality ladder; when every quality exceeds the target the smallest output is kept, per the superseding alpha-routed quality ladders note. A clean, single-frame 8-bit sRGB/sRGBA PNG, JPEG, or WebP within every normalization limit passes through byte-identically and retains content-addressed deduplication. GIF, animation, metadata, orientation, 16-bit PNG, and incompatible color spaces force conversion. The source and a converted output are each fully decoded once; the output must match its format, dimensions, depth, color space, and alpha facts before its digest enters the reference.
Batch admission prepares and verifies every normalized attachment once before publishing any member. Validation failure starts no writes. Publication uses those prepared bytes directly, so a large batch does not repeat full decoding and encoding during commit. A later storage failure returns no partial references; already published immutable objects may remain unreachable under the existing storage rule.
Deterministic request versions
AttachmentStore.readImageRequest derives a request version under route-owned total-pixel and encoded-byte budgets. Scaling is min(1, sqrt(maxPixels / (width * height))), with no enlargement, followed by inward integer rounding so the encoded raster never exceeds the total-pixel cap. DeepSeek V4 Flash Vision Exp uses 640,000 total pixels and a 1MiB raw encoded-byte target by default. Its catalog uses one imagePixelBudget field: a positive integer selects an exact total-pixel budget, low selects 512 by 512 total pixels, and omission selects the route default. A 2048 by 1024 normalized attachment projects to 1130 by 565 under the hard cap. Request encoding uses the same alpha routing and 85/75/60 quality ladder as normalization, executed lazily; a target no quality meets keeps the smallest ladder output (see the alpha-routed quality ladders note). The same derivation is used by normal agent turns, direct ctx.llm.stream calls, compaction, and other auxiliary streams.
The variantId and cache path cover the normalized attachment id, transform version, route pixel and byte budgets, and fixed encoder parameters. A new cache entry is fully decoded before publication. Cache hits use a header probe to check format, 8-bit sRGB/sRGBA facts, dimensions, and alpha without decoding the complete raster again; a mismatch regenerates the entry. DeepSeek Files and pi-ai inline base64 therefore use the same deterministic bytes for the same policy. Inline accounting uses the derived byte length after base64 expansion, not the normalized attachment byte count. Equal in-process variantId calls share one transform and cache write. Each caller can cancel its own wait; the shared transform is aborted only after every waiter has cancelled. Callers preserve order by applying Promise.all to singular readImageRequest calls. The local implementation runs normalization and request transforms through one FIFO limiter; imageCompressionConcurrency is configurable from 1 through 8 and defaults to 2. Batch publication remains sequential after every normalized attachment has been prepared.
Request-size offload follows the durable image-offload decision. Routes check retained occurrences against their request-version byte and image-count budgets and report IMAGE_OFFLOAD_REQUIRED with a removal count. The compaction plugin records exact oldest-first occurrences in image/offload; Session derivation applies the marks. Adapters prepare only retained attachments and render each marked occurrence as placeholder text with its identity and current execution-world access path, including nested tool-result images. Original message events retain the attachment references.
Stable handles
Every retained request image is preceded by its display name or complete attachment id and actual request dimensions. The attachment provider can supply its host object location; the LLM consumer combines it with the current filesystem mapping before adding an absolute read-only path, normalized dimensions, and media type. The descriptor states that normalization or request projection may have resized or re-encoded the upload, so the model cannot infer original upload properties from either representation. User messages, tool results, agent-loop requests, compaction, and direct ctx.llm.stream calls share this projection. The path is resolved independently from the deterministic request version and does not enter its variantId, the durable reference, or the session log.
DeepSeek Files lifecycle
The direct deepseek-official adapter uploads every retained request version through the Messages Files API and sends file-id references. A bounded inline fallback sends the same deterministic request versions when file resolution fails. Uploaded ids are indexed by Files endpoint and API-key scope plus variantId. Uploads request seven days by default. Messages metadata omits expiry, so local reuse uses the original upload time plus the configured lifetime without promising remote deletion. A mapping with no more than one hour of reuse remaining is replaced without a preceding retrieve call. The index never stores the API key.
An upload is indexed only after the response returns a valid native file object with a matching byte count and the client establishes the protocol’s reuse deadline. A missing or inconsistent response leaves no local mapping, so a later request uploads again. Concurrent upload resolution for one scoped variantId shares one provider operation; one waiter cannot cancel another, and the upload stops when every waiter has cancelled. A malformed upload index is an empty cache and is replaced on the next successful upload; filesystem I/O failures remain errors. If the model endpoint reports expired, deleted, missing, or invalid ids and names one or more ids used by the request, only those mappings are removed. A stale-file response without a specific id removes every mapping used by that model request. The affected request bytes are uploaded again and the model request is retried once. A second stale rejection clears the mappings identified by its response and returns the error without a third model request. One upload quota error first lists the configured number of oldest harness-owned dsh- files, then deletes that collected set and retries once; deleting after pagination keeps provider cursors valid. Public file operations expose list, retrieve, delete, one-variant release, and namespace-wide release. Every Files request carries the shared Harness User-Agent. The client enforces a 128MiB upload limit, a 32MiB request-image limit, 10,000-file and 25GiB quotas, and one-hour to 30-day expiry range.
Diagnostics
A 16-bit RGB or RGBA PNG is normal admitted input and converts to 8-bit sRGB/sRGBA. If local conversion fails, read_image names the path, detected 16-bit PNG, required normalized form, and manual conversion remedy. If DeepSeek rejects a normalized request version, the primary error names the attachment or display name, durable message and image position, normalized media type, 8-bit sRGB/sRGBA depth, dimensions, and provider message. An ambiguous multi-image rejection lists every candidate. The raw provider body remains the error cause rather than the only visible message.
Historical attachment objects that later disappear or fail integrity verification remain fail-loud. Durable quarantine and verified recovery require session events and are tracked by Quarantine unreadable historical attachments.
Alternatives considered
Use one 1MiB normalized attachment for storage and requests. This makes model resolution determine durable image detail and combines local storage, inline expansion, Files quota, and model pixels into one setting. Independent normalization and request policies keep those responsibilities explicit.
Reject images above provider dimensions or at the encoding quality floor. A provider limit is route-specific and future requests may use another model. Proportional normalization and request projection accept ordinary large images while bounding each later representation.
Treat PNG as a screenshot and reject 16-bit PNG. File format does not reveal pixel complexity, and 16-bit RGB/RGBA is a convertible sample depth rather than an unsupported image type. Pixel sampling and post-conversion probes give the required facts.
Keep DeepSeek data URLs as the primary transport. Inline base64 repeats bytes on every request and caps usable image history by request-body size. Files API references reuse uploaded deterministic request bytes with bounded local reuse and explicit deletion; the bounded fallback sends inline base64 only when file resolution fails.
Trust a locally indexed file id indefinitely. Remote expiry, deletion, and lost upload responses make local and provider state diverge. Response-directed invalidation and one re-upload recover without an unbounded retry loop; an ambiguous stale-file response must invalidate every file used by that attempt because it provides no safe exact target.
Refuse text-only model selection after any image. Durable history can outlive the model that first consumed it. Request-local placeholders keep the session usable without rewriting history.
Remove one image whenever a request crosses its limit. That changes an early request message after nearly every new upload. Quantized removed prefixes keep cache invalidation occasional while honoring the configured high bound.
Verification
Package tests generate 16-bit RGB and RGBA PNG fixtures, prove 8-bit conversion and clean 8-bit passthrough, retain alpha under byte pressure, stop lazy encoding after the first fitting candidate, keep the smallest ladder output above an unreachable byte target, cover square and wide 640,000-pixel projections, singleflight equal variants and uploads without shared-cancellation leaks, bound transform concurrency, preserve cache and upload identity, skip attachment reads for durably offloaded history, prepare batches once, reject inconsistent Files responses, refresh near-expiry ids without retrieve, recover once from single-id, multiple-id, and ambiguous stale responses, fall back to bounded all-inline requests after file resolution failure, paginate before quota deletion, normalize provider diagnostics, project text-only history, and share normal/compaction request bytes. Keyless assembled snapshots cover the real tool schemas and image request path. A credentialed test uses the built-in deepseek-official route and its configured endpoint, never a custom provider entry.
Consequences
Normalized attachments consume up to the independent local safety cap, while request caches and remote Files consume additional derived storage. Deterministic identities and singleflight make that work reusable across turns and sessions sharing the same DSH home. Two simultaneous transforms reduce batch latency while increasing peak RSS relative to serial execution; deployments with tighter memory can set the limit to one. Encoder or transform-version changes create new future identities without rewriting existing history. DeepSeek image requests prefer Files reuse; bounded stale-id recovery handles inconsistent remote state, while file-resolution failures use the smaller inline budget. Missing or corrupt durable attachments still require the separate quarantine design.
中文
Problem
持久图片历史、提供方分辨率、内联请求大小和远端文件复用有不同限制。过去把已接纳图片直接作为之后每次请求发送的字节，导致一个字节上限和一份光栅同时承担四种职责。普通大图会被拒绝；干净的 16-bit PNG 可以进入历史，之后才被 DeepSeek 拒绝；重复 base64 使长会话请求持续增长；提供方拒绝后，同一持久图片还会进入每次后续请求。
Decision
图片路径有两个显式版本。附件后端拥有提供方无关的持久规范化附件。每条支持图片的模型路由拥有确定性请求策略，附件后端从该附件派生并缓存确切请求版本。会话历史只包含规范化附件引用；内联字节和提供方文件 ID 都是瞬时请求投影。
提供方无关的规范化附件
每条消息最多准入 20 张图片，源图编码字节总量不超过 200MiB。每张源图会在可配置的 20MiB、64,000,000 像素和单边 8192px 限制内完整解码。规范化过程会应用 EXIF 方向，删除元数据和色彩配置文件，转换为 8-bit sRGB/sRGBA，并保持宽高比缩放进 normalizedImageMaxPixels 总像素预算（默认 2048×2048），随后受 normalizedImageMaxDimension 长边上限约束（默认 8192px）。缩放减小光栅时，originalDimensions 记录规范化之前、应用方向之后的输入宽高。
规范化附件有独立的 normalizedImageMaxBytes 编码字节目标，默认 4MiB。透明通道绝不铺平。编码路由只看解码出的 alpha 事实：透明输入编码为 WebP（effort 0），非透明输入编码为 JPEG，共用 85/75/60 质量阶梯；全部档位都超过目标时保留最小产物，见取代本节的按 alpha 路由的质量阶梯记录。处于全部规范化限制内的干净、单帧、8-bit sRGB/sRGBA PNG、JPEG 或 WebP 按字节原样直通，并保留内容寻址去重。GIF、动图、元数据、方向、16-bit PNG 和不兼容色彩空间都会触发转换。源图和转换输出各完整解码一次；输出的格式、尺寸、位深、色彩空间和透明通道事实通过校验后，其摘要才会进入引用。
批量准入在发布任何成员前，为每张图片各准备并验证一次规范化附件。校验失败不会开始写入。发布直接使用这些已准备字节，因此大批次不会在提交时重复完整解码和编码。之后发生的存储失败不会返回部分引用；按现有存储规则，已经发布的不可变对象可能保持不可达。
确定性请求版本
AttachmentStore.readImageRequest 按路由拥有的总像素和编码字节预算派生请求版本。缩放公式为 min(1, sqrt(maxPixels / (width * height)))，不会放大小图，随后向预算内取整，确保编码光栅不超过总像素上限。DeepSeek V4 Flash Vision Exp 默认使用总像素 640,000 和原始编码字节目标 1MiB。其 catalog 只使用一个 imagePixelBudget 字段：正整数选择确切总像素预算，low 选择总像素 512×512，省略时使用路由默认值。2048×1024 规范化附件在这个硬上限下会投影为 1130×565。请求编码与规范化共用同一套 alpha 路由和 85/75/60 质量阶梯，按需执行；没有任何档位达到目标时保留最小产物（见按 alpha 路由的质量阶梯记录）。普通 agent 轮次、直接 ctx.llm.stream 调用、压缩和其他辅助流都使用同一派生过程。
variantId 和缓存路径覆盖规范化附件 ID、变换策略版本、路由像素和字节预算及固定编码参数。新缓存条目在发布前会完整解码。缓存命中只探测文件头，校验格式、8-bit sRGB/sRGBA、尺寸和透明通道，不会再次完整解码光栅；不匹配时会重新生成。因此，同一策略下的 DeepSeek Files 和 pi-ai 内联 base64 使用相同的确定性字节。内联计量使用派生字节经过 base64 膨胀后的长度，不使用规范化附件字节数。同一进程内相同 variantId 的调用共享一次变换和缓存写入。每个调用方可以取消自己的等待；只有全部等待方都取消时，共享变换才会中止。调用方对单数 readImageRequest 使用 Promise.all 保持结果顺序。本地实现通过一个 FIFO 限流器运行规范化和请求变换，imageCompressionConcurrency 的可配置范围为 1 至 8，默认值为 2。全部规范化附件准备完成后，批次仍按顺序发布。
请求大小 offload 遵循持久图片省略决策。路由根据请求版本字节和图片数量预算检查保留的位置，通过 IMAGE_OFFLOAD_REQUIRED 报告所需移除数量。Compaction 插件在 image/offload 中记录明确的最旧图片出现位置，Session 派生时应用标记。适配器只准备保留的附件，将每个带标记的位置渲染为包含其身份和当前执行环境访问路径的占位文本，嵌套工具结果图片也适用。原始消息事件保留附件引用。
稳定句柄
每张保留请求图片前都有显示名称或完整附件 ID，以及实际请求尺寸。附件提供方可以给出宿主对象位置；LLM 消费方将其与当前文件系统映射组合后，再加入绝对只读路径、规范化尺寸和媒体类型。描述会说明规范化或请求投影可能缩小或重新编码上传图片，因此模型不能从任一版本推断上传图片原本的属性。用户消息、工具结果、agent loop 请求、压缩和直接 ctx.llm.stream 调用共享这套投影。路径独立于确定性请求版本解析，不进入其 variantId、持久引用或会话日志。
DeepSeek Files 生命周期
直连 deepseek-official 适配器通过 Messages Files API 上传每个保留的请求版本，并发送 file-id 引用。文件解析失败时，有界内联回退发送相同的确定性请求版本。上传 id 按 Files 端点、API-key 作用域及 variantId 建立索引。上传默认请求七天有效期。Messages 元数据不含过期时间，因此本地复用使用原始上传时间加配置的有效期，不承诺远端删除。剩余复用时间不超过一小时的映射会直接替换，不先调用 retrieve。索引从不存储 API 密钥。
只有上传响应返回有效的原生文件对象、字节数匹配，且客户端确定了该协议的复用期限时，上传结果才会写入索引。缺失或不一致的响应不会留下本地映射，后续请求会重新上传。同一作用域和 variantId 的并发解析共享一次提供方上传；单个等待方无法取消其他等待方，全部等待方取消时才会停止上传。格式损坏的上传索引按空缓存处理，并在下一次成功上传时替换；文件系统 I/O 失败仍是错误。如果模型端点报告 ID 已过期、删除、缺失或无效，并指出本次请求使用的一个或多个 ID，适配器只删除这些映射。如果响应只说明文件状态失效而没有指出具体 ID，适配器会删除该次模型请求使用的全部映射。受影响的请求字节会重新上传，模型请求只重试一次。第二次仍报告文件失效时，适配器会按响应清理映射并返回错误，不会发起第三次模型请求。一次上传配额错误会先列出配置数量的最旧 dsh- 文件，再删除收集到的文件并重试一次；分页完成后才删除，避免游标失效。公开文件操作提供列表、查询、删除、单个变体释放和整个作用域释放。每个 Files 请求都携带 Harness 的共享 User-Agent。客户端执行 Files 单次上传 128MiB、请求单图 32MiB、10,000 个文件、25GiB，以及一小时到 30 天有效期限制。
诊断
16-bit RGB 或 RGBA PNG 属于普通可接纳输入，会转换为 8-bit sRGB/sRGBA。本地转换失败时，read_image 会写明路径、检测到的 16-bit PNG、所需规范形式和手工转换方法。如果 DeepSeek 拒绝已规范化请求版本，主错误会写明附件 ID 或显示名称、持久消息和图片位置、规范化媒体类型、8-bit sRGB/sRGBA 位深、尺寸和提供方消息。多图片错误无法确定对象时会列出全部候选图片。原始提供方正文保留为错误 cause，不会成为唯一可见消息。
持久附件对象之后缺失或无法通过完整性校验时，系统仍会明确失败。持久隔离和经校验恢复需要新增会话事件，由隔离不可读历史附件继续跟踪。
Alternatives considered
使用一份 1MiB 规范化附件同时负责存储和请求。 这种做法让模型分辨率决定持久图片细节，并把本地存储、内联膨胀、Files 配额和模型像素合并成一个设置。独立的规范化和请求策略会明确区分这些职责。
拒绝超过提供方尺寸或达到编码质量下限的图片。 提供方限制属于具体路由，未来请求可能改用另一个模型。按比例规范化和投影请求版本可以接纳普通大图，同时约束每种后续表示。
把 PNG 当作截图，并拒绝 16-bit PNG。 文件格式不能说明像素复杂度，16-bit RGB/RGBA 是可转换位深，不是不支持的图片类型。像素采样和转换后探测能提供所需事实。
把 DeepSeek data URL 作为首选传输方式。 内联 base64 会在每次请求中重复字节，并按请求正文大小限制可用图片历史。Files API 引用会复用上传后的确定性请求字节，并限制本地复用期限、提供显式删除操作；有界回退只在文件解析失败时发送内联 base64。
永久信任本地索引中的文件 ID。 远端过期、删除和上传响应丢失会使本地与提供方状态不一致。按响应失效和一次重新上传可以恢复，同时避免无界重试；响应没有给出可安全使用的精确目标时，必须使该次请求使用的全部文件失效。
历史中出现图片后拒绝选择纯文本模型。 持久历史可能比最初读取它的模型存活更久。按请求生成的占位文本可以保持会话可用，无需改写历史。
请求每次越过上限就移除一张图片。 这种做法会在几乎每次新增图片后改写较早的请求消息。按固定步长递增的移除前缀会降低缓存失效频率，同时遵守配置的上限。
Verification
包测试会生成 16-bit RGB 和 RGBA PNG fixture，验证 8-bit 转换与干净 8-bit 字节直通、字节压力下保留透明通道、首个候选合规后停止编码、字节目标不可达时保留最小阶梯产物、正方形和宽屏 640,000 像素投影、相同变体与上传 singleflight 且不会共享取消、变换并发上限、缓存与上传身份、跳过已持久省略的历史附件读取、批量只准备一次、Files 响应不一致、进入刷新余量时不查询远端并更新 ID、单个 ID、多个 ID 和模糊失效响应只恢复一次、文件解析失败后回退到有界全内联请求、删除配额文件前完成分页、规范化提供方诊断、纯文本投影，以及普通请求与压缩共享请求字节。无需密钥的组装快照覆盖真实工具 schema 和图片请求路径。使用凭据的测试只使用内置 deepseek-official 路由及其已配置端点，不使用自定义提供方条目。
Consequences
持久规范化附件最多占用独立的本地安全上限，请求缓存和远端 Files 还会占用额外派生存储。确定性身份和 singleflight 使这些成本可以被共享同一 DSH home 的轮次和会话复用。同时执行两个变换会降低批次延迟，但峰值 RSS 高于串行执行；内存更紧张的部署可以把上限设为 1。编码器或变换策略版本变化会为未来内容产生新身份，不会改写已有历史。DeepSeek 图片请求优先复用 Files；有界的陈旧 ID 恢复会处理远端状态不一致，文件解析失败则使用较小的内联预算。缺失或损坏的持久附件仍需要单独的隔离设计。