🥮 mooncakes.io

    harness-self-evolution

    面向 DeepSeek Harness（DSH）的全盘自进化 stdio MCP 插件（MoonBit 原生）：16 个工具完成 扫描→监控→提案→审批→执行 闭环，含子 Agent 工厂与只读运行时快照；人工审批闸门；真实子进程执行在事务性 adapter 落地前 fail closed，失败只回滚提案状态。v3.0 起 DSH 单宿主，v3.1 启动路径可移植化，v3.2 注入面 + M7 安全闸门。

    harness
    evolution
    self-improvement
    matt-pocock
    mcp
    Download zip
    Version
    0.3.8
    License
    MIT
    Repository
    Last updated
    18 hours ago
    Downloads
    38

    Dependencies

    #Harness Self-Evolution Plugin

    A MoonBit-native plugin that scans, monitors, proposes, and validates evolutions for the DeepSeek Harness (DSH) plugin ecosystem. Version 3.2.3. MIT.

    This plugin targets DeepSeek Harness. Deployment hinges on how DSH launches the binary, where data is read and written, and which files (patches, manifests, panels) sit alongside the binary.

    #Where your data lives

    DSH resolves its home via $DSH_HOME (~/.dsh by default; a managed launcher may point it elsewhere). The plugin mirrors that: user-scope sub-agent definitions land under <DSH home>/skills/. Full install, verification, and recovery guidance: docs/deploy/deepseek-harness.md.

    Since v3.0.0 the plugin is DSH-only: the earlier claim of MiniMax Code compatibility was a over-declaration (DSH's own source has no such host concept) and has been removed.

    If you are writing or debugging a DSH plugin rather than deploying this one, read docs/dsh-plugin-integration.md: the host-half / client-half contract, the Lazy-CJS client bundle rule (one stray top-level export breaks every plugin in the combo), DSH_HOME routing, and the diagnosis order for "Failed to load plugins".

    #What it does

    • Scan installed plugins across the configured scan roots for the active host
    • Monitor call latency, success rate, token usage, retry count, and user feedback
    • Identify strong signals (user override, three consecutive failures, > 20% latency regression) and medium signals (repeated parameter misuses, loop detection, repeated preferences)
    • Propose benchmark-driven evolutions bound to Matt Pocock engineering principles
    • Approve is always a human step. auto_approve is false by default and stays false — the only gate against "code changes itself into a wall"
    • Execute with state machine pending → approved → executing → completed; real execution stays fail-closed until a transactional sandbox adapter exists, while failed validation returns the proposal to pending
    • Sub-agent factory persists Markdown + YAML frontmatter definitions; scope plugin lives under the plugin's data root, scope user lives under <DSH home>/skills/ (see the deploy guide)

    #Architecture in one paragraph

    The compiled binary (bin/harness-evolution.exe) is a stdio MCP server exposing sixteen tools. The DSH home resolves in src/store/paths.mbt::dsh_home: $DSH_HOME when set, else derived from the plugin's own install path (<X>/profiles/<name>/node_modules/...), else ~/.dsh. The user-scope directory is <DSH home>/skills (HARNESS_EVOLUTION_USER_DIR overrides it), and the default scan roots follow the same resolved home (src/scanner/scanner.mbt::default_scan_roots). Since v3.0.0 the multi-host abstraction is gone — DSH is the only target.

    See docs/code-architecture.md for the path-resolution detail.

    #Build

    .\build.ps1 -Task all    # check + test + build; one command rebuilds the binary

    Build prerequisites: MoonBit >=0.1.20260904, MSVC or Clang on Linux/macOS. The current binary in bin/harness-evolution.exe is Windows-native; rebuilding on the target platform produces a native binary for that platform.

    The complete workspace gate also runs panel/watcher builds and an isolated desktop coexistence smoke without opening a browser:

    $env:DSH_CLI = 'C:\\path\\to\\dsh-0.1.5-rc.2\\lib\\bin.js'
$env:DSH_REGRESSION_CLI = 'C:\\path\\to\\dsh-0.1.6-alpha.1\\lib\\bin.js'
node scripts/verify-workspace.mjs

    #Data and configuration

    The plugin stores proposals, metrics, signals, cache, and execution log under $HARNESS_EVOLUTION_HOME when set; otherwise it uses the current DSH tree at <DSH_HOME>/.harness-evolution/v2/ (with the host-compatible ~/.dsh fallback only when no DSH tree can be resolved). Override the env var to keep dev/test data separate.

    Configuration is read in this order, first file that exists wins:

    1. $HARNESS_EVOLUTION_CONFIG (explicit override)
    2. <cwd>/.dsh-plugin/plugin.json (self-manifest)
    3. .dsh-plugin/plugin.json (relative to plugin root)

    If none exist the plugin starts with built-in defaults — missing config is not a startup failure.

    Breaking change (unreleased v2.7.0): the legacy <cwd>/.zcode-plugin/plugin.json fallback is no longer read. A deployed instance that still carries that file must rename it to .dsh-plugin/plugin.json — the contents need no change.

    #Installing into the right tree (DSH_HOME)

    Installing is per DSH tree: dsh plugin add writes into $DSH_HOME/profiles/<name>, and $DSH_HOME decides which tree boots (~/.dsh by default — a managed launcher may point it somewhere else). Trees share nothing: bundles, node_modules, sessions, and skills are all per-tree, so an install verified in one tree stays invisible to a host booting another.

    Since v3.1 the shipped mount row is disabled: true and carries no machine path — a literal path for one machine points at a missing file everywhere else, so the plugin would simply never mount (DSH reports one warning line and keeps booting: mcp-harness-evolution is not in the host's requiredStartupEntryIds, so its activation failure is optional). The row is written per tree at install time:

    # resolves $DSH_HOME (or ~/.dsh), installs the bundle, and injects the mount row
# into that tree's profile patch layer; add -DryRun to inspect first
pwsh -File scripts/install-dsh.ps1 -Profile web

# then verify statically (no boot needed)
dsh --profile web --dump-config | Select-String 'mcp-harness-evolution'

    dsh.profile.bundles and the patch layers are read at boot, so the host must be restarted before the sixteen mcp__harness-evolution__* tools appear in a session. The plugin additionally derives its own tree from the install path, so create_sub_agent scope=user lands in <DSH home>/skills/ even if env.DSH_HOME were left unset.

    What is actually fatal, and what is not. failOnStartupError: true on the mount row does not abort the profile boot: it only rejects that one plugin's activation, and app-boot downgrades it to an optional entry warning (see the paragraph above). The genuinely fatal case is a patch layer DSH cannot parse — parsePatchList throws and the profile never boots. scripts/install-dsh.ps1 writes that file, so its output is pinned by a regression suite that parses it with the host's own js-yaml:

    pwsh -File scripts/test-install-dsh.ps1     # 7 scenarios; writes only to a temp tree

    Full install, verification, and recovery guidance: docs/deploy/deepseek-harness.md; the live-install practice (multi-home reality, .dsh-module-fallback pitfalls): docs/dsh-compatibility.md.

    #License

    MIT. See LICENSE.

    Powered by MoonBit

    Site sourceReport issuePackagesBuild queueSkillsStatistics

    © 2026 mooncakes.io