fix(ai): tighter agent timeline — thread the rail, fold MCP approval into its run-node

[ndemianc]

Two changes that make the agent activity timeline read as one clean thread — most visible on an MCP run, which is what prompted this.

1. Thread the rail across the row gap

The timeline rail is drawn per-row (.tl-rail::before spans one row), but #log stacks rows with a gap: 12px flex gap — so there was 12px of dead space between every row and the rail rendered as disconnected stubs ("cut in the middle"). Now a row that directly follows another timeline row extends its rail up to meet it:

#log > .tl + .tl > .tl-rail::before { top: -12px; }

A run of tool / approval / group nodes reads as one connected line (GitHub / Cursor / Claude style); a narration bubble in between still breaks it (intended). Scoped to #log's direct children, so rows nested inside a collapsed group keep their own contiguous rail.

2. Fold an approved MCP call into one row

A manually-approved MCP tool call rendered two rows — an Approved · server · tool chip and a separate 🔌 server · tool node. The run-node now folds into the approval chip so each call is one row (Copilot-style, one-row-per-action):

• agent.js tags the run-node kind: 'mcp'.
• addMcpApproval hands its row to the incoming node only when approved.
• addAgentLine folds a kind==='mcp' node into that pending row, and clears the window otherwise.

Edge cases preserved: allow-listed calls were already one row (no card); skipped calls keep their Skipped chip (nothing ran); only the approved / always-allowed pair folds. The folded row shows 🔌 server · tool (matching allow-listed calls) rather than the approval verb — the persistent allow-list effect lives in settings.

Tests

• Two new webviewCss static invariants: the rail-bridge offset stays equal to #log's gap (no drift), and the MCP fold's three-part handshake stays wired end to end (agent.js → addMcpApproval → addAgentLine).
• Verified both files parse (agent.js + the chat.html webview <script>); full extension suite 27 suites, 0 failures.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR refines the LevelCode AI webview’s agent activity timeline so MCP activity reads as a single, continuous “thread” (rail continuity + one-row-per-approved-MCP-call), and adds CSS/JS invariants to prevent regressions.

Changes:

• Extend the timeline rail across #log’s flex gap between consecutive top-level timeline rows.
• Fold an approved MCP tool call’s run-node into the approval row (so it renders as one timeline row instead of two).
• Add regression tests for the rail-gap offset sync and the MCP folding handshake; add a new design doc for chat sessions.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
extensions/levelcode-ai/media/chat.html	Adds the rail-bridging CSS rule; introduces mcpMergePending and folds approved MCP run-nodes into the approval row.
extensions/levelcode-ai/agent.js	Tags MCP tool run-node posts with kind: 'mcp' so the webview can detect and fold them.
extensions/levelcode-ai/test/webviewCss.test.js	Adds invariants to keep the rail gap/offset in sync and ensure the MCP fold wiring remains intact.
docs/levelcode-chat-sessions-design.md	Adds a proposed design document for persistent chat sessions (not referenced in the PR description).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (1)

extensions/levelcode-ai/media/chat.html:1876

• The MCP fold can incorrectly consume an agentTool row if mcpMergePending points at a detached element (e.g. the user triggers “New chat” / the host posts {type:'reset'} after approval but before the run-node arrives). In that case addAgentLine returns early after mutating a node that’s no longer in #log, so the tool row never renders. Guard the fold with log.contains(mcpMergePending) so a stale ref falls through to normal rendering and gets cleared.

    if (kind === 'mcp' && mcpMergePending){
      mcpMergePending.className = 'tl tl-tool';
      mcpMergePending.innerHTML = rowHtml;
      mcpMergePending = null;
      scrollIfStuck();