@kodax-ai/kodax 0.7.96-beta.4 → 0.7.96-beta.7

package/CHANGELOG.md

@@ -8,6 +8,103 @@ All notable changes to this project will be documented in this file.
 
 ---
 
+## [0.7.96-beta.7] - 2026-09-11
+
+The seventh beta pre-release of the v0.7.96 line: it carries every numbered-beta
+contract (sandbox-first permissions, trusted text, Windows setup generation
+11, context recovery, DeepSeek Anthropic-compat endpoint, compaction summary
+policy, wire-faithful prompt-cache diagnostics) plus the multimodal and
+interrupt-identity fixes below. npm publication remains a manual maintainer
+action.
+
+### Fixed
+
+- Preserve text/image tool results through direct and managed dispatch, constructed
+  Worker RPC, allowing guardrails, capacity recovery and MCP tool/resource adapters.
+  MCP images persist as attachments; image-capable OpenAI providers receive tool
+  images after valid paired responses, with matching prompt-cache diagnostics.
+- Carry local execution failures and Worker error codes through child results,
+  Actor metadata and Runtime/daemon status. Returned managed/MCP errors retain
+  their error classification instead of appearing successful or as provider faults.
+- Keep delivered interrupt entry references valid across context rewrites,
+  storage reads/saves, compaction, restart and conversation pagination. Preserve
+  explicit provenance, resolve physical ancestor aliases, reject competing claims,
+  and rebuild older conversation caches. Equal text or timestamps never create aliases.
+
+---
+
+## [0.7.96-beta.6] - 2026-09-11
+
+Sixth beta pre-release of the v0.7.96 line: a review-hardening pass that
+makes prompt-cache diagnostics describe the actual provider wire, retaining
+every beta.5 contract. npm publication remains a manual maintainer action.
+
+### Fixed
+
+- Prompt-cache diagnostics now mirror the provider wire: the Anthropic
+  diagnostic repair mirrors the provider's pairing semantics — a retained
+  orphan `tool_use` is answered with the interrupted-tool marker inside its
+  pairing scope, real non-adjacent answers are hoisted, foreign and
+  duplicate results are dropped, and other user blocks carry after the
+  answers. The OpenAI diagnostic repair gains the same full pairing scope,
+  and assistant projection keeps source order to match the
+  order-preserving serialization, so the diagnostic hash describes the
+  messages actually sent.
+- A new prompt-cache wire contract suite (224 cases) proves the projected
+  diagnostic envelope equals the real provider request across every
+  built-in and custom Anthropic/OpenAI provider, including retained
+  orphan calls that must stay distinct from a thinking-only placeholder.
+- The native SDDL auto-inherited flag normalization parses the flag field
+  instead of a blind string replace, preserving P/AR flags and S: sections,
+  with its own regression test (89 native tests). `@kodax-ai/llm` exports
+  `KODAX_INTERRUPTED_TOOL_RESULT_MARKER` as the shared wire-contract
+  constant.
+
+---
+
+## [0.7.96-beta.5] - 2026-09-10
+
+Fifth beta pre-release of the v0.7.96 line. Two change streams: the built-in
+`deepseek` provider moves to DeepSeek's official Anthropic-compatible
+endpoint with a new default model and replayed-history repair, and Windows
+sandbox setup advances to generation 11 with Codex-compatible profile/SSH
+ACL exclusions (Issue 333). All earlier beta.4 contracts are retained. npm
+publication remains a manual maintainer action.
+
+### Added
+
+- `deepseek-flash` (DeepSeek-V4.1-Flash, 2026-09-10): the new default on the
+  built-in `deepseek` alias, speaking the official
+  `api.deepseek.com/anthropic` wire with the `deepseek-v4-anthropic`
+  reasoning preset (thinking adaptive, `output_config.effort`, signed
+  thinking-block replay, count_tokens verification), 1M context, 384000 max
+  output tokens, and native image input. The legacy text-only
+  `deepseek-v4-pro` id stays selectable via `/model`. Cost rates use
+  converted official CNY idle rates with recorded peak-hour equivalents.
+
+### Fixed
+
+- Replayed restored-session histories no longer 400 on strict endpoints:
+  `repairToolCallHistory` answers an orphaned `tool_use` call with an
+  explicit interrupted-tool marker instead of dropping it, Anthropic
+  serialization keeps assistant blocks in their stored order while preserving
+  the required `tool_result`-first user grouping, and error classification
+  recognizes both wire dialects' pairing errors. Typed `AbortError`
+  cancellation now takes precedence over provider message matching; the
+  runner preserves the observed provider failure over a plain abort reason;
+  SDK Runtime aborts use typed `DOMException` aborts; and the REPL
+  distinguishes user cancellation from session-recovery cleanup.
+- Issue 333: Windows sandbox setup generation 11 matches Codex's profile and
+  SSH dependency ACL exclusions on read and write roots, keeping ordinary
+  home/tool reads and the existing fixed ACL plus per-command
+  token/concurrency mechanism. Generation-10 SSH ACE cleanup is setup-only,
+  removes provably owned entries, preserves owner and unrelated ACEs, and
+  retains the old SID/nonce/roots across interrupted retries. Generation-8/9
+  protocol-only in-place upgrades keep their behavior. See
+  `docs/test-guides/ISSUE_333_v0.7.96_REGRESSION_GUIDE.md`.
+
+---
+
 ## [0.7.96-beta.4] - 2026-09-08
 
 Fourth beta pre-release of the v0.7.96 line, retaining the beta.3 Windows WFP

package/README.md

@@ -367,7 +367,7 @@ If your custom provider's underlying model supports image input (vision), set `"
 
 `imageInput: true` forces `capabilityProfile.multimodalSupport: "image-input"` on every KodaX surface (provider instance, capability queries, policy gates), overriding an explicit `"none"`. The advanced alternative — a hand-written `capabilityProfile` block with `"multimodalSupport": "image-input"` — works too; see [Custom Providers](public_docs/configuration/custom-providers.md). Leave it unset for text-only models and image artifacts are rejected with `MODEL_INPUT_UNSUPPORTED` before the request is sent.
 
-Built-in vision-capable aliases (Anthropic, OpenAI, compatible aliases such as Kimi, Qwen, Zhipu, MiniMax, MiMo, Ark, plus Gemini-CLI via the CLI's `@<path>` file-include syntax) already ship with image input enabled. DeepSeek V4's default models (`deepseek-v4-flash` / `deepseek-v4-pro`) and Codex-CLI are text-only — on the built-in `deepseek` alias only `deepseek-v4-flash-vision-exp` takes images; custom providers need to opt in when their underlying model supports image input.
+Built-in vision-capable aliases (Anthropic, OpenAI, compatible aliases such as Kimi, Qwen, Zhipu, MiniMax, MiMo, Ark, plus Gemini-CLI via the CLI's `@<path>` file-include syntax) already ship with image input enabled. On the built-in `deepseek` alias the default `deepseek-flash` (DeepSeek-V4.1-Flash) takes images natively; the legacy `deepseek-v4-pro` id and Codex-CLI are text-only; custom providers need to opt in when their underlying model supports image input.
 
 The serializer layer (`packages/llm/src/providers/anthropic.ts:1431` for Anthropic-compat, `openai.ts:1496` for OpenAI-compat) forwards image blocks automatically through base-class inheritance — OpenAI-compatible endpoints receive standard `image_url` blocks. The flag only gates whether KodaX's policy layer pre-rejects multimodal requests — the model-level vision contract remains your upstream provider's responsibility. If the model is actually text-only, you'll see the real upstream API error instead of a KodaX-side rejection.
 
@@ -759,6 +759,23 @@ permission UI, and recovers stale prepared Session tails through an
 authoritative merge. Background persistence failures are surfaced as
 diagnostics rather than hidden.
 
+Beta.6 makes prompt-cache diagnostics mirror the provider wire: retained
+orphan `tool_use` calls are answered with the interrupted-tool marker inside
+their pairing scope and assistant projection keeps source order, so the
+diagnostic hash describes the messages actually sent. A 224-case wire
+contract suite proves the projected envelope equals the real provider
+request across built-in and custom Anthropic/OpenAI providers.
+
+Beta.5 moves the built-in `deepseek` provider to DeepSeek's official
+Anthropic-compatible endpoint (`api.deepseek.com/anthropic`) with new default
+model `deepseek-flash` (DeepSeek-V4.1-Flash, 1M context, native image input)
+and repairs replayed restored-session histories: orphaned `tool_use` calls
+are answered with an explicit interrupted-tool marker instead of being
+dropped, and typed `AbortError` cancellation takes precedence over provider
+message matching. Windows sandbox setup advances to generation 11
+(Issue 333), matching Codex's profile and SSH dependency ACL exclusions while
+keeping ordinary home/tool reads and the existing concurrency mechanism.
+
 Beta.4 shares manual and automatic compaction summary policy:
 `compaction.reasoning` applies to both paths independent of the main turn's
 effort, defaulting to disabled thinking where supported. Manual compaction
@@ -767,7 +784,15 @@ provider's model, REPL `/compact` clears the UI only after a durable save, and
 successful reports carry bounded `summaryRequests` plus `commitMs` per
 physical summary call.
 
-**v0.7.96-beta.4 release:** Windows shell admission now follows the
+**v0.7.96-beta.7 release:** The seventh beta pre-release of the v0.7.96 line. It
+adds multimodal result fidelity — text/image tool results survive direct and
+managed dispatch, Worker RPC, guardrails, capacity recovery, and MCP
+adapters (MCP images persist as attachments) — carries typed local failures
+through child results, Actor metadata, and Runtime/daemon status, and keeps
+delivered interrupt entry references valid across context rewrites,
+storage reads, compaction, restart, and conversation pagination.
+
+**v0.7.96-beta.6 release:** Windows shell admission now follows the
 Codex concurrency boundary: versioned setup performs legacy migration once,
 while ordinary admission accepts effective inherited normal-token access and
 only converges a missing exact-root restricted capability using `SET_ACCESS`
@@ -816,7 +841,7 @@ requests, preserves structured Child Agent Provider failures, honors exact
 run-scoped credential verification, and omits empty `tools` arrays for strict
 vLLM gateways (Issues 329-332). npm publication remains a
 manual maintainer action. See the
-[release checklist](docs/release.md#v0796-beta4-release-preparation).
+[release checklist](docs/release.md#v0796-beta7-release-preparation).
 
 **v0.7.96-alpha.3 release:** Provider credentials are lazy, scoped,
 revocable capabilities (ADR-068). The v2 credential broker keeps Provider
@@ -2010,7 +2035,7 @@ await runInkInteractiveMode({ provider: 'zhipu-coding', effort: 'auto' });
 | mimo | `MIMO_API_KEY` | Native | mimo-v2.5-pro (Xiaomi MiMo pay-per-token, Anthropic-compat) |
 | mimo-coding | `MIMO_CODING_API_KEY` | Native | mimo-v2.5-pro (Xiaomi Token Plan, Anthropic-compat) |
 | ark-coding | `ARK_CODING_API_KEY` | Native | glm-5.3 (Volcengine Ark Coding Plan — GLM-5.3 (1M ctx, 128K out) · GLM-5.2 (alias: `glm-latest`) · Kimi K2.7 Code / K2.6 · MiniMax M3 / M2.7 · DeepSeek V4 Pro / V4 Flash · Doubao Seed 2.0 Code / Pro / Lite · Doubao Seed Code) |
-| deepseek | `DEEPSEEK_API_KEY` | Native | deepseek-v4-flash (`deepseek-v4-pro` plus vision model `deepseek-v4-flash-vision-exp` with image input, via `/model`) |
+| deepseek | `DEEPSEEK_API_KEY` | Native | deepseek-flash (Anthropic-compat; DeepSeek-V4.1-Flash, 1M ctx, native image input; legacy text-only `deepseek-v4-pro`, via `/model`) |
 | gemini-cli | Provider CLI authentication (no KodaX API-key variable) | Prompt-only / CLI bridge | (via gemini CLI) |
 | codex-cli | Provider CLI authentication (no KodaX API-key variable) | Prompt-only / CLI bridge | (via codex CLI) |
 

package/README_CN.md

@@ -296,9 +296,28 @@ SDK 系统代码契约更新，但没有放宽 shell/sandbox 的 fail-closed 边
 `handleRuntimePermissionRequest()` 管理 SDK 权限 UI，并在 prepared Session 尾部遇到
 `data_changed` 时通过权威 delta 合并恢复；后台持久化失败会显示为诊断，不再静默丢失。
 
+**v0.7.96-beta.7 发布**：新增多模态结果保真——
+文本/图片工具结果在直接与托管分发、Worker RPC、guardrail、容量恢复与 MCP
+适配器中完整保留（MCP 图片持久化为附件）；本地执行失败以类型化错误贯穿
+child 结果、Actor 元数据与 Runtime/daemon 状态；已投递中断条目引用在上下文
+重写、存储读写、压缩、重启与分页中保持有效。
+
+beta.6 让 prompt-cache 诊断与真实 provider wire 对齐：保留的孤立 `tool_use`
+调用在其配对范围内以 interrupted-tool 标记应答，assistant 投影保持源顺序，
+诊断哈希因此描述实际发送的消息。新增 224 个 wire 契约用例，验证投影包络在
+全部内置与自定义 Anthropic/OpenAI provider 上等于真实 provider 请求。
+
+beta.5 将内置 `deepseek` provider 迁移到 DeepSeek 官方 Anthropic 兼容协议
+（`api.deepseek.com/anthropic`），新默认模型 `deepseek-flash`
+（DeepSeek-V4.1-Flash，1M 上下文，原生图片理解），并修复恢复会话重放时的
+历史校验：孤立的 `tool_use` 调用会以显式的 interrupted-tool 标记应答而不是
+被丢弃，类型化 `AbortError` 取消优先于 provider 文本匹配。Windows sandbox
+setup 升级到 generation 11（Issue 333），对齐 Codex 的 profile 与 SSH 依赖
+ACL 排除，同时保留普通 home/tool 读取与现有并发机制。
+
 beta.4 统一了手动与自动压缩的摘要策略：`compaction.reasoning` 同时作用于两条路径，与主回合 effort 无关（默认在支持时关闭思考）。手动压缩遵循生效的 Session provider/model，不会把切换前 provider 的 model 带入请求；REPL `/compact` 仅在持久化保存成功后清空 UI；成功报告按每次物理摘要调用记录有界 `summaryRequests` 与 `commitMs`。
 
-**v0.7.96-beta.4 发布**：Windows shell 准入现在遵循 Codex 的并发边界：
+**v0.7.96-beta.6 发布**：Windows shell 准入现在遵循 Codex 的并发边界：
 版本化 setup 只执行一次 legacy 迁移；普通准入在 capability 已存在时只读，只有缺失
 精确 root restricted capability 时才用 `SET_ACCESS` 与 DACL 回读收敛，且不等待任何
 跨进程目标互斥锁。稳定 filesystem capability SID 保留在对象上，每条命令的 token
@@ -337,7 +356,7 @@ scope 在 SDK、Agent 摘要、CLI 与 Runtime Worker 请求间共享，保留
 失败信息，精确遵循 run-scoped 凭据校验，并为严格 vLLM 网关省略空 `tools` 数组（Issues 329-332）。
 npm 发布仍由
 npm 发布仍由
-维护者手动执行。详见 [v0.7.96-beta.4 发布清单](docs/release.md#v0796-beta4-release-preparation)。
+维护者手动执行。详见 [v0.7.96-beta.7 发布清单](docs/release.md#v0796-beta7-release-preparation)。
 
 **v0.7.96-alpha.3 发布**：Provider 凭据成为惰性、受限、可撤销的能力（ADR-068）。v2
 credential broker 将 Provider 密钥保留在 OS keychain，按每次 wire call、为单一封闭
@@ -889,7 +908,7 @@ Verifier 的 `revise` / `blocked` 可执行消息；JSONL 输出使用同形
 
 `imageInput: true` 会在 KodaX 所有层面（provider 实例、能力查询、policy gate）强制 `capabilityProfile.multimodalSupport: "image-input"`，显式写了 `"none"` 也会被覆盖。进阶写法 —— 手写 `capabilityProfile` 块并设 `"multimodalSupport": "image-input"` —— 同样有效，详见 [Custom Providers](public_docs/configuration/custom-providers.md)。纯文本模型保持不设即可，图片 artifact 会在请求发出前被 `MODEL_INPUT_UNSUPPORTED` 拒绝。
 
-内置 vision-capable alias（Anthropic、OpenAI、Kimi、Qwen、Zhipu、MiniMax、MiMo、Ark，以及通过 CLI `@<path>` file-include 语法传图的 Gemini-CLI）已经默认开了图片输入。DeepSeek V4 默认模型（`deepseek-v4-flash` / `deepseek-v4-pro`）和 Codex-CLI 是纯文本 —— 内置 `deepseek` 只有 `deepseek-v4-flash-vision-exp` 这一个路由收图；自定义 provider 在底层模型支持图片输入时需要手动 opt-in。
+内置 vision-capable alias（Anthropic、OpenAI、Kimi、Qwen、Zhipu、MiniMax、MiMo、Ark，以及通过 CLI `@<path>` file-include 语法传图的 Gemini-CLI）已经默认开了图片输入。内置 `deepseek` 的默认模型 `deepseek-flash`（DeepSeek-V4.1-Flash）原生支持图片理解；兼容名 `deepseek-v4-pro` 和 Codex-CLI 是纯文本；自定义 provider 在底层模型支持图片输入时需要手动 opt-in。
 
 序列化层（Anthropic-compat 走 `packages/llm/src/providers/anthropic.ts:1431`，OpenAI-compat 走 `openai.ts:1496`）通过基类继承自动转发 image block —— OpenAI-compatible 端点收到的是标准 `image_url` 块。这个 flag 只控制 KodaX 自身是否预先拒绝多模态请求 —— 上游模型到底支不支持 vision 由 provider 自己决定。如果模型实际是 text-only，你会看到真实的上游 API 错误，而不是 KodaX 一侧的 `[Provider Policy] multimodal requests are unsupported` 预拦截。
 
@@ -1014,7 +1033,7 @@ dist/binary/linux-x64/
 | mimo | `MIMO_API_KEY` | Native | mimo-v2.5-pro（小米 MiMo 按量计费，Anthropic 协议） |
 | mimo-coding | `MIMO_CODING_API_KEY` | Native | mimo-v2.5-pro（小米 MiMo Token Plan，Anthropic 协议） |
 | ark-coding | `ARK_CODING_API_KEY` | Native | glm-5.3（火山方舟 Coding Plan — GLM-5.3（1M ctx、128K out） · GLM-5.2（别名 `glm-latest`） · Kimi K2.7 Code / K2.6 · MiniMax M3 / M2.7 · DeepSeek V4 Pro / V4 Flash · Doubao Seed 2.0 Code / Pro / Lite · Doubao Seed Code） |
-| deepseek | `DEEPSEEK_API_KEY` | Native | deepseek-v4-flash（可 `/model` 切换 `deepseek-v4-pro` 及视觉模型 `deepseek-v4-flash-vision-exp`，后者支持图片输入） |
+| deepseek | `DEEPSEEK_API_KEY` | Native | deepseek-flash（Anthropic 协议；DeepSeek-V4.1-Flash，1M ctx，原生图片理解；可 `/model` 切换纯文本兼容名 `deepseek-v4-pro`） |
 | gemini-cli | 由 Provider CLI 完成认证（无 KodaX API-key 环境变量） | Prompt-only / CLI bridge | （通过 gemini CLI） |
 | codex-cli | 由 Provider CLI 完成认证（无 KodaX API-key 环境变量） | Prompt-only / CLI bridge | （通过 codex CLI） |
 

package/config-templates/config.example.jsonc

@@ -7,10 +7,10 @@
   // Active provider alias and optional model override. Run `kodax setup` or
   // `/model` instead of editing these manually when possible.
   // "provider": "deepseek",
-  // "model": "deepseek-v4-flash",
+  // "model": "deepseek-flash",
   // Add or reorder model ids shown by `/model` for a provider.
   // "providerModels": {
-  //   "deepseek": ["deepseek-v4-flash", "deepseek-v4-pro", "deepseek-v4-flash-vision-exp"]
+  //   "deepseek": ["deepseek-flash", "deepseek-v4-pro"]
   // },
 
   // Interactive Runtime hosting: "embedded" (this process) or "daemon".

package/dist/chunks/{agent-MNDAK2RR.js → agent-MUNSGDNO.js}

@@ -1,2 +1,2 @@
 // @kodax-ai/kodax — bundled distribution. See docs/ADR.md ADR-022 + ADR-024.
-import{Aa as e,Ba as f,Ra as g,Ta as h,a as c,hb as i,ib as j,qb as k,za as d}from"./chunk-EFNU5LXL.js";import"./chunk-SR3TFWQX.js";import"./chunk-GGRVYK4Z.js";import"./chunk-WXWFHOHS.js";import"./chunk-XR4BBXXB.js";import{Ob as a,Pb as b}from"./chunk-7K7RYMWX.js";import"./chunk-VIFNL7HA.js";import"./chunk-HT65FRF4.js";import"./chunk-ONUPGMER.js";export{e as bucketProviderPayloadSize,g as buildAutoRepoIntelligenceContext,f as checkPromiseSignal,b as cleanupIncompleteToolCalls,h as describeTransientProviderRetry,c as emitResilienceDebug,d as estimateProviderPayloadBytes,k as runKodaX,j as saveRequiredSessionSnapshot,i as saveSessionSnapshot,a as validateAndFixToolHistory};
+import{Aa as e,Ba as f,Ra as g,Ta as h,a as c,hb as i,ib as j,qb as k,za as d}from"./chunk-POVUU6P2.js";import"./chunk-JA2DY4EC.js";import"./chunk-H6NNJO5M.js";import"./chunk-5KG5LAUO.js";import"./chunk-Y6VGGIKD.js";import{Ob as a,Pb as b}from"./chunk-F7JMG426.js";import"./chunk-WXBRPZMI.js";import"./chunk-V7ZTP5SV.js";import"./chunk-ONUPGMER.js";export{e as bucketProviderPayloadSize,g as buildAutoRepoIntelligenceContext,f as checkPromiseSignal,b as cleanupIncompleteToolCalls,h as describeTransientProviderRetry,c as emitResilienceDebug,d as estimateProviderPayloadBytes,k as runKodaX,j as saveRequiredSessionSnapshot,i as saveSessionSnapshot,a as validateAndFixToolHistory};

package/dist/chunks/argument-completer-6TSU5AEB.js

@@ -0,0 +1,2 @@
+// @kodax-ai/kodax — bundled distribution. See docs/ADR.md ADR-022 + ADR-024.
+import{Ma as a,Na as b}from"./chunk-LCGEVHAP.js";import"./chunk-NIGQTWGU.js";import"./chunk-YURCMKAL.js";import"./chunk-ZSJJECST.js";import"./chunk-HXH5TOAV.js";import"./chunk-POVUU6P2.js";import"./chunk-JA2DY4EC.js";import"./chunk-B4FK5XVU.js";import"./chunk-H6NNJO5M.js";import"./chunk-5KG5LAUO.js";import"./chunk-Y6VGGIKD.js";import"./chunk-F7JMG426.js";import"./chunk-WXBRPZMI.js";import"./chunk-V7ZTP5SV.js";import"./chunk-ONUPGMER.js";export{a as ArgumentCompleter,b as createArgumentCompleter};