vertex-palace 0.2.1 → 0.2.3

package/BUILD_WEEK.md

@@ -30,6 +30,22 @@ Version 0.1.5 adds `palace evaluate` and the `palace_evaluate` MCP tool. The eva
 
 This feature addresses a concrete limitation found during large-project use: a plausible-looking route and a high confidence value do not prove that the route included the files that ultimately mattered.
 
+### Cross-ecosystem release routing
+
+Version 0.2.3 adds a dedicated release task type after the tool misclassified
+its own 0.2.2 publication task as `unknown`. The frozen real-repository baseline
+found only 3/19 changed files with 0.25 route focus and 0.52 overconfidence
+error. The corrected route reached 12/19 coverage, 1.00 focus, and 0.02
+calibration error.
+
+A protocol committed before matrix execution then tested JavaScript monorepo,
+Codex plugin, Python/PyPI, and Chinese release scenarios plus publication
+failure, explanation, review, test-only, and deployment controls. The final
+matrix passed without lowering thresholds, and all scored routes reached 1.00
+focus. Rejected intermediate results remain in the repository research record.
+This study measures routing quality and does not claim lower total agent tokens
+or faster wall-clock time.
+
 ## Codex And GPT-5.6 Collaboration
 
 The submission-period work was developed with Codex using GPT-5.6. Codex was used to:
@@ -54,7 +70,7 @@ The Devpost submission will include the `/feedback` Session ID from the primary
 Install the public competition release, then run it in the repository you want to evaluate:
 
 ```bash
-npm install -g vertex-palace@0.2.1
+npm install -g vertex-palace@0.2.3
 palace --version
 palace context "improve route confidence calibration" --auto --format json
 palace evaluate "improve route confidence calibration" --changed-file packages/core/src/evaluation/evaluate-route.ts

package/CHANGELOG.md

@@ -2,6 +2,51 @@
 
 ## Unreleased
 
+## 0.2.3 - 2026-07-19
+
+### Added
+
+- A dedicated `release` task type with English and Chinese publication intent recognition.
+- Cross-ecosystem package-manifest routing for npm, PyPI, Cargo, Go, Maven/Gradle, Composer, Ruby, and .NET projects.
+- Reproducible release-routing matrix fixtures for JavaScript monorepos, Codex plugins, Python/PyPI, Chinese release tasks, and negative intent controls.
+- Opt-in structured test reporters for the frozen real-repository release task and the cross-ecosystem matrix.
+
+### Changed
+
+- Release routes reserve capacity for implementation, regression tests, package manifests, plugin metadata, release records, MCP, CLI, shared contracts, and CI surfaces.
+- Explicit route limits are treated as maxima instead of targets; unused capacity no longer pulls unrelated files into a release route.
+- Package-manifest selection is ecosystem-aware, and release test slots require release-related path evidence.
+- Release-route confidence remains conservatively capped and is evaluated against observed changed-file coverage.
+
+### Fixed
+
+- Prevented actual npm, Git tag, and plugin publication work from falling through to `unknown`.
+- Prevented explanation, review, test-only, and publication-failure tasks from being misclassified as releases.
+- Recognized `pyproject.toml` and other non-JavaScript package manifests as release surfaces.
+- Preserved the caller's explicit route limit for plugin distribution tasks.
+
+### Research
+
+- The frozen Vertex Palace release task improved from 3/19 changed-file coverage and 0.25 route focus to 12/19 coverage and 1.00 focus, with calibration error reduced from 0.52 to 0.02.
+- The preregistered cross-ecosystem matrix passed without lowering thresholds; all scored final routes reached 1.00 focus.
+- These routing results do not claim lower total agent tokens or faster wall-clock time.
+
+## 0.2.2 - 2026-07-19
+
+### Fixed
+
+- Preserve relevant, scoped memory when Adaptive mode selects `full-palace` instead of silently disabling the memory section.
+- Render Full Palace memory and its current-code-first guardrail in both Markdown and JSON while keeping the final serialized payload inside the selected budget.
+
+### Added
+
+- Regression coverage using the exact Aurora useful-memory task from the public 16-trial Adaptive v2.2 benchmark.
+- Memory-fidelity assertions for mode selection, delivered Markdown, delivered JSON, item counts, guardrails, and measured payload bytes.
+
+### Research
+
+- The public [Adaptive v2.2 benchmark](https://github.com/lohchanhin/benchmarks-ab-demo/blob/main/docs/research/ADAPTIVE_V2_2_FINAL.md) found the v0.2.1 omission in all four useful-memory trials. The frozen results remain unchanged; this release fixes the product behavior separately.
+
 ## 0.2.1 - 2026-07-19
 
 ### Fixed

package/README.md

@@ -37,10 +37,14 @@ The first floor also has an entrance pitfall board at `.palace/00-entrance/pitfa
 | --- | --- | --- |
 | `bypass` | A small repository and one explicit file make routing overhead unnecessary | No source content is packed; direct inspection is recommended |
 | `route-lite` | The task and route are focused with low cross-cutting risk | Loads Primary summaries or symbol snippets; keeps the rest as references |
-| `full-palace` | The task crosses layers, changes a public contract, or has uncertain routing | Loads Primary plus bounded Support context |
-| `guarded-memory-palace` | Prior decisions, stale behavior, or tenant isolation are relevant | Adds only scoped, recent, budgeted memory with contradiction warnings |
+| `full-palace` | The task crosses layers, changes a public contract, or has uncertain routing | Loads Primary plus bounded Support context and relevant scoped memory when available |
+| `guarded-memory-palace` | Prior decisions, stale behavior, or tenant isolation are explicitly relevant | Prioritizes scoped, recent, budgeted memory with contradiction warnings |
 
 Adaptive mode reports `contextBytes`, `contextEstimatedTokens`, route-tier counts, memory usage, and guardrails in Markdown and JSON. It is designed to reduce unnecessary preloaded context. It does **not** guarantee lower total agent tokens or faster wall-clock time on every task; those outcomes must be measured end to end with repeated, order-balanced trials.
+
+Version 0.2.2 preserves relevant memory when Adaptive selects `full-palace`. This fixes the omission reproduced in all four useful-memory trials of the public [Adaptive v2.2 benchmark](https://github.com/lohchanhin/benchmarks-ab-demo/blob/main/docs/research/ADAPTIVE_V2_2_FINAL.md); the frozen v0.2.1 research results remain unchanged. The [0.2.2 remediation record](https://github.com/lohchanhin/vertex-palace/blob/main/docs/research/ADAPTIVE_MEMORY_FIX_0_2_2.md) documents the root cause and clean-install evidence.
+
+Version 0.2.3 adds a dedicated release-task route across implementation, tests, package manifests, plugin metadata, MCP, CLI, shared contracts, CI, and release records. It recognizes npm, PyPI, Cargo, Go, Maven/Gradle, Composer, Ruby, and .NET manifests while keeping explain, review, test-only, and publish-failure intent separate. The frozen [baseline](./docs/research/RELEASE_ROUTING_BASELINE_0_2_2.md), [real-repository result](./docs/research/RELEASE_ROUTING_RESULT_0_2_3.md), and [cross-ecosystem matrix](./docs/research/RELEASE_ROUTING_MATRIX_RESULT_0_2_3.md) preserve both failed intermediate results and final evidence. This routing study does not claim lower total agent tokens or faster wall-clock time.
 
 ## Usage Guide
 
@@ -60,17 +64,17 @@ The agent should read the repository instructions, install the plugin or use the
 adaptive context (one call) -> inspect Primary -> expand Deferred only when needed -> implement -> test -> evaluate -> memory write
 ```
 
-Only `guarded-memory-palace` includes relevant entries from `.palace/00-entrance/pitfall-board.md`; unrelated or old memory is omitted.
+`full-palace` and `guarded-memory-palace` can include relevant entries from `.palace/00-entrance/pitfall-board.md`; guarded mode is selected when memory risk is explicit and adds stricter contradiction warnings. Unrelated or old memory is omitted.
 
 ### Manual Mode: Step By Step
 
 1. Install the plugin:
 
 ```bash
-codex plugin marketplace add lohchanhin/vertex-palace --ref v0.2.1
+codex plugin marketplace add lohchanhin/vertex-palace --ref v0.2.3
 ```
 
-Use `v0.2.1` or newer for the corrected Adaptive selector. Avoid `v0.1.0` and `v0.1.1`; those early tags had broken MCP install metadata.
+Use `v0.2.3` or newer for release routing, Full Palace memory fidelity, and the corrected Adaptive selector. Avoid `v0.1.0` and `v0.1.1`; those early tags had broken MCP install metadata.
 
 2. Or install the CLI directly from npm:
 
@@ -165,10 +169,14 @@ Vertex Palace 是一个面向 Codex 编程任务的本地上下文路由工具
 | --- | --- | --- |
 | `bypass` | 小型仓库且任务明确指向单一文件 | 不预载源码，建议直接检查目标文件 |
 | `route-lite` | 任务集中、路线明确、跨层风险低 | 只加载 Primary 摘要或符号片段，其余保留引用 |
-| `full-palace` | 跨前后端、公共契约变更或路线不确定 | 加载 Primary 与有限的 Support 上下文 |
-| `guarded-memory-palace` | 历史决策、旧行为或客户隔离确实相关 | 只加入有作用域、有时效、有预算并带矛盾检查的记忆 |
+| `full-palace` | 跨前后端、公共契约变更或路线不确定 | 加载 Primary、有限的 Support，以及可用时与任务相关的范围化记忆 |
+| `guarded-memory-palace` | 历史决策、旧行为或客户隔离明确相关 | 优先加入有作用域、有时效、有预算并带矛盾检查的记忆 |
 
 Adaptive Palace 的目标是减少不必要的“预载上下文”，不是承诺每个任务的总 Token 与总时间都一定下降。端到端效果必须用相同任务、重复执行、平衡顺序的 A/B 测试判断。
+
+0.2.2 修复了 Adaptive 选择 `full-palace` 时遗漏相关记忆的问题。这个缺陷由公开的 [Adaptive v2.2 基准研究](https://github.com/lohchanhin/benchmarks-ab-demo/blob/main/docs/research/ADAPTIVE_V2_2_FINAL.md)在四组 useful-memory trial 中稳定复现；冻结的 v0.2.1 研究结果不会被改写。[0.2.2 修复记录](https://github.com/lohchanhin/vertex-palace/blob/main/docs/research/ADAPTIVE_MEMORY_FIX_0_2_2.md)公开根因与 clean-install 证据。
+
+0.2.3 新增专门的发布任务路线，会同时覆盖实现、回归测试、package manifest、plugin metadata、MCP、CLI、shared contract、CI 和发布记录。它支持 npm、PyPI、Cargo、Go、Maven/Gradle、Composer、Ruby 与 .NET manifest，并把说明、审核、仅测试和发布失败任务与真正发布动作分开。冻结的[基线](./docs/research/RELEASE_ROUTING_BASELINE_0_2_2.md)、[真实仓库结果](./docs/research/RELEASE_ROUTING_RESULT_0_2_3.md)和[跨生态矩阵](./docs/research/RELEASE_ROUTING_MATRIX_RESULT_0_2_3.md)都保留失败中间结果与最终证据；这项路由研究不宣称总 Token 或总时间一定下降。
 
 ## 使用说明
 
@@ -188,17 +196,17 @@ Adaptive Palace 的目标是减少不必要的“预载上下文”，不是承
 adaptive context（单次调用）-> 先读 Primary -> 证据需要时再展开 Deferred -> 执行任务 -> 测试 -> evaluate -> 写入 memory
 ```
 
-只有 `guarded-memory-palace` 会带入 `.palace/00-entrance/pitfall-board.md` 中与当前任务相关且仍在时效内的记录。
+`full-palace` 与 `guarded-memory-palace` 都可以带入 `.palace/00-entrance/pitfall-board.md` 中与当前任务相关且仍在时效内的记录；当任务明确涉及记忆风险时会选择 guarded 模式，并加入更严格的矛盾检查。
 
 ### 勤劳用法：自己一步一步跑
 
 1. 安装插件：
 
 ```bash
-codex plugin marketplace add lohchanhin/vertex-palace --ref v0.2.1
+codex plugin marketplace add lohchanhin/vertex-palace --ref v0.2.3
 ```
 
-请使用 `v0.2.1` 或更新版本，以启用修正后的 Adaptive 模式选择器。不要再使用 `v0.1.0` 和 `v0.1.1`，这两个早期标签的 MCP 安装元数据有问题。
+请使用 `v0.2.3` 或更新版本，以启用发布任务路由、Full Palace 记忆一致性与修正后的 Adaptive 模式选择器。不要再使用 `v0.1.0` 和 `v0.1.1`，这两个早期标签的 MCP 安装元数据有问题。
 
 2. 打开 Codex，输入 `/plugins`，安装 Vertex Palace。
 
@@ -268,15 +276,15 @@ npx vertex-palace status
 Install the Codex plugin:
 
 ```bash
-codex plugin marketplace add lohchanhin/vertex-palace --ref v0.2.1
+codex plugin marketplace add lohchanhin/vertex-palace --ref v0.2.3
 ```
 
-Use `v0.2.1` or newer for the corrected Adaptive selector; `v0.1.0` and `v0.1.1` are obsolete because their MCP install metadata was broken.
+Use `v0.2.3` or newer for release routing, Full Palace memory fidelity, and the corrected Adaptive selector; `v0.1.0` and `v0.1.1` are obsolete because their MCP install metadata was broken.
 
 The plugin launches its MCP server through npm:
 
 ```bash
-npx -y -p vertex-palace@0.2.1 vertex-palace-mcp-stdio --stdio
+npx -y -p vertex-palace@0.2.3 vertex-palace-mcp-stdio --stdio
 ```
 
 Open Codex: