The browser e2e lane had been failing wholesale since this stack moved the
agent plane into presets, and nothing caught it: 34 of 48 files. Two of the
causes are product defects, not test breakage.
`bashEnv` goes back to the host plane. `apps/cli/src/web.ts` injects it to
publish `DSH_WEB_URL`/`DSH_WEB_MODE`, so the earlier note that "nothing outside
the agent plane injects bashEnv" was simply wrong — behind a preset's `shell`
realm those variables reached no shell at all, and a `dsh web` agent could not
find the address of its own interface. This is the same criterion that returned
`subagents`: a host row that injects a service resolves before any session
exists and has no agent to key by, so the service is host-plane. `tool-bash`
consumes the host registry from inside the preset, which works because an
agent context chains to the host; only the reverse is invisible.
`tool-subagent-report` goes back with it. It is not a tool this agent calls: it
registers a continuable SETUP on the host `subagents` singleton, and that list
is not scope-aware. One copy per mounted preset meant every child was handed
`report` once per live session, so the second registration threw and a cold
subagent resume failed with `subagent-not-resumable` — a diagnostic three
layers removed from the cause.
The lane's own composition facts follow. Skill roots resolve inside a preset
now, a subtree include patches cannot reach, so the scaffold pins the roots'
documented environment fallback for its whole lifetime rather than for the boot
— presets mount per session. Without it the developer's real `~/.dsh/skills`
enters replay requests and goldens while CI sees none. The `apps/cli`
composition test pins `storage-json` for the same reason: unpinned it wrote,
and then read back, the developer's own `~/.dsh/storages/`.
Three tests now address through an agent what they used to read off the root
context, because that is where the thing lives: the tool catalog, the skill
registry, and the token meter. The seeded-history projection baseline asserts
the opposite of what it did — a detached session yields a preset-plane
projection only from a durable checkpoint written while it was live, and this
seed was written straight to persistence and never ran.
Goldens re-recorded for the hero's preset chip and the settings nav entry.
`ui-question`'s node half called `ctx.tools.register` on the host context.
`ScopedLayers.merge()` combines the global layer with the agent's exact-scope
layer, and an unscoped registration lands in the global one — so the tool
reached every agent no matter which preset composed it. `core-web`, sold as
a two-tool benchmark surface, really presented three, and a locally authored
`bash-only` preset presented two.
Rendering a question is a host UI capability; having the tool is an agent
capability, and only a preset decides that. The node half is now empty and
the `tool-ask-user` row moved into the presets that want it. The TUI keeps
its own row, having no presets.
The web composition test now asserts the global tool layer is EMPTY, which
is the invariant that would have caught this: any tool outside a preset
reaches every agent.
Review follow-ups on the presentation intent.
The card claimed any single-question request declaring the intent, then sent one
of two labels — so a producer offering a third option, or a multi-select batch,
lost answers the generic flow would have shown. That contradicts the intent's
own contract, so `planReviewOf` now claims only a binary single choice and
leaves everything else to the flow that can express it.
`ask()` also rejects a plan-review intent on a question with no `detail`: the
intent declares detail IS the plan, and without one a honouring UI asks the
user to approve something invisible. The client keeps its own fallback — it sits
downstream of a wire boundary — but the misconfiguration now fails at the asker.
`planReviewOf` stops being a value export of the client contract face (client
export discipline: pure helpers stay internal; the tests already import it
relatively), and the ui-question README fallback list, both languages, now
states every condition the code enforces.
The Web composer now renders a request that declares the `plan-review` intent
as its own surface: the waiting-approval card language — amber "Plan review"
strip, the plan as the scrolling markdown body, the question as the card's
accessible name — over one decision row of Chat about it / Refuse / Approve.
Gone from that surface are the 1/1 pager, the numbered radio rows, the custom
answer row, and Skip/Submit, which made approving a plan read as sitting an
exam.
Approve and Refuse answer with the asker's own option labels and keep its
descriptions as tooltips; Chat about it cancels the request so the composer
returns and the user can simply say what they want. Copy is bilingual under the
existing `question` namespace.
The shape choice lives inside the single composer entry rather than a second
chain registration, so the two surfaces cannot race the same carrier, and
`planReviewOf` falls back to the generic flow for any request it cannot render
as a card — the client sits downstream of a wire boundary and every request
must stay answerable.
The pending question now owns exactly two surfaces: the redesigned
QuestionComposer takeover (footer pager, checkbox multi-select,
always-visible custom input, locale-injected bilingual chrome) collects
the answers, and a dedicated ask_user_question toolview row reports the
interaction outcome — waiting, N/M answered, cancelled (ASK_CANCELLED),
or interrupted with stopped semantics (ASK_ABORTED). PendingCard narrows
to approval waits only. Toolview leading icons and the hover chevron
unify on the tertiary label color, the checklist glyph matches the
14px figma extract, and dev-watch registers CSS modules so css-only
edits rebuild.
Regenerate the cordis/config/persistence catalogs and the module graph for
the ui-plan package and the plan projection unit; add the ui-plan Chinese
README and the Model Experience indirect sentence (allowlisted: the control
dispatches /plan lines, dsh-plan-mode owns the model surface); carry the
QuestionComposer Markdown/scroll improvements into the Chinese ui-question
README; re-record the three touched translation pairs.
