The tail dropped a whole kept chunk whenever a later chunk pushed the total
over the limit, so a flood that arrived as one read (Linux) left only the
bytes written after it. `StreamTail` trims the oldest kept chunk instead and
is unit-tested with both read patterns.
`disableBundle` now moves a package from `dsh.profile.bundles` to
`dsh.profile.disabledBundles`, and `reconcileInstalledBundles` enables only
a bundle named in neither list, so an unrelated pnpm run never re-enables a
bundle the user turned off while an update that gives a dependency
`dsh.bundle` still activates it (the built-bin e2e scenario the previous
"new dependencies only" rule broke). The record goes when its dependency
does.
Also: `visitIdentifiedRows` in patch-rows owns the per-patch restatement
tracking that compose-stack and external-bundles had cloned (duplication
gate), and the probe spec's flood fixture retries short and EAGAIN writes
because the child's stderr pipe is non-blocking once tsx touched
process.stderr (the test raced the parent's drain).
master now bounds a README Summary at 100 words; the boot group, the plugin
manager, and its host face say the same in fewer. The zh config catalog gains
the paragraph break the merge dropped.
The root entry now exports only what the CLI, the plugin inventory, the
plugin manager, and tool-cordis use, plus the parameter and result types of
those operations. The contained group and its registry, the external-layer
composition, the row-ownership helpers, the probe format number, and the
runtime option type stay module exports the package's own tests import.
The view attributes a disabled row through profileRuntime.userDisables, which
walks the groups holding the entry, and the manager spec stops passing the
runtime option the profile runtime no longer takes.
The inventory keeps dsh-app-boot an optional, type-only peer, so the walk
over an entry's owning groups moves behind ProfileRuntime.userDisables(): a
consumer with the runtime asks it, one without reports composition as before.
The module graph and the cordis catalog follow.
The probe compared two resolved URLs, which differ for one package under a
source launch (src against lib), across a symlink, and behind the proxy a
packaged executable writes; it now walks each URL up to the cordis manifest,
follows a dsh module proxy to its target, and compares real package
directories, with a probe test for the linked and the proxied case. The README
records that the probe child itself cannot run from a packaged executable.
The runtime's user-disabled row ids now come from the committed composition —
composeProfileStack collects them from the user layers it composed — so a
rejected or half-written patch file no longer makes the list disagree with the
running tree or fail to parse. The inventory attributes a row disabled through
a group the user disabled to the user, through the shared userDisablesEntry
walk, and lists a user-layer row the composition left out under the new user
trust instead of builtin.
The probe child now runs with the harness's credential-shaped variables
scrubbed — the one rule every spawner uses, now defined in
dsh-launch-environment and re-exported by dsh-subprocess — and its report is
the one message echoing a per-run token the child removes from its environment,
together with process.send, before importing anything. The failure text keeps
the last 16 KiB of stderr instead of every byte, and the probe settles only on
the child's close, so no pipe or channel outlives the call.
reconcileInstalledBundles now enables only the dependencies a run added, so
a bundle the user disabled keeps its state through an unrelated install or
removal. Row registration records the group each row is declared under: a
config override restates a row only under that group, while the same id twice
in one list or under another group is a duplicate — a contained bundle is left
out and a built-in layer fails loud, instead of the Loader rejecting the update
or moving the row at mount.
Every chunk now carries the directory pnpm ran in beside its argv, so a
renderer can say where the install lands. The generated event graph and the
cordis API catalog follow the chunk type.
The installer now decides pnpm's colouring per consumer: the Web host keeps
colours on for the install dialog's terminal, the CLI turns them on only when
stdout is a terminal, and a plain log strips what a colour=always pnpm still
prints. Every chunk carries the argv it came from, so a renderer can label
each pnpm run.
A contained row left pending is recorded as waiting for a service, but a
live recomposition that brings the service back — a sibling row switched
on again — activates the fiber in place, with no create() to clear the
record, so the registry kept reporting a wait that was over. The group now
watches fiber states once it has recorded a wait and clears the record
when that row's fiber comes to life.
The two `ts` fences declare their inputs and import what they use, so the
doc-typecheck gate compiles them against the workspace API instead of
failing on the names the surrounding prose supplied.
`src/index.ts` held the installer, the manager, the view fold, and every
helper in one 1100-line file. It is now the package API alone, re-exporting
from `installer.ts` (pnpm runs, the probe cache, post-install checks),
`manager.ts` (every operation over the booted profile and the mutex),
`view.ts` (the package view fold and the row-ownership walk the dependents
query shares), `modules.ts` (declared-module row naming and views), and
`helpers.ts` (the diagnostic prefix, tooling bounds, spawn seam, and manifest
readers). The public surface, the tests, and the failure codes are unchanged.
The `cordis.patch.yml` parser and the comment-preserving writer are the
Loader's format, not a mechanical primitive: they build on the include
plugin's dialect and the Loader's row types, which no other `util` package
does. They now live in `app-boot` as `src/patch-file.ts`, exported as
`@deepseek-ai/dsh-app-boot/patch-file` with its own bundle entry, so the
agent-preset roster and the plugin manager load them without the boot entry.
`dsh-patch-file` goes, with its README and catalog entries; `yaml` moves into
app-boot's dependencies.
A dozen plugin packages already depend on `dsh-app-boot`, so the roster
joining them costs nothing; the note records why the separate package was
tried and folded back.
What plugin management does now lives in `dsh-plugin-manager`, a boot-group
package beside `app-boot`: `PluginInstaller` runs pnpm, probes, and judges new
packages over the profile on disk, and `PluginManager` adds every operation
over the booted tree, each refusal a `PluginOperationError` with a `plugins/*`
code and no Remote protocol attached. The manager reads the profile runtime,
the preset roster, and the running-agent count through per-call readers, and
the roster only as `PresetLayers`, so the package depends on neither the
roster nor the agent registry.
`dsh-host-plugin-manager` keeps `pluginManager` and the `plugins` Remote as a
relay: one Remote method per manager method, and one exhaustive switch
turning a failure into the Remote error of the same code (`plugins/bad-request`
into `gateway/bad-request`). Its `./types` re-exports the manager's types, so
`api-remotes` and the client import one vocabulary unchanged.
`dsh plugin add` and `remove` now run the installer instead of a bare pnpm
forwarder: every new package is probed, a package that declares neither a
bundle nor a plugin module is removed again with its reason, and newly
installed bundles are enabled as before. Other pnpm verbs still forward.
Second review round on the contained-group composition:
- Ownership scanned only `insert` rows, so an external bundle could set a
built-in row's id as a child through `{ id: own, config: [...] }` and the
Loader re-parented that row under the bundle while `originOf` still named
the built-in layer. The rows an id-targeted patch sets as a group's config
now count as the layer's own: an external bundle that sets an id another
layer owns is left out whole, a built-in layer that takes another layer's
id fails loud, and a layer restating its own child is not a repeat.
- `recompose` read the committed profile before awaiting the update, so a
watcher firing while a bundle was being enabled composed the tree before
the enable and applied it after; the plugin vanished while `layers` kept
listing it. Recompositions now queue, each reading what the previous one
committed, and a rejection stays its own caller's outcome.
- A patch that replaced a built-in target's config removed the wrapper the
bundle had put there, yet later inserts still targeted the detached
wrapper and vanished. The composition forgets the wrapper on such an
override and creates a new one under the same id.
- Generated wrapper ids bypassed the duplicate check, so a bundle declaring
a row spelled like a wrapper produced two nodes with one id. Declared and
generated ids now go through one registration.
- Failure records outlived their rows when a bundle update dropped a failed
row without unmounting the group, or when a pending row sat in a plain
nested group whose id the record named. A contained group now drops the
records of rows its configuration no longer names after each update, and
the audit records every row under the contained group that isolates it.
- `applyEntryPatches` assigned override values by reference, so applying a
patch list carried a later insert's rows back into the patch object; the
vendored include now clones object-valued overrides as it does inserts.