yuta/twitter-lite
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: streamline conversation workflow and add Jev reply reviews

Browse Source
This commit is contained in:
yuta
2026-09-28 23:45:07 +09:00
parent 54d7a542f7
commit e51cafa570
14 changed files with 621 additions and 135 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/execution-support.md
+21 -2
View File
@@ -39,8 +39,9 @@ the new communication and journal prototypes do not consume those accounts.
`/` presents a greeting, a large input, and a Tiimo-inspired flexible day flow grouped as
Start here, Routines, and If there’s room. There are no fixed time slots.
Tasks can be reordered, completed, deferred, restored, or followed as routine
steps. Communications can be reviewed directly in a home dialog; changes also
appear in `/support`. Journal entries and corrections appear on the home, and
steps. Communications open their Messages page at
`/support?view=all&conversation=<id>`; browser Back returns to the home with
shared outcomes intact. Journal entries and corrections appear on the home, and
reader samples open for reading. Research remains a separate work surface.
Activity blocks expand in place into controls suited to their content: routine
@@ -79,3 +80,21 @@ The home greeting includes a short prose brief derived from current task and
communication state. Completed or deferred items leave the brief; the next
contact opens directly from the text. A current resumption note takes precedence
over the optional-task suggestion. This is local templated copy, not an LLM report.
## Jev reply review
The message thread keeps the reply composer at the bottom on wide screens;
summary and source context live in the side panel. Expand Reply suggestions to
compare candidate wording before editing.
Opening Review reply calls an authenticated server function using AI SDK
`experimental_evaluate` with `typesafe-ai/jev`. Set `AI_GATEWAY_API_KEY` on
the server. The draft, recipient role, and original sample messages are sent
through Vercel AI Gateway to TypeSafe AI. The model evaluates whether the reply
answers the request, stays grounded, and uses an appropriate tone. The UI shows
boolean probabilities, not a calibrated pass/fail verdict or generated advice.
Editing the draft invalidates its evaluation. Missing configuration and provider
failures show an unavailable state; no mock approval is substituted. Review
results do not trigger sending, which remains a simulation.
See [Vercel’s Jev integration](https://vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gateway).