The mid golden was racing the reasoning block's stream: captureStableAria
could freeze on the pre-render gap between steering acceptance and the
assistant step, pinning a snapshot without the Think row. Wait for the
[data-variant=think] row so the golden captures the complete assistant step.
Steering bubbles render as plain user bubbles; a mid-turn steer is
recognizable by its position in the flow. The runtime SteeringMessageNode
projection and pending-steering lifecycle are unchanged.
Partially supersedes the 2026-08-04 context-source and steer marks note;
the new simplification note owns the removal rationale.
A hand-damaged preset was silent until the worst moment. An unparsable
composition listed as an ordinary selectable row and failed only at the
next session start — set as default, every new session failed. A
directory whose composition file was deleted vanished from the roster
while still occupying its id: copy answered "delete the existing preset
first" while remove answered "not found", a dead end.
Discovery now owns health: every id-shaped directory is a roster slot,
broken when its composition is missing or unloadable, checked with the
loader's own entryListSchema dialect (!!js included) so health never
rejects what the loader accepts. `broken` rides AgentPreset, the
agentPreset.list entry, and the UI row; mount/recompose/standingKeyFor
refuse broken up front with the discovery-reported reason, while
resolve/read/remove still answer. The section renders marked red cards —
unselectable, uncopyable, deletable, location kept on custom rows — and
both pickers drop broken rows entirely.
The cordis preset's persona now forbids editing the shipped install
(corrupting cordis would disable the mode itself) and points authoring
at $DSH_HOME/.agent-presets; its skill teaches preset.yml metadata, the
copy-first workflow, the one-escalation sandbox reality, and honest
verification. Exercised live: asked to edit the shipped composition the
composed agent refuses citing both rules; asked for real presets (simple
and complex) it lands them under the user root with one approved
escalation each and self-checks with the loader dialect.
Copying an existing preset — or drafting one in Creator mode — replaces
the copy-minimal skeleton hint: the creator entry exists now, so the intro
names the two real paths instead of steering everyone at one source.
The settings section gains the conversational authoring entry beside
copying: a dashed add-card (the Models page's affordance) that stages the
self-referential cordis preset and starts a new session on it, closing the
settings panel through the shell's new section owner-prop `close` — the
one shell affordance a section receives, for flows that leave settings.
The seat learns the difference between picking and staging: `select()`'s
immediate apply meets the still-current running session and drops the
stage as unservable, so the entry uses `stage()` and leaves the apply to
the list-change applier that fires when the started session becomes
current. `load()` stops regressing the display when its reply lands after
an applied stage was consumed — staged pick first, then the composition
the current session already carries, then the deployment default.
The authoring lane connects a workspace and drives the gesture to a
composed host session; the section goldens gain the entry.
The web YAML editor is gone. agentPreset.write (arbitrary composition
text) became agentPreset.copy { from, agentPreset, name? }: a host-side
whole-directory copy of ids the host resolves itself — symlinks
dereferenced, modes re-tightened to owner-only with owner-execute kept,
metadata rewritten to keep the source's description but never its name or
roster order. No composition text or path crosses the wire in either
authoring direction, and the entryListSchema/!!js concern dissolves with
assertComposition itself.
The settings section becomes: a read-only viewer over shipped
compositions, a copy dialog (id + optional display name) as the only
create entry, delete for custom rows, and a location action leading into
the preset's own files — agentPreset.openDocument { agentPreset } resolves
the directory host-side and opens it natively, or answers
{ opened: false, path } for the row to show as text where the deployment
has no desktop. agentPreset.list reports hasDocument beside authorable;
the gateway's nativeOpen config pins the capability where
canOpenNativePath platform detection would mislead. The privileged set is
now read/copy/openDocument/remove.
With files as the only composition editor, standing mounts grew
stamp-keyed generations: ensureStanding compares the composition file's
mtime+size and starts the next generation for later sessions, while every
joined session keeps the generation it runs on.
New keyless web lane (agent-preset-authoring, overlay pins
nativeOpen: false so goldens render one branch on every platform) drives
view/copy/reveal/delete end to end; the real-composition CLI e2e switches
to copy semantics.
20 files, each +2 lines: the label icon and 标准模式 in the banner — the DOM
the label fix makes appear for sessions created in-tab, which these goldens
recorded the absence of.
The composer's single primary action now stops a running continuable
child: Session.cancel() routes a continuable address through
subagent.interrupt (ordinary sessions keep session.cancel; one-shot
addresses stay uncancellable), and InputBar switches the same Send button
to Stop while a continuable child runs. A running parent-offline child
keeps the default composer with disabled input so that Stop stays
reachable; the read-only takeover returns once it stops.
Refs #1535
Authoring meets standing mounts: write() and remove() drop the standing
pointer so the NEXT session composes the edited roster, while every session
already joined keeps the generation it runs on — a superseded generation is
never disposed while the process lives. The settings-dialog golden re-records
with this layer's Agent Preset nav entry, which the incoming layer-3 record
had overwritten.
The composer's preset seat stays removed here — this layer moved it to the
hero chip and the session-header action — so only the model seat takes the
`modelSeatLocked` narrowing master introduced.
Conflicts:
apps/web/tests/snapshots/*/*.expected.md
packages/client/connection/README.md
packages/client/connection/README.zh.md
packages/client/ui-conversation/src/client/skeleton/InputBar.tsx
packages/host/apiproxy/src/fetch/client.ts
The Client API carrier's `agentPresets` member was the one member of its class
without an `IApiClient[...]` annotation. Inferring it inlined `AgentPresetEntry`
into the emitted declaration by the specifier TS picks — the host `index.ts` —
dragging the whole gateway, and with it the host `Context` merges, into every
Client program importing the carrier. Annotated like its siblings.
`ApiRemoteAgentOptions.setup` now takes the inspected session rather than its
header alone: this layer resolves a resumed session's preset from the LOG,
because a session that switched while blank ran its turns under the newer
composition and the header is written once at creation.
Conflicts:
apps/web/tests/snapshots/*/*.expected.md
packages/client/ui-conversation/src/client/skeleton/InputBar.tsx
packages/host/apiproxy/src/api-proxy.ts
scripts/doc-budgets.manifest.json
Invocation is an ordinary session.prompt again: the pre-step gesture
boundary makes it deterministic host-side for every front end, so the
dedicated RPC (handler, wire schema, error codes, client face, fixtures)
and ui-skill's claim machinery are net deletions. The menu keeps decision
21 exactly — a pick lands literal /name text — plus the user-only marker
from skill.list's modelInvocable flag.
The policy scenario now expects the user-only quadrant in the menu with
its marker (riding the description — the hint field is claim-state ghost
text, which the menu never renders), and a new skill-user-invoke scenario
drives /name args through the composer against the real host: the claim
lands skill.invoke, the transcript shows the dedicated card with the
collapsed <skill_content> body, and a paced replay answers the injected
turn deterministically.
A model entry's reasoningEfforts dict declares its selectable thinking
levels — key = offered level, value = the wire spelling dispatch sends;
only off may leave the value empty (supported, send nothing). false
strips reasoning from a catalog model; every level is materialized
explicitly into pi-ai's thinkingLevelMap so nobody has to know pi-ai's
asymmetric absent-key defaulting. compat.thinkingFormat and
compat.supportsReasoningEffort become configurable on the route and per
model (model > route > catalog entry > pi-ai's URL-derived guess),
openai-completions only, so a private gateway speaking the DeepSeek
reasoning dialect no longer depends on its URL being recognizable.
Record-typed drift gates pin both enums to pi-ai's, and an unserviceable
declaration is refused at the write that produced it, naming route,
model, and level.
Master added this scenario while this branch was open, so its golden froze
the composer seat's "Select model" fallback. The scaffold's route-only
adapter (added here for fixture-less scenarios) makes the seat resolve the
model those scenarios actually route to, which is what the other eight
goldens on this branch already show. Only the two seat lines move.
**Reasoning effort leaves the provider cards entirely.** It is a per-MODEL
capability and the models under one provider disagree about which levels
they accept: setting `anthropic` to `max` made six of its eight models
throw UNSUPPORTED_REASONING_EFFORT, and because the catalog build catches
per provider, the whole provider vanished from the picker behind one error
row. A provider-scoped control can only ever be set to a value some of its
models reject.
The composer's model picker already offers each model its own levels, and
a switch there now records provider, model, and effort together as the
next session's default — so the setting has a better home at the right
granularity. The profile field stays in `settings.yaml` for a deployment
that knows its route; only the control is gone, from both cards and both
adapter families. Two `components.spec` cases used the control as the
vehicle for their op assertions and now use `baseURL`, which is what they
were actually testing.
**A rejected Provider ID now reads as a fault.** It shared the neutral
hint paragraph with the field's guidance, so the copy telling the user
what they got wrong looked like advice. Reuses the existing `.error`
style, matching the split the key field already makes.
The lower layer names the preset in the session header; this one moves the
choice to the new-session chip and leaves the header a read-only label, so the
merged goldens carried a label these screens no longer show.
Every changed golden gains only the preset label and the settings row's copy —
78 insertions, no deletions — which is this layer's surface appearing where it
was always going to appear.
Turtle
creatixchu
imccyu
_Kerman
Yichen Jiang
Tianyi Cui
Hypatia May