QwenLM · yiliang114 · Sep 1, 2026 · Aug 5, 2026 · Aug 5, 2026 · Aug 5, 2026
diff --git a/docs/design/experimental-session-plan-review.md b/docs/design/experimental-session-plan-review.md
@@ -8,11 +8,9 @@ and the existing permission lifecycle.
 
 ## Rollout
 
-`experimental.sessionWorkflow` is disabled by default. When disabled, the Web
-Shell keeps the existing Todo list and Plan Mode behavior but does not render
-the Workflow DAG or rename Plan Mode. The setting changes presentation only;
-it does not register tools, alter Todo semantics, or create another approval
-mode.
+`experimental.sessionWorkflow` is disabled by default. When disabled, Core,
+the ACP session, and the Web Shell retain their existing Todo, Agent, approval,
+and split-view behavior. No Workflow marker or revision context is created.
 
 When enabled, the existing `plan` mode is presented as **Plan & Review**. Plan
 Mode remains the execution gate: read-only investigation is allowed, mutating
@@ -40,11 +38,51 @@ approving exits Plan Mode.
   Todo list.
 - Reuse the existing plan ID lineage so later snapshots and Agent executions
   continue updating the same Workflow without another store.
+- Mark Todo output only while the experimental setting and Plan/revision
+  context are both active. The Web Shell ignores ordinary Todo snapshots.
+- After approval, require top-level Agent calls to use a `todo_id` from the
+  approved revision. Nested Agents and ordinary sessions keep existing rules.
 - Fall back to the existing text-only approval when no matching snapshot is
   available.
 
+### Phase 3: current-session Workflow
+
+- Add Workflow to the existing right-panel tab system. Clicking the floating
+  Todo summary or the Session header opens this inspector without replacing
+  Chat.
+- Reuse the active Todo snapshot, daemon task polling, linked Agent tools, and
+  the existing artifact panel instead of introducing another workflow model.
+- Keep a matching `exit_plan_mode` approval in Chat. After approval, the user
+  can open Workflow for observation at any time.
+- Put the progress summary, attention items, ordered steps, selected-step
+  context, recent Agent activity, and deliverables in the inspector. Agent and
+  artifact links open sibling tabs in the same right-panel controller.
+- Keep the dependency graph out of the narrow inspector. An explicit **Expand
+  dependency graph** action opens the existing full-page canvas while the
+  inspector becomes the selected-node detail surface.
+- Keep Chat mounted while either Workflow surface is visible so observation
+  does not interrupt execution or discard composer state.
+- Let a selected step show its upstream and downstream relationships plus the
+  linked Agent's latest activity and runtime metrics. Opening an Agent continues
+  into the existing transcript and artifact panel.
+- Show Session artifacts from the existing artifact store and keep current
+  permission decisions in the existing Chat approval component.
+- Keep the Workflow entry available after completion, later chat turns, and
+  session resume by reading marked Todo snapshots from the transcript. Opening
+  the Session normally restores the entry when a marked snapshot exists.
+- Preserve an active Todo's existing dependencies when an update for the same
+  ID omits `blockedBy`; an explicit empty array removes dependencies.
+
 ## Boundaries
 
 The Workflow remains observational. It does not schedule dependencies, retry
 Agents, propagate completion, or add a Workflow store. `blockedBy` and
-`todo_id` remain optional for sessions outside Plan & Review.
+`todo_id` remain optional for sessions outside Plan & Review. Skill versions
+and a durable decision ledger are not claimed because the current transcript
+does not provide them.
+
+The standalone cockpit mock remains a product reference rather than a second
+application embedded through an iframe. The Web Shell uses a context-preserving
+inspector for routine checks and reserves the full-page surface for the DAG.
+DataWorks-specific scheduling, retry, and approval queues remain with their
+owning product.
diff --git a/docs/design/session-workflow-cockpit-demo.md b/docs/design/session-workflow-cockpit-demo.md
@@ -0,0 +1,57 @@
+# Session Workflow Demo
+
+## Purpose
+
+This branch turns the collaboration-cockpit reference into a native Web Shell workflow inspector and dependency canvas without introducing another workflow engine. Both surfaces are projected from the persisted session transcript and daemon task snapshot.
+
+## Data and control flow
+
+- A `todo_write` snapshot marked inside an enabled Plan/revision context supplies stable Todo IDs, status, content, and `blockedBy` dependencies. Ordinary Todo snapshots are ignored by Workflow.
+- Agent calls are linked to a Todo through `todo_id`; daemon task snapshots supply live status, activity, usage, and persisted transcript/output paths.
+- `exit_plan_mode` remains the execution gate. When the experimental Session Workflow setting is enabled, its revision-bound approval stays in the existing Chat approval flow.
+- Approval uses the existing permission API. Workflow does not schedule, pause, or persist tasks itself.
+- The floating Todo summary and Session header open Workflow as a tab in the existing right panel. Chat remains visible and usable.
+- The inspector shows the summary, items needing attention, ordered steps, selected-step dependencies, Agent activity, and deliverables. Agent and deliverable actions open sibling tabs in the existing panel.
+- **Expand dependency graph** opens the dedicated canvas. Graph selection stays synchronized with the inspector's selected step.
+- `?view=cockpit` keeps the dependency canvas directly addressable and browser navigation returns to Chat.
+- A completed Session restores its Workflow entry from the marked Todo snapshot when the Session is opened normally.
+
+## Demo
+
+From the repository root, start the daemon and Web Shell in separate terminals. Generate the daemon token in the first terminal and copy the printed value for the browser URL below:
+
+Terminal 1:
+
+```bash
+export QWEN_SERVER_TOKEN="$(openssl rand -hex 32)"
+printf 'Demo token: %s\n' "$QWEN_SERVER_TOKEN"
+npm run dev -- serve --port 4293 --workspace "$PWD" --no-web
+```
+
+Terminal 2:
+
+```bash
+QWEN_DAEMON_URL=http://127.0.0.1:4293 npm run dev --workspace @qwen-code/web-shell -- --host 127.0.0.1 --port 5294
+```
+
+Enable **Experimental → Session Workflow Plan & Review**, enter Plan & Review mode, and send:
+
+> Prepare a five-bullet repository orientation. Before doing any inspection, call todo_write with exactly these pending steps and dependency IDs: inspect-readme; inspect-package; compare-findings blocked by inspect-readme and inspect-package; write-summary blocked by compare-findings. Immediately call exit_plan_mode. After approval, launch exactly two Explore subagents in parallel: one reads only README.md and returns at most three bullets; the other reads only package.json and returns at most three bullets. Pass todo_id inspect-readme and inspect-package to the matching Agent calls. Do not run shell commands or edit files. Update the Todo statuses as each phase completes and return at most five bullets.
+
+The cockpit is also directly addressable after the Session exists:
+
+```text
+http://127.0.0.1:5294/session/<session-id>?token=<copied-token>&view=cockpit
+```
+
+Expected flow:
+
+1. The pending `exit_plan_mode` request shows the existing in-flow plan approval in Chat.
+2. Approval starts execution; clicking the Todo summary opens the Workflow inspector without interrupting the conversation.
+3. Selecting a step shows its dependencies and linked Agent; selecting the Agent opens its persisted transcript in a sibling right-panel tab.
+4. **Expand dependency graph** opens the full canvas, where selecting a graph node updates the inspector detail.
+5. Returning to Chat keeps the inspector available. The header or Todo summary reopens the completed workflow later, and `?view=cockpit` opens the graph directly.
+
+## Deliberate boundary
+
+The reference design's synthetic organization-wide queues, policy engine, scheduler, Skill version catalog, and durable decision ledger are not reproduced. “待我处理” is derived from real failed/cancelled Agent tasks; permission requests remain in Chat, so the page does not claim data the daemon does not provide.
diff --git a/packages/acp-bridge/src/status.ts b/packages/acp-bridge/src/status.ts
@@ -177,6 +177,7 @@ export const SERVE_CONTROL_EXT_METHODS = {
   workspaceAgentGenerate: 'qwen/control/workspace/agents/generate',
   workspaceGenerationStart: 'qwen/control/workspace/generation/start',
   workspaceGenerationCancel: 'qwen/control/workspace/generation/cancel',
+  workspaceSessionWorkflow: 'qwen/control/workspace/session-workflow',
   workspaceMemoryRememberAvailability:
     'qwen/control/workspace/memory/remember/availability',
   workspaceMemoryRemember: 'qwen/control/workspace/memory/remember',

diff --git a/packages/acp-bridge/src/transcript-replay.test.ts b/packages/acp-bridge/src/transcript-replay.test.ts
@@ -1422,6 +1422,7 @@ describe('createTranscriptReplayMachine', () => {
           resultDisplay: {
             type: 'todo_list',
             planId: 'plan-1',
+            sessionWorkflow: true,
             todos: [
               {
                 id: 'ship',
@@ -1447,6 +1448,7 @@ describe('createTranscriptReplayMachine', () => {
         },
       ],
       _meta: {
+        qwenSessionWorkflow: true,
         stats: {
           promptTokens: 5,
           candidateTokens: 3,

diff --git a/packages/acp-bridge/src/transcript-replay.ts b/packages/acp-bridge/src/transcript-replay.ts
@@ -141,6 +141,7 @@ export interface TranscriptTodoItem {
 
 export interface TranscriptTodoPlan {
   readonly planId?: string;
+  readonly sessionWorkflow?: boolean;
   readonly todos: TranscriptTodoItem[];
 }
 
@@ -903,6 +904,14 @@ class DefaultTranscriptReplayMachine implements TranscriptReplayMachine {
             ...meta,
             planToolCallId: callId,
             todoPlanId: plan.planId,
+            ...(plan.sessionWorkflow
+              ? {
+                  extra: {
+                    ...meta.extra,
+                    qwenSessionWorkflow: true,
+                  },
+                }
+              : {}),
           }),
         );
       }
@@ -1534,6 +1543,9 @@ function extractTodoPlanFromDisplay(value: unknown): TranscriptTodoPlan | null {
           ...(typeof value['planId'] === 'string'
             ? { planId: value['planId'] }
             : {}),
+          ...(value['sessionWorkflow'] === true
+            ? { sessionWorkflow: true }
+            : {}),
           todos: normalizeTodos(value['todos']),
         }
       : null;
@@ -1548,6 +1560,9 @@ function extractTodoPlanFromDisplay(value: unknown): TranscriptTodoPlan | null {
           ...(typeof parsed['planId'] === 'string'
             ? { planId: parsed['planId'] }
             : {}),
+          ...(parsed['sessionWorkflow'] === true
+            ? { sessionWorkflow: true }
+            : {}),
           todos: normalizeTodos(parsed['todos']),
         }
       : null;