feat(assistant): improve document workflows and request feedback

CHANGELOG.md

@@ -22,6 +22,18 @@ notes. Entries describe what users experience, not internal refactors.
 
 ### Added
 
+- **Start with a task for the document in front of you.** Assistant suggestions
+  follow the editor and selection. Choose a task to prepare an editable request,
+  then send. If the document scope changes, your draft is kept for you to confirm
+  again.
+- **Read drafts and review changes more clearly.** Document drafts and incomplete
+  responses have distinct labels. Change reviews emphasize Apply and Cancel,
+  with editing, refinement, copying and location in More actions.
+
+- **See what the assistant is doing while you wait.** Request understanding,
+  generation and review show their actual stage. A compact elapsed-time hint
+  clears when content arrives or you stop, while your next draft stays editable.
+
 - **Find earlier assistant conversations.** Search titles and messages locally,
   browse recent chats by date and see the current conversation at a glance.
   Compact row actions confirm which conversation will be deleted; history

docs/explorations/2026-10-14-assistant-workspace.md

@@ -0,0 +1,105 @@
+# Contextual assistant workspace
+
+## Product direction
+
+The assistant should help complete the current document task. A generic welcome,
+undifferentiated replies and a row of equally prominent review buttons made the
+next step harder to identify. This iteration improves task discovery, reading
+and the existing review decision together.
+
+Contextual tasks stage an editable request. Explicit Send remains the execution
+boundary, and Apply remains the document-change boundary. A host-selected task
+can carry its known intent, avoiding a separate classification request when its
+text, session and original editor scope still match. Arbitrary typed requests
+continue through normal routing.
+
+Result identity follows host metadata: ordinary answers, document drafts and
+interrupted output retain their separate meaning. Presentation labels never
+become part of copied or applied source content. Review gives Apply and Cancel
+priority, with Edit, Refine, Copy and Locate in an accessible disclosure.
+
+## Architecture and boundaries
+
+The panel owns submission, preparation, consent and controller/session state.
+Task entry owns transient staging provenance; existing bounded context capture
+and SDK event subscriptions remain authoritative. ChatView remains independent
+of the editor. ActionPreview retains target validity, confirmation and truthful
+sent/verified receipts.
+
+No provider default, persistent history format or editor mutation contract is
+replaced. Local inference continues to use existing WebGPU support. Visual
+feedback uses CSS/SVG and real application state; no simulated reasoning or
+decorative GPU renderer is introduced.
+
+## Verification
+
+Specification and code-quality reviews approved each functional chunk. The
+integration checks passed 5,448 unit tests in 187 files, `pnpm run lint:ts`
+(including workspace and application TypeScript), and `pnpm run build`.
+
+Reverse regressions established that the tests detect missing task staging,
+extra classification on trusted tasks, duplicate context capture, incorrect
+Chinese writing intent, missing artifact identity and invisible keyboard focus.
+Review tests fail against the earlier primary/secondary action layout. Progress
+tests detect a wall-clock jump, a timer restarting after renderer-only clear,
+and direct task answers retaining the planning phase.
+
+Browser acceptance covers the actual OnlyOffice editor with controlled provider
+and speech events. It checks shared leading edges, compact controls, light and
+dark themes, 320/432/520-pixel sidebars, narrow Chinese feedback, reduced motion,
+keyboard focus, draft preservation, Stop/retry, review and native Undo. Progress
+phase changes move the composer by at most one pixel; dictation checks stable
+composer and message geometry. The clock is outside live regions so a screen
+reader receives phase changes without per-second announcements.
+
+The initial combined run passed 34 cases; four failed while saving screenshots
+because the disk was full. After scoped cleanup of ignored generated outputs,
+all four reruns passed. The final combined Chromium run passed all 40 cases
+with one worker, including Stop/resume, history, dictation, writing, context
+events, layout and new process feedback. The repository format check passed;
+35 final screenshots were preserved and inspected across the changed surfaces.
+
+Physical microphone behavior, manual screen-reader operation, high browser zoom
+and model quality or first-token latency on the user's hardware are not verified
+by these controlled checks. No measured model speedup is claimed.
+
+## Browser-local latency follow-up
+
+Source inspection establishes that ordinary non-native requests first await
+`resolveRequestIntent` and then start the answer or document task. Trusted staged
+tasks avoid that first inference. The ordinary chat controller already forwards
+streamed text; its local preflight supplies no native tools, so the runtime's
+tool-prose deferral does not explain this path's first-content delay.
+
+`captureRequestContext` attaches readable current-document data, capped at 8,000
+characters, even for ordinary questions. Writing transformations await a complete
+JSON body and validation before publishing the review. These are concrete places
+to measure classification time, input processing and first visible output. This
+inspection does not establish the user's actual model, GPU/CPU backend, prompt
+token count or latency.
+
+The requested process feedback identifies actual host phases and elapsed waiting;
+it does not claim faster model computation or invent a reasoning transcript.
+Prompt compression, selective context and validated streamed writing remain
+separate performance interventions requiring quality and scope regressions.
+
+## Overall experience assessment
+
+The next action is clearer at each boundary: choose a document task, edit and
+send its request, read the answer or draft, review the change, then apply and
+undo if needed. Task selection does not execute automatically. Interrupted
+content cannot acquire completed-draft authority. Existing history and voice
+controls remain compact and preserve the editing surface while their state
+changes.
+
+The largest remaining opportunity is time to useful content for ordinary
+browser-local questions. Measure intent classification, document input
+processing and answer generation separately before changing defaults. A scoped
+read-only path could avoid repeated inference, but it must explicitly disable
+mutation tools and preserve target, consent and stream validity checks; simply
+routing document tasks as ordinary chat would lose those guarantees.
+
+A smaller recovery opportunity is an expired staged task after changing model
+or document scope. The current behavior safely preserves the draft and requires
+confirmation. A later explicit recovery action could retain that draft and make
+reselection easier without silently retargeting or sending it.

docs/superpowers/plans/2026-10-14-assistant-workspace.md

@@ -0,0 +1,90 @@
+# Assistant Workspace Implementation Plan
+
+> **For agentic workers:** REQUIRED: Use superpowers:subagent-driven-development. Steps use checkbox syntax for tracking.
+
+**Goal:** Make document tasks discoverable and make generated results and reviewed changes clearer, within one new PR.
+
+**Architecture:** The panel retains submission and consent ownership. An app-owned task module stages bound request intent, ChatView improves generic result rendering, and ActionPreview improves its existing review composition. No persistence or model-provider contract is replaced.
+
+**Tech Stack:** TypeScript, ranui builders/components/tokens, Vitest/jsdom, Playwright and the existing OnlyOffice editor bridge.
+
+**Specification:** `docs/superpowers/specs/2026-10-14-assistant-workspace-design.md`.
+
+## Chunk 1: Contextual task entry and routing
+
+Files: new `lib/agent-plugin/ui/task-entry.ts`; integration in `panel.ts` and
+`panel-context.ts`; task-entry styles in `styles/assistant.css`; locale tables
+and `i18n/types.ts`; unit tests and `test/e2e/agent-workspace.spec.ts`.
+
+- [x] Write failing tests for format/selection suggestions, no automatic sends,
+      draft preservation, expired scope and edited-request fallback.
+- [x] Run focused Vitest; verify behavioral failures before implementation.
+- [x] Build the contextual task entry using existing bounded context reads.
+      Stage `{ text, intent, target, session }` only on a user task click; consume
+      after explicit Send and fresh scope checks. Preserve preparation/consent.
+- [x] Integrate known intent into routing without altering arbitrary typed
+      requests, native tools, literal commands, cancellation or refinement.
+- [x] Verify a ready staged writing task makes zero classification calls and
+      one writing call. Add real Word flow through native change and Undo.
+- [x] Commit as `feat(assistant): add contextual tasks and direct request routing`.
+
+## Chunk 2: Conversation and result hierarchy
+
+Files: `packages/chat-ui/src/{chat-view,styles,types}.ts`; existing chat-view
+unit suites; relevant browser result/streaming tests.
+
+- [x] Write failing result tests: artifact identity, interrupted/restored
+      semantics, original source for copy/apply, accessible actual activity.
+- [x] Run red tests, then improve typography, document-draft framing, compact
+      actions and truthful streaming/activity states through optional labels.
+- [x] Preserve editor independence and all existing public method behavior.
+      Coordinate any new localized labels with the task-entry owner.
+- [x] Verify long Markdown/table/code, keyboard focus, draft editing, Stop and
+      reader scroll position. Honor reduced motion and narrow layouts.
+- [x] Commit as `feat(chat-ui): clarify drafts and conversation results`.
+
+## Chunk 3: Review and completion
+
+Files: `lib/agent-plugin/ui/action-preview.ts`, review styles only in
+`styles/assistant.css`, action-preview units and content-review E2E.
+
+- [x] Add failing tests for clear primary/secondary action hierarchy and
+      accessible secondary controls; retain all existing review behaviors.
+- [x] Improve target/comparison composition and action grouping using ranui.
+      Use actual review data and actual execution outcomes. Preserve expiration,
+      cancellation, refinement, current-target checks and draft editing.
+- [x] Verify Apply/Cancel/Edit/Refine/Copy/Locate, selection invalidation,
+      sent versus verified outcomes, error states and restored focus.
+- [x] Run native Word review/application/Undo E2E and screenshot validation.
+- [x] Commit as `feat(assistant): refine change review and completion`.
+
+## Chunk 4: Immediate process feedback
+
+The user requested visible process while browser-local inference is waiting.
+ChatView owns presentation and elapsed-time cleanup; the host supplies the
+actual phase and translated labels. Existing provider and history contracts stay
+unchanged. A classification phase is only shown when classification runs.
+
+- [x] Add failing tests for immediate phase feedback, measured elapsed waiting,
+      first streamed content, review resting state and Stop/disposal cleanup.
+- [x] Add an explicit routing phase and localized labels. Keep elapsed text out
+      of repeated live announcements and preserve stable composer geometry.
+- [x] Validate a held endpoint request before and after its first content delta,
+      Stop/retry and focus. Do not claim real-device model latency improvement.
+- [x] Commit as `feat(assistant): show immediate request progress`.
+
+## Chunk 5: Integration and delivery
+
+- [x] Spec compliance review, then code-quality review for each functional unit.
+      Fix verified findings and check final consumers/state owners.
+- [x] Integrate functional commits; keep root and subagent checkouts isolated.
+- [x] Inspect final light/dark/narrow screenshots and measure shared alignment
+      and stable composer geometry; record unverified acceptance surfaces.
+- [x] Update changelog and concise exploration record. Run reverse regressions.
+- [x] Run full unit suite, format, lint/typecheck and production build, then
+      selected real-editor E2E including Stop, history and dictation regressions.
+- [ ] Fetch latest main, verify clean checkout and ahead/behind, push the feature
+      branch, create/attach one new PR, and inspect checks for the pushed SHA.
+
+This checklist is committed before publication. The PR records the final branch
+revision and subsequent CI status.

docs/superpowers/specs/2026-10-14-assistant-workspace-design.md

@@ -0,0 +1,104 @@
+# Contextual document assistant workspace
+
+## Approved direction and evidence
+
+The user approved implementing the overall UX/design direction discussed in the
+conversation: document-aware entry points, clearer context, stronger answer and
+review hierarchy, less uninformative waiting, and direct completion in the editor.
+Use one new PR from current main, with functional commits. PR #268 is merged.
+
+Current code facts:
+
+- `panel-context.ts` and `composer-context.ts` already own bounded selection reads
+  and event-driven refresh. The welcome view currently offers text hints only.
+- `panel.ts` routes ready non-native requests through `resolveRequestIntent`
+  before generating an answer or writing proposal. Exact host-selected tasks
+  can carry their known intent without this extra model classification.
+- `ActionPreview` already owns target validity, comparison, confirmation,
+  cancellation, refinement and verified/sent receipts. Improve its composition;
+  retain these execution safeguards.
+- `ChatView` is editor-independent. The host marks document artifacts; generic
+  rendering must never infer write permission from prose or show interrupted
+  results as completed artifacts.
+
+No claim is made about actual model latency or answer quality without measurement.
+
+### Follow-up: immediate process feedback
+
+The user reports long waits before browser-local answers and requests visible
+process feedback. Show actual host phases immediately: request classification,
+reading content and generating an answer. Keep human review a resting state.
+Elapsed waiting time is measured locally, does not announce every tick to screen
+readers, and ends on completion, Stop, failure, view disposal or a new turn.
+Streaming text still appears as it arrives. Do not expose simulated reasoning,
+invent percentages or claim that process visibility speeds up model computation.
+Keep provider settings, inference and history contracts unchanged.
+
+## Experience
+
+1. Open the assistant to see the current document scope and a compact set of
+   useful tasks for Word, Excel, PowerPoint or PDF. Selected text changes the
+   suggestions. Empty, oversized/unavailable selection and read-only contexts
+   must remain honest about scope and supported operations.
+2. Clicking a task stages an editable request and focuses the composer. It does
+   not invoke a model, download a model, overwrite a nonempty draft, or queue work.
+   Explain the staging interaction briefly. Explicit Send starts the operation.
+3. A staged task is host-owned metadata, consumed once only if the request and
+   captured session/editor scope still match. Edited requests use normal routing;
+   stale selection/context gives visible guidance and preserves the request.
+   Never skip classification solely by matching arbitrary user-authored text.
+4. Preserve current preparation, endpoint consent and local writing consent.
+   Ready staged writing tasks avoid exactly one routing model call. Prove this
+   with request counts, not an invented percentage or timing promise.
+5. Distinguish ordinary answers, document drafts, interrupted output and activity
+   with restrained typography, surfaces and labels. Main content remains readable;
+   copy and apply retain the original source and completion guards.
+6. A pending review shows its actual target and original/proposed changes, with
+   one clear Apply decision, Cancel, and a discoverable secondary-actions area.
+   Preserve Edit, Refine, Copy and Locate. Completion shows only the actual
+   verified/sent outcome, location and existing Undo guidance.
+
+## Scope and ownership
+
+Allowed: assistant UI modules and integration in `lib/agent-plugin/ui/panel.ts`,
+`styles/assistant.css`, generic presentation in `packages/chat-ui/src`, seven
+locale tables/types, relevant tests, changelog and design/verification records.
+
+Ownership: panel coordinates submission/consent/session state; a dedicated task
+entry module owns staged-task provenance and UI; existing context reads remain
+single-owned; ChatView owns rendering; ActionPreview owns review state.
+
+Compatibility boundaries: preserve storage formats, provider defaults, explicit
+Send, experimental local-writing consent, cloud consent, read-only behavior,
+selection bounds, model/tool cancellation, target validity and truthful receipts.
+No vendor edits, new renderer dependency, GPU visual framework, background model
+calls, synthetic thinking, fabricated audio waveforms or automatic document writes.
+Existing browser-local inference continues using WebGPU where supported.
+
+## Visual acceptance
+
+Use the repository's ranui components/tokens. GPUI-native APIs do not apply to
+this TypeScript browser UI. Common leading edges, a stable composer and compact
+secondary chrome give the document/result space priority. Use CSS/SVG for state
+transitions, honor reduced motion, and retain keyboard operation/focus restoration.
+
+Check real editor layouts at 320px sidebar width and default/wide widths, light
+and dark themes, long/CJK content, empty/loading/error/read-only states. Review
+actual screenshots. Avoid generic welcome screens repeated for every format.
+
+## Acceptance
+
+- Context suggestions update through existing SDK events and perform no model
+  call until explicit Send; drafts survive setup, Stop and state changes.
+- Host-selected writing request: classification count decreases from one to
+  zero; writing still executes once, then waits for review. Edited/stale/session
+  switched requests cannot reuse a privileged or stale task hint.
+- Real Word: select text, stage polish, Send, review old/new, Apply, verify native
+  text and Undo. Exercise a follow-up refinement and Stop/retry.
+- Excel/PPT/PDF suggestions describe only available scope; read-only/oversized
+  selection does not promise unsupported mutation.
+- Copy/apply/interrupted/restored result semantics and history/speech regressions
+  remain passing. Reverse-disable new behavior to prove regressions fail.
+- Full unit suite, format, lint/typecheck, production build and appropriate real
+  editor E2E pass before push/PR. Physical model quality and microphone accuracy
+  remain separate acceptance surfaces.