@kodax-ai/kodax 0.7.88 → 0.7.89

package/CHANGELOG.md

@@ -6,7 +6,61 @@ All notable changes to this project will be documented in this file.
 
 ## [Unreleased]
 
-<!-- Keep this section empty between releases. -->
+---
+
+## [0.7.89] - 2026-08-16
+
+### Fixed
+
+- Managed compaction context replacement no longer makes a topology-valid
+  ordinary conversation ambiguous or exposes physical branch/copy duplicates.
+  Conversation page-cache v3 is invalidated so affected Sessions rebuild the
+  corrected projection after upgrade, and the incremental append fast path
+  passes managed-context envelopes through without projecting them, so
+  managed-heavy Sessions keep warm page caches.
+- Managed-context transparency now keys on the `_source` tag alone (matching
+  the compaction stripping side) and also covers the lineage-unavailable
+  fallback projection.
+
+### Added
+
+- Host Tools first-class visibility (`FEATURE_294`): host tools bound to a
+  run by lease now materialize into the agent tool table as run-scoped
+  definitions (name + schema) on both the SA and managed-agent paths, so
+  models see them without calling `mcp_search`. The cached capability
+  catalog gains a `## Host Capability Provider (run-bound)` line
+  advertising the bound server entry, tool names, and a content-hash
+  revision that stays stable across runs. The `runBoundHostTools`
+  capability bumps to `2` (`materializedAgentTools`).
+- Resilient zero-service web search (`FEATURE_293`) now tries DuckDuckGo HTML,
+  Bing RSS, and Bing HTML in a bounded order, preserving truthful fallback
+  diagnostics, normalized direct URLs, freshness metadata, and isolated custom
+  endpoint behavior without adding a hosted service or runtime dependency.
+
+### Changed
+
+- Materialized host tool dispatch routes through the run's capability
+  channel (`executeCapability('mcp', 'host:<leaseId>:<name>')`), reusing the
+  reverse-bridge timeout/idempotency/invocation state machine; results
+  render through the `mcp_call` retrieval pipeline. Host tool descriptors
+  map conservatively: `none` → readonly + plan-mode allowed,
+  `idempotent`/`non_idempotent` → mutates-state + plan-mode blocked
+  (fail-closed). Revoking a lease removes the materialized tools and the
+  catalog line from subsequent turns.
+- Run bindings whose host tool names collide with registered tools are
+  rejected up front with `invalid_params` (before any host-tool run
+  record), and host tool lease ids must match
+  `[A-Za-z0-9][A-Za-z0-9_-]{0,127}` (colon stays reserved for capability
+  ids).
+- Remote managed agents (A2A) can now authorize host capability ids: the
+  `authorizeMcp` id pattern accepts `host:` leases and role tool policy
+  carries host tool authorization semantics.
+
+### Documentation
+
+- Updated the release checklist, product/architecture/design baselines,
+  feature tracker, public guides, `kodax_manual`, and v0.7.89 regression
+  guides for FEATURE_293, FEATURE_294, and Issue 293.
 
 ---
 

package/README.md

@@ -692,6 +692,15 @@ the `glm-latest` alias. Coding Plan model IDs are sent verbatim without a
 synthetic context suffix, and GLM-5.3 `off` / `none` intent is lowered to
 `low`. See the [v0.7.88 release checklist](docs/release.md#v0788-release-preparation).
 
+**v0.7.89 release:** this release ships Issue 293's topology-transparent managed
+context projection and v4 conversation page cache, FEATURE_293's bounded
+zero-service web search fallback (DuckDuckGo HTML → Bing RSS → Bing HTML), and
+FEATURE_294's run-scoped Host Tools. Host tools appear in the leased run's model
+surface and capability catalog, dispatch registry-first, revoke fail-closed, and
+never enter the global registry or unrelated CLI runs. The web-search custom
+endpoint remains isolated, and no shell/sandbox system code changed in this
+release. See the [v0.7.89 release checklist](docs/release.md#v0789-release-preparation).
+
 **v0.7.87 GLM provider release:** `zhipu-coding` defaults to `glm-5.3` and
 retains `glm-5.2` as an explicit rollback route. `zai-coding` retained both
 models but defaulted to `glm-5.2` until the overseas Coding Plan rollout

package/README_CN.md

@@ -249,6 +249,8 @@ v0.7.93。详见 [v0.7.83 发布清单](docs/release.md#v0783-release-preparatio
 
 **v0.7.88 发布**：本版本包含 Actor settlement convergence v2 持久化边界、受界定的启动/恢复工作、guardrail classifier reason 诊断，以及首次提交查询后自动收起过期 learning-recovery 提示的 REPL 修复。这些都是明确的 Runtime、Agent、LLM 与 REPL 系统代码改动。`zhipu-coding`、`zai-coding` 与 `ark-coding` 均默认使用 `glm-5.3`，同时保留 `glm-5.2` 作为显式路由；Ark 继续保留 `glm-latest` 别名。Coding Plan model ID 原样发送，不附加虚假的 context 后缀；GLM-5.3 的 `off` / `none` 会降为 `low`。详见 [v0.7.88 发布清单](docs/release.md#v0788-release-preparation)。
 
+**v0.7.89 发布**：本版本包含 Issue 293 的 managed context 拓扑透明投影与 v4 conversation page cache、FEATURE_293 的有界零服务 web search fallback（DuckDuckGo HTML → Bing RSS → Bing HTML），以及 FEATURE_294 的 run-scoped Host Tools。绑定到 Run 的 Host Tools 会出现在该 Run 的模型工具表和 capability catalog 中，按 registry-first 规则 dispatch，撤销后 fail-closed，不进入全局 registry，也不会泄漏到其他 CLI Run。自定义 web-search endpoint 仍保持隔离；本版本没有修改 shell/sandbox 系统代码。详见 [v0.7.89 发布清单](docs/release.md#v0789-release-preparation)。
+
 **v0.7.87 GLM Provider 发布**：`zhipu-coding` 默认使用 `glm-5.3`，并保留 `glm-5.2` 作为显式回退路由。`zai-coding` 当时保留两个模型但默认使用 `glm-5.2`，直到 v0.7.88 的海外 Coding Plan 路由切换。Coding Plan model ID 原样发送，不附加 context 后缀。GLM-5.3 的思考不可关闭，因此 `off` / `none` 会降级为 `low`，不会发送上游不支持的 disabled-thinking 请求。Issue 256 剩余的 Worker owner lease 边界在 v0.7.87 后仍保持 Open，且本版本不指定新的替代目标。详见 [v0.7.87 发布清单](docs/release.md#v0787-release-preparation)。
 
 Windows workspace Shell 还会保留大小写不敏感的 `PATH`/`Path` 与 `PATHEXT` 约定，

package/dist/chunks/{agent-ZVZPWX5T.js → agent-AKCXT573.js}

@@ -1,2 +1,2 @@
 // @kodax-ai/kodax — bundled distribution. See docs/ADR.md ADR-022 + ADR-024.
-import{Ka as g,Ma as h,Za as i,_a as j,a as c,sa as d,ta as e,ua as f,yc as k}from"./chunk-OG4ML65E.js";import"./chunk-O2ZV775M.js";import"./chunk-6MGD2A7Z.js";import"./chunk-W5RPVR3O.js";import"./chunk-PROPMIT4.js";import"./chunk-UNEC2WXJ.js";import{Jb as a,Kb as b}from"./chunk-72H6JNPN.js";import"./chunk-5247AKGN.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{Ka as g,Ma as h,Za as i,_a as j,a as c,sa as d,ta as e,ua as f,yc as k}from"./chunk-A63MGLH6.js";import"./chunk-O2ZV775M.js";import"./chunk-5UBP47VJ.js";import"./chunk-W5RPVR3O.js";import"./chunk-PROPMIT4.js";import"./chunk-UNEC2WXJ.js";import{Jb as a,Kb as b}from"./chunk-72H6JNPN.js";import"./chunk-5247AKGN.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-BJGY57H2.js

@@ -0,0 +1,2 @@
+// @kodax-ai/kodax — bundled distribution. See docs/ADR.md ADR-022 + ADR-024.
+import{Ha as a,Ia as b}from"./chunk-CQHLFQSU.js";import"./chunk-SOIIBADK.js";import"./chunk-4Y6OSDZE.js";import"./chunk-S74F7MFN.js";import"./chunk-IFD6TUQU.js";import"./chunk-A63MGLH6.js";import"./chunk-O2ZV775M.js";import"./chunk-BOMSVUJS.js";import"./chunk-5UBP47VJ.js";import"./chunk-W5RPVR3O.js";import"./chunk-PROPMIT4.js";import"./chunk-UNEC2WXJ.js";import"./chunk-72H6JNPN.js";import"./chunk-5247AKGN.js";import"./chunk-ONUPGMER.js";export{a as ArgumentCompleter,b as createArgumentCompleter};