pi-dsh-mimic 0.1.0

package/LICENSE

@@ -0,0 +1,22 @@
+MIT License
+
+Copyright (c) 2026 Utopia-V
+Portions Copyright (c) 2026 DeepSeek
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/NOTICE

@@ -0,0 +1,16 @@
+pi-dsh-mimic is a Pi implementation that reproduces the DSH Minimal request
+surface for one task-bearing bootstrap, then returns to Pi-native execution.
+
+Its Minimal persona, bash description, editor description, and function
+schemas reproduce the public DeepSeek Harness Minimal request protocol from:
+
+  https://github.com/deepseek-ai/deepseek-harness
+  commit 47f943859bef60e4160492346772ded9b24f765a
+
+DeepSeek Harness is distributed under the MIT License:
+
+  Copyright (c) 2026 DeepSeek
+
+The retained permission notice is included in LICENSE. Historical experiments
+and research sources are documented separately from the package's source-code
+provenance.

package/README.md

@@ -0,0 +1,155 @@
+[Repository index](../../README.en.md) · [简体中文](README.zh-CN.md) · [Experiments and design](docs/advanced.md) · [Project2 evidence ledger (Chinese canonical)](docs/project2-evidence.md)
+
+# pi-dsh-mimic
+
+`pi-dsh-mimic` reproduces the DeepSeek Harness Minimal environment for the
+first model request in [Pi](https://github.com/earendil-works/pi). DeepSeek V4
+Pro begins from a verified high-capability trajectory; after the first valid
+response, execution returns immediately to Pi with its complete tool catalog
+and the user's other plugins.
+
+One session gets both:
+
+- the V4 Pro engineering trajectory established by a DSH Minimal first request;
+- Pi's mature tools, session management, and plugin ecosystem.
+
+Users do not need to install or run the complete DSH harness, and the session
+does not remain restricted to Minimal's two tools.
+
+## How it works
+
+For a fresh target session, the real task itself becomes the first model
+request:
+
+1. request #1 uses the Minimal persona and exposes only DSH-style `bash` and
+   `str_replace_editor`;
+2. after the first successful assistant response or durable tool call, request
+   #2 restores Pi's native provider payload and complete current tool catalog;
+3. the Minimal persona remains for the session, while Pi's larger generated
+   system prompt is not replayed as an extra user message.
+
+The extension adds no identity warm-up or model round, makes no hidden copy of
+the task, does not inject `We need`, and does not proxy the API.
+
+## Project2 results
+
+Default Pi scored 92. Four final runs of the same one-shot request flow scored
+**98, 96, 96, and 98**, all with F3 at 16/16:
+
+| Run | Provider | Score | Key dimensions |
+| --- | --- | ---: | --- |
+| Default Pi baseline | Official DeepSeek API | 92 | F3 11, F6 10, F8 6 |
+| Early Minimal-mimic prototype | Official DeepSeek API | **98** | F3 16, F6 10, F8 7 |
+| Same-flow replication | Official DeepSeek API | 96 | F3 16, F6 10, F8 5 |
+| First packaged implementation | Official DeepSeek API | 96 | F3 16, F6 8, F8 7 |
+| Current independent implementation `0.1.0` | OpenCode Go | **98** | F3 16, F6 10, F8 7 |
+
+The results show that a DSH Minimal first request can establish a strong V4 Pro
+trajectory inside Pi and that restoring Pi's full tool catalog does not break
+it. See [Experiments and design](docs/advanced.md) for the rejected alternatives
+and the evidence behind each implementation choice. The
+[Project2 evidence ledger](docs/project2-evidence.md), maintained in Chinese,
+contains evaluator IDs and sources.
+
+## Install and use
+
+Requirements: Node.js 22.19 or later, Pi 0.84.2 or later, and either official
+DeepSeek API or OpenCode Go configuration. Pi 0.84.2 includes both provider
+routes.
+
+Try the current checkout directly:
+
+```bash
+pi -e ./packages/pi-dsh-mimic \
+  --provider deepseek \
+  --model deepseek-v4-pro
+```
+
+For OpenCode Go, configure its key through Pi's expected `OPENCODE_API_KEY`
+environment variable:
+
+```bash
+pi -e ./packages/pi-dsh-mimic \
+  --provider opencode-go \
+  --model deepseek-v4-pro
+```
+
+Install it for the current user:
+
+```bash
+pi install ./packages/pi-dsh-mimic
+```
+
+Use `pi install -l ./packages/pi-dsh-mimic` for project-local settings. The
+package has not been published to npm.
+The manifest already includes the `pi-package` keyword required by the
+[Pi Package Catalog](https://pi.dev/packages); the catalog can discover it only
+after publication to npm. The published install command will be
+`pi install npm:pi-dsh-mimic`.
+
+Select V4 Pro before starting a new session. Switching to V4 Pro inside an
+existing conversation does not forge a new bootstrap.
+
+## Request lifecycle
+
+| | Bootstrap request | Execution requests |
+| --- | --- | --- |
+| System persona | `You are a helpful software engineer assistant.` | unchanged |
+| User history | Original task, including Pi-native images | unchanged |
+| Provider-visible tools | `bash`, `str_replace_editor` | Pi's complete current catalog, including plugin tools |
+| Provider payload | DSH Minimal shape; Pi cache fields omitted | Pi-native encoding and ordering; cache follows provider configuration |
+
+The extension never calls `setActiveTools` and never overrides Pi's native
+`bash`, `read`, `edit`, or `write`. It registers one executable
+`str_replace_editor` only when a fresh target session arms. From request #2,
+tools contributed by other packages appear in the complete catalog.
+
+Provider errors and aborted responses do not consume the bootstrap. A small Pi
+custom entry persists the phase, and durable assistant/tool history repairs a
+crash-stale state on resume.
+
+## Offline verification
+
+TypeScript typechecking and all 13 automated tests pass. Twelve cover package
+behavior and its manifest; one verifies the session-statistics helper. Package
+behavior coverage includes:
+
+- original task, persona, field order, cache omission, and two-tool schemas in
+  request #1;
+- Pi-native payload, complete catalog, and plugin-tool restoration in request
+  #2;
+- images, API-error retry, text promotion, session resume, and crash-stale
+  recovery;
+- existing-conversation isolation, both providers, and non-target models;
+- editor create, view, replace, insert, and ambiguous-replacement rejection.
+
+OpenCode Go bash-first and editor-first offline loopbacks also pass through Pi's
+real package loader and provider path with `realModelCalls=0`. Request #2 carries
+a real tool result and restores `read/bash/edit/write/str_replace_editor`. After
+normalizing the user task, the new request #1 matches the historical 98-point
+capture in messages, key order, schemas, `strict:false`, and serialized
+structure.
+
+## Development
+
+```bash
+cd packages/pi-dsh-mimic
+npm install
+npm run check
+npm run pack:check
+```
+
+## Security, cost, and sources
+
+This package runs with the Pi process's filesystem permissions;
+`str_replace_editor` can create and modify files. The task, context, and tool
+results are sent to the selected DeepSeek or OpenCode provider. The package
+does not read, store, log, or transmit API keys and adds no model round. See
+[SECURITY.md](SECURITY.md).
+
+The Minimal persona and two tool-protocol texts come from DeepSeek Harness's
+[public Minimal protocol](https://github.com/deepseek-ai/deepseek-harness). See
+[LICENSE](LICENSE) and [NOTICE](NOTICE). The current implementation is based on
+that public protocol, request captures made by this project, and the Project2
+experiments. It is not affiliated with or endorsed by Pi, DeepSeek, OpenCode,
+or OpenAI.

package/README.zh-CN.md

@@ -0,0 +1,133 @@
+[仓库索引](../../README.md) · [English](README.md) · [实验与设计](docs/advanced.zh-CN.md) · [Project2 证据账本](docs/project2-evidence.md)
+
+# pi-dsh-mimic
+
+`pi-dsh-mimic` 在 [Pi](https://github.com/earendil-works/pi) 的第一次模型请求中复现
+DeepSeek Harness Minimal 环境，让 DeepSeek V4 Pro 从已验证的高能力轨迹起步；第一次
+有效响应后，它立即回到 Pi 原生执行，恢复完整工具目录和用户安装的其他插件。
+
+它把两件事放进同一个 session：
+
+- DSH Minimal 首请求带来的 V4 Pro 工程能力；
+- Pi 成熟的工具、session 管理和插件生态。
+
+用户无需安装或运行完整 DSH harness，也不必长期停留在两个工具的 Minimal 环境中。
+
+## 工作方式
+
+新建目标 session 的真实任务直接成为第一次模型请求：
+
+1. 请求 #1 使用 Minimal persona，只暴露 DSH 风格的 `bash` 与
+   `str_replace_editor`；
+2. 第一个成功 assistant response 或 durable tool call 后，请求 #2 恢复 Pi 原生
+   provider payload 和当下完整工具目录；
+3. Minimal persona 在整个 session 保持，Pi 较长的自动 system prompt 不会作为额外
+   user message 重放。
+
+扩展不增加身份 warm-up 或模型轮次，不复制原任务，不注入 `We need`，也不代理 API。
+
+## Project2 结果
+
+默认 Pi 得到 92。相同 one-shot 请求流程的四次最终结果为 **98、96、96、98**，均保持
+F3 16/16：
+
+| 运行 | Provider | 分数 | 关键分项 |
+| --- | --- | ---: | --- |
+| 默认 Pi baseline | DeepSeek 官方 API | 92 | F3 11，F6 10，F8 6 |
+| 早期 Minimal 模拟原型 | DeepSeek 官方 API | **98** | F3 16，F6 10，F8 7 |
+| 同流程复现实验 | DeepSeek 官方 API | 96 | F3 16，F6 10，F8 5 |
+| 首次 package 化实现 | DeepSeek 官方 API | 96 | F3 16，F6 8，F8 7 |
+| 当前独立实现 `0.1.0` | OpenCode Go | **98** | F3 16，F6 10，F8 7 |
+
+这些结果表明：DSH Minimal 首请求可以在 Pi 中稳定建立高能力轨迹，而恢复 Pi 完整工具
+目录不会破坏它。实验过程、被否决的设计和每项实现选择的依据见
+[实验与设计](docs/advanced.zh-CN.md)；完整 evaluator IDs 与来源见
+[Project2 证据账本](docs/project2-evidence.md)。
+
+## 安装和使用
+
+要求：Node.js 22.19 或更新版本、Pi 0.84.2 或更新版本，并已配置 DeepSeek 官方 API 或
+OpenCode Go。Pi 0.84.2 已内置两条 provider 路径。
+
+直接试用当前 checkout：
+
+```bash
+pi -e ./packages/pi-dsh-mimic \
+  --provider deepseek \
+  --model deepseek-v4-pro
+```
+
+使用 OpenCode Go 时，把 key 配置在 Pi 期望的 `OPENCODE_API_KEY` 环境变量中：
+
+```bash
+pi -e ./packages/pi-dsh-mimic \
+  --provider opencode-go \
+  --model deepseek-v4-pro
+```
+
+安装到当前用户：
+
+```bash
+pi install ./packages/pi-dsh-mimic
+```
+
+项目级安装使用 `pi install -l ./packages/pi-dsh-mimic`。当前 package 尚未发布到 npm。
+Manifest 已包含 Pi gallery 要求的 `pi-package` keyword；发布到 npm 后才会出现在
+[Pi Package Catalog](https://pi.dev/packages)，安装命令将是
+`pi install npm:pi-dsh-mimic`。当前 checkout 不会被市场索引。
+
+先选择 V4 Pro，再新建 session。在已有对话中途切换到 V4 Pro 不会伪造新的 bootstrap。
+
+## 请求生命周期
+
+| | Bootstrap 请求 | 执行阶段请求 |
+| --- | --- | --- |
+| System persona | `You are a helpful software engineer assistant.` | 保持不变 |
+| User 历史 | 原始任务，包括 Pi 原生图片 | 不变 |
+| Provider 可见工具 | `bash`、`str_replace_editor` | Pi 当下完整目录，包括其他插件工具 |
+| Provider payload | DSH Minimal 形状；移除 Pi cache 字段 | Pi 原生编码与排序；cache 随 provider 配置 |
+
+扩展不调用 `setActiveTools`，也不覆盖 Pi 原生 `bash`、`read`、`edit` 或 `write`。目标
+新 session 武装时才注册一个可执行 `str_replace_editor`；请求 #2 起，其他 package
+提供的工具会自然出现在完整目录中。
+
+Provider error 或 aborted response 不消耗 bootstrap。阶段通过小型 Pi custom entry
+持久化，恢复时也能根据 durable assistant/tool 历史修正 crash-stale 状态。
+
+## 离线验证
+
+TypeScript typecheck 已通过，13 项自动化测试全部通过。其中 12 项覆盖 package 行为与
+manifest，1 项验证 session 统计工具。Package 行为测试覆盖：
+
+- 首请求原任务、persona、字段顺序、cache 移除和两工具 schema；
+- 请求 #2 恢复 Pi 原生 payload、完整目录和其他插件工具；
+- 图片、API error 重试、文本回答晋升、session resume 和 crash-stale 恢复；
+- 已有对话隔离、DeepSeek/OpenCode 双 provider 与非目标模型隔离；
+- editor 的 create、view、replace、insert 和歧义替换拒绝。
+
+OpenCode Go bash-first 与 editor-first 离线回环还通过 Pi 的真实 package loader 和 provider
+路径验证了两个请求，`realModelCalls=0`。请求 #2 携带真实工具结果，并恢复
+`read/bash/edit/write/str_replace_editor`。归一 user task 后，新请求 #1 与历史 98 分
+capture 在消息、字段顺序、schema、`strict:false` 和序列化结构上逐项相等。
+
+## 开发
+
+```bash
+cd packages/pi-dsh-mimic
+npm install
+npm run check
+npm run pack:check
+```
+
+## 安全、费用与来源
+
+本 package 以 Pi 进程的文件系统权限运行；`str_replace_editor` 可以创建和修改文件。
+任务、上下文和工具结果会发送给所选 DeepSeek 或 OpenCode provider。package 不读取、
+保存、记录或传输 API key，也不增加额外模型轮次。完整说明见
+[SECURITY.md](SECURITY.md)。
+
+Minimal persona 与两项工具协议文本来自
+[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 的公开 Minimal 协议；
+对应许可声明见 [LICENSE](LICENSE) 和 [NOTICE](NOTICE)。当前实现依据该公开协议、本项目
+捕获的请求和 Project2 实验完成，与 Pi、DeepSeek、OpenCode 或 OpenAI 均无隶属或官方
+背书关系。

package/SECURITY.md

@@ -0,0 +1,41 @@
+# Security
+
+## API keys
+
+`pi-dsh-mimic` does not read, save, log, or transmit API keys. Provider
+credentials remain under Pi and the selected provider's configuration. Never
+put a DeepSeek or OpenCode key in a prompt, repository file, issue, screenshot,
+or captured request.
+
+If a credential has been exposed, revoke or rotate it with the provider before
+continuing.
+
+## Provider data boundary
+
+For target sessions, the task, conversation messages, tool schemas, and later
+tool results are sent through either `deepseek/deepseek-v4-pro` or
+`opencode-go/deepseek-v4-pro`. The extension changes the first request's
+persona and tool surface, but it is not an API proxy and does not create a
+confidential channel.
+
+The package adds no model round. The user's ordinary first task is the
+bootstrap request and is billed normally by the selected provider.
+
+## Filesystem and tool ownership
+
+The extension runs with the Pi process's filesystem permissions. Its
+`str_replace_editor` can view, create, and modify files. Review the package
+before installation and apply the same workspace and permission controls used
+for any other Pi extension.
+
+The package does not call `setActiveTools` and does not override Pi's native
+`bash`, `read`, `edit`, or `write`. Another extension that registers the same
+`str_replace_editor` name creates a real ownership conflict and should not be
+combined without review.
+
+## Reports
+
+Security reports and redacted repository issues should follow the root
+[security guidance](../../SECURITY.md). Do not submit full provider payloads
+until task content, headers, credentials, and unrelated source have been
+removed.

package/docs/advanced.md

@@ -0,0 +1,182 @@
+[简体中文](advanced.zh-CN.md) · [Back to the user guide](../README.md) · [Complete evidence ledger (Chinese canonical)](project2-evidence.md)
+
+# Start V4 Pro with DSH Minimal, then return to Pi
+
+`pi-dsh-mimic` reproduces only the most useful part of DSH: the first model
+request. The real task enters a Minimal-persona environment with
+`bash/str_replace_editor`, establishing a verified high-capability V4 Pro
+trajectory. After the first valid response, requests return to Pi-native form
+and the model regains Pi's complete tool catalog and other plugins.
+
+There is no need to reproduce the complete DSH harness. DSH Minimal supplies
+the first-request training interface; Pi continues to own sessions, tool
+execution, file editing, and plugin composition. The result combines the
+96–98 Project2 band observed in these experiments with Pi's mature agent
+runtime.
+
+## Experimental basis
+
+Local runs use Project2 V4.1b at frozen commit
+`04255b55f16c4439e538239fb9783070c4165081`. When a run resumed after an
+interruption, only its final submission score entered the comparison. A wrong
+project root, a zero-token configuration failure, and duplicate evaluation do
+not count as new samples.
+
+`We need`, `Let me`, and reasoning-block counts identify trajectory changes.
+Final judgment comes from the F3 ambient-session boundary, F6 migration, F8 ESP
+contract, hidden tests, and complete delivery. Evaluator IDs and exclusions are
+recorded in the [evidence ledger](project2-evidence.md).
+
+## DSH Minimal established that the capability is reachable
+
+External results provided a clear calibration. Under the same WSL/max/build
+conditions, DSH Minimal scored 99 and 96, while Standard scored 91 and PTC
+scored 92. Later Anchored Standard runs restored a 25-tool catalog after the
+first tool call and scored 98 and 99.
+
+These results establish two facts: the first request's persona and schemas can
+change V4 Pro's engineering trajectory, and a complete tool catalog does not
+automatically break the trajectory after it has formed. `pi-dsh-mimic` brings
+both properties into Pi.
+
+## The decisive Pi comparison
+
+A matched local pair narrowed the main difference to the first request:
+
+| Run | First request and later execution | Score | Key result |
+| --- | --- | ---: | --- |
+| Default Pi baseline | Full Pi prompt and five tools; native throughout | 92 | F3 11, F6 10, F8 6 |
+| Early Minimal-mimic prototype | Real task in a Minimal two-tool request; Pi-native after the first tool call; no Pi-context replay | **98** | F3 16, F6 10, F8 7, hidden 44/45 |
+
+Around assistant response 107, the 98-point trajectory reversed its earlier
+ambient fallback. Sensitive context had to carry an explicit `session_id`, and
+the voice path first resolved the current session before passing it onward.
+That self-correction directly explains F3 16/16 and is more meaningful than a
+first-line phrase.
+
+## Negative results narrowed the implementation
+
+Later runs tested the remaining explanations directly:
+
+| Question | Experiment | Result | Implementation choice |
+| --- | --- | ---: | --- |
+| Is an identity warm-up required? | Zero-tool Whoami | 93 | Put the real task in request #1 |
+| Is tool scarcity sufficient? | Pi-native `bash/read` | 93 | Use the DSH Minimal schemas |
+| Should the model be forced to emit the fingerprint first? | Wire Think | 94; F3 16, F6 6 | Do not inject or optimize for `We need` |
+| Should DSH wire normalization persist? | Persistent DSH wire | 94; F3 11 | Return to Pi-native requests after the first valid response |
+| Should Pi system context be replayed after promotion? | Full context replay | 90; F6 8 | Keep the Minimal persona without replaying full Pi context |
+| Is an approximate Minimal surface sufficient? | Approximate Minimal → Pi-native | 93; first line returned to `Let me` | Use the observed 98-point first request as the implementation baseline |
+
+A same-flow replication scored 96 with F3 still at 16/16. Most of its variation
+was in ESP static completeness.
+
+## From experimental prototype to current implementation
+
+The same request flow then reached complete Project2 endpoints in two clean
+implementations:
+
+| Implementation | Provider | Score | F3 | F6 | F8 |
+| --- | --- | ---: | ---: | ---: | ---: |
+| First packaged implementation | Official DeepSeek API | 96 | 16 | 8 | 7 |
+| Current independent implementation `0.1.0` | OpenCode Go | **98** | 16 | 10 | 7 |
+
+The initial OpenCode Go request hit an account opt-in 403 before producing a
+token. The same session resumed after opt-in, stopped naturally, and scored 98.
+The run verifies that the current implementation reaches the same high band
+through OpenCode Go.
+
+Both the official DeepSeek API and OpenCode Go reached 96–98. A weaker OpenCode
+model is no longer needed to explain the observed gap.
+
+## Why Pi plugins remain available
+
+| | Bootstrap request | Execution requests |
+| --- | --- | --- |
+| System persona | `You are a helpful software engineer assistant.` | unchanged |
+| User history | Original task, including Pi-native image blocks | unchanged |
+| Provider-visible tools | `bash`, `str_replace_editor` | Pi's complete current catalog, including plugin tools |
+| Provider payload | DSH Minimal shape; Pi cache fields removed | Pi-native encoding and ordering; cache follows provider configuration |
+
+The extension never calls `setActiveTools` and never overrides Pi's `bash`,
+`read`, `edit`, or `write`. It registers one executable `str_replace_editor`
+only when a fresh target session arms. From request #2, Pi again owns the
+provider payload and active catalog, so tools from other packages appear
+naturally.
+
+That is the boundary between a mimic and a complete harness port: the extension
+establishes a DSH Minimal first request, while Pi remains the execution runtime.
+
+## Product behavior and evidence
+
+| Current behavior | Main evidence |
+| --- | --- |
+| Arm only fresh sessions on the target provider/model pairs | An existing conversation cannot reproduce a real bootstrap; isolation tests cover target and non-target models |
+| Put the original task in request #1 | Early prototype 98; Whoami 93 |
+| Use the Minimal persona and DSH schemas in request #1 | Pi-native `bash/read` 93; approximate Minimal 93 |
+| Return to Pi-native execution after the first valid response | One-shot 96–98; persistent DSH wire 94 |
+| Keep the Minimal persona without replaying full Pi system context | Context replay 90 |
+| Retry bootstrap after provider error or abort | Error-retry, durable-stage, and crash-stale recovery tests |
+
+## Trajectory and tool scale
+
+With case-insensitive matching, the default baseline contained 156 occurrences
+of `let me` across 78 reasoning blocks, while the 98-point prototype contained
+one across 135 blocks. Exact case-sensitive counts are 115 and one. This quickly
+shows whether the first request changed the model's reasoning habit.
+
+The [evidence ledger](project2-evidence.md) now records assistant, reasoning,
+`We need / Let me`, tool-call, tool-name, and stopReason aggregates for all 11
+runs. Tool volume alone does not determine ability: persistent DSH wire made
+225 tool calls and scored 94, while the early prototype made 148 and scored 98.
+The current implementation used `bash/read/edit/write` 193 times and scored 98.
+
+The persistent-wire run also contained very few `Let me` occurrences, and Wire
+Think began with `We need` but scored 94. The extension observes these phrases
+without injecting them into the prompt.
+
+## Implementation verification
+
+TypeScript typechecking and all 13 automated tests pass. Twelve cover both
+providers, images, error retry, text promotion, session resume, crash-stale
+recovery, existing-conversation isolation, and non-target model isolation; one
+verifies the session-statistics helper. Editor tests cover create, view, unique
+replacement, insertion, relative-path rejection, and ambiguous-replacement
+rejection.
+
+OpenCode Go bash-first and editor-first offline loopbacks capture two requests
+through Pi's real package loader and provider path with `realModelCalls=0`.
+Request #2 restores `read/bash/edit/write/str_replace_editor` and carries a real
+tool result. After normalizing the user task, the new request #1 matches the
+historical 98-point capture in messages, key order, schemas, `strict:false`, and
+serialized structure.
+
+## Score details and reopening criteria
+
+The 98-point prototype and the verifiable DSH Minimal 99 both passed hidden
+44/45 and F3 16/16, missing the same one-point F12 rejection-reason string. The
+prototype's remaining loss is a grader function-name mismatch even though the
+implementation checks Wi-Fi, UID, room, and bed. The published score remains
+the evaluator's original score.
+
+The current request flow has four Project2 results in the 96–98 band. A new
+task, model or service version, material provider-payload change, or new complete
+task result would justify another paid run.
+
+## Sources and attribution
+
+The Minimal persona, bash description, editor description, and function schemas
+come from DeepSeek Harness's
+[public Minimal protocol](https://github.com/deepseek-ai/deepseek-harness). The
+exact commit and license notices are in [LICENSE](../LICENSE) and
+[NOTICE](../NOTICE). External Project2 calibration comes from
+[`xiaobright/modeltest`](https://github.com/xiaobright/modeltest) and
+[`xiaobright/dsh-anchored-standard`](https://github.com/xiaobright/dsh-anchored-standard).
+
+The current source was implemented from the DeepSeek Harness Minimal protocol,
+request captures made by this project, and this project's experiments. It did
+not reference or port source from
+[`kxh4892636/pi-deepseek-anchor`](https://github.com/kxh4892636/pi-deepseek-anchor).
+An early exploration loaded that extension to test the two-stage path in Pi, so
+the evidence ledger retains that run provenance. The current implementation,
+automated tests, offline loopbacks, and Project2 scores are produced by this
+project.