Thinking Path

Issue #503 identifies that the cost/token display attached to the last assistant message shows cumulative session totals (S.session.input_tokens, S.session.output_tokens). This means the number attached to message 5 is the total cost of the entire conversation, not just that turn — which reads as misleading.

The fix: compute a per-turn delta from the cumulative usage delivered in the done SSE event, store it on the message object, and render it individually on each assistant bubble.

What Changed

static/messages.js

• At the done SSE event handler, when d.usage arrives (cumulative totals), compute the delta vs the previous session totals (S.session.input_tokens etc.)
• Store the delta as _turnUsage on the last assistant message: {input_tokens, output_tokens, estimated_cost}
• Only sets _turnUsage if values actually increased (skips no-op turns)

static/ui.js

• Replaced the "cumulative usage on last assistant bubble" block with a per-turn loop
• Iterates over S.messages, finds assistant messages with _turnUsage, and renders inline usage on each one's footer
• Uses a separate ai counter to correctly map message indices to .assistant-turn DOM rows (since user messages don't create assistant-turn rows)
• The context ring / composer footer still shows cumulative totals via _syncCtxIndicator() — no information is lost

Backwards compatibility

• Messages loaded from old sessions won't have _turnUsage — they simply won't show inline usage, which is correct (we don't have per-turn data for them)
• The _showTokenUsage setting still gates the feature

Closes #503