The reconnect supervisor used connection lost for every transition into backoff, including an initial startup attempt that never established a connection and later retry attempts that also failed. That wording implied a previously healthy generation and obscured whether any tools had ever been registered.
Capture whether the generation had reached the established state before scheduling recovery. Established disconnects retain connection lost/reconnecting; startup and retry failures now report connection failed/retrying. The reconnect-disabled diagnostic uses the same distinction while preserving its concrete manual-recovery guidance.
Unit assertions cover established loss, initial failure, retry failure, and both reconnect-disabled branches. Focused package coverage remains 100%, and the bilingual Agent Note records the observable state vocabulary.
Strict startup intentionally wraps connection and synchronization failures with the server-qualified activation diagnostic while retaining the original error in Error.cause. The prior assertion checked only the wrapper text, so the causal chain could regress unnoticed and erase the actionable transport failure.
Assert the full wrapper message and object identity of the original connection error in cause. This keeps operator-facing context and the underlying SDK diagnostic independently stable without changing production behavior.
The failure-cap path already appends tool disposal to syncChain, but the existing tests only covered give-up after settled discovery. They could not detect a future change that disposed the old set immediately and then allowed a blocked re-sync to publish a new leaked generation.
Hold a list_changed fetch open, drive the reconnect budget to exhaustion, then release a different tool list. The test proves final cleanup runs after that in-flight swap and removes both the previous and late-published tool names while creating no attempt beyond the configured cap.
The supervisor selected strict startup registration with a shared isFirstSync flag. Because the MCP SDK may deliver tools/list_changed before connect() resolves, that notification could enter the sync queue first, consume the strict option inside its contained handler, and leave the actual activation sync non-fatal.
Pass startup intent explicitly to connectGeneration(). Only the plugin activation attempt receives the failOnStartupError registration policy; notification-driven syncs and later reconnect generations always use contained runtime semantics. Queue arrival order can no longer redefine startup behavior.
A regression test injects list_changed from inside connect(), keeps a foreign namespace squatter in place, and proves activation still rejects after the notification's contained sync. Focused package coverage remains 100%, and the bilingual reconnect note records the ownership rule.
The MCP SDK starts a fire-and-forget close when initialization fails. Its stdio transport clears its process field before that close finishes, so our second Client.close() could return immediately and the reconnect timer could launch a replacement while the original child was still alive.
Track the transport onclose signal for every client generation and gate failed-attempt backoff on both Client.close() settlement and that signal. Use the same barrier during plugin disposal. If the SDK's bounded stdio termination window expires without onclose, fail closed and report incomplete shutdown instead of risking overlapping server processes.
Regression coverage models the SDK's early-returning second close, delayed and missing close signals, pending-connect disposal, close rejection, and the terminal timeout path. The reconnect Agent Note and Chinese counterpart now record the quiescence contract.
Replaces the patch round trip. An app's entrypoint resolves the command
line into a service, and the rows it configures read that service from
their own config — port: !!js ctx.get('webStartup')?.port ?? 3080 — so the
resolved value beats the value written beside it and nothing is written
back into a row or handed to the launcher.
A bundle names the entrypoint row in its manifest (dsh.bundle.entrypoint),
which is what lets the boot mount in two passes: entrypoints alone, then
the whole composition. That ordering is required, not cosmetic — a row's
config expressions are evaluated when the include applies the row, and a
strict ctx.get only answers for a service whose providing fiber is already
active.
What this removes: ctx.appPatches and the launcher-owned patch layer, the
disable/re-enable recycle and its in-flight-mount barrier, overrideConfig,
and the reload hazard they existed for. A live config edit now re-applies
the second pass against services that are still up, so a served port
survives by construction.
What it adds: ctx.appReady, because Loader settlement no longer means the
app is up — a row mounted in the second pass can observe a settled tree
while that pass is still running, or already rolling back. The web URL line
waits for it, so a boot that fails in the second pass announces nothing.
dsh-web-app owns --host/--port/--dev/--workspace-root/--trusted-host and
its --help in a web-startup row; the rows it configures wait for the
webStartup service, and the client-plugin HMR receiver now ships disabled
so --dev is a row toggle rather than a runtime insert (the Loader cannot
resolve a row inserted from inside a mounting plugin).
dsh-headless owns the task positional and rejects a missing task as its own
usage error. Its runner ships disabled, not merely waiting: the schema
requires the task, and a row's config is validated when its fiber is
created, before the startup row can supply one. A composition has exactly
one command-line owner, so the patch disables the web startup row and this
one provides webStartup too, leaving the web rows on their composed
one-shot values.
The keyless web scaffold provides the same three values with no arguments,
which is what an embedding host with no command line does.
A launcher provides three values before the tree mounts: ctx.cmdlineArgs
(get() is the whole interface) carrying everything after its own flags,
ctx.appExit for a bounded exit, and ctx.appPatches for decisions a later
recomposition must keep. An app's startup row injects cmdlineArgs and calls
runStartup() with its own commander program.
Rows the app configures inject its startup service, so they wait until the
startup row has resolved their values and provided it; --help prints,
disables those rows, and exits without the app ever starting.
A changed row is recycled — disabled, then re-enabled with its new values —
because a row's config is resolved when the Loader creates its fiber, while
the row is still waiting. Recycling never touches inject: an inject update
restarts the row from its unwrapped callback and loses the plugin's own
static injections. A mount still in flight is allowed to settle first, so
the disable has a fiber to dispose instead of racing one into existence.
Integrate master at 8b4ef532 before addressing review feedback so each bot issue remains isolated in its own descendant commit. Resolve the vendor package rescope by moving the PR's new MCP supervisor imports to @deepseek-ai/cordis and @deepseek-ai/schemastery, preserve the reconnect implementation and e2e coverage, and regenerate the merged config and module catalogs with fresh bilingual pairing records.
Machine-produced by `pnpm run rescope-vendor --apply` plus the regeneration it
prints: `pnpm install` for the lockfile, `pnpm run gen-third-party-notices`,
`verify-translation-pairing --write` for the touched bilingual pairs,
`gen-doc-graphs`, and one typert snapshot whose ids embed character offsets.
`pnpm run rescope-vendor --check` verifies the result.
Renames nine vendored packages (cordis, cosmokit, schemastery and the six
@cordisjs plugins) and every reference that resolves them: manifest names and
dependency keys, module specifiers including declare-module merges, cordis.yml
plugin names, tsconfig paths, every Markdown fence, and `docs/` prose.
Directory names, upstream versions, and dependency ranges are unchanged, so
vendor/README.md still reads as an upstream snapshot; its manifest table gains
an upstream-name column so THIRD_PARTY_NOTICES keeps MIT attribution pointed
at each fork's origin.
The tutorial tier follows the rename end to end: its yaml fences named plugins
the Loader can no longer resolve, its `ts ignore-check` fences disagreed with
the compiled fences beside them, and its prose quoted both. The contracts that
told readers to keep upstream names — the root convention and the vendoring
cookbook's tree comment and manifest invariant — now say to rescope instead.
Two rules read `@deepseek-ai/` as "another workspace plugin": the client bundle
purity gate now names the vendored libraries a browser bundle inlines, and the
files where a bare `cordis` is an agent-preset id keep that product data.
Two sites reached a registry for the vendored framework, which the rescope
turns from a silent second copy into a hard failure.
Live-link mode relinked only the root manifest, so a generated workspace member
— `plugins/*/package.json` — resolved its own dependencies from the registry
and installed upstream cordis beside this repository's vendored copy.
`LinkWorkspace.relinkNestedManifest()` relinks every nested generated manifest;
`peerDependencies` keeps its range because package managers reject a link spec
there.
The sandbox publish-path rehearsal installs this repository's vendored cordis
and cosmokit tarballs instead of naming a registry version.
A restriction was compiled against the global tool layer alone: only
global-layer tools were tested against `admits()`, and every chain-layer
tool was overlaid unfiltered afterward. That read the exempt set as "the
global layer" when what it means is "what this scope registers itself" —
two descriptions of the same set only while every model-facing tool sat in
the host composition.
Moving those rows onto the agent plane separated them. A preset's tools are
an ANCESTOR contribution to a joined agent, so a subagent's `toolFilter`
stopped constraining anything it was given; and with the global layer empty
`restrict()` rejected every name it received as unknown, failing the child
outright. With the same tools in the global layer the filter still admits
and applies normally, which is what makes this a regression of the move
rather than a standing limitation.
`view()` now filters everything a scope inherits — the global layer and
every ancestor layer on its chain — and exempts only the layer the scope
owns. That exemption is load-bearing rather than incidental: the delegation
runtime registers a child's `report` and structured-output tools into the
child's own layer, and a filter naming the capabilities the child may use
must not strip the machinery it answers through. Tool order, and with it
prefix-cache reuse, is unchanged: inherited names keep their global-then-
ancestor position and own-layer names still come last.
The diagnostic said "unknown global tool" while listing what is really the
inherited surface; it now names the surface it checks and says why an
own-layer name is not restrictable.
Fixes#2185
