dsh-continual-evolve 0.1.0

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 dsh-continual-evolve contributors
+
+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/README.md

@@ -0,0 +1,290 @@
+# dsh-continual-evolve
+
+[中文](README.zh.md) | English
+
+[![awesome · DSH plugin](https://awesome-dsh-plugin.com/badge.svg)](https://awesome-dsh-plugin.com)
+[![CI](https://github.com/ZK-Andy/dsh-continual-evolve/actions/workflows/ci.yml/badge.svg)](https://github.com/ZK-Andy/dsh-continual-evolve/actions/workflows/ci.yml)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
+[![Node](https://img.shields.io/badge/node-%5E22.19%20%7C%7C%20%3E%3D24-339933)](package.json)
+[![Tests](https://img.shields.io/badge/tests-184%20passing-brightgreen)]()
+[![Status](https://img.shields.io/badge/status-all%20phases%20complete%20%C2%B7%20maintenance-ff69b4)]()
+
+Continual self-evolution for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness): a versioned, auditable, rollback-safe layer of harness state — prompt notes, memories, skills, and subagent specs — refined from session trajectories.
+
+> **Status: all phases complete; in long-term maintenance.** Phases 1–3
+> shipped the full evolution loop: the pure-core engine, model tools and
+> the `/evolve` command, the automatic review gate (turn-interval +
+> compaction checkpoints, human approval for global edits), real
+> system-prompt injection (prompt notes + delegation specs, zero token
+> cost when empty), and the benchmark-driven validation loop (code-owned
+> scoring, non-regressive acceptance, rubric ACL). Since then the plugin
+> keeps growing with usage-driven enhancements — the memory layer (ranked
+> injection, trajectory citations, archive), per-installation rubric keys,
+> and plugin-owned file logging. See the Roadmap for the full shipped and
+> candidate lists.
+
+## Background
+
+This project started as a research question: *can a harness improve itself,
+and what would a production-grade version look like?* Three lines of evidence
+shaped the answer:
+
+- **penguin-harness** demonstrated the concept (benchmark → evaluate →
+  optimize → accept/rollback) but with **zero code-level enforcement** — every
+  guarantee was a prompt contract. Its report (`docs/research/`) became the
+  hardening checklist this project implements.
+- **prime-agent `/refine`** proved the engineering shape: versioned harness
+  entries, atomic persistence, optimistic concurrency, inverse-op rollback.
+  This package is an original implementation of that shape on the DSH plugin
+  surface.
+- Academic work (Self-Harness, AHE, HarnessOpt-Bench) supplied the discipline:
+  frozen evaluation runtime, code-owned aggregation, non-regressive
+  acceptance.
+
+The result: **the model proposes, the code guarantees.** Every mechanical
+safety property (schema validation, snapshots, versioning, audit trail,
+acceptance decisions) is enforced in code — never by asking the model to
+behave.
+
+## Why
+
+Agents accumulate reusable experience in every session — repeated failures, durable facts, reusable procedures — and then forget it at the next turn or session. This plugin makes that experience first-class persistent state:
+
+- **Versioned entries** keyed by kind (`prompt` / `memory` / `skill` / `subagent`), each with a recorded provenance and version
+- **Evidence trail**: every refinement appends an event carrying `trigger / changes / evidence / outcome`
+- **Deterministic rollback**: inverse edits are generated from applied results — no LLM re-guessing
+- **Code-enforced safety**, not prompt discipline: schema validation, atomic writes, corrupt-file degrade, optimistic concurrency, immutable base system prompt
+- **Local (session) and global (cross-session) scopes** with merge semantics
+
+## Design provenance
+
+Inspired by three bodies of work (see [`docs/design.md`](docs/design.md)):
+
+- **prime-agent `/refine`** (MIT): the state model, atomic persistence, optimistic concurrency, per-edit validation, and inverse-op rollback this package implements — annotated reference source in [`docs/research/prime-agent-refinement.ts`](docs/research/prime-agent-refinement.ts). The code here is an original implementation, written for the DSH plugin surface.
+- **penguin-harness** (Apache-2.0): the benchmark-driven evolution loop — research report in [`docs/research/penguin-harness-self-evolution.md`](docs/research/penguin-harness-self-evolution.md); its prompt-only contracts are the anti-pattern this package hardens.
+- Academic: Self-Harness (arXiv 2606.09498), AHE (arXiv 2604.25850), HarnessOpt-Bench (arXiv 2608.06301).
+
+## Tech stack
+
+| Layer | Choice |
+|---|---|
+| Language | TypeScript (strict, ES2024, ESM) |
+| Runtime | Node `^22.19.0 \|\| >=24.0.0` (matches DSH) |
+| Plugin seam | `@deepseek-ai/cordis` (`name` / `apply` / `inject` entry) |
+| Package manager | pnpm (DSH ecosystem standard) |
+| Build | `tsc` → `lib/` (main `lib/index.js`, types `lib/index.d.ts`) |
+| Tests | Vitest |
+| Lint | oxlint (DSH official repo convention) |
+| License | MIT |
+
+## Project layout
+
+```
+dsh-continual-evolve/
+├── package.json          # exports / files / engines / scripts + dsh.bundle manifest
+├── cordis.patch.yml      # bundle patch (dsh plugin add activates on install)
+├── tsconfig.json / .oxlintrc.json / .editorconfig / .gitignore
+├── LICENSE / README.md / README.zh.md
+├── docs/
+│   ├── design.md               # full design doc (incl. hardening matrix)
+│   └── research/               # penguin-harness report + prime-agent reference source
+├── src/
+│   ├── index.ts          # cordis plugin entry (service mount + wiring)
+│   ├── types.ts          # HarnessState / entry / edit / result types
+│   ├── state.ts          # atomic persistence, corrupt degrade, merge, concurrency
+│   ├── validate.ts       # code-enforced edit validation
+│   ├── apply.ts          # per-edit apply pass with optimistic locking
+│   ├── rollback.ts       # deterministic inverse-op rollback
+│   ├── plan.ts           # proposal JSON parsing (truncation-aware)
+│   ├── tool.ts           # evolve_* model-facing tools (5)
+│   ├── command.ts        # /evolve command (incl. benchmark subcommands)
+│   ├── planner.ts        # ctx.llm planner
+│   ├── render.ts         # bounded prompt rendering
+│   ├── inject.ts         # dynamic system-prompt section (prompt notes + delegation specs, ranked injection)
+│   ├── source.ts         # trajectory citations (sessionId + event seqs of distilled entries)
+│   ├── auto.ts           # auto-review gate (turn/compaction triggers + audit, global-aware view)
+│   ├── notify.ts         # gate visibility — follow-up notice after an approved auto-refine
+│   ├── goal.ts           # goal-driven evolution rounds (/evolve goal)
+│   ├── review.ts         # gate LLM judgment (declines local duplicates of globally covered topics)
+│   ├── approval.ts       # human approval for global edits
+│   ├── skill.ts          # skill materialization ($DSH_HOME/skills/)
+│   ├── mount.ts          # hot-mounted skill plugins (loader.create + boot restore)
+│   ├── benchmark.ts      # benchmark store
+│   ├── rubric.ts         # rubric ACL (AES-256-GCM envelopes, auto-generated local key)
+│   ├── logfile.ts        # plugin-owned file logging (JSONL exporter + rotation)
+│   ├── score.ts          # code-owned aggregation + acceptance rule
+│   ├── evaluate.ts       # evaluation matrix runner (structured-output subagents)
+│   ├── pool.ts           # bounded-concurrency worker pool for evaluation runs
+│   ├── store.ts          # store layout + snapshots + result history
+│   └── service.ts        # evolution engine (onApplied hook)
+└── test/                 # 20 files, 184 tests
+```
+
+## In-session usage (after restart)
+
+```
+/evolve                  help + current local store
+/evolve list [global]    list entries
+/evolve history          applied refinements (ids for rollback)
+/evolve rollback <id>    deterministically revert a refinement
+/evolve plan [msg]       LLM planner against the current store
+/evolve archive <id>     hide an entry from injection (data kept, restorable)
+/evolve unarchive <id>   restore an archived entry
+/evolve log [tail N] [session <id>]  show the recent plugin log (default 50 lines; optional per-session filter)
+/evolve export <path>    backup the local store to JSON
+/evolve import <path>    restore a store from an export file
+/evolve mount <skillId>  hot-mount a skill entry as a live cordis plugin (tool: skill_<name>)
+/evolve mount list       list hot-mounted plugins (restored on boot)
+/evolve unmount <id>     remove a hot-mounted plugin
+/evolve goal             show the evolution goal (round-driven auto-review)
+/evolve goal <objective> create/update the evolution goal — while active, the review gate runs EVERY round
+/evolve goal done        complete the evolution goal
+```
+
+Model-facing tools: `evolve_list`, `evolve_add`, `evolve_update`, `evolve_delete`, `evolve_rollback`.
+
+## Memory layer
+
+Beyond the persisted store itself, three features keep injected memory
+"understanding you" as entries grow (gap analysis vs. Mem0 / Letta / Zep /
+LangMem; no external services — everything is pure functions):
+
+- **Ranked injection** — when a kind holds more than the 6-entry cap, the
+  injected block no longer shows the fixed first six: entries are scored by
+  relevance to the agent's most recent direct user messages (keyword/BM25
+  level: title hits weigh 2×) and then by recency (`updated_at`, 30-day
+  half-life), so the freshest *and most relevant* entries fill the cap. The
+  empty-store zero-token behavior is unchanged.
+- **Trajectory citations** — every newly created entry records
+  `metadata.sourceSession` + `metadata.sourceSeqs` pointing at the direct
+  user messages it was distilled from (DSH sessions are event-sourced with
+  contiguous seqs, so the citation expands back into the durable session
+  log). Listings show `src=<sessionId>:<seqs>`; old entries are not migrated
+  and never error.
+- **Archive** — `/evolve archive <id>` hides an entry from injection
+  (`metadata.archivedAt`, data kept, rollback-compatible) and
+  `/evolve unarchive <id>` restores it. Archived entries are marked
+  `[archived]` in `evolve_list` and skipped by injection; the overflow count
+  excludes them.
+- **Global-aware gate** — the auto-review gate and planner judge the merged
+  global + local state with every entry's real scope labeled, so a topic
+  already covered by a global entry is declined instead of being re-sedimented
+  as a local duplicate.
+
+## Logging
+
+Plugin-owned file logging: every cordis log message (from this plugin or any
+other) is appended to `<dshHome>/evolve/plugin.log` as JSONL (0600, rotated to
+`plugin.log.1` past `logMaxBytes`). It works no matter how `dsh web` is
+launched — no extra component to install, no startup-script dependency.
+View the tail with `/evolve log [tail N]`, or read the file directly:
+
+```bash
+tail -f ~/.dsh/evolve/plugin.log          # live
+/evolve log 100                            # last 100 lines in the chat
+```
+
+For live output in a foreground terminal, the official
+`@deepseek-ai/cordis-plugin-logger-console` plugin can be added to the
+profile (optional; the file log remains the baseline that always exists).
+
+## Benchmark-driven validation (Phase 3)
+
+```
+/evolve benchmark new <title> [runs]                   create a benchmark (runs = repeats per case, default 1)
+/evolve benchmark add-case <bid> <title> <statement> <rubric>
+/evolve benchmark list                                 list benchmarks
+/evolve benchmark reset <bid>                          clear the scoreboard (re-run reference)
+/evolve benchmark status <bid>                         scoreboard + decisions
+/evolve benchmark run <bid>                            evaluate current state → reference
+/evolve benchmark run <bid> candidate <refinementId>   evaluate post-refinement state → decide
+```
+
+The loop: freeze a reference score → evolve a candidate (`/evolve plan`) →
+run the same case × run matrix against the post-refinement state → the
+**code-owned** acceptance rule keeps the candidate only if the overall mean
+strictly improves with no case regressing (Self-Harness style). The model
+produces raw per-cell scores only; aggregation and decisions live in
+`src/score.ts`. Rubric isolation is by construction (the planner never sees
+rubric files); a rejection is recorded in the scoreboard and the refinement
+is rolled back automatically (`autoRollbackOnReject`, on by default).
+
+### Real recorded run (ACCEPT)
+
+A live `dsh web` session, one case, one candidate — the first genuine
+acceptance:
+
+| Step | Command | Outcome |
+|---|---|---|
+| reference | `/evolve benchmark run lint_convention` | **90** — the evaluator agent actually grepped the harness store and reported *"lint/ruff/eslint/mypy appear in zero entries"* |
+| candidate | `/evolve plan 记住：写代码前必须先运行适用的 lint 检查` | creates `memory:convention_lint_before_code` |
+| re-evaluate | `/evolve benchmark run lint_convention candidate <id>` | **100** — evaluator ran `evolve_list`, hit the memory, quoted it verbatim |
+| decision | — | `overall: 90 → 100` · `lint_knowledge: 90 → 100` · **DECISION: ACCEPTED** |
+
+The evaluator does not grade model common sense — it inspects the actual
+harness state under test (grep, `evolve_list`) and scores against it, so a
+harness change measurably moves the score. Earlier runs in the same session
+produced honest `REJECTED` decisions (0 → 0 placeholder cases, and 100 → 100
+where the baseline was already perfect).
+
+## Configuration
+
+| Key | Default | Meaning |
+|---|---|---|
+| `baseDir` | resolved DSH home | root for the `evolve/` stores |
+| `sectionOrder` | 118 | system-prompt section order |
+| `autoReview` | `false` | enable the automatic review gate (costs a cheap model call per interval) |
+| `reviewIntervalTurns` | 6 | gate runs when this many turns passed since the last review |
+| `maxReviewInputChars` | 40000 | trajectory slice handed to the gate |
+| `reviewBudgetTokens` | 4096 | output budget for the gate call |
+| `notifyOnAutoReview` | `true` | after an approved gate run that applied edits, queue a visible follow-up notice in the session (persisted entries + rollback command) |
+| `requireGlobalApproval` | `true` | cross-session (global) edits ask the user for "批准" before applying |
+| `skillsDir` | `<dshHome>/skills` | root where skill entries materialize as SKILL.md bundles |
+| `rubricKey` | auto-generated local key file (`<dshHome>/evolve/rubric.key`, 0600) → dev fallback | passphrase for AES-256-GCM rubric encryption (benchmark rubrics never touch the disk in plaintext). When unset, the plugin generates a random per-installation key file on first use — every install gets its own key, no setup needed; `DSH_EVOLVE_RUBRIC_KEY` is the environment-variable override |
+| `logToFile` | `true` | write all cordis log messages to `<dshHome>/evolve/plugin.log` (JSONL, 0600) — plugin-owned logging works with any launch method, no extra component to install |
+| `logLevel` | `1` | file log level: 0=error, 1=info, 2=warn, 3=debug |
+| `logMaxBytes` | 5 MiB | rotate the log to `plugin.log.1` when it exceeds this size |
+| `autoRollbackOnReject` | `true` | after a benchmark decision rejects a candidate, roll the refinement back automatically (same engine path as `/evolve rollback` — deterministic, snapshotted, audited) |
+
+Example (profile `cordis.patch.yml`):
+
+```yaml
+- insert:
+    - id: continual-evolve
+      name: 'dsh-continual-evolve'
+      config:
+        autoReview: true
+        reviewIntervalTurns: 6
+```
+
+## Development
+
+```bash
+pnpm install        # install dev deps
+pnpm build          # tsc -> lib/
+pnpm test           # vitest run
+pnpm lint           # oxlint src test
+```
+
+Hit a wall? See [`docs/FAQ.md`](docs/FAQ.md) — real failure/fix records (service planes, schema DSL, structured output, gate counting, verifying prompt injection).
+
+
+## Roadmap
+
+**Shipped**
+
+- **Phases 1–3 (done)**: pure-core engine (state model, validation, apply, rollback, proposal parsing) → `evolve_*` tools + `/evolve` command + `ctx.llm` planner → auto-refine review gate (turn-interval + compaction checkpoints, visible follow-up notices), global-scope human approval, executable skills, real system-prompt injection (prompt notes + delegation specs, inherited by subagents), benchmark-driven validation loop (code-owned scoreboard, non-regressive acceptance, rubric isolation by construction), hot-mounted skill plugins, goal-driven evolution rounds.
+- **2026-08 maintenance wave (done)**:
+  - **memory layer** — ranked injection (relevance + recency scoring fills the per-kind cap), trajectory citations (`metadata.sourceSession` + `sourceSeqs`, shown as `src=session:seqs`), archive/unarchive (`/evolve archive <id>`, injection skips archived entries), global-aware gate (declines local duplicates of globally covered topics)
+  - **per-installation rubric key** — auto-generated local key file (`<dshHome>/evolve/rubric.key`, 0600); no more publicly known dev key
+  - **plugin-owned file logging** — every cordis log message lands in `<dshHome>/evolve/plugin.log` (JSONL, 0600, rotated), viewable via `/evolve log`; works with any launch method, no extra component to install
+  - **trajectory-grounded planning** — `/evolve plan` (and every planner call, including the gate's refine step) now reads the session trajectory: the caller's recent direct user messages are extracted from the session log and fed to the planner as a `<session_trajectory>` block, so proposals are grounded in what the user actually said (explicit `trajectory` overrides; empty trajectory is omitted at zero cost)
+  - **gate-proposed archiving** — stale entries are a first-class refine target: the planner can emit `action: "archive"` (kind + id only), which stamps `metadata.archivedAt` through the normal apply path — snapshot, version bump, audit event, and a deterministic rollback inverse that restores the pre-archive state. Archive hides from injection but never deletes; re-archiving an archived entry is rejected, and the base system prompt stays immutable
+  - **automatic rollback on benchmark rejection** — the acceptance loop is closed: when the code-owned decision rejects a candidate, the refinement is reverted automatically through the same engine path as `/evolve rollback` (deterministic inverse edits, snapshotted and audited; configurable via `autoRollbackOnReject`, on by default). Failures report the manual fallback instead of throwing
+  - **per-session log filtering** — `/evolve log [tail N] [session <id>]` keeps only the lines mentioning a given session id (exact token match, drawn from the rendered message and raw args); gate records now carry the session id in their log line
+
+The planned/candidates list is empty for now — future work is driven by real usage.
+
+## License
+
+MIT. Independent project — not affiliated with DeepSeek.

package/README.zh.md

@@ -0,0 +1,240 @@
+# dsh-continual-evolve
+
+[English](README.md) | 中文
+
+[![awesome · DSH plugin](https://awesome-dsh-plugin.com/badge.svg)](https://awesome-dsh-plugin.com)
+[![CI](https://github.com/ZK-Andy/dsh-continual-evolve/actions/workflows/ci.yml/badge.svg)](https://github.com/ZK-Andy/dsh-continual-evolve/actions/workflows/ci.yml)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
+[![Node](https://img.shields.io/badge/node-%5E22.19%20%7C%7C%20%3E%3D24-339933)](package.json)
+[![Tests](https://img.shields.io/badge/tests-184%20passing-brightgreen)]()
+[![Status](https://img.shields.io/badge/status-all%20phases%20complete%20%C2%B7%20maintenance-ff69b4)]()
+
+[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness)（`dsh`）的持续自进化插件：一套**版本化、可审计、可回滚**的 harness 状态层——提示词补充、记忆、技能、子代理规格——从会话轨迹中沉淀而来。
+
+> **状态：全部阶段完成，进入长期维护。** Phase 1–3 交付了完整进化闭环：纯核心引擎、模型工具与 `/evolve` 命令、自动 review 门禁（回合间隔 + 压缩检查点、全局写入人工审批）、真实系统提示词注入（prompt 补充 + 委派规格，空 store 零 token 成本）、benchmark 驱动验证闭环（代码所有计分、非退化接受、rubric ACL）。此后插件随真实使用持续增强——记忆层（排序注入、轨迹引用、归档）、每安装实例独立的 rubric 密钥、插件自带文件日志。已交付与候选清单见"路线图"。
+
+## 背景
+
+这个项目始于一个研究问题：*harness 能自我改进吗？生产级版本长什么样？* 三条证据线塑造了答案：
+
+- **penguin-harness** 证明了概念（benchmark → 评估 → 优化 → 接受/回滚），但**代码层零强制**——所有保证都是提示词契约。它的研究报告（`docs/research/`）成了本项目的硬化清单。
+- **prime-agent `/refine`** 证明了工程形态：版本化 harness 条目、原子持久化、乐观并发、逆操作回滚。本包是在 DSH 插件表面上对该形态的原创实现。
+- 学术工作（Self-Harness、AHE、HarnessOpt-Bench）提供了纪律：冻结评估运行时、代码所有聚合、非退化接受。
+
+结果：**模型提议，代码保证。** 每一项机械化安全属性（schema 校验、快照、版本、审计、接受决策）都由代码强制——从不要求模型自觉守规矩。
+
+## 为什么
+
+Agent 在每个会话里积累可复用经验——重复失败、持久事实、可复用流程——然后在下个回合或下个会话忘掉。本插件把这些经验变成一等公民的持久状态：
+
+- **版本化条目**：按 `prompt` / `memory` / `skill` / `subagent` 分键，每条带来源与版本
+- **证据链**：每次进化追加一条携带 `trigger / changes / evidence / outcome` 的事件
+- **确定性回滚**：逆操作由已应用的结果生成——不需要 LLM 再猜
+- **代码强制安全**，而非提示词纪律：schema 校验、原子写、损坏降级、乐观并发、基础系统提示词不可变
+- **局部（会话内）与全局（跨会话）双作用域**，带合并语义
+
+## 设计来源
+
+受三方面工作启发（见 [`docs/design.md`](docs/design.md)）：
+
+- **prime-agent `/refine`**（MIT）：本包实现的状态模型、原子持久化、乐观并发、逐条校验与逆操作回滚——参考源码在 [`docs/research/prime-agent-refinement.ts`](docs/research/prime-agent-refinement.ts)。代码为原创实现，面向 DSH 插件表面编写。
+- **penguin-harness**（Apache-2.0）：benchmark 驱动的进化循环——研究报告在 [`docs/research/penguin-harness-self-evolution.md`](docs/research/penguin-harness-self-evolution.md)；其"纯提示词契约"正是本包要硬化的反面教材。
+- 学术：Self-Harness（arXiv 2606.09498）、AHE（arXiv 2604.25850）、HarnessOpt-Bench（arXiv 2608.06301）。
+
+## 技术栈
+
+| 层 | 选择 |
+|---|---|
+| 语言 | TypeScript（strict、ES2024、ESM） |
+| 运行时 | Node `^22.19.0 \|\| >=24.0.0`（与 DSH 一致） |
+| 插件接缝 | `@deepseek-ai/cordis`（`name` / `apply` / `inject` 入口） |
+| 包管理 | pnpm（DSH 生态标准） |
+| 构建 | `tsc` → `lib/`（main `lib/index.js`，types `lib/index.d.ts`） |
+| 测试 | Vitest |
+| Lint | oxlint（DSH 官方仓库惯例） |
+| License | MIT |
+
+## 项目结构
+
+```
+dsh-continual-evolve/
+├── package.json          # exports / files / engines / scripts + dsh.bundle manifest
+├── cordis.patch.yml      # bundle patch（dsh plugin add 安装即激活）
+├── tsconfig.json / .oxlintrc.json / .editorconfig / .gitignore
+├── LICENSE / README.md / README.zh.md
+├── docs/
+│   ├── design.md               # 完整设计文档（含硬化对照表）
+│   └── research/               # penguin 研究报告 + prime-agent 参考源码
+├── src/
+│   ├── index.ts          # cordis 插件入口（服务挂载 + 接线）
+│   ├── types.ts          # HarnessState / 条目 / 编辑 / 结果类型
+│   ├── state.ts          # 原子持久化、损坏降级、合并、乐观并发
+│   ├── validate.ts       # 代码强制编辑校验（基础提示词不可改、skill 契约）
+│   ├── apply.ts          # 逐条应用 + 乐观锁
+│   ├── rollback.ts       # 确定性逆操作回滚
+│   ├── plan.ts           # 提案 JSON 解析（截断诊断）
+│   ├── tool.ts           # evolve_* 模型工具（5 个）
+│   ├── command.ts        # /evolve 命令（含 benchmark 子命令）
+│   ├── planner.ts        # ctx.llm 规划器
+│   ├── render.ts         # 有界提示词渲染
+│   ├── inject.ts         # 动态系统提示词段（prompt 补充 + 委派规格，打分排序注入）
+│   ├── source.ts         # 轨迹引用（沉淀条目的 sessionId + 事件 seq）
+│   ├── auto.ts           # 自动 review 门禁（回合/压缩触发 + 审计，global 感知视图）
+│   ├── notify.ts         # 门禁可见性——approved 自动沉淀后发送可见通知
+│   ├── goal.ts           # goal 驱动的进化轮次（/evolve goal）
+│   ├── review.ts         # 门禁 LLM 判断（拒绝 global 已覆盖主题的 local 重复沉淀）
+│   ├── approval.ts       # 全局写入人工审批
+│   ├── skill.ts          # 技能物化（$DSH_HOME/skills/）
+│   ├── mount.ts          # 技能热挂载插件（loader.create + 启动恢复）
+│   ├── benchmark.ts      # benchmark 存储
+│   ├── rubric.ts         # rubric ACL（AES-256-GCM 密文信封，自动生成本地密钥）
+│   ├── logfile.ts        # 插件自带文件日志（JSONL exporter + 轮转）
+│   ├── score.ts          # 代码所有聚合 + 接受规则
+│   ├── evaluate.ts       # 评估矩阵执行器（结构化输出子代理）
+│   ├── pool.ts           # 评估运行的有界并发工作池
+│   ├── store.ts          # store 布局 + 快照 + 结果历史
+│   └── service.ts        # 进化引擎（onApplied 钩子）
+└── test/                 # 20 个文件，184 个测试
+```
+
+## 会话内用法（安装后）
+
+```
+/evolve                       帮助 + 当前局部 store
+/evolve list [global]         列出条目
+/evolve history               已应用的 refinement（回滚用 id）
+/evolve rollback <id>         确定性回滚某个 refinement
+/evolve plan [msg]            LLM 规划器
+/evolve archive <id>          归档条目——不再注入（数据保留，可恢复）
+/evolve unarchive <id>        恢复已归档条目
+/evolve log [tail N] [session <id>] 查看最近插件日志（默认 50 行；可加会话过滤）
+/evolve export <path>         备份局部 store 为 JSON
+/evolve import <path>         从导出文件恢复 store
+/evolve mount <skillId>       热挂载 skill 条目为实时 cordis 插件（工具：skill_<name>）
+/evolve mount list            列出热挂载插件（重启自动恢复）
+/evolve unmount <id>          移除热挂载插件
+/evolve goal                  查看进化 goal（轮次驱动自动 review）
+/evolve goal <objective>      创建/更新进化 goal——active 时 review 门禁每轮触发
+/evolve goal done             完成进化 goal
+```
+
+模型工具：`evolve_list`、`evolve_add`、`evolve_update`、`evolve_delete`、`evolve_rollback`。
+
+## 记忆层
+
+在持久 store 之外，四项增强让注入的记忆在条目增多时依然"懂你"（对照 Mem0 / Letta / Zep / LangMem 的差距分析；不引入外部服务——全部是纯函数）：
+
+- **打分排序注入**——某类条目超过 6 条封顶时，注入块不再固定取前 6 条：先按与 agent 最近直接用户消息的相关度打分（关键词/BM25 级别，标题命中权重 2×），再按新鲜度排序（`updated_at`，30 天半衰期），让"最新 + 最相关"的条目填满封顶。空 store 零 token 行为不变。
+- **轨迹引用**——每条新沉淀条目都会记录 `metadata.sourceSession` + `metadata.sourceSeqs`，指向它蒸馏自的直接用户消息（DSH 会话是事件溯源、seq 连续，引用可展开回持久会话日志）。列表显示 `src=<sessionId>:<seqs>`；旧条目不迁移也不报错。
+- **归档**——`/evolve archive <id>` 让条目不再注入（`metadata.archivedAt`，数据保留、与快照/回滚兼容），`/evolve unarchive <id>` 恢复。归档条目在 `evolve_list` 中标记 `[archived]`，注入跳过，溢出计数不含它们。
+- **global 感知门禁**——自动 review 门禁与规划器评审的是合并后的 global + local 状态，每条条目标注真实 scope；global 已覆盖的主题会被 declined，不再重复沉淀为 local 条目。
+
+## 日志
+
+插件自带文件日志：所有 cordis 日志消息（本插件或其他插件）追加写入 `<dshHome>/evolve/plugin.log`（JSONL、0600，超过 `logMaxBytes` 轮转到 `plugin.log.1`）。与 `dsh web` 的启动方式无关——无需安装额外组件、不依赖启动脚本。查看方式：
+
+```bash
+tail -f ~/.dsh/evolve/plugin.log          # 实时跟随
+/evolve log 100                            # 在对话里看最近 100 行
+```
+
+前台终端想要实时输出时，可（可选）在 profile 加官方 `@deepseek-ai/cordis-plugin-logger-console` 插件；文件日志始终是默认存在的基础。
+
+## benchmark 驱动验证（Phase 3）
+
+```
+/evolve benchmark new <title>                         创建 benchmark
+/evolve benchmark add-case <bid> <title> <statement> <rubric>
+/evolve benchmark list                                列出 benchmark
+/evolve benchmark status <bid>                        查看计分板 + 决策
+/evolve benchmark reset <bid>                         清空计分板（重跑参考线）
+/evolve benchmark run <bid>                           评估当前状态 → 参考线
+/evolve benchmark run <bid> candidate <refinementId>  评估进化后状态 → 决策
+```
+
+闭环：冻结参考分 → 进化候选（`/evolve plan`）→ 用同一 case × run 矩阵复测进化后状态 → **代码所有**的接受规则只在总体均值严格提高且无 case 退化时保留候选（Self-Harness 风格）。模型只产出原始细胞级分数；聚合与决策都在 `src/score.ts`。rubric 隔离靠构造（规划器的提示词永远不含 rubric 文件）；拒绝会记录进 scoreboard 并自动回滚该 refinement（`autoRollbackOnReject`，默认开）。
+
+### 真实运行记录（ACCEPT）
+
+一次真实的 `dsh web` 会话，一个 case、一个候选——第一次真正的接受：
+
+| 步骤 | 命令 | 结果 |
+|---|---|---|
+| 参考线 | `/evolve benchmark run lint_convention` | **90**——评估子代理真的 grep 了 harness store，报告"lint/ruff/eslint/mypy 在所有条目中零出现" |
+| 进化候选 | `/evolve plan 记住：写代码前必须先运行适用的 lint 检查` | 创建 `memory:convention_lint_before_code` |
+| 复测 | `/evolve benchmark run lint_convention candidate <id>` | **100**——评估器跑 `evolve_list` 命中记忆并逐字引用 |
+| 决策 | — | `overall: 90 → 100` · `lint_knowledge: 90 → 100` · **DECISION: ACCEPTED** |
+
+评估器评的不是模型常识，而是**被测 harness 状态本身**（grep、`evolve_list` 检查）——所以 harness 的改动会真实地反映在分数上。同一会话早些时候还产生过诚实的 `REJECTED` 决策（0→0 占位符 case、100→100 满分基线无法超越）。
+
+## 配置
+
+| 键 | 默认值 | 含义 |
+|---|---|---|
+| `baseDir` | 解析后的 DSH home | `evolve/` store 的根 |
+| `sectionOrder` | 118 | 系统提示词段落顺序 |
+| `autoReview` | `false` | 启用自动 review 门禁（每间隔一次廉价模型调用） |
+| `reviewIntervalTurns` | 6 | 距上次 review 满这么多回合时触发门禁 |
+| `maxReviewInputChars` | 40000 | 交给门禁的轨迹切片 |
+| `reviewBudgetTokens` | 4096 | 门禁调用的输出预算 |
+| `notifyOnAutoReview` | `true` | 门禁 approved 且实际应用了编辑后，在会话中排一条可见通知（沉淀条目 + 回滚命令） |
+| `requireGlobalApproval` | `true` | 跨会话（全局）编辑需用户批准"批准"后才应用 |
+| `skillsDir` | `<dshHome>/skills` | 技能条目物化为 SKILL.md 包的根目录 |
+| `rubricKey` | 自动生成的本地密钥文件（`<dshHome>/evolve/rubric.key`，0600）→ dev 兜底 | rubric 加密（AES-256-GCM）口令：benchmark rubric 明文永不着盘。未配置时插件首次使用自动生成随机密钥文件——每台安装实例一把独立密钥，零配置；`DSH_EVOLVE_RUBRIC_KEY` 为环境变量覆盖项 |
+| `logToFile` | `true` | 所有 cordis 日志消息写入 `<dshHome>/evolve/plugin.log`（JSONL、0600）——插件自带日志，与启动方式无关、无需安装额外组件 |
+| `logLevel` | `1` | 文件日志级别：0=error、1=info、2=warn、3=debug |
+| `logMaxBytes` | 5 MiB | 超过该大小轮转到 `plugin.log.1` |
+| `autoRollbackOnReject` | `true` | benchmark 决策拒绝候选后自动回滚该 refinement（与 `/evolve rollback` 同一引擎路径——确定性、快照、审计） |
+
+示例（profile `cordis.patch.yml`）：
+
+```yaml
+- insert:
+    - id: continual-evolve
+      name: 'dsh-continual-evolve'
+      config:
+        autoReview: true
+        reviewIntervalTurns: 6
+```
+
+## 安装
+
+```bash
+# 从 npm 安装（安装即激活，自带 bundle patch）
+dsh plugin --profile web add dsh-continual-evolve
+
+# 或从源码安装
+dsh plugin --profile web add /path/to/dsh-continual-evolve
+```
+
+## 开发
+
+```bash
+pnpm install        # 安装开发依赖
+pnpm dev            # tsc --watch
+pnpm build          # tsc -> lib/
+pnpm test           # vitest run
+pnpm lint           # oxlint src test
+```
+
+遇到问题先看 [`docs/FAQ.md`](docs/FAQ.md)（真实踩坑记录：服务平面、schema DSL、结构化输出、门禁计数、注入验证等）。
+
+## 路线图
+
+**已交付**
+
+- **Phase 1–3（完成）**：纯核心引擎（状态模型、校验、应用、回滚、提案解析）→ `evolve_*` 工具 + `/evolve` 命令 + `ctx.llm` 规划器 → 自动 review 门禁（回合间隔 + 压缩检查点、approved 后可见通知）、全局人工审批、可执行技能、真实系统提示词注入（prompt 补充 + 委派规格，子代理沿父链继承）、benchmark 驱动验证闭环（代码所有计分板、非退化接受、rubric 构造性隔离）、技能热挂载插件、goal 驱动的进化轮次。
+- **2026-08 维护期增强（完成）**：
+  - **记忆层**——排序注入（相关度 + 新鲜度打分填满每类封顶）、轨迹引用（`metadata.sourceSession` + `sourceSeqs`，显示为 `src=session:seqs`）、归档/恢复（`/evolve archive <id>`，注入跳过归档条目）、global 感知门禁（拒绝 global 已覆盖主题的 local 重复沉淀）
+  - **每安装独立 rubric 密钥**——自动生成本地密钥文件（`<dshHome>/evolve/rubric.key`，0600）；不再有全世界公开的 dev 键
+  - **插件自带文件日志**——所有 cordis 日志消息写入 `<dshHome>/evolve/plugin.log`（JSONL、0600、自动轮转），`/evolve log` 查看；与启动方式无关、无需安装额外组件
+  - **轨迹接地规划**——`/evolve plan`（及所有规划调用，含门禁 refine 步骤）现在读取会话轨迹：从调用方会话日志提取最近直接用户消息，作为 `<session_trajectory>` 块喂给规划器，提案以用户真实说过的话为依据（显式 `trajectory` 覆盖；空轨迹省略、零成本）
+  - **门禁提议归档**——过时条目是一等 refine 目标：规划器可输出 `action: "archive"`（仅需 kind + id），代码经正常 apply 通道盖 `metadata.archivedAt` 戳——快照、版本 +1、审计事件、以及恢复归档前状态的确定性回滚逆编辑。归档隐藏于注入但绝不删除；重复归档被拒绝；基础系统提示词保持不可变
+  - **benchmark 拒绝自动回滚**——接受闭环已闭合：代码所有决策拒绝候选时，refinement 经与 `/evolve rollback` 相同的引擎路径自动撤销（确定性逆编辑、快照 + 审计；`autoRollbackOnReject` 配置，默认开）。失败时给出手动回滚提示而不是抛错
+  - **日志按会话过滤**——`/evolve log [tail N] [session <id>]` 只保留提及指定会话 id 的行（精确 token 匹配，取自渲染消息与原始 args）；门禁记录的行现在携带会话 id
+
+规划中/候选清单暂时为空——后续工作随真实使用驱动。
+
+## License
+
+MIT。独立项目——与 DeepSeek 无关联。

package/cordis.patch.yml

@@ -0,0 +1,9 @@
+# dsh-continual-evolve bundle layer.
+#
+# Installed via `dsh plugin --profile <name> add dsh-continual-evolve`, this
+# patch inserts the plugin row so it activates without manual profile editing.
+# Conservative defaults: the auto-review gate is OFF (opt in via your profile
+# cordis.patch.yml config) and cross-session edits require human approval.
+- insert:
+    - id: continual-evolve
+      name: 'dsh-continual-evolve'

package/lib/apply.d.ts

@@ -0,0 +1,24 @@
+/**
+ * The apply pass: turn a validated proposal into state changes with full
+ * per-edit accounting. Every failure is recorded per edit — a bad edit never
+ * invalidates the whole proposal — and optimistic-concurrency checks reject
+ * edits whose target entry changed while planning was in flight.
+ */
+import type { EntrySource, HarnessScope, HarnessState, RefinementProposal, RefinementResult } from "./types.js";
+import { entryChangedSince } from "./state.js";
+export interface ApplyOptions {
+    id: string;
+    scope?: HarnessScope;
+    rollbackOf?: string;
+    /** State captured before planning; used to reject conflicting edits. */
+    baselineState?: HarnessState;
+    /**
+     * Trajectory citation stamped into newly created entries' metadata
+     * (sourceSession + sourceSeqs); updates keep whatever the entry already
+     * carries. Omitted entirely when the caller cannot determine it.
+     */
+    source?: EntrySource;
+}
+export declare function applyRefinementProposal(state: HarnessState, proposal: RefinementProposal, options: ApplyOptions): RefinementResult;
+export { entryChangedSince };
+//# sourceMappingURL=apply.d.ts.map