* Overhaul dashboard UX with two-pane rooms and live turn persistence
- Two-pane room board (Discord-style list + detail) with sender role
colors, inbox attention badges, and per-role left-border sectioning
- Settings page: nickname + language extracted to /#/settings; sidebar
freed up for mini stats (rooms, queue, inbox) and nav badges
- react-markdown + remark-gfm for message body (tables, code, links,
proper line breaks); unified timeline merges outputs and human messages
- Live turn elapsed counter ticks every second; paused state preserved
across service restarts via new progress_text column on paired_turns
(migration 015) so in-flight context survives reboots
- Batched /api/rooms-timeline endpoint with server-side stale-while-
revalidate cache + gzip; cold ~3s -> warm ~10ms
- Optimistic UI for outgoing messages with rollback on failure;
nickname propagated to sender_name on send
- Density sweep: tighter paddings, radii, font sizes; tokens align
across panel/card/sidebar/timeline
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Strip redundant LIVE/intent/state labels from turn header
The left dot already signals LIVE and the role chip already shows the
turn role; the standalone LIVE label, intentKind em, and state pill
were repeating that information. Keep only role chip, time, elapsed,
and a "중단됨" label when paused.
Also picks up Prettier-formatted long lines in i18n and backend files
left over from the previous commit.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Convert sidebar to lucide-react icon rail and fix optimistic UI
- Replace 168px text-label sidebar with 56px flush-left icon rail
(lucide-react icons: Gauge/Inbox/Activity/MessageSquare/Clock/Settings)
Discord/Slack pattern: rail against left edge, no centered shell margin
- Optimistic UI now survives polling: pending messages tracked in a
separate state and merged into the timeline render; an effect prunes
pending entries once the server snapshot confirms them by content match
- LIVE elapsed counter shows seconds at all granularities (formatLiveElapsed
uses m+s/h+m+s instead of dropping seconds after 60s)
- Refresh button shows spin animation while busy
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Make Rooms the default landing view
Rooms is the primary surface (chat-style live agent flow), so reorder
nav (Rooms first) and switch DEFAULT_VIEW from inbox to rooms. The
topbar fallback label also changes from 'usage' to 'rooms'.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Submit room message on Enter, allow Shift+Enter for newline
IME composition (event.nativeEvent.isComposing) is excluded so that
mid-Hangul Enter does not trigger send.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Viewport-lock rooms view to remove double scroll
The rooms view now fills the viewport (height: 100vh - topbar) with
list and detail each scrolling in their own column. The page itself
no longer scrolls when the detail thread grows long, so the wheel
scrolls whichever column the cursor is over.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Flatten rooms view, single-line compose with icon send button
Drop nested card chrome on rooms view: outer panel now has no padding,
border, or background so the rail / list / detail sit flush against
the page. Internal panes (rooms-list, rooms-detail) lose their own
borders and rely on a single divider line for separation.
Compose textarea is single-row by default (rows=1) and grows up to
~5 lines; the send button is icon-only (Send from lucide-react),
fixed 36x36, pinned to the bottom of the detail pane. The redundant
'메시지' h4 and the 세부 fold are removed — the room title in the
header already establishes context.
Detail pane is now a flex column where only the timeline scrolls;
the head sticks at top and the compose pins at bottom (messenger
pattern), so the input is always reachable.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Fix rooms layout: viewport lock, restore panel tone, ensure thread scrolls
- Lock body to viewport height (overflow:hidden) only when rooms view is
active, via body:has(.view-stack.view-rooms). Section nav becomes
static so the layout fits exactly in the viewport instead of being
clipped under the sticky header.
- Restore subtle panel background and border on .rooms-list and
.rooms-detail so chat content has tonal contrast against the page bg
(pure black was unreadable). Outer panel chrome stays flat.
- Make the thread section the only flex:1 child of the pinned card so
it actually scrolls when content overflows; everything else (head,
meta, current turn, watcher fold, compose) stays flex:none.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Make rooms-detail itself scroll with sticky head and compose
Drop the brittle flex-column-with-1fr-child setup; just let .rooms-detail
be overflow-y:auto and pin .room-card-head to top and
.room-compose-section to bottom via position:sticky. The head and
compose stay visible while the thread between them scrolls naturally.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Match compose-section background to panel
The sticky compose bar was using var(--bg) (near-black) which jarred
against the surrounding panel tone. Use var(--panel) so it sits
continuously with the room detail.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Fix compose full-width and hide internal IPC payloads
- The compose section's full-width grid was being clobbered by the
later .room-section { display: grid } rule (same specificity, source
order won). Bump specificity to .room-section.room-compose-section
with display: block so the inner form/textarea fills the panel.
- Filter out internal protocol payloads from the timeline:
isInternalProtocolPayload() detects {"author":..."recipient":...}
shaped agent IPC messages and excludes them from humanMessages and
outputEntries. These are routing metadata, not human-readable chat.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Tighten rooms width: explicit 100% on detail/card, responsive list
- Make rooms-list responsive (clamp(180px, 16vw, 240px)) so on smaller
screens the chat area gets more room
- Add explicit width: 100% to rooms-twopane, rooms-detail, and the
pinned room-card-v2 so nothing collapses to content-size
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Remove timeline left indent and broaden IPC payload filter
- Drop the 3px colored left border + 12px left padding on each
timeline item; messages were starting 29px from the panel edge
(3 border + 12 item padding + 14 thread padding). Role colors are
already conveyed by the role chip, so the border is redundant.
- Tighten thread-section padding from 14px to 12px to match compose.
- Broaden isInternalProtocolPayload to also catch <sub-agent-prompt>,
<tool-call>, <internal> wrapped payloads — these were leaking into
the visible thread when the agent emitted them as message content.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Restore role color borders and tighten left padding
- Bring back the role-colored left border on timeline items (2px now,
was 3px) so owner/reviewer/arbiter/human are visually distinct
- Tighten left-side spacing across the chat panel: thread section
padding-left 12 -> 4, item padding-left 12 -> 8, card head
padding-left 14 -> 8, compose-section padding-left 12 -> 4. Content
now starts ~10px from the detail panel's inner edge instead of ~30px
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Unclamp output text in unified timeline + add measure debug toggle
- The legacy .room-output-item p / .room-message-item p / .room-latest-output p
rules used -webkit-line-clamp:2/4 to truncate previews. The unified
timeline doesn't use those classes so it shouldn't be clamped, but
add a defensive override on .room-timeline-body .parsed-body that
unclamps and clears max-height in case any sibling rule leaks in.
- Add ?debug=1 (outline panels) and ?measure=1 (overlay container x/w
in viewport coords) toggles via App.tsx so we can diagnose layout
issues without npm-installing devtools.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Tighten rooms layout left/right margins
Drop view-panel and section-nav left margin from 12px to 8px. Removes
the visible empty band between icon-rail and rooms-list (was 12px,
now 8px).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Remove internal left padding in rooms-detail panel
The detail panel had 4-8px of left padding inside the thread and
compose sections, leaving an empty band on the left. Drop padding-left
to 0 on .room-thread-section and .room-compose-section. Timeline items
keep their own 8px padding-left (for the role border + text gap), and
the textarea inherits the same flush-left starting position as
messages — they align vertically now.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Fix 26px left empty band: zero out .room-expanded padding in pinned mode
Root cause located via pixel scan + DOM dump: the legacy
.room-expanded wrapper (used by the click-to-expand layout) has
padding: 14px 0 4px 26px to align expanded content with the head's
pulse-dot column. In the pinned twopane layout it was leaving a
26px empty band between the panel border and the chat content.
Override .room-expanded inside .room-card-v2.is-pinned to zero
padding/margin/border-top/gap, so thread-section and compose-section
align flush at the panel's inner left edge.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Add Playwright + dash-inspect helper script
Playwright (chromium) is installed so future layout debugging can use
real DOM/computed-style queries instead of pixel scans on screenshots.
Helper commands:
bun scripts/dash-inspect.ts shot [route]
bun scripts/dash-inspect.ts measure '<selector>'
bun scripts/dash-inspect.ts eval '<expr>'
bun scripts/dash-inspect.ts dom '<selector>'
bun scripts/dash-inspect.ts click '<selector>'
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Auto-scroll chat to bottom on send and on new messages
- When the user sends a message, scroll the rooms-detail panel to the
very bottom over 4 frames (covers the post-render layout settle).
- Add a near-bottom tracker: a ref flips false when the user scrolls
more than 80px from the bottom, true otherwise. The auto-follow
effect (triggered when messages/outputs/pending counts change) only
scrolls if the user was already near the bottom — so reading scrollback
doesn't get yanked.
- Initial room selection and any new content added while at-bottom
always shows the latest message.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Add Settings page: model role config + multi-account management
Backend:
- src/settings-store.ts: filesystem reader for Claude (~/.claude,
~/.claude-accounts/N) and Codex (~/.codex, ~/.codex-accounts/N)
account directories. Returns safe metadata only (expiresAt, scopes,
subscriptionType, planType, accountId from JWT) — never raw tokens.
Atomic .env writer for OWNER/REVIEWER/ARBITER MODEL/EFFORT pairs.
Account directory deletion (refuses index 0). Claude account add via
pasted OAuth JWT (decodes exp/sub from payload, writes new
~/.claude-accounts/N/.credentials.json).
- src/web-dashboard-server.ts: new endpoints
GET /api/settings/accounts
GET /api/settings/models
PUT /api/settings/models (PATCH alias)
POST /api/settings/accounts/claude (body: { token })
DELETE /api/settings/accounts/{provider}/{index}
Frontend:
- apps/dashboard/src/api.ts: typed clients for the new endpoints.
- apps/dashboard/src/App.tsx: SettingsPanel split into 일반 / 모델 / 계정.
Model section: editable owner/reviewer/arbiter model+effort fields
with dirty tracking and a save button. Account section: per-provider
list with delete buttons (default index 0 protected) and a token-paste
field for adding a Claude account. Both sections expose a "스택 재시작"
button that calls the existing /api/services/stack/actions endpoint
with a confirmation prompt — changes only take effect after restart.
UX scope: phase B (model edit + account delete) + phase D (account add
via paste) per the recommendation; full OAuth flow not included.
Restart is gated behind a confirm() since it kills running agents.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Format: Prettier follow-up on settings PR
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Show cron / external-trigger messages inline (separate from CI watcher)
trusted_external_bot messages (Sentry Cron, GitHub webhook triggers,
etc.) were lumped into the CI-watcher fold and hidden by default.
Split them: WATCHER_RE now only matches actual CI status messages
(Build/Deploy/Lint/Release/[CI]/[Watcher]/GitHub Actions on bot kind),
so the watcher fold stays focused on CI noise.
trusted_external_bot messages now get rendered inline in the unified
timeline as cronMessages (sourceKind preserved). senderRoleClass
returns 'role-cron' for senders matching cron/sentry/webhook so the
chip and left border show in muted gray to visually distinguish from
agent and human messages.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Live progress fallback: scan TASK_STATUS messages for current turn only
When paired_turns.progress_text is empty (codex runner doesn't always
populate it in time), fall back to scanning the room's messages table
for ones starting with the TASK_STATUS_MESSAGE_PREFIX (the invisible
3-char prefix used on Discord live-edit messages). Strip the prefix
and the trailing elapsed indicator (e.g. "\n\n12초"), then render in
the LIVE timeline entry.
Filters:
- timestamp must be >= turn.createdAt (don't show pre-turn history)
- sender role must match the current turn's role (don't show a stale
reviewer status when owner is now speaking)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Server fallback: derive live progressText from messages when column empty
Keep paired_turns.progress_text as the primary source (populated by
recordTurnProgress in MessageTurnController). When the column is empty
on an active turn — happens for codex turns where the controller path
isn't fully wired, or any path where ingestion lags — scan the messages
table for the most recent TASK_STATUS_MESSAGE_PREFIX-prefixed message
that belongs to the current turn (sender role match + timestamp >=
turn.createdAt) and use that as progressText.
This makes the dashboard converge with Discord without requiring the
column to be reliably written: if Discord shows a live edit, the
ingest pipeline writes it to messages, and we surface it.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Poll room timelines every 1.5s on rooms view (was every 15s)
The rooms-detail thread was reusing the 15s overview refresh tick,
which made the chat lag noticeably behind Discord. Add a dedicated
setInterval(1500ms) fetch on the rooms view (cleared when leaving),
with an in-flight guard so overlapping ticks don't pile up. Backend
already has a 2s stale-while-revalidate cache so this stays cheap.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Real-time rooms updates via Server-Sent Events
Backend: GET /api/stream emits text/event-stream with 'rooms-timeline'
events whenever the server's roomsTimelineCache rebuilds (key change).
A 1.5s tick checks the cache; if builtAt advanced, push the new JSON,
otherwise send a SSE comment heartbeat ': ping ts\n\n' to keep proxies
from idle-closing the connection. AbortSignal handler closes the
ReadableStream when the client disconnects.
Frontend: replace setInterval(1500ms) on rooms view with EventSource.
Initial fetch primes the data so first paint isn't blank, then SSE
takes over for instant updates. EventSource auto-reconnects on
disconnect; while it's flapping, a 2s polling fallback runs so data
keeps flowing. When SSE recovers (next 'rooms-timeline' event) the
polling fallback is cleared.
Latency: previously 0–1.5s polling lag; now ~tick latency (50–150ms)
once cache rebuilds.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Add breathing room between timeline header and body
Increase room-timeline-item gap from 2px to 6px so the role chip /
timestamp line has visible breathing room above the parsed body —
header → body relationship reads more clearly. Bump vertical padding
from 6px to 8px and item separator from 1px to 2px to match.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Hide the 2px processing bar in pinned twopane to free role-color borders
.room-card-v2.status-processing::before draws a 2px green bar at the
card's left edge. In pinned twopane mode it sits right on top of each
timeline item's role-colored left border (owner blue / reviewer
orange / arbiter green / human yellow), masking them. The header's
pulse dot already conveys "processing", so suppress the bar in pinned
mode and let the role colors show through.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Polish room compose: tighter pill input + override global button min-height
The 48px button min-height global rule was forcing the compose grid track to
48px, pushing the textarea content visibly down inside the pill. Adding
min-height: 0 on the compose button lets the track collapse to ~29px so the
placeholder lines up with the send button vertically.
Also slimmed paddings (form 3px vert, textarea 4px vert) and added bottom
breathing room on the compose section (8px → 14px) at user request.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
* Format styles.css to multi-line transition values
Prettier 3.8.1 prefers multi-line transition shorthand. CI was failing
format:check on these single-line transitions in the room-compose rules.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
- /stop now resets the active paired task to completed so the next
user message routes to the owner instead of the stuck reviewer
- Owner completion with sawOutput=false (e.g. interrupted by /stop)
no longer auto-triggers the reviewer — treated as interrupted
- Clear pendingProgressText when intermediate updates arrive on an
existing progress message, preventing flushPendingProgress from
overwriting latestProgressText with stale buffered content
The alreadyVisible check compared progress text with final text and
skipped deliverFinalText when they matched. This caused legitimate
final outputs to vanish when the progress message happened to
accumulate the same text through edits.
When intermediate text was displayed as a progress message and the final
result had the same content, both the progress message and a separate
final message were sent to Discord. Now detects when the progress message
already shows the final text and finalizes in-place without sending a
duplicate.
- Remove UNIFIED_MODE legacy: single service manages all 3 Discord bots
- Add OWNER_AGENT_TYPE / REVIEWER_AGENT_TYPE env vars for configurable agent selection
- Fix Discord channel routing: reviewer output goes through correct bot (discord/discord-review)
- Fix channel name assignment: explicit names prevent agentTypeFilter collision
- Remove silent output suppression system: harness-level protections replace prompt workarounds
- Add reviewer approval detection: DONE marker on first line stops ping-pong
- Include user's original message in reviewer prompt for context
- Fix round_trip_count auto-reset on new user message
- Fix session ID conflict: reviewer doesn't reuse owner's session
- Fix pending progress text flush in runner on close sentinel
- Promote buffered intermediate text to final result when result event has no text
- Remove legacy files: .env.codex.example, .env.codex-review.example, migrate-unify.cjs
- Update CLAUDE.md: server-side build deployment, unified architecture docs
Previously, intermediate text was sent as separate messages when
subagents were active, causing message floods. Now always edit
the progress message regardless of subagent state.
- New token-rotation module: stores multiple tokens from
CLAUDE_CODE_OAUTH_TOKENS env var (comma-separated)
- On rate-limit, rotates to next healthy token before falling
back to Kimi provider
- Also fixes: intermediate text routes to progress message
when no subagents active (prevents message flooding)
- Track lastIntermediateText to prevent duplicate Discord delivery
when same text arrives as both intermediate and final
- Filter task_progress descriptions >80 chars (AI summaries)
- Mark completed subagents with ✅ instead of removing
- Update claude-agent-sdk 0.2.76→0.2.81, codex 0.115→0.116
Track each subagent separately via agentId from SDK task events.
Single subagent shows detailed view with activity sub-lines,
multiple subagents show compact one-line-each format.
Add 'intermediate' phase for assistant text between tool calls.
These are sent as separate Discord messages without finalizing
or disrupting the active progress message.
Buffer intermediate assistant text in the runner. When result arrives
with matching text, discard the buffered progress instead of emitting
it. This prevents the duplicate from ever reaching Discord, eliminating
the flash-then-revert behavior.
- Show subagent description from task_started immediately (no 30s wait)
- Remove AI summary (agentProgressSummaries) — task_started is sufficient
- Revert progress heading to previous when final text matches current
- Track previousProgressText in direct-update path
- Ensure progressTicker runs during tool-activity updates for time refresh
- Last activity line uses └ instead of ├
- Summary (📋) and description shown together when both present
- Description indented as sub-item under summary
Exclude is_bot_message from implicit continuation trigger check.
Previously, bot A's response would trigger bot B's continuation
window, causing infinite back-and-forth responses.
Also: tool activity sub-lines use description over summary,
and max activities reduced to 2.
Display subagent tool descriptions as sub-lines under the main
progress text, updating in-place via Discord message edit.
Shows up to 5 recent tool activities (e.g., "Reading index.ts").
Progress updates are now buffered: each new progress flushes the
previous one. The latest pending progress is discarded when final
arrives with matching text, preventing duplicate display in Discord.
When the final result text is identical to the last progress update,
edit the progress message back to the previous progress text to avoid
showing the same content twice in Discord.
- Use system/task_progress subtype (not top-level type) for subagent events
- Extract intermediate assistant text between tool calls as progress
- Only show SDK-generated summaries for subagent progress (skip noisy per-tool updates)
- Truncate progress messages to Discord 2000 char limit
- Remove verbose debug logging from production
- Auto-suspend tasks after 3 consecutive quota/auth errors with
retry-after date parsing and Discord notification
- Add tool_progress and tool_use_summary streaming from Claude Agent SDK
- Prefix progress messages with invisible marker so bots ignore them
- Fix misleading provider label in codex service logs
