38 plugins on cordis: what a downstream family leans on (and where we hit the edges)

[PerryLink]

Hello - I maintain a family of 38 published npm packages (github.com/PerryLink) that all run as cordis plugins inside an agent harness. Sharing a few notes in case they are useful for the roadmap.

What we lean on daily:

• ctx.effect / ctx.on for lifecycle, with every registration returning its disposer;
• inject plus service resolution across plugin trees;
• schemastery-backed Config schemas for every plugin;
• cordis.yml profiles composed from a base plus include patch overlays (insert / disabled / config overrides);
• HMR while iterating on plugins.

Where we hit the edges (all reported upstream):

• include: a row inserted by one patch cannot be addressed by later patches in the same list #112: a row inserted by one patch cannot be addressed by later patches in the same list (fix plus regression test open as PR fix(include): apply patches to rows inserted by earlier patches #116);
• the debounced include write path on Windows (fix(include): serialize config writes #47, plus a follow-up on transient rename failures that I am preparing);
• loader error messages for default-exported plugin modules can point at the wrong cause, which cost us a few debugging sessions.

What we can offer:

• a downstream compatibility matrix: we can validate the full family against rc.9 (or future releases) in an isolated environment and report per-package results;
• docs review from a heavy user's perspective (patch semantics, lifecycle rules, loader diagnostics);
• triage help once our first fixes land, if the maintainers want it.

Thanks for cordis - the composition model is the reason a family this size stays maintainable.

大家好 - 我维护着一个 38 个已发布 npm 包的插件家族 (github.com/PerryLink), 它们全部作为 cordis 插件运行在一个 agent harness 里。分享几点, 希望对路线图有用。

我们每天依赖的能力:

• ctx.effect / ctx.on 生命周期管理, 每次注册都返回 disposer;
• inject 与跨插件树的服务解析;
• 每个插件都有 schemastery 支撑的 Config schema;
• cordis.yml profile: 基础层 + include patch 覆盖层组合 (insert / disabled / 配置覆盖);
• 迭代插件时的 HMR。

我们撞到的边界 (都已上报上游):

• include: a row inserted by one patch cannot be addressed by later patches in the same list #112: 同一 patch 列表中, 前一条 insert 插入的行无法被后续 patch 定位 (修复 + 回归测试已作为 PR fix(include): apply patches to rows inserted by earlier patches #116 开放);
• Windows 上 include 防抖写入路径 (fix(include): serialize config writes #47, 以及我正在准备的瞬时 rename 失败跟进);
• loader 对 default 导出插件模块的报错可能指向错误原因, 让我们多花了几次排查时间。

我们能提供的:

• 下游兼容矩阵: 可以在隔离环境里把整个家族对照 rc.9 (或未来版本) 验证一遍, 逐包回报结果;
• 重度使用者的文档 review (patch 语义、生命周期规则、loader 诊断);
• 等我们首批修复落地后, 如果维护者需要, 可以帮忙 triage。

感谢 cordis - 组合模型是这个体量的家族还能维持可维护性的原因。

[PerryLink]

rc.10 refresh (2026-09-09, registry-direct): the 48-symbol export surface
is unchanged (0 added / 0 removed). Two of the seven known rc.9-vs-fork
structural diffs have now closed upstream: Events 'internal/update' and
Fiber.update both return Awaitable, matching what the harness fork
already shipped. Five remain open (Context.baseUrl | undefined, Events
symbol index signature + missing 'internal/config', EventsService
register/unregister visibility, LoggerLevel numeric order, symbols.caller).

Re-probed the 45-file cordis.patch.yml corpus against the rc.10 type chain:
43/45 PASS, identical to the rc.9 run (the same two template placeholders).
Six-repo typecheck remap: 5/6 zero errors; dsh-mask's two errors are the
fork-side include exports, unchanged.

New this round: a first runtime mount regression in a sandbox (rc.10 +
loader 1.0.0-rc.7 + include 1.1.0). File level 42/45 PASS; all 38 primary
repo patches mount cleanly (39 fibers ACTIVE, 2 PENDING on inject services
the synthetic environment does not provide). The 3 failures are fixture
files importing packages that are not on the registry - genuine import
failures, recorded as such. Loader rc.7 breaking rename noted:
EntryTree.write() -> commit(change).

#124 (NodeNext export surface) and #126 (tarball src) both remain timely on
rc.10 - the entry d.ts still has 7 extensionless re-exports and files still
lacks src. First benchmark numbers for the paper's section 5.3 future work
are recorded separately (effect pair ~9.2-9.9 us, small fiber
mount/unmount ~131-138 us on this machine; first baseline, not comparable).

Method and raw outputs are kept in a local evidence tree; happy to share
any specific file on request.