Widget-native assisted drafts

Advisory conversations that patch declared host drafts without acquiring submission, DOM, or domain authority.

Details

Authority and opt-in

The mounted Dashboard widget owns the recoverable current draft. Conversations owns transcript and driver lease; the assistance broker owns immutable per-turn snapshot/patch evidence and host acknowledgements. It is not a second draft store, task/job writer, or proposal authority.

A widget declares a normal drafts entry and adds assistableDrafts: [assistedDraftDeclaration("task-create")] (or job-create). Its renderer calls useAssistedDraft(draftName, handle, options), renders AssistDraftButton, and spreads fieldProps(["field"]) on the actual controls. The full identity includes profile, workspace, app, view, widget instance/type, draft name, and scope key. Non-assistable metadata remains in the host draft.

The sole field authority is work_buddy/dashboard/assistance/form_schemas.json, read by Python and imported directly by TypeScript. It declares types, bounds, descriptions, disclosure, required/optional fields, default values, operations, patch behavior, and submitPolicy: user_only. There is no parallel DOM-field map. Remove resets an optional field to its declared empty/default value; required fields cannot be removed. Batch lines and proposal bindings are not assistable fields. Jobs parameter JSON refuses credential-shaped keys before disclosure.

Conversation and disclosure

The entry action is the shared duotone sparkle icon and AI help. Its non-modal side panel sits beside the view's grid content, below page headings and navigation. Drag the shared orange divider, use its Left/Right arrow keys, or double-click to restore default widths. The view remembers the chosen width. Narrow or short workspaces switch between Form and AI help while keeping both mounted. Resizing or closing preserves the form, conversation and unsent message. Closing returns focus to the originating action; asynchronous replies do not steal focus from the form.

The existing Dashboard Hover Help mode also covers the panel heading, close action, model/context controls, Launch and resize divider. Optional explanations stay in help; required disclosure, errors and recovery remain visible. The panel keeps its controls reachable when a long draft or recovery message needs more space than the viewport. Copy and Close are icon actions in the panel's top header with their explanations in Hover Help. Copy uses the shared Chat action and exports plain Speaker: message text without timestamps, model details or controls. The composer paints its complete focus indicator inside the control so a clipped side panel cannot cut it off.

The dock composes the same ConversationChat, ChatPanelState, composer, message/choice rendering, execution hook and picker used by Co-work. It is not a separately placeable chat widget and does not import Co-work document state. Form context, explicit lifecycle actions and patch receipts use narrow host accessories. The conversation provider remains stable across lifecycle and model changes, preserving the transcript and unsent composer draft.

System → Dashboard AI owns the wb.dashboard.assistance opt-in and wb.dashboard.chat-execution-default for new chats. A prepared conversation pins its own provider/model pair; its picker never changes the global default. Internal inference tiers are not interactive model choices. Claude Code and Codex are registered interactive providers; a local inference profile is not a local agent driver, and there is no hidden provider fallback.

Opening AI help prepares metadata and the canonical conversation, but sends no form context and starts no model. Explicit Launch authorizes the displayed provider/model revision and freezes the allowlisted initial fields, draft revision and hash before any asynchronous work. The agent consumes the canonical form purpose, schema and frozen values through assisted_draft_context_get, then sends a useful initial greeting through the ordinary conversation tool. No user-authored message is fabricated and the user need not retype form context. While that promised first message is pending, the shared transcript shows an accessible animated ellipsis without disabling the composer or model picker; reduced-motion mode renders the same status without animation.

Launch occupies the composer's normal Send position until authorization succeeds; there is no separate startup button or idle readiness paragraph. Launching… prevents duplicate activation. An uncertain attempt offers Retry Launch against its frozen disclosure, with Launch with current fields as a separate explicit authorization. Launch never sends or clears an unsent message. The same composer returns to ordinary Send after launch.

If the hosted driver exits during or after Launch, the broker stops exactly that generation and advances the lifecycle revision once. Current fields, transcript, unsent composer text, receipts, conditional Undo, and normal human submit remain available. A known Claude Code authentication failure gives a sign-in instruction; unknown or legacy diagnostics map to fixed safe copy. Passive polling, reopen, and exact-request replay never launch a replacement. After correcting the provider, the user must choose Launch explicitly; that creates fresh request and snapshot identities from the then-current fields.

Subsequent authored turns keep the house message ID. Before transport, the host stages one immutable snapshot under that ID; an uncertain acknowledgement retains the same prepared envelope. Later manual edits cannot enlarge that turn's disclosure and remain protected by normal patch conflicts.

A hosted agent receives only source-free binding identifiers at launch. Its non-overridable tool set contains assisted_draft_context_get, assisted_draft_reference_search, assisted_draft_propose_patch, and the existing send, ask, poll, receive and acknowledge conversation tools. A form manifest decides which reference scopes exist. Jobs may search the same registered capability/workflow names, slash aliases, one-line descriptions and reduced parameter schemas shown by its picker; Tasks exposes no such scope. Reference results are untrusted data, never instructions, and grant no authority to execute a capability, start a workflow, browse, read generic documentation, submit the form, or create/schedule a job. Every call checks the exact session, conversation, consumer, generation, pinned model and applicable policy gates. Initial context consumption precedes conversation access; each user turn must be received and its exact snapshot consumed before edits or acknowledgement. Only finite-choice/boolean questions use inline answer tools; ordinary composer messages do not answer a pending question.

Exact content releases are Sources-backed and recorded in the shared worker disclosure manifest before return. Output binds the ordered input manifest and durable producer identity. A possibly-sent release is never automatically replayed or retried under a new generation. Reference receipts additionally bind the active Start, generation, exact consumed turn, consumption receipt, stable request ID, reference kind and normalized query; a cross-turn reuse conflicts. Replies and patches have stable identities; errors are sanitized. Unbound generic conversation routes/tools cannot bypass the form session.

The protected HTTP flow is metadata-only POST /api/assistance/sessions, GET/PATCH execution selection, explicit Launch with initialSnapshot, and snapshot preparation followed by canonical conversation response. Responses to inline questions include their exact in_reply_to. Stop and permanent End are separate authority-reducing actions.

Model changes fence the old driver and return to a prepared state. Explicit Launch is required before the new recipient receives current fields or history. A fresh Launch supersedes unfinished prior work and old pending questions while preserving the transcript; its frozen current snapshot is the new working base. Start and scoped Stop compare the server-owned integer controlRevision. Start checks it before and after provider validation; an exact durable Start retry never probes or launches again. Scoped Stop carries its request identity, expected revision and, when applicable, exact pending Start request. A delayed Start or stale Stop retry cannot restart or cancel a successor generation.

Stop requires a fresh explicit Launch, while End permanently revokes that binding. Cleanup remains available after opt-out, expiration, read-only or a Source Foundation restore fence. Older session protocols require an explicit new session; their transcript and receipts remain inspectable. Ended and expired sessions also return a read-only recovery projection without resolving defaults or probing providers, so a reload preserves history and conditional Undo.

Patch and recovery laws

A patch binds assistant session, conversation, full draft identity, schema, base revision, SHA-256 base snapshot, stable patch ID, and allowlisted path-array set/remove operations. Malformed, oversize, secret, unknown, root, wildcard, or prototype-polluting operations reject the entire patch. There are no submit, DOM selector, arbitrary callback, or navigation operations.

After validation, the host flushes queued saves and takes a fresh synchronous snapshot. A stale patch is evaluated field-by-field: unchanged and unfocused fields apply in one local revision; changed or focused fields become visible suggestions. The normal draft repository CAS fences another tab's saves. Applied and pending fields appear in a durable acknowledgement, visual non-color marks, and a polite announcement. Each receipt is interleaved with its exact durable assistant reply; until that reply is visible it stays beside the source turn, while genuinely unanchored legacy evidence remains in compact field-change history. Receipt accessories are not conversation messages and are omitted from Copy chat. Applying suggestions is an explicit user review.

Only fields changed by the latest successful assistant mutation receive the applied highlight. A newer mutation replaces the emphasized set and the mark clears after about four seconds; a human edit clears that field's emphasis immediately. The highlight is presentation state and never returns after reload, remount, or reset. Pending-review outlines remain until Apply or Dismiss because they communicate durable work still awaiting a decision. Undo restores a field only while it still contains the assistant's applied value and is not focused.

The host retains bounded patch receipts and inverses in tab-scoped session storage keyed by full draft identity and assistant session. It journals the planned inverse before CAS and records completion after flush. A remount restores the journal before polling; an interrupted local write is inspected against current fields, never replayed blindly. A missing server acknowledgement is retried without applying the patch again. These receipts are evidence, not another current draft. A panel close preserves the conversation and Undo. Ending the session preserves the form and receipts but permanently revokes its active assistant binding on the server.

Leaving an active form host durably queues a scoped Stop, including a pending Start, while preserving the conversation binding and unsent composer draft. Reopening requires fresh explicit Launch. Simply closing the still-mounted dock preserves its session. Cleanup acknowledges only a typed missing-session response as already absent; authentication, network and other failures remain visible and retryable.

An accepted host draft.clear() or draft.reset(value, { ifRevision }) synchronously notifies reset subscribers. Assistance invalidates its editing generation, closes the dock, and removes the saved session binding before a late reply can cross into the next editing lifetime. A durable local revoke intent retries permanent server cleanup without resurrecting the old binding. Reset preserves explicitly chosen fields with one normal repository CAS replacement save; it never deletes the old durable draft before saving the replacement. Clear retains its existing delete behavior. A failed revision check does not revoke the session, and a failed replacement save leaves the previous durable record recoverable.

Operate-only checks exist both in the host and in the server-owned dashboard read-only callback, including before model handoff/result publication. Arrange, Preview, or read-only transitions pause sends and patches. Unmount/reset fences late operations. Startup/storage/model failure never disables ordinary form editing or replaces the form's normal human submit intent.

Manual form editing remains available in Preview through the existing host's forked draft repository. Normal task/job submissions stop at the host's declared Preview effect policy before reaching a provider. Proposal saves, acceptance, dismissal and retries are separately Operate-only, including before pending request metadata is allocated. Leaving Preview restores the original draft; Arrange and read-only access continue to lock the form.

Verification

Isolated HTTP and tool tests exercise contextual startup, model selection, exact questions, binding denial, source disclosure, idempotency, generation fences, policy gates, cleanup, malformed patches and host receipts. React tests exercise canonical Chat composition, preserved draft state, focus, conflicts, conditional Undo and reset/recovery. The disposable browser host uses the real registry, broker, Sources and conversation/form tools with deterministic interactive provider doubles; no production tasks, jobs or external model calls are required.

Children

• Assisted Draft Context Get
• Assisted Draft Propose Patch
• Assisted Draft Reference Search