wireSchemas() already advertised only run_code under mode: 'code', but the
executor resolved every call through get(), which returns the full visible
map plus the reserved transport. A model could name a native tool directly
and bypass run_code entirely. Route the execution-path lookups through a
new private resolveExecution() that applies the mode collapse at the
operation boundary: model-direct calls under 'code' may only name run_code
(UNKNOWN_TOOL otherwise), while SDK sub-dispatches (parent token set) keep
every visible tool. get()/schemas() public semantics are unchanged.
The denial happens at createExecution, before the extensible policy
pipeline — pre-execute listeners, approval ask, and guards never observe
a call that is deterministically denied. A collapsed call honors the
pre-dispatch cancellation contract, routes aborted results through the
visible tool's finalizeContent, and captures the finalizer before
argument materialization.
Under code mode, a system-prompt/assemble listener filters out tool:*
guidance sections that told the model to call native tools directly.
The tools:sdk section and SDK types remain so programs can still use
all tools through run_code.
Fixes#1815
The previous pass over-corrected: presenting both roots as unknowable made the
guidance vaguer without making it truer. Every `dsh` launcher puts authored
presets under `$DSH_HOME/.agent-presets`, so the skill and persona state that
as the answer to "where do my presets live" and reserve the read-back for the
path an agent actually opens or edits, which `copy()` already reports.
Neither root is a property of the preset layer. `Config.roots` defaults to `[]`;
the two the shipped agent sees are patched in by the dsh CLI, whose user root is
`dshHomePath('.agent-presets')` and whose system root ships beside the install.
`writableRoot()` then takes the first `user` root from whatever configuration
supplied it.
The skill and the persona stated `${DSH_HOME:-$HOME/.dsh}/.agent-presets/<id>/`
as where authored presets live, which is only the default install's answer, and
no call reports either root: `authorable` says whether a writable one exists,
and `list()` cannot reveal a user root that is still empty. Both now keep that
path as what to tell a user asking where to look, and direct every path an agent
acts on to `list()` or `resolve()` — which is also what `copy()` already relies
on, since it chooses the writable root itself and reports the file it created.
Review found two errors in the first pass.
The skill quoted `service "<name>" has been registered at <Owner>` as the
rejection for a service published into the root realm. That message only
appears when the loose name collides with one the host already supplies. A
preset publishing its OWN new service name reaches the root realm successfully
and is rejected by the mount audit instead, with `row(s) published
process-global service(s) [...]` — which is the shape a forgotten realm
actually takes. Both messages are now listed, verified by mounting one preset
of each kind.
The Agent Note claimed every harness package publishes exactly three file
patterns. 139 do; 28 also publish `lib/client.js`, 22 publish
`lib/types/**/*.js`, and `dsh-message-feedback` publishes `src`. The note now
states only the fact it needs: no package's `files` includes its README.
Also from review: the copy step no longer repeats the display name it already
set, the id rule spells out `[a-z0-9][a-z0-9-]*`, and the sandbox step says
reads need no escalation.
apps/web/tests/README.md records why these e2e type-check in the Host aggregate
and why importing a Client package there pulls its project tree into the Host
build graph, with mirroring as the standing answer. The Agent Note drops the
directory-picker face split (assessed and declined) and the grep-level gate in
favour of that README.
docs: regenerate the catalogs and retarget the moved declarations
The forwarded-event change moved three owner packages' cordis `Events`
declarations and their branded types into client-safe `./types` modules, and
the settings-scope split moves the shell spec into ui-settings-general. Point
the type-equivalence manifest and the affected Agent Note at those homes,
register the new `remote/*` event scope and the `ctx.settingsScope` service in
the catalog partition, and re-run the generators.
`$on` joins the documented `TypeRTClientRemote` surface, and the two Agent Note
fences that quote a bare member signature are marked `ignore-check`: they are
declaration fragments, not compilable units.
refactor(client): make ui-settings the settings domain's base layer
The settings-namespace transport lived in client/runtime, where every feature
could value-import it because runtime is a platform module. It belongs to the
settings domain, but moving it into ui-settings as a shared function fails
twice: the client bundle purity gate forbids cross-plugin value imports, and
ui-settings reached ui-sidebar for its shell, so any feature depending on it
closed a cycle through ui-layout and ui-theme.
Both halves move. `ctx.settingsScope` is now a cordis service — the
collaboration shape the purity gate prescribes, and the service proxy binds
`this.ctx` to the caller, so a bound scope's disposer belongs to the calling
fiber. The shell ui-settings used to own (the `sidebar.settings` occupant, its
navigation, and the nav-row projection) moves to ui-settings-general, which
already owns the chrome and the General section. What stays in ui-settings is
what carries no `ui-*` dependency: the scope service and the canonical settings
slot types, `settings.general.item` included. That type was parked in the locale
package precisely because the declarer was unreachable without a cycle; every
registrant now depends on this base layer, so it comes home.
The scope CONTRACT stays in client/runtime: a feature service accepts a scope
through its own signature without depending on the surface that binds it.
The forwarded settings invalidation replaces the deleted client-side
`settings/changed` event, so the transport reads `ctx.remote.$on`. It reaches
`$on` through the gateway's Client half plus the allowlist's type-only subpath
rather than api-remotes' Client face: that face imports a Host-tsdown-generated
artifact, and this package is reachable from the Host build graph through its
callers.
refactor(client): reach the settings transport through ctx.settingsScope
Every feature that owns a preference row switches from value-importing a shared
binder to the settings domain's service, and declares the two injections that
binding needs: `settingsScope` for the transport and `remote` for the forwarded
invalidation it subscribes to on the caller's own context.
The rows stay with the features that own the preferences — Language with locale,
Appearance with ui-theme, Composer Enter with ui-conversation. Only their route
to the transport changes, so no settings surface moves and no feature gains a
dependency on the shell.
The `settings.general.item` slot type now arrives from ui-settings, the base
layer every registrant already depends on, which retires the re-export outlet
ui-theme kept and the parked declaration in the locale package.
client/runtime drops its settings-form and schemastery dependencies with the
transport that used them.
test(client): bind the settings transport in the specs that boot a preference row
Every bench that activates a plugin owning a preference row now supplies the two
services that plugin injects: the forwarded-event port and the scope service.
Specs that exercise no settings path get the minimal doubles; the ones that do
drive their refresh chains through `remote/host-event`, the same signal
client/runtime republishes from a forwarded frame, replacing the deleted
client-side `settings/changed` event.
Also fixes a publication defect the built-invariant gate catches once it runs:
api-remotes' invariant companion shared the allowlist module with the package
index, so rolldown hoisted it into a third chunk beside the two bundled entries
— a file the mechanically derived publication list does not carry, leaving an
installed companion unable to import it. The companion now reads the allowlist
through this package's own published `./types` subpath, which the bundle keeps
external, so each entry stays self-contained.
The dynamic-subscription cast in apiproxy is gone: after the vendored cordis
rescope, `on` accepts the rest-parameter handler directly, and the allowlist's
shape assertion still carries the safety argument.
fix(client): carry the settings-scope move across the release manifests
Rebasing onto the publishable release set replaced every manifest's dependency
block, so the packages this change touches restate their additions in the
workspace-protocol form: the base layer's own transport dependencies, and the
`ui-settings` plus `remote` edges each preference-row owner now needs.
ui-settings-general takes clsx with the shell it received, and client/runtime
drops the settings-form and schemastery dependencies that left with the
transport.
fix(api-gateway): give each $on subscription its own registration and containment
Two defects in the forwarded-event subscription table, both raised in review:
A set keyed on listener identity stored one entry when two callers subscribed the
same function object to the same event, so the first frame reached it once instead
of twice and either disposer silenced the surviving registration. Subscriptions are
now records addressed by registration, which is what "the disposer belongs to the
calling fiber" requires.
A listener declared void may still be `async`, and the synchronous `try/catch`
could not see its rejection: the promise was dropped and surfaced as an unhandled
rejection outside the documented containment. Delivery now attaches a rejection
handler when a listener returns a promise, so both failure modes are logged and
isolated alike.
Delivery also iterates a snapshot, so a listener that subscribes or disposes during
a frame no longer changes who receives that frame, and production matches the
TestRemote double instead of relying on live Set iteration order.
Both fixes are pinned by tests that fail against the previous implementation. The
double gains its own spec for the `$mount` refusal and the unsubscribed-name drop —
per-file coverage reaches it — plus a note that it propagates a throwing listener
where production contains one, so no spec mistakes it for the containment guarantee.
Three prose corrections: `assertJsonArgs` states where its throw actually surfaces
(the emitter's listener containment, not load or emit time), the browser e2e README
names every standing Client import rather than claiming one exception, and two
comments and a test title state the forwarded event instead of the deleted
client-side one.
refactor(remote): deliver forwarded frames through ctx.remote.$dispatch
The carrier used to relay each decoded frame over an internal
`remote/host-event` cordis event so the delivery port could stay off the Remote
contract. The relay was the wrong shape twice over: it put a client-face event
into a scan whose subject is the Host vocabulary, forcing a walk exemption for
something that is not a Host event at all, and it made a direct handoff between
two Client plugins look like a broadcast any plugin participates in.
`TypeRTClientRemote` now carries both roles of one surface — consumers subscribe
with `$on`, and whoever owns the Host frame sink hands frames over with
`$dispatch` — so client/runtime calls the Remote service directly and the event
declaration is gone. A cordis service method is the collaboration shape the
client bundle purity gate prescribes, and it needs no relay to satisfy it.
The trade is that the handoff is now developer-visible: any plugin holding
`ctx.remote` can synthesize a forwarded event. That is the exposure the relay
already had — `ctx.emit` was equally reachable — stated in the contract instead
of hidden behind a private subscriber.
runtime reaches `ctx.remote` through the gateway's Client face rather than
api-remotes': that face imports a Host-tsdown-generated artifact, and this
project sits in the Host build graph.
refactor(api-remotes): keep the allowlist value out of types.ts
`src/types.ts` carries only types by package convention, but it held the
forwarded-event array, so the type-only subpath published runtime code. The
array moves to `src/remote-events.ts` and `types.ts` derives its projection from
it; both compiler faces list both files, so the Host forwarding loop and the
consumer key face still read one declaration and the package's exports are
unchanged.
The invariant companion returns to an empty installer. Its dispatch-shape check
was the only reason the companion imported the allowlist, which made the two
bundled entries share a module: rolldown hoisted it into a third chunk that the
mechanically derived publication list does not carry, so an installed companion
could not import it. Dropping the check retires that coupling along with the
subpath-import and bundle-external workarounds it needed, and the shape the
check enforced at runtime is the part the Host face's `TypeRTForwardableEvent`
assertion already refuses at compile time.
test(ui-task): bind the locale plugin's new injections in its bench
The bench boots the real locale plugin, which now injects the settings-scope
service and the forwarded-event port, so it stayed pending and left `ctx.locale`
undefined. Supplies both doubles like the other benches that boot a plugin
owning a preference row.
docs: close the documentation gates for the forwarded-event surface
Regenerates the two graph catalogs and re-records every bilingual pair this
branch edited. Several pairs needed real work beyond the record:
- The generators write only the English side, so the Chinese sides of
`event-producer-consumer` and `module-graph` had drifted: the former still
listed the three deleted client-face events and pointed at declaration sites
this branch moved into `types.ts` modules, and the latter carried a stale
dependency graph.
- `TypeRTClientRemote`'s documented declaration gains `$dispatch` on both sides.
- The pairing contract requires both sides to link the same target, so the
apiproxy README and the design note now link the English note from both
languages, and the note's code blocks are byte-identical across the pair
(a translated comment inside a fence counts as divergence).
- `apps/web/tests/README.md` gains its Chinese counterpart; the browser e2e lane
documents a discipline reviewers apply, so it belongs in the bilingual corpus
rather than in the pairing exemption list.
- Four fences in the design note are marked `ignore-check`: each quotes a member
signature, a union arm, or a snippet that names symbols it does not import, so
none is a compilable unit.
docs(agent-note): transition the forwarded-event note to implemented
The design shipped in this PR, so the pair moves into `implemented/` and takes
that folder's skeleton: `## Proposal` becomes a present-tense `## Decision`,
and `## Acceptance criteria` plus `## Risks` fold into `## Verification` (what
pins the behavior) and `## Consequences` (what the shipped shape costs).
Facts that moved after the proposal are corrected rather than preserved: the
allowlist value now lives in `remote-events.ts` beside a type-only `types.ts`,
the delivery port is `$dispatch` rather than an internal cordis event, and the
invariant companion is an explained empty installer. `Verification` states the
two `$on` defects the review found — independent registration identity and
async-rejection containment — since those are now the properties tests pin.
Supersession is partial, so five active notes stay active and gain a
cross-link each: `web-config-plane`, `web-client-session-scope`,
`config-plane-boundaries`, `versioned-gui-welcome-onboarding`, and
`permission-default-for-new-sessions` each described a frame this change
replaced. Only the mechanism sentence is annotated; every conclusion those
notes own is untouched, and `host/models-changed` remains apiproxy's own
derived frame in all of them.
Also pins the disposer's idempotence: calling one `$on` disposer twice must not
splice a surviving twin registration out from under its owner.
fix: docs
fix: test
The client-hmr row joins the web bundle as an ordinary always-on roster
row: without a rebuild watcher rewriting client bundles it polls
unchanged files and stays idle. This deletes the --dev flag, the web
runtime's mode config, the mode-forked prompt contract, the DSH_WEB_MODE
bash variable, and the post-settlement row-creation machinery the
conditional row required. dsh web + pnpm run dev:web remains the
development loop.
The `cordis` preset's `editing-cordis-compositions` skill is the only guidance
an agent has when it authors a preset, and four of its statements were false.
`tool-bash` was named as the worked example of a row that hides a service; it
provides nothing and injects `bashEnv` from the host's own `bash-env` row, so
following that advice strands the row behind its realm and the preset fails to
mount. The `isolate` example composed `tasks-local` with `tool-tasks`, which
the shipped compositions' own comments say breaks `run_in_background`. A string
realm label was described as pooling one instance; labels join realms and
`provide()` still throws on the second registration. Rows were to be checked
against a package README, which no harness package publishes.
Verification is now the agent's own: `standingKeyFor(id)` runs the same mount a
session start performs and rejects an unresolvable package, an invalid config, a
service in the root realm, and a row that never activated. The skill states that
`list()`'s `broken` field is a shape check that every one of those passes, ships
the `cordis_mount` plugin that reaches the roster service, and names `copy()` as
the authoring write. The prohibition on touching the shipped install is promoted
to its own section and extended to the host composition.
Fixes#2266
The bubble clamped horizontally and did nothing vertically, so a long
label under an anchor low on the page ran off the bottom. Preset cards
made that the common case: custom presets sit at the end of the roster
and carry the longest descriptions.
A `top` or `bottom` bubble now flips to the other side, and only into a
side that genuinely fits — an anchor with room on neither keeps the
requested placement rather than oscillating. Sliding the bubble
vertically instead would cover the text being read.
A preset publishes its own description, of any length, and `.cards` sizes
rows with `grid-auto-rows: 1fr` — which makes every implicit row the same
height, not just the row holding the tall card. One long description
therefore set the height of the whole roster.
The description now clamps to four lines and offers the rest through the
shared Tooltip, attached only while the element actually overflows. Card
height stays derived: with the description bounded, `grid-auto-rows: 1fr`
already equalizes, and a card carrying the broken-preset reason or a
revealed path still sizes itself.
Tooltip gains an optional `maxWidth`; its default half-viewport cap
renders a description wider than the settings dialog it belongs to.
The failed-shape-check badge reads "Failed to load" rather than "Broken":
discovery reports a composition that is missing, unreadable, or malformed,
which overstates as damage.
Fixes#2238
list_agents called a storage-only continuable child 'complete', inviting
parents to poll it for a result that will never appear there. Rename the
status to 'ready' — resumable, not terminal — and rewrite the tool
description to say a parent is told when a child finishes, so listing is
for recall, not polling.
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
`session.create` also adopts an already-live session, and the preceding
commit newly allows adopting one under the preset it switched to while
blank. Its response still echoed `header.agentPreset`, so that adoption
answered with the preset the session had just left — contradicting the
request it had accepted and the row `session.list` serves for the same
session from `resolveSessionPreset()`. The echo now resolves the same way.
The `assertPresetUnchanged` parameter doc said `existing` was the preset
the session was created under; both callers now pass what it runs.
`composeFrom()` was documented as "infallible" and "cannot fail" beside two
`@throws`. It has no composition failure mode — no roster read, no mount, no
file — but it does reject a caller error, and the wording now says which.
The package-level "switched preset" test re-linked to the same preset id,
so it could not tell reading the parent's live scope chain from reading its
creation header. A second fixture preset makes the switch real.
The Web browser lane's subagent goldens gain the preset badge a child now
shows, which is the visible consequence of recording its composition. That
lane runs only under DSH_EXAMPLE_MODE=lib and was missed before.
The Agent Note records two limits found in review: a cold-resumed
continuable child joins its parent's current composition rather than the one
its header names, and `toolFilter` does not constrain a joined child. The
latter is a regression from the agent-plane move rather than anything this
change introduces — with the same tools in the global layer the filter
applies normally — and is tracked in #2185.
Refs #2185
Tool and prompt-section visibility is inherited along dsh-scope's parent
chain, and an agent's scope key is minted with no parent. Per-session agent
presets moved every model-facing row onto the agent plane and made
AgentPresets.mount() the one thing that binds that link, from the api-proxy's
session create, resume, and fork paths. The two in-process subagent drivers
installed only the per-child persona and tool filter, so a child's scope chain
had length one and its registry view resolved the global layer alone — which
is empty wherever a preset roster is composed. One-shot children reached the
model with no tools, continuable ones with only the host-plane `report`, and
neither carried its parent's persona, workspace context, or skill catalog.
AgentPresets.composeFrom() joins one agent to the standing composition another
already runs on. It is a bind, not a mount: the child gets its parent's exact
generation, so a composition edited since the parent started cannot fork it
onto another one, and it is synchronous, which is what lets a child creation
window use it. applyChildComposition() now takes the parent and performs the
join first, making a child composed without it unrepresentable at the call
sites. childSessionMeta() records the joined id so a cold read rebuilds the
composition the child actually ran under.
The audit that followed found two api-proxy readers on the wrong authority:
presenterScopeFor() and the live-agent branch of assertPresetUnchanged() both
read header.agentPreset, which goes stale the moment a blank session switches
preset. A switched session's cold transcript resolved presenters in the older
composition's layer and silently degraded to generic cards, and the gateway
refused to adopt a live session under the preset it actually runs while
accepting the one it left. Both now resolve through resolveSessionPreset(),
matching the resume branch fifteen lines above. The owning architecture Agent
Note carried the stale claim that the header records what a session runs; it
is corrected to name the header/log pair and its three readers.
Fixes#2165
Address ds-review-bot on #2127:
- assistant-output: the rule has ONE implementation, the incremental
AssistantOutputFold (push/pushText/collect); finalAssistantOutput folds a
complete suffix, the SDK backend folds notification events, and the ACP
backend folds raw chunk text into the same streamed fallback.
- subagent/end.lastAssistantMessage: 'no output' is encoded once — absent,
never [], on both lifecycle shapes (observeRun now omits empty output).
- tool-subagent: a non-completed foreground result stays isError but appends
the child's preserved partial text after the stop-reason headline.
- Authored keyless snapshot scenario subagent-max-tokens-partial pins the
assembled transcript: the child's committed log carries the usage-only
empty message and the parent's tool result carries the partial answer.
- Rule-boundary sentence (message wins over later streamed text) and the
consumer half recorded in the Agent Note; comments trimmed to pointers.
