workspaces/chat: usability polish (rolling) by luokerenx4 · Pull Request #183 · TraderAlice/OpenAlice

luokerenx4 · 2026-05-13T12:46:57Z

Summary

Rolling PR for the next wave of Workspaces + Chat usability iterations
following beta.5, plus retirement of pre-workspace subsystems whose
problems the workspace architecture has dissolved (Brain, Diary,
OpenClaw + browser tool). Per-session contributions append below as
new commits land — don't expect this body to stay stable.

Per-session contributions

2026-05-13 — Retire OpenClaw + bundled browser tool

The src/openclaw/ subtree (188 files, 28,151 lines, frozen legacy)
existed solely to back one AI tool: browser, wrapping the OpenClaw
Browser Relay Chrome extension. Same architectural pattern as the
Brain / Diary retirement earlier today — workspaces dissolve the need
for OpenAlice to bundle a heavy in-process subsystem when the
workspace itself can host the capability per-task via MCP.

Users who want browser control now install @playwright/mcp (or
equivalent) into the workspace where they actually need it. Scoped to
the task, not a global tool on every Alice surface.

Removed:

Entire src/openclaw/ subtree (browser/, gateway/, agents/,
sessions/, terminal/, infra/, plugins/, channels/, routing/, cli/,
process/, media/, security/, config/, logging/, utils/, shared/)
src/tool/browser.ts (Vercel AI SDK wrapper)
main.ts registration + SettingsPage 'browser' tool-group label
8 production deps (chalk, json5, undici, file-type, sharp, express,
tslog, playwright-core) + @types/express dev-dep — confirmed
openclaw-only via strict import grep. Net: 21 packages out of the
lockfile
All anachronistic "OpenClaw" references in CLAUDE.md /
project-structure.md / mcp-ask-connector.md / source comments
(mcp-export, media, agent-event, mcp-ask-plugin). The multi-modal
AgentToolResult wire format lives on past openclaw and is no longer
branded as such

BREAKING (MCP): AI tool browser (16 actions + 11 act sub-actions

Chrome Extension Relay) no longer exposed via the MCP server.
Migration path: install @playwright/mcp into the workspace that
needs browser control. The workspace's MCP config is per-spawn, so
the capability is scoped to the task that actually needs it instead
of being a global Alice tool.

Key commits: 0304c6d, d42aac7, 91a08d2
Net diff: −28,613 lines

2026-05-13 — Retire Brain (frontal-lobe) + Diary

The frontal-lobe note system existed to maintain self-consistency
across compacted chat sessions; the Diary dashboard existed as the
single-Alice-loop status feed. Both predate workspaces. With workspaces
now the primary substrate — each workspace owns its own scrollback,
transcript, working tree, and CLAUDE.md — both problems dissolve:
"want consistency? write a markdown file in the workspace"; "want to
see what Alice is doing? open the workspace's transcript". The single-
self assumption that made the de-se principle coherent doesn't hold
under N parallel workspaces.

Removed:

src/domain/brain/, src/tool/brain.ts, src/webui/routes/brain.ts,
ui/src/components/BrainSidebar.tsx, ui/src/api/brain.ts
src/webui/routes/diary.ts + spec, ui/src/pages/DiaryPage.tsx,
ui/src/api/diary.ts, the kind: 'diary' ViewSpec entry, the Diary
row in the chat sidebar, /diary route, /api/diary mount
main.ts brain wiring + formatRelativeAge helper; getInstructions
simplified to persona-only
SettingsPage brain tool-group label

Retained: data/brain/persona.md and data/brain/heartbeat.md (still
read by workspace bootstrap + persona route + heartbeat default
override). data/brain/ directory stays — it's now just the home for
those two override files.

Heartbeat itself is untouched. Its real downstream is the notifications
system (notify_user → NotificationsStore), not Diary. The
SessionStore('heartbeat') JSONL keeps being written by
agent-work-listener — orphan but harmless, separate decision.

Migration 0006_retire_brain deletes the two leftover brain-state
files (data/brain/{commit.json,frontal-lobe.md}) on next boot.
ENOENT-tolerant + idempotent.

Follow-up fix: bumped Zustand workspace store version: 2 → 3 so users
with a Diary tab open in localStorage discard the stale state on next
load rather than crashing on an unknown kind: 'diary'.

BREAKING (MCP): three brain tools no longer exposed via the MCP
server — getFrontalLobe, updateFrontalLobe, getBrainLog. External
integrations calling these need a migration path, typically a
workspace-local markdown file edited via the CLI's own read/write tools.

Key commits: 25f9532, 0a10b3f, 756d1bf, b75781c

2026-05-13 — Paused-session view: faux-TUI backdrop

The right pane for a paused session was a bare "Resume" button on a
panel — clear if you know what "resume" means, opaque otherwise.
Replaced with a two-layer composition:

Faux-TUI backdrop: heavily blurred (5px) + dimmed mock of an
agent-CLI terminal. Titlebar, prompt boxes, bullet responses,
* Brewed for Ns status lines, bottom prompt + statusbar. Structural
Unicode characters (›, •, *, ▶▶, ●) partially survive the blur and
carry the "this is a terminal" tell; body text blurs into soft
horizontal bands. Static placeholder for every session — we can't
show real content because agent-CLI scrollback isn't persisted on
pause (the CLI re-renders its own transcript on resume).
Centered Resume card: radial vignette pushes focus to the
middle. Copy flipped from Resume to Continue conversation with
a MessageSquare icon. Meta (agent / created / transcript hint)
tucked under a divider below the CTA.
Shell-template sessions keep the existing scrollback-restore hint.

Key commits: 852f1d5

Full commit log

91a08d2 chore(docs): scrub openclaw references after subtree removal
d42aac7 chore(deps): drop openclaw-only dependencies
0304c6d chore: retire openclaw + browser tool
b75781c fix(tabs): bump persist version after retiring diary kind
756d1bf chore(migration): 0006 — prune brain commit artifacts
0a10b3f chore: retire Brain — frontal-lobe + tools + dashboard
25f9532 chore: retire Diary surface
852f1d5 feat(workspaces/ui): paused-session view — blurred faux-TUI backdrop

Test plan

tsc --noEmit clean (backend + UI) after each commit
pnpm test passes — 88 files, 1676 tests
Backend smoke test: tool-center registers 46 tools (down from 50
pre-retirement — Brain 3, Browser 1, leaving 46), all subsystems
(snapshot / cron / news-collector / listener-registry) start
cleanly. Net dep tree shrinks by 21 packages.
Paused codex/claude session: backdrop visible behind centered card
User to confirm visual feel at full viewport size
Frontend renders cleanly after Diary/Brain removal (regression
fix b75781c verified by user)

🤖 Generated with Claude Code

The paused-session right pane used to be a bare "Resume" button on a panel — useful if you already know what "resume" means, opaque if you don't. Replace with a two-layer composition that telegraphs "paused conversation" visually: 1. **Faux-TUI backdrop** — a heavily-blurred, dimmed mock of an agent-CLI terminal (titlebar, prompt boxes, bullet responses, `* Brewed for Ns` status lines, bottom prompt + statusbar). It's static placeholder content for every session; the structural Unicode characters (›, •, *, ▶▶, ●) partially survive the 5px blur and give the "this is a terminal" tell, while the body text blurs into soft horizontal bands. We can't show the real conversation — agent-CLI scrollback isn't persisted across pause (the CLI re-renders its own transcript on resume) — so this is about telegraphing the *shape* of what's paused, not the content. 2. **Resume card** centered on top, with radial vignette pushing focus inward. Copy: "Continue conversation" + MessageSquare icon. Meta (agent / created / transcript hint) tucked under a divider below the button so the CTA reads first. For shell-template sessions the hint about scrollback restore is preserved. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The Diary dashboard existed as the single-Alice-loop status feed — "what did the heartbeat cycle do" rendered against `SessionStore('heartbeat')` + `agent.work.*` filtered by source=heartbeat. With workspaces now the primary substrate, each workspace owns its own transcript and the cross-cut Diary view became redundant: the notifications system handles the actual push surface, and per-workspace scrollback handles the inspect-state surface. Diary itself was passive polling that fired roughly never in practice. Removed: - src/webui/routes/diary.ts + spec - src/webui/plugin.ts: /api/diary mount - ui/src/pages/DiaryPage.tsx, ui/src/api/diary.ts, api/index registration - ui/src/tabs/types.ts: `kind: 'diary'` from ViewSpec - ui/src/tabs/registry.tsx: DiaryPage import + diaryModule + VIEWS entry - ui/src/tabs/UrlAdopter.tsx: /diary route - ui/src/components/ChatChannelListContainer.tsx: Diary sidebar row - Diary mentions in App.tsx / ActivityBar.tsx / MarkdownContent.tsx / agent-work-listener.ts comments - tabs store spec: diary scenarios → news (same shape) Note: `SessionStore('heartbeat')` continues to be written by the agent-work-listener — that's a separate decision (orphan but harmless). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The frontal-lobe system existed to maintain self-consistency across compacted chat sessions: Alice wrote de-se notes to herself that survived compaction and got re-injected into the next round's system prompt with a "written Nh ago" cue. With workspaces as the primary substrate, the problem dissolves — each workspace owns its own scrollback + transcript + working tree, so consistency comes from "write a markdown file in the workspace" rather than a single global note threading across sessions. The single-self assumption that made the de-se principle coherent doesn't hold under N parallel workspaces. Removed: - src/domain/brain/ (Brain class, types, index) - src/tool/brain.ts (getFrontalLobe / updateFrontalLobe / getBrainLog tools — MCP-exposed, see PR description) - src/webui/routes/brain.ts + /api/brain mount - ui/src/components/BrainSidebar.tsx, ui/src/api/brain.ts - src/main.ts: brain wiring, BRAIN_FILE / FRONTAL_LOBE_FILE constants, formatRelativeAge helper, brain tool registration. `getInstructions` is now just `readFile(PERSONA_FILE)` — no more frontal-lobe injection - ui/src/pages/SettingsPage.tsx: 'brain' tool-group label (the group no longer registers) - TODO.md: "Brain commits" line from the TimeView replay-asset inventory `data/brain/persona.md` and `data/brain/heartbeat.md` stay where they are (still read by main.ts, workspace bootstrap, and the persona route). The leftover `data/brain/commit.json` and `data/brain/frontal-lobe.md` files are cleaned up by migration 0006 (next commit). BREAKING (MCP): three brain tools are no longer exposed via the MCP server. Any external integration calling getFrontalLobe / updateFrontalLobe / getBrainLog should plan a migration path — typically a workspace-local markdown file edited via the CLI's own read/write tools. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Cleans up the two leftover brain-state files after the Brain subsystem was retired in the preceding commit: - data/brain/commit.json (BrainExportState — commit history) - data/brain/frontal-lobe.md (latest frontal-lobe snapshot) persona.md and heartbeat.md stay where they are — both still in active use by main.ts (persona injection), workspace bootstrap, and the persona route. Idempotent via ENOENT-tolerance: re-running on a clean tree is a no-op. Affects files outside data/config/ so this uses raw fs/promises (same pattern as 0004); ctx helpers are config-scoped. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The previous commit removed `kind: 'diary'` from ViewSpec but didn't bump the Zustand persist version. Any user who had a Diary tab open (everyone who used the feature) would crash on next load: localStorage restores a tab with `spec.kind === 'diary'`, getView('diary') returns undefined, UrlSync's `view.toUrl(spec)` throws. The store's own comment calls this out — "Schema bumps clear stored state (no migrate function) — loud-fail beats silent migration when shape changes." Bumping version 2 → 3 triggers Zustand's discard path: on load, version mismatch with no migrate function → start fresh. Cost: users lose their open-tab restoration once. Better than a hard crash for anyone who had Diary open at the moment of the upgrade. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The OpenClaw subsystem (28,151 lines / 188 files in src/openclaw/) was frozen legacy that only existed to back a single AI tool — `browser` — via the OpenClaw Browser Relay Chrome extension. With workspaces as the primary substrate, the browser-tool capability is now better served by per-workspace MCP integrations: users who want browser control install @playwright/mcp (or equivalent) into the workspace they care about, scoped to that task, without OpenAlice bundling a heavy browser stack into the trading core. This is the same architectural pattern as the Brain / Diary retirement: workspaces dissolve a problem that previously required a bespoke in-process subsystem. Removed: - `src/openclaw/` — entire 188-file subtree (browser/, gateway/, agents/, sessions/, terminal/, infra/, plugins/, channels/, routing/, cli/, process/, media/, security/, config/, logging/, utils/, shared/) - `src/tool/browser.ts` — Vercel AI SDK wrapper - main.ts: createBrowserTools import + toolCenter.register('browser') - SettingsPage: 'browser' tool-group label Retained: - Comment-level "OpenClaw format" mentions in mcp-export / media / agent-event / mcp-ask-plugin — they describe a wire-format shape that lived on past openclaw and is now just "the tool-result format we accept". Cleaned up in a follow-up commit. BREAKING: - AI tool `browser` (16 actions + 11 act sub-actions + Chrome Extension Relay support) no longer exposed via MCP. Migration path: install @playwright/mcp into the workspace that needs browser control. The workspace's MCP config is per-spawn, so this is scoped to the task that actually needs it instead of being a global tool on every Alice surface. Deps that became orphaned are dropped in the next commit. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

After the openclaw subtree was deleted, the following deps had zero remaining consumers in the codebase. Confirmed by strict import grep (every `from '<dep>'` site was inside src/openclaw/). Production: - chalk (openclaw/logging console formatting) - express (openclaw/browser/server.ts HTTP bridge) - file-type (openclaw/media MIME detection) - json5 (openclaw config parsing) - playwright-core (openclaw/browser Playwright driver) - sharp (openclaw/media image-ops) - tslog (openclaw/logging primary logger — OpenAlice uses pino) - undici (openclaw HTTP outbound) Dev: - @types/express Net: 21 packages removed from the lockfile. pnpm.overrides for express-rate-limit / path-to-regexp / follow-redirects left in place — they're declarative and inert when the named package isn't in the tree, so removing them is a cleanup for a separate pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

CLAUDE.md / project-structure.md: drop the openclaw/ tree entry and the now-defunct tool/brain.ts + tool/browser.ts lines. Retag data/brain/ description as "Persona + heartbeat user-override prompts" (more accurate now that the Brain commit/frontal-lobe artifacts are gone). docs/mcp-ask-connector.md: rebrand the "OpenClaw integration" example as a generic External Agent integration; OpenClaw was the original prototype caller but the connector is meant to serve any orchestration agent that speaks MCP. Also fix the tool inventory ("trading, analysis, brain" → "trading, analysis, market" — brain tools are gone). Source comments: drop OpenClaw-as-source-of-format references in mcp-export.ts, media.ts, agent-event.ts, mcp-ask-plugin.ts. The wire format (`{ content: [...], details: {...} }`) lives on past openclaw and is just our multi-modal AgentToolResult shape now. TODO.md heartbeat-as-template warning: keep the warning, drop the "OpenClaw-era" branding (the point is structural, not historical). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Workspace landed (PR #182, #183) and is now reshaping the chat surface end-to-end — the legacy single-session Traditional Chat is being demoted to a fallback for connector-driven flows, the new Workspace + Inbox path is the recommended one, and the inbox_push MCP route gave AI agents a structurally clean way to push back to the user. None of that is "patch-level" — it's a generation shift from 0.10 → 0.20. Skipping 0.11–0.19 deliberately so the version reflects how much of the original 0.10 chat design has been replaced rather than how many incremental changes shipped. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Ame and others added 8 commits May 13, 2026 20:42

luokerenx4 merged commit ed7d47a into master May 13, 2026
2 checks passed

marrrrrcus mentioned this pull request May 23, 2026

feat(auto-trading): restore Phase 1 snapshot scheduler after UTA-split #207

Open

3 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

workspaces/chat: usability polish (rolling)#183

workspaces/chat: usability polish (rolling)#183
luokerenx4 merged 8 commits into
masterfrom
dev

luokerenx4 commented May 13, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

luokerenx4 commented May 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Per-session contributions

2026-05-13 — Retire OpenClaw + bundled browser tool

2026-05-13 — Retire Brain (frontal-lobe) + Diary

2026-05-13 — Paused-session view: faux-TUI backdrop

Full commit log

Test plan

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

luokerenx4 commented May 13, 2026 •

edited

Loading