Stream chat messages (Rebased)

[kevinthai256]

Summary

• wire the tutoring agent + provider clients for token response streaming with fallback to non-stream responses
• update the frontend chat flow to consume SSE incrementally, render live assistant text, and finalize with persisted metadata (artifacts, actions, widgets)

Test plan

• Start backend and frontend locally
• Send a normal message and verify tokens stream progressively in the chat UI
• Confirm final message replaces the placeholder and includes widgets/artifacts when returned
• Simulate a slower first token and verify the stream stays open (no premature disconnect)
• Verify non-stream fallback still works when SSE content type is not returned

[Youdahe123]

Really solid work Kevin. The streaming-with-fallback pattern in client.py is well architected — handles partial output, falls back gracefully, and logs appropriately. The SSE parsing on the frontend is clean too.

Two things:

1. Merge conflicts — needs a rebase onto main.
2. Minor: missing newline at EOF in client.py.

This is a good-to-merge once conflicts are resolved. Nice work on the fallback handling especially — that's production-quality thinking. 👍

[Youdahe123]

Code Review

Cannot merge — all 4 CI checks are failing. Like PR #27, the frontend failures are mostly pre-existing issues unrelated to this PR's changes. The backend changes here are solid.

What's good

• Clean streaming implementation in client.py with proper AsyncIterator return types
• Correct fallback logic: if streaming fails after partial output, it doesn't retry (avoids duplicate tokens); only falls back to non-stream when nothing was emitted
• on_token callback in agent.py is a clean hook — doesn't force callers to care about streaming internals
• Fixing response.choices[0].message.content or "" and the Anthropic content block join are good defensive fixes

Issues to fix

1. Conflict with PR #27 (markdown renderer)

This PR adds MarkdownLine and renderInline inline inside MessageBubble.tsx. PR #27 moves those same functions into frontend/src/lib/markdown.tsx as a shared module. These two PRs will conflict at merge time.

Recommended order: merge PR #27 first, then update this PR to remove the inline MarkdownLine/renderInline definitions and import from ../lib/markdown instead. That way the streaming work lands cleanly without re-introducing the duplication that #27 fixed.

2. Missing newline at end of client.py

The diff shows EOF without a trailing newline on the last line of client.py. Ruff will flag this. Add a newline after the last line.

3. The existing stream() method in client.py

There's already a stream() method in the client. The new chat_stream() appears to overlap in purpose. Before this lands, confirm whether stream() is still used anywhere or if chat_stream() replaces it. If stream() is dead code, remove it to avoid confusion.

4. Pre-existing CI failures

Same pre-existing frontend lint issues as PR #27. These aren't from this PR but block the merge.

Action items

1. Wait for PR Lightweight Markdown Renderer for Chat Messages (Rebased) #27 to merge, then rebase and switch MessageBubble.tsx to import from ../lib/markdown
2. Add trailing newline to client.py
3. Clarify / clean up the stream() vs chat_stream() situation
4. Re-push so CI re-runs

The backend streaming work is well-written — just needs the ordering sorted out with PR #27.

[Youdahe123]

Status

CI: All 4 checks failing. No reviewer assigned.

This PR needs:

1. A human review
2. A rebase on main once fix: backend-lint and backend-test CI (#35) #36 and fix: frontend lint tests ci #38 land — CI failures here are pre-existing lint issues resolved in those PRs

Not mergeable until reviewed and CI is green.