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.
`dsh-global-tool-mask` wrapped `tools.restrict()` as a composition row so a
preset's user patch layer could hide a host tool. No shipped preset, the
plugin manager, and the settings UI compose such a row, so the package had a
README, a test, and no consumer. It goes, with its `apps/cli` dependency,
its `tsconfig.host.json` reference, the README example that named it, and the
generated catalogs. `tools.restrict()` stays: subagent tool filters call it.
Hiding a host tool from one preset is a per-preset setting for the
scope-resolved settings of the next layer, not a hand-written row; the Agent
Note records the removal under alternatives considered.
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.
The failover runbook's Windows switch scope returns to the native Windows jobs:
remove the Python runtime eligibility sentences, the pool Python prerequisite,
and the proposal links. The #3629 self-hosted Python runtime proposal is
retired (deleted as obsolete after its unmeasured-throughput run), and a new
implemented note records that the Windows x64 runtime lane stays on
GitHub-hosted Windows with the evidence and alternatives. Bilingual sidecars
re-recorded.
The failover runbook's Windows switch scope returns to the native Windows jobs:
remove the Python runtime eligibility sentences, the pool Python prerequisite,
and the links to the proposal. The self-hosted Python runtime proposal is
moved to rejected with its unmeasured-throughput verdict, and a new
implemented note records that the Windows x64 runtime lane stays on
GitHub-hosted Windows with the reasons. Bilingual sidecars re-recorded.
Windows x64 runtime builds resolve their hosted matrix.runner unconditionally
again (windows-2025 for pull-request CI). Remove the DSH_CI_FAILOVER_WINDOWS
selector, job-private Python toolchain, self-hosted dependency install and
post-step cleanup, the private setup script, and the routing spec introduced
in #3629. The Windows failover switch again covers only the native Windows
jobs in ci.yml.
Consolidate the existing notes-only PR into one review baseline. Preserve the exact reviewed tree; subsequent corrections remain separate commits with their own rationale and regression evidence.
Reflect the review feedback on the reader list: assistantStreamFirstTokenTime and the has-visible readers stop at the first qualifying member, while lastAssistantStreamChunk, assistantStreamChunks, and joinAssistantStreamText scan the whole stream; assembleAssistantStream feeds a BlockAssembler one joined delta per run. Record-level readers trust the static record type; expandAssistantStream is the validating path.
Session format v2 embeds each attempt's compact stream in
assistant/message and assistant/attempt, but Host and client consumers
still expanded it into per-member TimedStreamChunk arrays and did
per-member work; expandAssistantStream materializes the full array before
find/toReversed/break can answer. Session Stats (the projection phase of
every Session open), the token meter's usage and provider-assembly folds,
the subagent output fold, and the Session Controller image lookup still
paid O(members) allocation and time per settlement.
The Chat and Trajectory definitions were already settled from
message.content on master; the remaining per-member folds stay.
dsh-llm now exports record-level readers (first token, visible content,
visible text, last raw chunk of a type, raw chunks of a type, joined
text, run-aware assembly, per-run first-token/first-visible times) that
scan the compact records once with early exit. Session Stats reads
assistantStreamFirstTokenTime, the token meter reads
lastAssistantStreamChunk(stream, 'usage') and assembles through
assembleAssistantStream, the subagent output fold appends
joinAssistantStreamText, and the Session Controller scans
assistantStreamChunks(stream, 'block-end').
expandAssistantStream is deliberately not memoized: retaining expansions
costs roughly ten times the compact stream for the Session's lifetime.
It remains the validating path at durable boundaries.
Synthetic 200-turn v0 migration benchmark, median of five: first-open
projection 28.0 ms -> 5.4 ms, first-open total 76.9 -> 50.0 ms, peak RSS
137.2 -> 94.9 MB; reopen projection 17.8 -> 5.6 ms; all phase budgets and
the 128 MB heap constraint keep passing.
Yichen Jiang