A message's attachments persist as metadata (messaging.Message.Media) and
cross the gateway wire, but the dashboard transcript rendered text only, so
a photo was invisible to the operator while the model had seen it.
The transcript view now carries the persisted metadata, and the browser
labels each attachment in the placeholder vocabulary the channels already
write into a message's text ('[photo]', '[document: report.pdf]'), with the
surviving metadata (mime, name, size, dimensions, duration) in the chip's
tooltip.
Bytes are never rendered: they are turn-scoped and stripped before a record
is stored, so there is nothing to fetch and no broken image or 404 link. The
platform file id is deliberately not served either -- it is a download
handle, not an address -- so the chip links out only when the attachment
itself carried a URL. A message with no attachments omits the media key, so
transcripts written before this change keep the payload they had.
Neutral warm greys with one swappable sage accent replace the violet canvas,
washes and brand gradient. IBM Plex Sans/Mono replace Geist. The light theme
keeps the same rules with a darker sage. No layout changes.
The bright violet passed the contrast ratio on paper and failed in the
hand: black-on-purple vibrated at button and FAB sizes. The primary surface
is now deep enough for white labels, and link text on the canvas uses a
dedicated lighter tint, since one violet cannot be both a dark-enough
surface and a light-enough text colour. Also drops an empty-state
description that restated its title.
The transcript scrolled itself with a watcher that set
`scrollTop = scrollHeight` on every new message and streamed token, reaching
into ScrollArea's internals through `[data-slot="scroll-area-viewport"]` to find
the element that actually scrolls.
That follows the tail and nothing else. It has no notion of the reader having
scrolled away, so every streamed token yanked them back, and no way down for
someone who had deliberately scrolled up.
shadcn-vue's MessageScroller is the component for exactly this, so the surface
is composed from it instead: the provider owns autoscroll, the viewport reports
its edges, the content is a `role="log"` region, and the item carries the scroll
anchor. It buys the part the watcher could not express: once the reader
scrolls, autoscroll stops rather than fighting them, and a "Scroll to end"
button appears.
Verified in the browser: at the tail the root reports `data-scrollable="start"`
and the button is inert; after a wheel scroll it reports `data-scrollable="end"`,
autoscroll disengages, and the button is visible at opacity 1 and reachable. A
programmatic scrollTop does not count as intent, which is right: autoscroll
reclaims the tail unless a person asked for something else.
Every page from the deleted Preact dashboard, rebuilt in Vue 3 + shadcn-vue.
The navigation restructure landed earlier on this branch; this is the surfaces
behind it, and the two are now one app.
Dashboard, Tasks (list and detail), Configuration split six ways, Logs,
Inspector, Mappings, Bindings, Workflows, Skills, Curators, Channels, chat.
Components, not files: docs/prds/dashboard-port-brief.md requires one
component, one job, one file, with shared state in a module beside it. So a
page file composes and holds nothing, and a shape repeated across panes becomes
its own component (LogRow, PageHeader, PanelLoading, WorkflowBar).
Behaviour that had to survive, because each was a fix:
- Live activity's detail is clamped to one line with the payload on the title.
Unclamped, one row made the dashboard 7,481px tall.
- Each health card takes its own height. Equalising left a mostly-empty box
taking a quarter of the first screen once setup was done.
- Tasks: `tab`/`attempt` live in the query string and are written with
router.replace, so a tab click adds no history entry and does not remount.
The list's deep-link reveal honours prefers-reduced-motion.
- LOG_LEVELS is a wire contract (the server matches that CSV server-side), used
unchanged by both log panes.
- The chat panel's Phase 5 list: a FAB that is the only toggle, a morph whose
scale, radius and position interpolate together, no scrim, no close button,
Escape returning focus to the launcher, Stop only while streaming.
- Unconfigured is not broken. A 501 on /api/version or the log stream renders
what is true ("Not configured on this deployment") rather than blank, and a
read-only Configuration says so instead of offering controls that cannot work.
Three things the port decided rather than carried:
- Button gains an `attention` variant and Badge the five status tones, in the
project's own cva. The pill's dot did not come across: the Badge's shape is
the affordance, and the same tone now covers a trend indicator.
- The icon module is not ported. Its name-to-glyph lookup is the pattern
shadcn-vue's rules call out as wrong, so icons come from @lucide/vue at the
call site. The five bespoke glyphs have no consumer yet and become components
if a page draws them.
- The theme control left the topbar for /system/appearance, per the approved
navigation design. ThemeToggle.vue and Placeholder.vue are deleted rather
than orphaned: every route now resolves to a real page.
Also fixes the bundle freshness check failing on the first build after every
change. ui/dist is committed, so nothing ignores it and Tailwind's content
detection scanned it, making each build an input to the next.
The Preact tree's ui/src/base/ was deleted with the dashboard. This rebuilds
what depended on it, on the shadcn layer rather than as new bespoke markup.
Badge and Card come from the shadcn-vue CLI, so they match the already-vendored
components, and the composed pieces sit on top of them:
- Badge gains the five status variants the pills used, over the palette's status
tokens. The pill's dot did not come across: the Badge's own shape is the
affordance, and the same variant now covers a trend indicator, where a dot
reads as noise.
- base/StatTile.vue uses full Card composition instead of dumping everything
into one content slot, with base/Sparkline.vue and the goodDirection logic
that says whether "up" is good news.
- base/Gauge.vue and base/SegmentBar.vue, both of them: the original exported
two components from one file and budget composition uses the second.
- base/LogRow.vue over lib/log.ts, which holds the LOG_LEVELS wire contract
(the server matches that CSV server-side, so it is not a display choice).
- lib/status.ts, so status colour has one home: a task state, a gate result, a
log level and a trend all draw from the same five kinds.
Two deliberate departures:
The icon module is not ported. Its name-to-glyph string lookup is the pattern
shadcn-vue's rules call out as wrong, and nothing consumes the five bespoke
glyphs anyway: the navigation is text with tooltips. Icons come from @lucide/vue
at the call site, and the five become components if a page draws them.
The gauge's aria-label announces the rounded percentage. It announced the raw
one, so a screen reader read 71.38461538461539% where the screen showed 71%
(bead 86au).
Also lands the four modules the checklist marks copy-verbatim, which encode
bugs fixed just before the delete: dashboard/activity-detail, chat/
command-scroll, workflows/workflow-rows and logs/logs-empty.
Sam owns this codebase and reads Vue fluently, not Preact; a UI its maintainer
cannot read is the liability that outweighs the port. shadcn-vue is Vue-only,
so adopting it means the framework moves with it.
Scaffolded per the official Vite guide: `npm create vite --template vue-ts`,
Tailwind v4 via @tailwindcss/vite, `shadcn-vue init` (Reka UI, Lucide, neutral
base, CSS variables). Button and DropdownMenu are added, the latter being what
the pending navigation restructure needs.
Two deviations from the guide, both forced:
- The guide sets `baseUrl` alongside `paths`. TypeScript 6 deprecates
`baseUrl` and errors on it, and this project is heading for TS 7 native, so
`paths` stands alone and resolves relative to the config file. Not a
downgrade: the scaffold's typescript ~6.0.2 is untouched.
- vite.config.ts keeps `base: "./"` and unhashed asset names. ui/embed.go
embeds a committed dist/, served from a path the build does not know;
hashed names would also orphan a file in git on every rebuild.
The Preact app and its 43 test files are deleted rather than ported, on Sam's
instruction: tests return when there is behaviour worth pinning rather than a
port to re-assert. `task test:ui` is now a vue-tsc typecheck, which is what
would actually catch a broken component; an empty `npm test` would pass
vacuously.
ui/src/router/index.ts carries the route table and is the new source of truth
for TestDashboardPagesRegistryCoversEveryRoute, which parsed main.jsx before.
Keeping that contract alive matters: it is what caught the stale /chat entry
earlier today. Every route resolves to a named placeholder view, so an unported
page is visibly unported rather than blank.
golangci-lint now excludes ui/node_modules. The dashboard toolchain pulls in
`flatted`, which ships a Go implementation beside its JavaScript, and the gate
was failing on complexity in a vendored third-party file.
The shadcn MCP server is registered in the root .mcp.json. `shadcn-vue mcp init`
writes it to the directory it runs in, which for a nested ui/ is a path Claude
Code never reads.
Remediation from testing the launcher (archie-core-ifay), plus the morph Sam
asked for.
The panel grows out of the launcher rather than fading in beside it: closed, it
is collapsed to a pill the size of the button and translated over it; open, it
is the full panel. Scale, border-radius and position interpolate together, and
the contents fade in once the shape has mostly arrived, so the button reads as
becoming the panel. The launcher stays visible throughout and is the only
control that opens and closes it, so the panel carries no close button.
- Header shows the open conversation's name; ChatApp owns the header now,
because the title and the actions that change it belong to the same
component.
- New chat became the Lucide message-circle-plus glyph and moved into the
header. Added to the icon set, which is the same 24-grid stroke system.
- The commands accordion is gone. The composer's slash palette was already
the command surface and floats above the input; having both was two
answers to one question.
- Arrow keys scroll the palette so the selection stays visible. The decision
is pure arithmetic in command-scroll.js, including the case of an option
taller than the list, which aligns to its top so the command name is not
the part scrolled out of sight.
- The settings popover closes on any pointerdown or focus outside it.
- Stop renders only while a reply is streaming.
- The composer keeps focus and stays writeable during a send, so the next
message can be typed while one streams. It was `disabled`, which also made
the focus() in the send's finally block a silent no-op. sendMessage already
refuses a second concurrent turn via isSendingRef.
- Composer actions sit at the right of the input row, the resize handle is
gone, and the Enter/Shift+Enter hint is removed.
- Message bubbles dropped their "Archie"/"You" labels; side and colour
already say who spoke.
- Fixed an empty band under the composer: the panel still declared two grid
rows from when the launcher owned the header.
Four P0 defects from the frontend audit, plus the one-line theme fix that was
making every one of them look worse.
Chat is a floating launcher and the panel it opens, not a route
(archie-core-yuvg, archie-core-hesx). /chat rendered an empty page with a
drawer over it, and that drawer hosted the full-page chat workspace: a page
title, a conversations column, a four-control action row and a command-help
column, all inside 440px. Measured at 1440x900, .chat-workspace was 174px wide
holding 270px of content and the composer ran off the viewport.
- The Chat nav entry, the /chat route and the topbar chat icon are gone;
the launcher is the single way in.
- The panel anchors to the launcher and opens leftward from it, sharing its
bottom edge, so the button never moves. No scrim and no scroll lock: the
page stays readable and clickable, which is the point of asking Archie
about the page you are on.
- Conversations became a switcher in a compact bar; personality, provider
and model moved into a disclosure beside it; the empty state and composer
dropped to panel type scale.
- The launcher renders outside .shell because .shell's backdrop-filter is a
containing block for fixed positioning, which otherwise pinned the button
to the scrolling shell instead of the viewport.
- Update-check, Dangerous actions and Reload Telegram left the panel; their
api.* calls are untouched for the rehome to Configuration
(archie-core-tf20).
The topbar no longer clips its utility cluster (archie-core-mkb3). brand, nav
and topbar-end were all flex:none with no wrap above 1220px, and .app sets
overflow:hidden, so between 1221px and ~1582px the avatar, theme toggle and
documentation button were cut off with no way to reach them. The nav is now
the one region allowed to shrink, the search collapses to its icon below
1500px, and the nav wraps to its own row below 1270px where it stops fitting.
Live activity clamps its detail column (archie-core-qo7f). An agent_finish
payload rendered raw turned one row into six viewports and the dashboard into
7,481px. The cell shows the first meaningful line; the full text stays on the
title.
color-scheme is bound to the active theme (archie-core-nm0y), so scrollbars
and native controls stop rendering light inside the dark UI.
* fix(ui): make the served lifecycle vocabulary actually load
base/task-meta.jsx documents the backend as the source of truth: /api/task-meta
returns the statuses (label, pill severity, "needs you") and the operator
actions, and loadTaskMeta() replaces the freeze-dried defaults with that
catalog. Nothing called it. Every pill, status filter and action control read
the snapshot, so a catalogue change on the server never reached the browser --
the module's first paragraph was simply not true.
main.jsx now asks for the catalog once at boot, alongside the capabilities read
it already does for the same reason: the shell paints immediately and upgrades a
moment later.
The two vocabularies the dashboard mirrors rather than serves -- FILE_STATUS
(ids owned by internal/domain/workflow/task/changes.go) and CONFIG_SCHEMA
(events.ConfigCapturedSchema) -- are now documented as the deliberate second
tier, with their two-sided pins named, so the tree stops telling two stories
about where a vocabulary comes from. Both degrade to a raw id or the verbatim
payload rather than rendering something wrong, which is why mirroring is
tolerable there and not for a presentation catalog.
RED before (loader never called):
x booting the dashboard replaces the freeze-dried vocabulary with the served catalog
actual: 'custom_status', expected: 'Custom status'
Also repairs a drifted LAW asset. A fresh build of pristine main dirties both
ui/dist/assets/index.css and index.js: 4c484887 (#889, memory unification slice
5) changed ui/src without rebuilding ui/dist, and task check runs task ui --
which repairs dist instead of verifying it -- so a stale dist is invisible to
the gate. Rebuilding here restores the invariant; a dist-freshness gate is filed
as follow-up.
* fix(ui,gateway): restore the search focus the port dropped, and the page registry truth
An adversarial review of the port (5787287e) found three things worth acting on.
1. Topbar search -- a regression the port introduced. The el() version focused
the input when the narrow-layout toggle opened it, and blurred it after a
successful jump; the port kept neither. That layout hides the field with
width:0/pointer-events:none until .is-open or :focus-within, so without
focus() the caret was not in the field the operator had just opened, and
without blur() :focus-within left it expanded across the topbar after the
jump. Both calls are back, with the ref they need.
2. dashboard_tools.go advertised a page that does not exist and omitted two that
do. The registry calls itself "the single source of truth for what the
dashboard exposes", but it is a hand-copy of the route table in
ui/src/main.jsx: /memory stayed after the page was deleted (that deletion
also left ui/dist stale -- archie-core-1vhh), and /bindings and /curators
were never added. So the agent could hand the operator a chip to a route that
now falls back to the dashboard, and could not answer "where do I go for
playbook bindings?".
3. TestDashboardPagesRegistryCoversEveryRoute did not test that. It asserted
three hardcoded paths and passed while the registry listed a deleted page. It
now reads the route table back out of ui/src/main.jsx and compares in both
directions, which is what makes the registry's claim holdable. RED before:
route "/bindings" is a dashboard page the registry does not list, so the agent cannot point at it
route "/curators" is a dashboard page the registry does not list, so the agent cannot point at it
registry lists "/memory" (Memory), which is not a dashboard route: the agent would send the operator somewhere that does not exist
Also deletes the dead .chat-page rules the deleted chatPage() bridge left
behind: nothing renders .chat-page any more (chat.jsx renders
.chat-page-content, and the drawer sizes it from layout.css).
Not changed: the unreachable `soon` path, which predates the port and was
reported rather than acted on.
* fix(build): verify ui/dist is fresh instead of never looking at it
ui/dist is committed because the Go build embeds it, so it is a generated
source file. Nothing in the gate read it: task check does not run task ui, and
nothing else compares the bundle with ui/src. So when 4c484887 (#889) changed
ui/src without rebuilding, the committed bundle went on serving the Memory page
whose source had been deleted -- with the gate green throughout.
task ui:check (tools/uidist.sh) rebuilds into a temporary directory and compares,
never rewriting the committed artifact. That is the rule docs:check already
states for the contract reference -- a gate that silently rewrote the artifact
would bless an unintended change with no reviewer ever seeing it -- and it
matters more now that several agents work in their own worktrees, where each
could otherwise pass while producing a different bundle.
Evidence:
fresh tree exit 0, matches
hand-staled bundle exit 1, names the differing file
the #889 drift replayed exit 1, names BOTH index.css and index.js
(pre-#889 bundle, post-#889 sources)
git status after a run unchanged: the check never writes ui/dist
Cost: ui:check declares ui:deps like ui and test:ui do, so the gate runs npm
install once per dependent. Task's sources/generates up-to-date check does not
apply to a node_modules directory (measured: it ran on every invocation), so the
target does not claim otherwise; guarding that is follow-up work.
Gate: task check exit 0, UI suite 230/230, all Go packages ok.
* fix(ui): make the served vocabulary reach an already-painted page
Greptile's finding on the first cut of this fix, and it was right: loadTaskMeta()
resolves after a page has painted and only mutated module state. The route tree
renders into its own root, so App re-rendering cannot reach it, and a painted
page kept the freeze-dried labels it read on the way up.
show() is split so the page render can be repeated on its own: renderRoute()
mounts the current route into the outlet without the teardown/chrome side
effects, and the catalog effect calls it once the catalog lands. The key is
unchanged, so Preact re-renders the mounted page in place and keeps its state.
Deeper instance of the same defect, found while fixing that one:
task-filters.jsx captured the vocabulary in module-level constants at import
(ATTENTION_STATUSES/TASK_STATUSES), so the task filter could never see a served
status, no matter when or how correctly the catalog landed. Both are read at
call time now.
Tests, and what they actually cover:
- the task-filter test has teeth: with the source fix stashed it fails, with
the fix it passes.
- task-meta-late-catalog.test.js guards the boot fetch end to end (it fails
when loadTaskMeta is never called), but it does NOT isolate the re-render:
a page also re-renders for its own reasons, so the assertion holds without
renderRoute too. Said plainly here rather than implied as coverage -- the
fix makes the re-render deterministic instead of incidental, and this
harness cannot tell those apart.
- it lives in its own file because main.jsx boots once per process; a second
test in the same file would assert against a detached tree.
UI suite 232/232, task check exit 0.
The dashboard was a hybrid: a Preact page per route rendered into a shell
built by a hand-rolled el() helper. Every route was Preact and the shell was
not, which is why nothing failed -- the last of the vanilla layer was the
part that composed the app.
Removed:
- ui/src/base/dom.jsx -- el/append/mount, the unused pill/empty pair, and a
statusKind re-export that duplicated task-meta for one caller
- ui/src/chat/chat-tools.jsx and the el-based bubble builders in
chat-render.jsx: nothing in src/ called them, only their own tests did
- the chatPage() DOM bridge, which existed so an imperative shell could embed
a VNode
Ported:
- base/icons.jsx: icon() -> <Icon>; the glyph path table stays data
- base/gauge.jsx, base/statTile.jsx, base/log-row.jsx: DOM builders ->
components, with inline SVG instead of innerHTML
- base/format.jsx: ago/compact, which never were DOM
- chat/markdown.jsx: renderMarkdown() -> <ChatMarkdown>, so a streamed reply
is diffed rather than rebuilt block by block
- main.jsx: shell, nav, jump-to search, theme and chat drawer are one App;
pages still render into one long-lived outlet
Routing stays wired at import rather than from an effect. A page installs its
own hash-sensitive state as it mounts, and task detail writes its tab
selection back with replaceState as it settles: if the outgoing page is still
mounted when the hashchange lands it canonicalises the URL back to its own
route and the navigation is silently lost. The outlet and the two
navigation-owned pieces of chrome are published through a callback ref, which
runs at commit time, before any page effect.
Tests: dom.test.js (el) and chat-tools.test.js are gone. The statTile cases
in dom.test.js were not dropped -- they pinned the zero-value headline bug
that made fresh installs look broken, and moved to test/statTile.test.js
against the component. log-row and chat-markdown render their components.
ui/dist rebuilt: index.js 146977 -> 146010 bytes.
Closes archie-core-v9gv.2 and archie-core-v9gv.
Five files survived the Preact chat cutover with no production consumer:
- ui/src/chat/chat-view.jsx, chat-catalog.jsx, chat-panels.jsx: zero importers
anywhere in ui/src or ui/test.
- ui/src/chat/chat.html: referenced only by chat-view.jsx, itself orphaned.
- ui/src/chat/chat-stream.jsx: imported only by ui/test/chat-markdown.test.js.
Its updateStreamingReply helper swaps a rendered Markdown node into the DOM --
the imperative hyperscript pattern the Preact port replaced.
Corroborating evidence that these were already outside the build graph: a full
"task ui" rebuild after deleting them produced NO change to the committed
ui/dist bundle, so tree-shaking had already excluded all five.
ui/test/chat-markdown.test.js loses only the case that exercised the deleted
module. Its two renderMarkdown cases stay, because renderMarkdown is live --
ui/src/chat/chat-render.jsx and chat.jsx both call it.
Not touched: chat.jsx, chat-render.jsx, chat-retry.jsx, chat-state.jsx,
chat-tools.jsx, chat-update-status.jsx, markdown.jsx -- each still has at least
one importer, so none of them is an orphan.
Three dashboard mutations were refused 415 on every use. archied's
authorizeTaskMutation runs before the handler body and answers
"Content-Type must be application/json" unless the request declares a JSON
body -- including when there is no body to send, because that requirement is
what makes a simple cross-origin form post inexpressible. bindingApprove,
mappingDelete and bindingDelete sent Accept and X-Archie-CSRF but no
Content-Type, so approving a binding or deleting a mapping or binding never
worked from the dashboard.
Reproduced against the real handler with the browser's exact headers: 415,
where the same request plus Content-Type returns 200 and arms the binding.
The Go tests missed it because doJSONWithHeaders sets Content-Type
unconditionally, even for a nil body, so no test ever exercised the shape the
client actually sends.
Every api.* method now goes through one request() (and one send()), which
sets Content-Type and the CSRF marker on every mutation rather than only
where a payload happens to exist. The module already claimed to be "the
single place that knows how to talk to archied" while 22 call sites built
their own fetch, headers, timeout and error handling inline and chat.jsx
built a 23rd; that claim is now true, and chat.jsx's stream goes through
api.chatStream so only the URL, headers and failure shape are shared while
chat.jsx keeps its abort/timeout and SSE parsing.
Behaviour deliberately preserved: 204 deletes still parse no body, the
install timeout stays 120s, and 401 still classifies as session-expired. One
intentional change -- a 401 message is now the server's own text ("open the
dashboard URL archied logged at startup") rather than a bare "unauthorised",
since errorMessage already prefers the actionable reason and the classification
is derived from the status either way.
ui/test/api-client.test.js asserts the browser mutation contract for every
method, that reads carry no mutation marker, that a 204 resolves without
parsing, that api.jsx holds exactly one fetch call, and that every api method
is classified by the suite so a new one cannot be added uncovered. ui/dist is
rebuilt, being a committed asset.
I had claimed this work was needed to fix a 401 defect. It was not: all 22 raw
call sites passed res.status to ApiError, so classifyActionError already mapped
401 to session-expired, and chat.jsx never classified at all. The missing
Content-Type was the real defect.
The web chat is now a persistent slide-over drawer available on any page
(not just the /chat route), and the agent is told which dashboard page the
operator is on so it can point them to the relevant page with a clickable chip.
Backend (gateway):
- Message gains a per-message Page field; the system prompt renders it in
the <env> block (Current page), so Archie knows where the operator is.
- New dashboard_tools.go: a DashboardPages registry plus two web-only tools
— page_index (lists every page and what it shows) and dashboard_navigate
(validates a path, returns {path,label}; rejects unknown/empty paths).
- The web chat sink emits a dedicated navigate stream event when the model
calls dashboard_navigate, independent of ShowToolCalls.
Frontend (webui):
- chat sends currentPage() (location.hash) with each stream request; a
dashboard_navigate result renders as a clickable link chip (href=#/path).
- The chat is mounted once in a slide-over drawer with a topbar launcher,
scrim and Esc-to-close; /chat opens the drawer instead of a second page.
- ui/dist rebuilt (committed) with the bundled changes.
- .gitignore: ignore .impeccable/ (impeccable tool critique snapshots).
New capability (per scope discipline): designed in
docs/prds/dashboard-aware-chat-navigation.md.
Tests: dashboard_tools_test (registry, resolve/reject, web-only gating),
prompt page-line cases, webui Message.Page + navigate-frame capture, and
chat-tools chip tests. task check is green.
Arrow-key navigation in the chat composer's command menu updated
aria-selected but never scrolled the option into the visible area, so
selecting past the viewport with the keyboard left the highlighted
item hidden. Mouse scrolling already worked since it doesn't rely on
selection state.
chatStreamEvent gained a Failed bool, set from ToolCallEvent.Err !=
"" -- the field the rest of the codebase already treats as
authoritative -- rather than the browser sniffing summary.startsWith
"failed:" on rendered text. A successful tool whose own output starts
with that literal string (a grep hit, a log line read back) was being
styled as a failed step.
Telegram streaming callbacks now enqueue cancellable rendering rather than performing network I/O while draining the authoritative model stream. Final edit failures fall back to normal split delivery.
Completed tool events now correlate inputs by call ID and expose only bounded, conservative parameter shapes. Tool-call visibility is snapshotted per reply so reloads cannot change an in-flight turn.
Chat returned only the final answer, so a wrong tool choice, a
hallucinated action, or a footgun like `systemctl stop` against
`restart` was invisible from the conversation. Only text deltas were
threaded through the turn, and tool calls executed inside Generate with
nowhere to report.
Replace the delta callback with gateway.TurnStream, which carries text
fragments and completed tool calls on one ordered path from the model
loop to the channel. The sequence is the point: a tool line that
overtakes the sentence it interrupted describes a turn that never
happened. Text-only callers use gateway.DeltaFunc and are unaffected.
A tool is reported on its result, never on its call -- the call part
carries no outcome yet, so reporting both would list every tool twice,
once with nothing to say.
Telegram renders it behind [chat.telegram] show_tool_calls, off by
default: this is debugging visibility, and it narrates every internal
step to provide it. The dashboard always shows it, being the operator's
own observability surface; its tool lines are live-only and are not
persisted with the conversation.
The Telegram live reply changes mechanism, and with it what a frame
MEANS. It used to stream through sendMessageDraft: drafts are ephemeral
30-second previews, so a tool call longer than that outlived the
preview, and successive frames were observed concatenated with their
cursors intact. It now sends one real message and edits it. An update
used to be "the next preview of the answer"; it is now "the whole
answer as it currently stands, replacing what is there". The cursor is
appended at render time and never stored, so exactly one exists and
always at the end, and a frame too long for one message is cut to its
tail rather than rejected outright.
A stopped or failed turn now edits the cursor away instead of leaving
it blinking on a message nothing will finish. That cleanup runs on a
live context because turnCtx is already dead, and it runs on both
failure branches: ai-sdk raises an aborted stream as ErrAborted with
ctx.Err() formatted in rather than wrapped, so errors.Is cannot be
trusted to tell a /stop from a fault.
Closes#467
Fixes#444
.chat-page previously used min-height: calc(100vh - 5rem), which is a
floor, not a cap. Any dynamic content pushed above .chat-layout (an
error banner, extra status pills, the "Reload Telegram"/"Check
updates" row wrapping to two lines) grew total page height past the
viewport, so the whole page scrolled instead of just
.chat-transcript.
Turn .chat-page into a fixed-height flex column (height, not
min-height, plus overflow: hidden) so it can never grow past its
viewport budget. .chat-page-content, .chat-topbar, .chat-update-panel
and .chat-dangerous-panel become flex children that size to their own
content, and .chat-layout switches from a min-height floor to
flex: 1 1 auto; min-height: 0 so it absorbs whatever space is left
after the topbar/banners, mirroring the minmax(0, 1fr) pattern
.chat-workspace already used internally. .chat-sidebar gains
overflow-y: auto so a long session list scrolls in place rather than
pushing the layout taller.
refreshUpdate() rendered the fetch error's raw message text into a
visible banner on every page load, including a plain "501: Not
Implemented" when chat.telegram.update_check_command is unset -- a
normal "feature not configured" state, not a failure.
Add updateUnavailableMessage() (chat-update-status.js) to branch on
ApiError.status: 501 renders nothing (matching how Channels shows a
calm "not configured" state for an unconfigured channel), any other
error gets a fixed human message instead of the raw fetch text.
Fixes#443
Composer previously required Ctrl/Cmd+Enter to send. Plain Enter now
sends (still yielding to the command-menu's own Enter handling when
it's open); Shift+Enter is left unhandled so the textarea's default
newline behavior applies. Updated the visible hint to match.
send.onclick = sendMessage bound the DOM click handler directly, so every
click invoked sendMessage(event) -- the MouseEvent became `retry`, which
is truthy but has no .turn, and retry?.turn.text threw immediately,
before any UI state changed. The click silently did nothing.
Extracted the retry/composer-text resolution into chat-retry.js's new
resolveTurn(), which only treats an argument as a genuine retry when it
actually carries a turn.text string -- a stray event (or any other
non-conforming truthy value) now falls back to the composer text
instead of crashing. Fixed the call site itself
(send.onclick = () => sendMessage()) as the actual root cause.
Fixes#447
Brings the chat surface to the Web UI: /api/chat with its own handlers
and tests, a self-contained ui/src/chat feature folder, and the shared
plumbing to reach the same router Telegram uses.
Includes a dangerous-command approval path (internal/gateway/dangerous.go,
internal/webui/dangerous.go, and the Telegram counterpart) so /rollback
and /stop are gated the same way on both surfaces, persona handling for
web sessions, and release-update formatting split out of the Telegram
service so both can render it.
web_parity_test.go pins that the two surfaces expose the same commands.
Authored in a parallel session; committed here so it lands with the rest
rather than sitting uncommitted while other work builds on top of it.
task check passes with it in place.
Screenshots used while building are gitignored (*.png) rather than
committed, matching the call made on the earlier design reference.