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:

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:

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:

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.