Files
lobe-chat/src/features/DevPanel/RenderGallery/fixtures/lobe-agent-management.ts
T
Arvin Xu 373b5e90b2 style(device): run remote CC on a configured device (#15343)
*  feat(device): run remote CC on a configured device with cwd + device context

Make `claude-code`/`codex` dispatched to an `lh connect` device (executionTarget
='device') run in the user's configured directory with a device-appropriate
system context, instead of inheriting the cloud-sandbox setup.

3a — server cwd passthrough:
- resolve the run cwd in the useDevice branch: topic-level workingDirectory
  override > the bound device's `defaultCwd` (read from DB via DeviceModel; the
  gateway only knows live connections, not the user-owned cwd), and pass it to
  dispatchAgentRun.

3b — device-specific systemContext, end to end:
- new `buildRemoteDeviceHeteroContext` — strips the cloud-sandbox boilerplate
  (ephemeral /workspace, pre-cloned repos, commit-or-lose warnings) that would
  mislead an agent on the user's own persistent machine; keeps agent static
  context + resumed conversation history + a minimal cwd note.
- thread `systemContext` through the contract: AgentRunRequestMessage,
  GatewayHttpClient.dispatchAgentRun, deviceProxy.dispatchAgentRun.
- desktop: spawnLhHeteroExec now injects systemContext as the first text block
  of a content-block array on stdin (mirrors spawnHeteroSandbox); previously it
  wrote only the bare prompt, so any context was silently dropped.

The gateway relays unknown fields transparently (`...runParams`), so no gateway
change is needed.

Tests: buildRemoteDeviceHeteroContext unit (6) + GatewayConnectionCtr forwards
cwd/systemContext. type-check clean; existing device/desktop/pkg suites green.

Part of LOBE-9579 (Step 3a/3b). Old ephemeral boundDeviceId migration (3d) and
the web cwd picker (3c) are out of scope here.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

*  feat(device): optimistic device cwd persistence (defaultCwd + recentCwds)

Foundation for the device-scoped cwd picker (executionTarget=device): persist a
working-directory pick to the bound device's registry record so the server's
hetero dispatch (which reads device.defaultCwd) stays in sync and the picker can
offer recent dirs.

- nextRecentCwds: pure most-recent-first / dedupe / cap-20 list builder (the
  server stores recentCwds verbatim, so the client owns this) — unit tested.
- useUpdateDeviceCwd: optimistic `device.updateDevice` — patches the listDevices
  cache in onMutate for instant UI, invalidates onSettled to re-sync truth (self-
  corrects a failed write without manual rollback).

Not yet wired into a picker — the target=device recentCwds-list + manual-input
picker mode that consumes this is the next step.

Part of LOBE-9579 (Step 3c, data layer).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

*  feat(device): gate send on bound-device online for device-targeted hetero

Extend the pre-send device guard from remote-only agents (openclaw / hermes) to
any hetero agent whose run dispatches to a device — i.e. claude-code / codex with
executionTarget='device'. If the bound device is offline (or none is bound), the
send button is disabled and a guard alert is shown, instead of letting the run
fail at dispatch time.

- new selector currentAgentExecutionTarget
- isDeviceExecution = remote-typed OR executionTarget==='device'; drives the
  guard's enabled flag, the blocked state, and the alert.
- device execution no longer requires cloud credentials (it doesn't use the
  cloud sandbox), so the cloud-not-configured gate now exempts it.

The guard hook already handled non-remote types (online check only, no platform
capability probe), so no hook change is needed.

Part of LOBE-9579 (Step 3, device online guard).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 style(tool-render): flatten nested-background tool renders into single-layer surfaces

Remove the card-in-card look across builtin tool renders by dropping the outer
colorFillQuaternary container fill (the framework tool card already provides the
surface) and keeping at most one delineated inner box.

- claude-code AskUserQuestion: rebuilt as a flat Question / divider / Selected
  layout; add i18n keys (question/selected/reply/noAnswer)
- claude-code Skill, local-system WriteFile: flat container + single previewBox
- agent-management CreateAgent/GetAgentDetail: flat container, keep outlined
  systemRole block
- web-onboarding SaveUserQuestion: drop the redundant inner value box

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 📝 docs(builtin-tool): document single-layer surface rule for tool renders

Add §0.8 "stay single-layer — don't nest filled cards": the framework tool
card is already the surface, so the Render's outer wrapper carries no fill and
at most one filled box delineates real content. Cross-link from §2 Render rules
and the diagnostic table, and note the deliberate outlined-panel exception
(TodoWrite / Task).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 📝 docs(builtin-tool): consolidate fragmented UI shared-style rules

The §0 shared rules had drifted into 8 one-line subsections (0.1–0.8). Fold the
five mechanical "every file looks like this" rules ('use client', memo +
displayName, BuiltinXProps generics, t('plugin'), store reads) into a single
annotated component skeleton (0.1), merge the two styling rules into 0.2, and
keep the single-layer surface rule as 0.3. Update the §0.8 cross-references in
§2 and the diagnostic table to §0.3.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 📝 docs(builtin-tool): split UI reference into a per-topic ui/ folder

The single 770-line ui.md had grown unwieldy. Break it into references/ui/
with a README index and one file per topic: principles, shared-rules, the six
surfaces (inspector/render/placeholder/streaming/intervention/portal),
composition, and diagnostics. Convert in-doc §-number cross-refs to cross-file
links and repoint SKILL.md + tool-design.md at the new folder.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

*  feat(device): device-scoped cwd picker for executionTarget=device

When a hetero run is bound to a remote device, the device's filesystem isn't
browsable from here, so the local folder picker doesn't apply. Add
DeviceWorkingDirectory — a self-contained bar item (chip + popover) sourced from
the bound device's recentCwds plus a manual path input.

- Picking/typing a cwd pins it to the active topic (override) and persists it to
  the device via useUpdateDeviceCwd (optimistic defaultCwd + recentCwds), which
  is exactly what the server's device-dispatch branch reads back.
- Same per-cwd CC-session-reset confirm as the local picker.
- WorkingDirectoryBar routes to it when executionTarget==='device' (both web —
  replacing CloudRepoSwitcher — and desktop, replacing the local picker +
  GitStatus); local/sandbox paths are unchanged.
- Reuses existing i18n keys (recent / noRecent / placeholder).

Completes LOBE-9579 Step 3c. type-check clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 style(tool-render): flatten ToolResultCard + de-duplicate Read header

ToolResultCard was the card-in-card shared component (colorFillQuaternary
wrapper around a colorBgContainer box) behind CC Read/Grep/Glob/Write/WebSearch/
WebFetch. Flatten it to single-layer (flat wrapper, one colorFillTertiary
content box) so all consumers stop stacking fills inside the framework tool card.

CC Read header showed the filename strong-label and then dumped the full
absolute path whose tail repeated the same basename, end-truncated so the
meaningful suffix was hidden. Show the directory only (filename stays the
strong label), and drop the conflicting word-break so the dir ellipsizes on one
line.

Note ToolResultCard in the skill as the canonical single-layer header+content
card to reuse.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 fix(device): mark current device, native cwd browse, fix edit Save button

Settings → Devices page polish:
- Badge the row for the machine you're on ("This device"), resolved from the
  desktop gateway's own deviceId (web has no current device → no badge).
- For the current device, the edit modal's Default working directory gains a
  native folder picker (electronSystemService.selectFolder) next to the manual
  input — you can't browse a remote device's filesystem, only your own.
- Edit modal footer now uses real Button components (Cancel + primary Save)
  instead of the base-ui Modal's default okText, which rendered with the wrong
  (non-primary) color.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 fix(device): neutral current-device tag + per-channel tags

- "This device" badge uses the default neutral tag instead of success green.
- Show each live connection's channel as a small tag (desktop / cli) so a
  multi-channel device's connections are individually legible.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

*  feat(devtools): add API jump-list column to the render gallery

The render gallery stacked all of a toolset's API previews in one scroll column
(67 for Claude Code), making any specific render slow to find. Add a middle
column listing the toolset's apiNames: clicking scrolls the matching preview
card into view (landing below the sticky lifecycle bar via scroll-margin), and
an rAF-throttled scrollspy highlights the API the reader is on and keeps that
item visible in the list. A leading dot marks APIs that ship a Render. The
content area now owns its own scroll so the list stays pinned.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 fix(devtools): make the API jump-list readable + deep-linkable

The jump-list was a wall of identical `mcp__claude_ai_Linear__…` truncations and
the active item barely differed from hover. Show just the trailing action for
mcp__ tools (full id in a title tooltip + the preview card header), render names
in monospace, and give the active item a primary left-accent so it reads as
selected. Clicking now pins a `#api-<name>` hash (deep-linkable / shareable) and
loading a hashed URL jumps straight to that card below the sticky bar.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

*  feat(devtools): add an Aggregate message-flow preview tab

The gallery only previewed each API in isolation. Add a View tab (By API /
Aggregate): Aggregate stitches every render-bearing API into one compact
content + tool message flow, so renders can be judged in conversational context
across any lifecycle mode. Inspector-only MCP tools are dropped to keep the
thread about the renders, and the API jump-list column hides in this view.

Extract the Inspector/Body surface rendering out of ToolPreview into shared
ToolInspectorSlot / ToolBodySlot (toolSurfaces.tsx) so both tabs derive props
identically and never drift. View choice persists to localStorage.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 fix(devtools): densify API list + keep mcp prefix visible

The earlier "shorten mcp names" change solved the wrong problem and hid the
`mcp__` prefix, so MCP tools no longer read as MCP. The actual complaint was row
height. Restore the full identifier and instead middle-elide it
(`mcp__claude_ai_Li…get_diff`) so both the muted `mcp` namespace and the
distinguishing trailing action stay visible; full id remains in the title
tooltip. Drop row height to a fixed dense 22px (flex-shrink:0 so it scrolls
instead of squishing) to fit far more APIs per screen.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* ♻️ refactor(devtools): render Aggregate tab through the real Conversation renderer

The hand-rolled MessageList only approximated the chat. Replace it with the
actual shipping renderer: seed a `ConversationProvider` (skipFetch) with fixture
`assistantGroup` messages and map each render-bearing API to a real tool
payload, then render the real `MessageItem` for each. Tool state is driven
purely by the message shape — `result` → success, `result.error` → error,
`intervention.pending` → intervention, unterminated `arguments` JSON →
streaming — so the preview is byte-for-byte what users see in chat. Skips the
virtualized `ChatList` (and its data fetches) by mapping `MessageItem` directly.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

*  feat(device): device detail drawer (channels + recent dirs + config)

Clicking a device row now opens a right-side detail drawer instead of a small
edit modal:
- Connections: render every live connection from the `channels` array, each
  with its channel tag (desktop / cli) + connected-since.
- Name + default working directory (native folder browse on the current
  device); saving a default cwd also seeds the recent list.
- Recent directories: list `recentCwds`, click to reuse, × to remove — this is
  where you can see and manage the recent list (previously not surfaced).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

*  feat(device): record recentCwds on the local device picker

Local-mode runs execute on this machine, but the local working-directory picker
only persisted to a desktop-local recents store — the dir never reached the
device registry, so the settings detail view (and a future device-mode picker)
couldn't see it.

- WorkingDirectory.selectDir now also records the chosen dir into the current
  device's recentCwds (resolved from the gateway's own deviceId).
- useUpdateDeviceCwd gains a { setDefault } option so local mode records
  recentCwds without repointing the device's defaultCwd.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🩹 fix(devtools): thread Aggregate preview messages via parentId

Each fixture turn was an orphaned message with no parentId, so the renderer saw
a pile of disconnected messages rather than one conversation. Chain every turn
onto the previous one (`parentId` = prior message id) so they read as a single
linear thread.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* ♻️ refactor(devtools): seed flat messages so conversation-flow groups the Aggregate

The previous version hand-built `role: 'assistantGroup'` messages, bypassing the
real grouping. Seed the flat DB-shaped messages instead — an `assistant` message
carrying the tool_use plus a linked `role: 'tool'` result message per API — and
let conversation-flow's `parse()` synthesize the assistantGroup exactly as it
does in chat. The consecutive tool turns now collapse into one real workflow
group (one avatar, N content+tool blocks) instead of N hand-rolled groups.
Lifecycle state rides the tool message the same way production carries it
(content/pluginState = success, pluginError = error, pluginIntervention = pending).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 refactor(device): inline master-detail device settings; drop uppercase labels

Per feedback:
- Replace the floating edit Drawer with an inline right-hand detail panel —
  the devices page is now a master-detail layout (device list on the left,
  selected device's detail on the right), like the rest of settings.
- Drop the ALL-CAPS section labels (no more text-transform: uppercase /
  letter-spacing) — labels use natural case + a muted color.

DeviceItem becomes a selectable list row (no own modal); DeviceDetailPanel
renders the detail inline (connections per channel, name, default cwd + browse,
recent dirs). Keyed on deviceId so the form resets on selection change.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 refactor(device): detail panel opens on click, not by default

Per feedback — mirror the memory-preferences master-detail pattern:
- No device is selected by default; the right detail panel only renders once a
  row is clicked (clicking the selected row again closes it). Panel has its own
  close (×).
- List flexes to fill when nothing is selected; the detail appears as a right
  column on selection.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🐛 fix(devtools): bind render gallery to viewport height so columns scroll

The page root used height:100%, which only resolves when an ancestor route
provides a bounded height — under mounts that don't, the whole page grew to
content height and the API list never scrolled internally. Bind the root to
100dvh directly and add min-height:0 to the flex chain (main + the API list)
so the scroll container engages regardless of how the route is mounted.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

*  feat(devtools): add WebFetch / WebSearch fixtures so they render

Both APIs had no fixture, so the gallery fell back to schema-sampled args with no
content and the renders drew empty (just the icon). Add fixtures with realistic
args + content: WebFetch (url + prompt + markdown answer), WebSearch (query +
allowed_domains + results), plus their apiList descriptions.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 fix(device): render connections straight from device.channels[]

Drop the device.online-based synthetic single-channel fallback — the connection
rows now come purely from the device.channels[] array (one row per live
connection), with offline = empty array.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🐛 fix(hetero): distinguish CC server throttle from user quota limit

A 429 "Server is temporarily limiting requests (not your usage limit)"
was classified as a user rate_limit, rendering the misleading "Claude
Code usage limit reached" reset-time guide. Key the rate_limit vs
overloaded decision on the structured rate_limit_event reset window
(resetsAt / rateLimitType) instead of the HTTP status, so 429/529 with
no quota signal fall through to the overloaded (retry) UX.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 fix(devtools): loosen the API list density

22px rows at 12px overcorrected into a cramped sidebar. Relax to 30px rows,
13px label, a small inter-row gap, and a touch more vertical padding so the
jump-list reads comfortably.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 fix(device): align connection rows in the list item (drop 30px indent)

The connection rows had a 30px inline-start padding that pushed them right of
the cwd line; align them with the rest of the device info.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 fix(device): move connection status dot to the first line

The online/offline status now sits as a dot next to the device name + badges
(with the connected / last-active time as a tooltip), instead of a separate
third line. Per-channel connection detail still lives in the detail panel.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 feat(devtools): show the Aggregate preview as "Lobe AI"

The seeded preview conversation resolved its avatar/name through an agentId that
wasn't in the agent store, so every turn fell back to the unresolved-agent
"Unnamed Assistant" / UN avatar. Seed agentMap with a Lobe AI meta
(DEFAULT_INBOX_AVATAR + title) for the devtools agentId, shared via
DEVTOOLS_AGENT_ID / DEVTOOLS_AGENT_META so MessageList's context and the store
seed stay in sync. Restored on unmount.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🐛 fix(devtools): carry tool result state in BuiltinInspectorProps

The Aggregate preview passes `result.state` to inspectors, matching the
real runtime, but the canonical `result` type omitted `state` — failing
type-check. Add `state?: any` so devtools and runtime agree.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* 🐛 fix(device): pin topic cwd and add hetero-tracing toggle

- Prefer the topic's own `metadata.workingDirectory` over the device
  default when dispatching, so an existing topic keeps its pinned cwd
- Add `heteroTracingEnabled` store flag to trace CLI raw streams in
  packaged builds (Help menu checkbox)
- Reorder the connection status dot ahead of badges in DeviceItem

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

*  feat(device): add Help-menu toggle to record hetero-agent CLI traces in production

Packaged builds previously never wrote hetero-agent (CC / Codex) CLI traces,
so production issues couldn't be captured. Add a persisted `heteroTracingEnabled`
toggle in the Help menu (all 3 platforms) plus an "Open HeteroAgent Directory"
entry. Dev still always traces to `cwd/.heerogeneous-tracing`; packaged builds,
when enabled, centralize traces under `<appStoragePath>/heteroAgent/tracing`
(sibling to the existing files cache) via shared dir constants.

Closes LOBE-9828

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 📝 docs(skills): fold stacked-prs guidance into the pr skill

Merge the standalone `stacked-prs` skill into `pr` as a supplementary section
(ordering rule, file placement, git split recipe, dependency verification,
Linear bookkeeping, gotchas) and absorb its triggers into the pr description,
rather than keeping a separate skill.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🐛 fix(devtools): chain RenderGallery previews into one assistantGroup

Unfinished tool states (streaming / loading) now emit a paired tool result
message with `LOADING_FLAT` content instead of none, and every assistant turn
chains onto the previous message's id. The tool_use → tool_result link is what
lets conversation-flow merge the turns into one assistantGroup; without it the
unfinished modes rendered as one orphaned group per tool.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* ♻️ refactor(device): key hetero trace location off the toggle, not isPackaged

`resolveTraceRootDir` now centralizes traces under
`<appStoragePath>/heteroAgent/tracing` whenever `heteroTracingEnabled` is on,
instead of gating on `isPackaged`. Packaged behavior is unchanged (it only
traces when the toggle is on), and a dev who opts in now also gets the
centralized dir reachable from the Help-menu entry. Plain dev runs keep
writing to `cwd/.heerogeneous-tracing`.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 🐛 fix(device): move hetero dir consts to a side-effect-free module

Importing the new `HETERO_AGENT_*` constants from `@/const/dir` dragged that
module's load-time `app.getPath()` / `app.getAppPath()` calls into the menu and
controller import graphs, breaking menu/controller suites whose electron mocks
or partial `@/const/dir` mocks didn't anticipate it. Relocate the pure path
segments to `@/const/heteroAgent` (no electron import) and point the controller
+ all three menu impls there. Also add the now-required `storeManager.get/set`
to the menu test app mocks (the Help-menu tracing checkbox reads it at build).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* 💄 style(devtools): refine RenderGallery surfaces and fix local-system fixtures

- flatten the active ApiList item (drop accent bar) and the ToolPreview card shadow
- give the Aggregate thread a white container surface
- hide deprecated lobe-notebook toolset and legacy *Local* aliases from the gallery
- re-key local-system fixtures to current API names + add missing call args
- backfill agent-management call args so inspectors render their argument rows

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

*  test(desktop): default global electron mock so import-time app access is safe

`@/const/dir` reads `app.getAppPath()` / `app.getPath()` at module load — fine
in production (app is ready), but it forced every test that transitively imports
it to stub those basics, which is the real root of the recent breakages.

Register a default `electron` mock in the global vitest setup, giving every
suite a ready `app` (paths + readiness) plus light stubs for the common
namespaces. Suites that need specific behavior still declare their own
`vi.mock('electron', …)`, which overrides this per-file. This keeps production
free to use plain value-style path constants instead of lazy getter functions.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
2026-05-31 15:11:55 +08:00

116 lines
3.3 KiB
TypeScript

'use client';
import { defineFixtures, single } from './_helpers';
export default defineFixtures({
identifier: 'lobe-agent-management',
fixtures: {
callAgent: single({
args: {
agentId: 'agent_workspace_helper',
instruction:
'Review the `/devtools` route and list any preview cards that still need richer fixtures.',
},
}),
createAgent: single({
args: {
description: 'Internal helper for preview and QA workflows.',
model: 'gpt-5.4',
plugins: ['lobe-web-browsing', 'lobe-local-system'],
provider: 'openai',
systemRole: 'You help engineers verify UI changes quickly and carefully.',
title: 'Preview QA Agent',
},
}),
duplicateAgent: single({
args: {
agentId: 'agent_workspace_helper',
newTitle: 'Workspace Helper Copy',
},
pluginState: {
newAgentId: 'agent_preview_clone',
sourceAgentId: 'agent_workspace_helper',
success: true,
},
}),
getAgentDetail: single({
args: {
agentId: 'agent_preview_specialist',
},
pluginState: {
config: {
model: 'gpt-5.4',
plugins: ['lobe-web-browsing', 'lobe-cloud-sandbox'],
provider: 'openai',
systemRole: 'Focus on frontend verification and fast local feedback loops.',
},
meta: {
avatar: '🧪',
backgroundColor: '#EEF6FF',
description: 'Specialized in preview harnesses and UI regression checks.',
tags: ['preview', 'qa'],
title: 'Preview Specialist',
},
},
}),
installPlugin: single({
args: {
agentId: 'agent_preview_specialist',
identifier: 'lobe-cloud-sandbox',
source: 'official',
},
pluginState: {
installed: true,
pluginId: 'lobe-cloud-sandbox',
pluginName: 'Cloud Sandbox',
},
}),
searchAgent: single({
args: {
keyword: 'preview',
source: 'all',
},
pluginState: {
agents: [
{
avatar: '🧪',
backgroundColor: '#EEF6FF',
description: 'Preview route and fixture maintainer.',
id: 'agent_preview_specialist',
isMarket: false,
title: 'Preview Specialist',
},
{
avatar: '📚',
backgroundColor: '#FFF7E8',
description: 'Keeps internal docs and issue writeups tidy.',
id: 'agent_doc_partner',
isMarket: true,
title: 'Documentation Partner',
},
],
},
}),
updateAgent: single({
args: {
agentId: 'agent_preview_specialist',
config: JSON.stringify({
model: 'gpt-5.4',
systemRole: 'Prioritize maintainable developer tooling and preview coverage.',
}),
meta: JSON.stringify({
description: 'Expanded to cover internal tooling previews.',
title: 'Workspace Preview Partner',
}),
},
}),
updatePrompt: single({
args: {
agentId: 'agent_preview_specialist',
prompt:
'When asked for a visual check, prefer building a reusable preview harness before taking a screenshot.',
},
}),
},
});