@maxiaochao/pi-toolkit 0.9.4 → 0.10.0

package/CHANGELOG.md

@@ -1,3 +1,21 @@
+## 2026-10-08 - v0.10.0
+
+- bundle herdr-subagents v0.2.0: replace Tiny Subagent with visible interactive Herdr agents and durable batch workflows
+- replace Tiny Subagent's hidden print/shell execution with Herdr Subagents: visible interactive Pi Agents, one exclusive on-demand pane per task, Fleet keyboard navigation, durable blocked/history continuation, parent-session cancellation, and focus-safe batch tab cleanup
+- fix Herdr Subagents Fleet state after blocked responses, preserve the active task widget while later background runs queue, and share ownership, pane-allocation, and status-reduction logic across recovery paths
+
+## 2026-10-08 - v0.9.6
+
+- bundle codex-edit v0.1.9: install the pinned Pi runtime for clean-runner publication tests
+- install the pinned Pi test runtime on clean publication runners before executing Codex Edit integration checks
+
+## 2026-10-08 - v0.9.5
+
+- bundle codex-edit v0.1.8: fix repeated same-path updates and publish reproducible benchmark evidence
+- harden bundled Codex Edit for repeated updates to one path and add mandatory native-vs-Codex and released-vs-local comparisons over a growing regression corpus
+- generate and bundle the latest Codex Edit Markdown and self-contained HTML test report on every test run
+- commit the completed 100-task benchmark evidence and make report generation deterministic on clean checkouts
+
 ## 2026-09-30 - v0.9.4
 
 - retain Codex Edit in the root toolkit while restoring `@maxiaochao/pi-codex-edit` as an independently installable child package; both distributions share the implementation under `packages/codex-edit`

package/README.md

@@ -25,7 +25,7 @@ The root package embeds the child-package source snapshots listed below. Install
 | [pi-todo](packages/todo/README.md) | Session plans, automatic task advancement, `/todos` | `pi install npm:@maxiaochao/pi-todo` | `pi update npm:@maxiaochao/pi-todo` |
 | [pi-cache-export](packages/cache-export/README.md) | Tau-aligned interactive cache dashboard for Pi sessions | `pi install npm:@maxiaochao/pi-cache-export` | `pi update npm:@maxiaochao/pi-cache-export` |
 | [pi-codex-edit](packages/codex-edit/README.md) | Codex-harness `apply_patch` routing for selected GPT models | `pi install npm:@maxiaochao/pi-codex-edit` | `pi update npm:@maxiaochao/pi-codex-edit` |
-| [pi-tiny-subagent](packages/tiny-subagent/README.md) | One focused task in a fresh child Pi context | `pi install npm:@maxiaochao/pi-tiny-subagent` | `pi update npm:@maxiaochao/pi-tiny-subagent` |
+| [pi-herdr-subagents](packages/herdr-subagents/README.md) | Interactive Herdr Pi Agents with bounded concurrency, direct navigation, and recovery | `pi install npm:@maxiaochao/pi-herdr-subagents` | `pi update npm:@maxiaochao/pi-herdr-subagents` |
 | [pi-worktree](packages/worktree/README.md) | Worktree CLI, completion, installer, and agent skill | `pi install npm:@maxiaochao/pi-worktree` | `pi update npm:@maxiaochao/pi-worktree` |
 
 Each child README owns its feature description, motivation or upstream alignment, examples, screenshots/demos, configuration, and test documentation.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maxiaochao/pi-toolkit",
-  "version": "0.9.4",
+  "version": "0.10.0",
   "description": "Personal pi coding agent toolkit: /todos, /btw, /cache_export and apply_patch extensions, browser automation, and bundled theme/skills.",
   "type": "module",
   "workspaces": [
@@ -15,7 +15,7 @@
     "packages/todo",
     "packages/cache-export",
     "packages/codex-edit",
-    "packages/tiny-subagent",
+    "packages/herdr-subagents",
     "packages/worktree",
     "themes",
     "README.md",
@@ -34,7 +34,7 @@
       "./packages/cache-export/extensions/cache-export.ts",
       "./extensions/btw/index.ts",
       "./packages/codex-edit/extensions/codex-edit.ts",
-      "./packages/tiny-subagent/extensions/tiny-subagent.ts"
+      "./packages/herdr-subagents/extensions/herdr-subagents.ts"
     ],
     "skills": [
       "./skills",
@@ -57,8 +57,8 @@
     "ws": "^8.19.0"
   },
   "scripts": {
-    "test": "node tests/package-layout.mjs && npm --prefix packages/todo test && npm --prefix packages/cache-export test && npm --prefix packages/codex-edit test && npm --prefix packages/tiny-subagent test && npm --prefix packages/worktree test && node --experimental-strip-types extensions/btw/tests/history.mjs",
-    "load-test": "pi -ne -e ./packages/todo/extensions/todo.ts && pi -ne -e ./packages/cache-export/extensions/cache-export.ts && pi -ne -e ./extensions/btw/index.ts && pi -ne -e ./packages/codex-edit/extensions/codex-edit.ts && pi -ne -e ./packages/tiny-subagent/extensions/tiny-subagent.ts",
+    "test": "node tests/package-layout.mjs && npm --prefix packages/todo test && npm --prefix packages/cache-export test && npm --prefix packages/codex-edit test && npm --prefix packages/herdr-subagents test && npm --prefix packages/worktree test && node --experimental-strip-types extensions/btw/tests/history.mjs",
+    "load-test": "pi -ne -e ./packages/todo/extensions/todo.ts && pi -ne -e ./packages/cache-export/extensions/cache-export.ts && pi -ne -e ./extensions/btw/index.ts && pi -ne -e ./packages/codex-edit/extensions/codex-edit.ts && pi -ne -e ./packages/herdr-subagents/extensions/herdr-subagents.ts",
     "eval:agents": "node evaluations/agents-md/run.mjs",
     "eval:agents:test": "node evaluations/agents-md/tests/run-tests.mjs",
     "postinstall": "node scripts/install-agents.mjs"

package/packages/codex-edit/CHANGELOG.md

@@ -1,3 +1,18 @@
+## 2026-10-08 - v0.1.9
+
+- install the pinned Pi runtime for clean-runner publication tests
+- install the pinned Pi test runtime on clean publication runners before executing native comparison and extension-load checks
+
+## 2026-10-08 - v0.1.8
+
+- fix repeated same-path updates and publish reproducible benchmark evidence
+- apply repeated plain `Update File` blocks for one path in order while retaining conflict rejection for add/delete/move targets
+- expand regressions and testing policy for repeated regions, empty chunks, stale context, malformed arguments, and project-scale patches
+- add mandatory real-native-edit versus local and released-v0.1.7 versus local comparisons over one growing scenario corpus
+- generate a Markdown test report and self-contained HTML report on every package test run
+- interleave the frozen Neutral/Challenge schedule at 4:1 and add five production-incident regression entries so budget-limited prefixes still exercise reported failure shapes
+- complete the frozen 100-task three-arm benchmark and commit sanitized evidence so clean checkouts reproduce the reports without local cache data
+
 ## 2026-09-30 - v0.1.7
 
 - match Codex `apply_patch` semantics by selecting the first matching region at or after the current hunk cursor

package/packages/codex-edit/README.md

@@ -63,10 +63,15 @@ The tool accepts raw Codex patch text between `*** Begin Patch` and `*** End Pat
 - `Update File`
 - `Move to`
 - multiple update hunks with `@@` context
+- repeated `Update File` blocks for the same path, applied in patch order
 - EOF-oriented insertions
 
 Before writing, it parses the complete patch, resolves paths, rejects workspace escapes and unsafe symlink/parent paths, reads all source files, and computes every update in memory. Only after preflight succeeds does it mutate files. It returns per-file diff details and reports committed paths if a later filesystem operation fails.
 
+Repeated plain updates to one file are coalesced into one file operation and each block sees the result of the previous block. Conflicting reuse of a path by add/delete/move operations is still rejected. Empty update chunks, missing `patch` arguments, and stale expected lines fail without guessing or partially applying the patch.
+
+This repeated-header tolerance is a deliberate Pi-side extension: current Codex verified invocation rejects multiple operations resolving to one path and expects multiple `@@` chunks under one `Update File` header. Codex Edit accepts repeated headers only when they are plain updates to the exact same raw path, because this malformed-but-unambiguous shape has occurred in real model output. Resolved aliases such as `src/a.ts` and `src/./a.ts`, moves, adds, and deletes remain conflicting.
+
 ## Configuration
 
 The extension defaults to a model-ID allowlist in `config.json`: `"gpt-5.6-*"`. Provider does not affect routing. Each entry is a `*` glob, so this matches every `gpt-5.6-` variant. The route additionally requires:
@@ -116,8 +121,10 @@ Edit `config.json` in the standalone installed package, or `packages/codex-edit/
 
 ## Documentation
 
-- `summary.md`: benchmark summary and decision record.
+- `summary.md`: historical 20-case benchmark and decision record.
 - `codex-edit-explainer.html`: self-contained interactive architecture and configuration explainer.
-- `TESTING.md`: deterministic test flow, regression fixtures, and latest baseline.
+- `TESTING.md`: deterministic test flow, mandatory native-vs-Codex and released-vs-local comparisons, growing regression corpus, and latest baseline.
+- `docs/test-summary.md`: current short test summary.
+- `docs/test-report.html`: current self-contained report; its authoritative source is `docs/test-report.md`.
 
-The benchmark found equal final correctness in the fair 20-case comparison, with first editor-call success improving from 85% to 95%. Overall tool-call count and cache-inclusive token use were effectively unchanged, so the route stays limited to the GPT/Codex allowlist rather than replacing `edit` universally.
+The historical 20-case benchmark found equal final correctness, with first editor-call success improving from 85% to 95%. The completed frozen 100-task three-arm benchmark (`explicit-edit-100-sol-low-v3-complete`) ran all 300 chains: native and released v0.1.7 scored 91.25%, while the local version scored 92.00%. All arms reached 92% Final Exact, all eight final failures were shared, and there was no local-only final regression. The sanitized run evidence is committed under `benchmark/results/`; the generated report contains the full methodology, paired results, efficiency, and safety evidence.

package/packages/codex-edit/TESTING.md

@@ -2,19 +2,50 @@
 
 ## 流程
 
-1. `npm --prefix packages/codex-edit test`：运行 parser、matcher、路由和包结构测试。
+1. `npm --prefix packages/codex-edit test`：运行 parser、matcher、路由、工具契约、包结构、下面两组强制对比、入口加载和 tarball 检查；随后生成 `docs/test-report.md` 并从该 Markdown 转换为自包含 `docs/test-report.html`。这是秒级反馈环，修改 parser 前必须先让对应回归用例变红。
 2. `pi -ne -e ./packages/codex-edit/extensions/codex-edit.ts`：加载权威入口；独立子包与根包 manifest 都指向该实现。
 3. `npm pack ./packages/codex-edit --dry-run`：检查独立 tarball。
-4. 确认 README 的协议示例、真实 parser 运行输出 `docs/screenshot.png` 和 `docs/demo.gif` 随包发布。
+4. 在根目录运行 `npm test`、`npm run load-test` 和 `npm pack --dry-run`，确认主包内嵌的同一份实现没有退化。
+5. 确认 README 的协议示例、真实 parser 运行输出 `docs/screenshot.png` 和 `docs/demo.gif` 随两个 tarball 发布。
 
-## 用例数据
+`docs/test-report.md` 是报告内容的权威中间产物，HTML 是每次测试生成的交付视图。提交或发布 Codex Edit 改动时必须包含两者，且 HTML 顶部 Overall 必须为 PASS。
 
-- Add/Delete/Update/Move、多 hunk、BOM、CRLF、EOF 和 workspace escape。
-- 与 Codex 一致的 cursor 顺序匹配：重复 HTML section 与 Markdown `---`。
-- 多文件结构重构 patch 的解析，以及重复目标拒绝。
-- `gpt-5.6-*` glob、Responses API 路由和 disabled 配置。
-- 自定义精确规则与 `*` glob；相同 model ID 在不支持的 API 上必须回退到 Pi `edit`。
+## 两组强制对比
+
+1. `npm --prefix packages/codex-edit run test:compare-native`：使用当前 Pi 公开导出的真实 `createEditToolDefinition` 和本地 extension 实际注册的 `apply_patch`，将两个真实工具跑在同一组临时 workspace 与期望文件快照上；失败 case 还要求 Codex Edit 不产生部分写入。它比较编辑协议能力和确定性结果，不替代真实模型质量 benchmark。
+2. `npm --prefix packages/codex-edit run test:compare-released`：将冻结的 npm `v0.1.7` parser 语义与本地 parser 跑在同一个 corpus 上，明确显示旧版失败而本地修复成功的 case，并锁住旧有成功/安全失败行为。
+
+共享语料在 `tests/comparison-cases.mjs`。每次出现新的生产报错，先把最小复现加入此文件，并分别声明 native、released、local 的预期；不要改写旧 case 来让测试变绿。冻结基线 `tests/legacy-parser-v0.1.7.mjs` 不得随本地实现修改。若以后发布新基线，应新增版本文件并保留旧基线。
+
+这套结构参考 OpenAI Codex 的 [`apply-patch` scenario tests](https://github.com/openai/codex/blob/9a8730f3/codex-rs/apply-patch/tests/suite/scenarios.rs)：以输入文件、patch 和期望最终文件系统为真值；parser 变更还应做旧/新差分。上游在统一 streaming parser 时另用 24 小时 2,788,059 个真实 payload 做过差分验证（[`e26f734`](https://github.com/openai/codex/commit/e26f734f919819bdd802b1b67ad653863c7a90e5)）。本仓库没有该生产 corpus，因此不能宣称同等覆盖；这里只采用可持续扩充的方法。
+
+当前 Codex verified invocation 会拒绝解析后指向同一路径的多个 operation（见 [`invocation.rs`](https://github.com/openai/codex/blob/main/codex-rs/apply-patch/src/invocation.rs) 和 [duplicate-path change](https://github.com/openai/codex/pull/37867)）。本地只对“完全相同 raw path + 无 move 的纯 update”提供顺序容错，这是针对已观察模型输出的明确扩展；解析路径别名及 add/delete/move 冲突仍由执行层拒绝。
+
+## 回归矩阵
+
+| 形态 | 预期 | 锁定位置 |
+|---|---|---|
+| Add/Delete/Update/Move、多 hunk、BOM、CRLF、EOF | 正确解析并保留文件格式 | `parser.test.mjs` |
+| 重复 HTML section、Markdown `---` | 从当前 cursor 起选择第一个匹配，不再报 ambiguous | `parser.test.mjs` |
+| 同一路径出现多个纯 `Update File` | 按块顺序执行，后一块可匹配前一块的输出 | `parser.test.mjs` |
+| 同一路径混用 add/delete/move，或 move 目标冲突 | 在写文件前拒绝，避免覆盖 | `parser.test.mjs` |
+| 空 update chunk | 明确拒绝，不猜测意图 | `parser.test.mjs` |
+| expected lines 已过期 | 明确拒绝并带路径，不做模糊替换 | `parser.test.mjs` |
+| `{}` / 缺少 `patch` 参数 | 由必填 TypeBox schema 在 execute 前拒绝 | `package.test.mjs` + Pi load test |
+| 项目级多文件重构、重复 SKILL 分隔符、重复 tiny/release 文件块 | 解析为唯一文件操作并得到确定性结果 | `parser.test.mjs` |
+| workspace escape | 拒绝 | `parser.test.mjs` |
+| `gpt-5.6-*`、Responses API、disabled、自定义 glob | 仅允许规则内模型使用 `apply_patch`，其余回退 `edit` | `package.test.mjs` |
+
+空参数是模型/提供方生成的无效工具调用，Pi 会在扩展 `execute` 之前拦截；parser 无法也不应为缺失 patch 猜内容。`Failed to find expected lines` 表示 patch 基于旧文件内容，安全行为是失败后重新读取并生成新 patch，而不是放宽到可能改错位置。
 
 ## 最近验证结果
 
-当前基线为 12 个确定性测试全部通过；真实触发过 `Ambiguous ... regions found` 的重复 HTML/Markdown 形态已经加入回归集。发布前必须验证子包 tarball 和根包 tarball 都包含该入口。
+当前基线为 22 个 Node 确定性测试及 12 个共享对比 case 全部通过。原生 edit 为 5 pass / 2 fail / 5 unsupported，本地 Codex Edit 为 7 pass / 5 正确安全拒绝；发布版 v0.1.7 为 6 pass / 6 fail，其中 5 个是正确安全拒绝，另 1 个是本地版已修复的同路径顺序 update。另有 5 个生产事故条目冻结真实问题来源，其中 4 个可进入 Agent Benchmark，1 个仅验证指定协议形态。真实触发过的 runner 类型重命名 ambiguous、重复同路径、空 chunk、stale context、解析路径别名冲突、工作区外绝对路径和项目级多文件形态均已进入回归集。发布前必须验证子包 tarball 和根包 tarball 都包含并成功加载该入口。
+
+正式 Agent Benchmark 的预算门禁使用仓库内 `benchmark/run-with-budget.mjs` 包装上游 runner。不要修改 `/tmp` 中的 benchmark clone，也不要允许 `apply_patch` 写出当前工作区；工作区外路径必须作为 Safety Regression 正确拒绝。包装器在 `--stop-cost` 停止调度，并以 `--max-cost` 检查最终硬上限。
+
+冻结清单使用 `benchmark/schedule.mjs` 按 4 Neutral + 1 Challenge 交错排序。任何 5-task 完整前缀都必须保持 80/20；当前调度测试同时断言第 47 个任务处应已覆盖 9 个 Challenge，防止预算停止再次只留下 Neutral。
+
+`summary.md` 的 20-case 报告是历史证据。当前正式运行 `explicit-edit-100-sol-low-v3-complete` 已完成 300/300 条 chain，实际成本 $12.388；本地版 First Exact 92%、Final Exact 92%、Score v2 92.00%，满足全部正式门槛。原生与发布版均为 First 91%、Final 92%、Score v2 91.25%；8 个最终失败为三臂共同例外，本地独有最终回归为 0。
+
+公开、去敏后的正式证据提交在 `benchmark/results/explicit-edit-100-sol-low-v3.json`。正常测试只读取该文件，因此干净 checkout 不依赖 `.cache`，且 Markdown 与 HTML 可逐字节重建。导入新运行时，使用 `benchmark/export-evidence.mjs <run-directory> <output.json>`；只有在确认运行完整性、配置 identity 和无敏感字段后才替换正式证据。预算必须在每次运行前声明 scheduling stop 与 hard cap，预算停止的前缀一律标为 INCOMPLETE，禁止选择性补跑后宣称完整。任何模型路由、prompt 或 tool schema 变化仍需重跑独立模型基准，不能用 parser 测试替代。