Rebuild Bot Mode on the app's design system

[OutThisLife]

Summary

Bot Mode arrived as a 16,291-line plugin.js that reimplemented most of the desktop app instead of using it: its own scroll container, color palette, status dots, empty states, buttons, time formatting and cron surface, none of which could follow the theme. It also shipped a second "Bots" workspace tab that duplicated the sidebar, and a workspace-scoping filter that could take the main zone off screen entirely.

This rebuilds it on the primitives the app already ships, and widens the plugin SDK with the ones it was forced to reinvent so the next plugin doesn't have to.

• Bot Mode is now 44 focused modules (plugin.tsx is imports and register()). A bot row renders the same SessionStatusDot, swatches and age labels as the session row beside it; the UI settles on "bot" as the noun, with model-facing prompt text still saying "agent".
• It speaks every language the app does. The ~170 Bot-Mode-specific strings live in a plugin locale bundle registered through ctx.i18n, shipped in en / ja / zh / zh-hant using core's own nouns (ゲートウェイ / 网关 / 閘道, プロファイル / 配置文件 / 設定檔) so the rail doesn't read like a second translation of the app. Vocabulary core already ships in every locale — generic verbs, weekday names, Daily/Hourly — resolves against core rather than being duplicated into a second, worse translation. That includes the intro a new bot is born with, which used to greet every user in English and pull the bot's reply into English with it.
• The SDK exports what was being hand-rolled — the Panel master-detail family, SessionStatusDot, ColorSwatches + PROFILE_SWATCHES, RowButton, DisclosureCaret, formatAgo, translateNow, and the unread store behind the dot — each with a docblock naming the hand-rolled shape it replaces.
• The main zone can no longer vanish. Workspace mode goes back to being a signal consumers adapt to, never a decision about whether a pane renders, and a structural floor keeps the registered main pane's subtree alive whatever happens to its tabs. Bot chats are ordinary tabs in the main strip. The duplicate "Bots" workspace tab and its auto-open/close signal machine are gone (-1,623 lines).
• Bot chats share core's unread store rather than keeping a parallel map that could disagree with the session dot. Hidden sessions have no listed row to read a profile from, so the writers take an explicit profile hint.
• An empty bot chat is titled with the bot. It used to be blank space: core's intro splash belongs to a fresh draft and stands down for any session that exists. Rather than teach core about bots, the transcript gained a chat.empty contribution area (resolved like transcript.directives), and the splash's lettering became a Wordmark primitive both callers share — so the bot's face and name render in the same face as HERMES AGENT without forking it.
• The composer adapts instead of forking. A bot chat has no repo, so the git branch/PR row stands down for it — the same composer, one condition, no second implementation.
• A profile created without a clone gets a usable model block. Creating a bot from the desktop dialog built the profile tree but no config.yaml, so it resolved no provider and its first turn died with "No LLM provider configured" — every bot made that way was dead on arrival.

The row geometry has one owner now

chrome.tsx calls itself "shared, content-agnostic sidebar chrome" and warns that owning row height anywhere else makes rows float 1–2px off sessions — but its measurements were private consts, so anything wanting to line up copied the literals. The lead cell alone appeared verbatim in six files, four of them core's. That copy is the mechanical reason a bot row can drift from a session row however carefully it's styled.

The measurements moved to row-geometry.ts with the public names still re-exported from chrome.tsx, and the lead cell and ConnectionGlyph now reach the plugin through the SDK. Six copies became two owners — one for transcript lines, one for sidebar rows — and the plugin's GatewayKindGlyph, a fork of ConnectionGlyph down to the icon set, wraps the real one.

The harness is gone, not ported wholesale

The 84 .mjs files under tests/ read plugin.js as text and sliced it with vm/regex — the pattern AGENTS.md bans outright, and pointed at a file that no longer exists. They are replaced by real vitest renders against the real SDK, and check:test:plugins is deleted along with them. The contract tests AGENTS.md names by path are all present and green.

Three latent bugs surfaced during the split and are fixed with regression tests: groupChatSyncMemberKey keyed on a field descriptors never carry, so every member hashed to '' and the round engine saw one member where there were several; botMetaWriteAt grew an entry per write and never dropped one; aliasRouteIndex replaced the whole map on rebuild, so a slower rebuild finishing second clobbered a newer one. Two caches that grew for the window's life (decoded pet sprites over a 4,500-pet gallery, relay agent rows) are bounded by a shared LruCache, which also replaces the private one inside katex-memo.ts.

Review round

Five findings from review, each with the test that pins it:

• The adopt-on-conflict open skipped its staleness probe. It runs a full extra round-trip past the point its two siblings are probed at — registry miss, mint, title conflict, re-consult — so it was the likeliest of them to land after the user clicked another bot, and the only one that navigated unguarded. Adopting the winner still settles identity; only the workspace steal is gated.
• The bot-meta v1→v2 migration suite was dropped, not ported. migrateBotMeta and commitBotMetaV2 kept every branch — the topology gate, the marker that makes v2 authoritative, the rollback around a failed commit — while the only surviving coverage was the happy-path write ordering asserted incidentally by bot-delete. Nine cases restored, mutation-checked: deleting the marker rule fails three of them.
• The composer's bot-chat flag subscribed to one store while reading three. isBotChatSession resolves a live id to the stored one it's filed under, so it reads $sessionTiles and $sessionStates as well as the scope set. It looked correct only because $sessionStates republishes per streaming token and overwrote the stale answer within a frame. useStoresSelector now sits beside useStoreSelector for the general case.
• chat.empty mounted only the first contributor. Ownership is per session and resolves asynchronously, so first-wins let a plugin that declines a session permanently suppress the one that owns it — for whichever users' installed plugins registered in that order. Every contributor mounts and answers for itself; two claiming one session render both, a visible conflict instead of a silent drop.
• A bot row click returns to the tabs that bot already has open instead of re-opening a Bot Chat the user closed (ported from main onto the split modules, with the explicit "Open Bot Chat" menu item that re-opens it on purpose).

The "one passive re-front attempt, never a remount strobe" invariant is structurally moot here: the strobe was a close+open of the Bots home tab, and this branch deletes the home — 11 references on main, zero here. The surviving pane sync is a guarded register/unregister, not a close+open cycle.

Second review round

• The kickoff was the first line of the forever-chat, in English (the remaining half of Bot Mode: canonical Bot Chat is born with a hardcoded English kickoff prompt attributed to the user #91827). It resolves through the bundle now. Attribution — it arrives on the user's side of the conversation — is the half that can't close here: prompt.submit is the user-turn API, so an unattributed birth needs the lazy or silent path Bot Mode: canonical Bot Chat is born with a hardcoded English kickoff prompt attributed to the user #91827 proposes, a gateway contract change rather than a rename. The intro itself is deliberate and contracted in AGENTS.md.
• The residual owned strings are swept as a class, not as the cited lines. Every user-visible string the group room and the bot-scoped cron pane own is in the bundle. The frequency and weekday option lists stopped being module consts, which froze whichever locale happened to load first. What stays hardcoded is now stated in the bundle header with its reason: prompts addressed to a model (the room-picture prompt, the routine instruction), cron syntax and React keys, and 'You' — a marker that is persisted into the room log and compared as a sentinel, so localizing it needs the marker split from its rendering rather than a rename.
• The round cap is ceded, not re-buried. GROUP_CHAT_MAX_ROUNDS and its four siblings already sit in one exported block at the values plugin.js shipped, so no rebase inherits a behavior change on top of a rewrite. Making them configurable is live contributor work — #92213 for per-room limits, #96842 for config plus a token budget — and both want that same seam, so the block says so where the constants are instead of this PR adding a config hook it has no consumer for.
• The first-open git rail from the live smoke run is fixed. A bot created from the dialog opened its chat with the workspace fields omitted, because they were spread only when the caller passed a staleness probe — and the create path is the one caller without one. The composer reads that scope to stand its branch rail down, so a just-created bot showed the rail until the next click reopened the row scoped. The probe answers whether to navigate; what the session is never depended on it. Carried over from plugin.js, not introduced here. With the gate gone all four openers in that file are the same call, so they became one.

The #97008 session contracts (follow_profile_config, room_plumbing) were restored on the rebuilt modules in d22e8e402 with contract-shape tests.

Deliberate behavior changes, called out

Everything else is parity — the plugin registers the same five contribution areas, and a function-level diff against the deleted plugin.js accounts for all thirteen removals: eight are the Bots-home tab and the Active-Now strip removed on request, two folded into core's ConnectionGlyph, and three were provably dead (generatedSessionTitle had no callers, isGenericTitle only had it, groupChatSyncDeletedRevision survived only in a comment). Notification count is 44 before and after. These are the intentional differences:

• "+" in the Bots workspace no longer refuses. It used to notify and do nothing whenever the sidebar's selection had no exact route — a group chat's blocked target, an orphaned roster row. Since the main strip carries plain session tabs alongside bot chats, a dead "+" there is worse than a fallthrough, so it aims at the selected bot's profile and falls through to an ordinary session otherwise. This is the "silent UX change" from review, now not silent.
• The Bots home tab and the Active-Now strip are gone, on request — a second tab named Bots that previewed one session and opened another, and a status chip row that duplicated SessionStatusDot.
• A bot row click opens in place rather than as a new tab, which is what stopped a click producing two chats.
• The composer stands its branch rail down in a bot chat. Bots aren't working sessions and the git context was leaking into their replies.
• An empty bot chat gets the bot's own intro through the new chat.empty area, instead of the app's generic one.
• Three latent bugs are fixed rather than carried: groupChatSyncMemberKey keyed on a field descriptors never carry, so it was always ''; botMetaWriteAt grew forever; aliasRouteIndex's whole-map replace could clobber a concurrent rebuild.
• The Routines pane starts collapsed, and the noun is "bot" throughout rather than alternating with "agent".

Pre-existing edge, noted not fixed

Bot chats open into the main pane, so the composer's model pill counts as touchesPrimary and selectModel scopes its write --global — model.default + model.provider on the active gateway profile. For a bot on the active connection that is the right profile: ensureBotAgent moves the gateway onto it first, so the pill writes the bot's own config.yaml, the same destination as Edit Profile's picker.

For a bot on another connection the routed path deliberately does not activate ("Desktop's active connection does not move"), so $activeGatewayProfile is still the local default while the remote bot's chat is on screen. If that holds at runtime, changing the model there rewrites the local profile's default instead of the bot's. This is code-reading, not a reproduction, and it predates this branch — recorded here so it isn't rediscovered as a regression.

Test plan

• npx tsc --noEmit and npm run lint clean across apps/desktop, ui-tui, apps/shared, web
• npx vitest run --project ui — 677 files, 6,681 tests green
• Merged main; its Bots-home flash fix guards a surface this branch deletes, and its bot-row-click fix is ported onto the split modules
• scripts/run_tests.sh tests/hermes_cli/test_profiles.py
• Click a bot: one chat opens, the main zone stays put, the roster dot matches the session dot
• Close a bot chat and delete its profile: the main zone still renders, the sidebar doesn't stretch
• Scheduled jobs arrives as the right-edge rail tab; expanding it survives a reload
• Open a bot chat with no history: face + name centered, correct after changing the avatar or renaming
• Switch the app to 日本語 / 简体中文 / 繁體中文: the Bots rail, dialogs and group room follow

The middle four were driven live rather than by hand: a real Electron build against a real backend and real model turns on Linux (CDP), and the same matrix on a windows-latest runner (Playwright _electron), including no-clone bot creation. The one red there is #87761 — Windows refuses to remove a profile directory whose state.db the backend still holds open — which this branch neither causes nor touches.

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 21cc469 — Merge remote-tracking branch 'origin/main' into bb/bot-mode-}

⚠️ Warnings

OSV vulnerability scan · View job

6 known vulnerabilities found in pinned dependencies.

• CVE-2026-67213 in website/package-lock.json
• CVE-2026-71554 in uv.lock
• CVE-2026-70608 in package-lock.json
• CVE-2026-56876 in package-lock.json
• CVE-2026-70606 in package-lock.json
• CVE-2026-71554 in uv.lock

How to fix:

Review the findings in the Security tab. Update the affected dependencies if a patched version is available.

---

debug info

CI timings

CI timings · View report · View job

Wall time 4m2s vs 4m14s (-4.7%). 8 job(s) slower, 5 faster,

• OS-specific tests / Windows-only tests: -48.0s
• Check contributors / check-attribution: +33.0s
• JS & TS checks / JS & TS checks: -25.0s
• Python tests / Run tests: +23.0s
• OSV scan / Scan lockfiles / osv-scan: +9.0s