Chat suggestions are a fixed list per page. They don't follow from the reply just given, so they're often irrelevant or repeat what was already asked. The current stopgap only hides exact repeats.
Wanted
After each assistant turn finishes, a separate model call (its own context, not the chat agent) proposes follow-up suggestions.
Input: the last user message and the last assistant reply (trimmed, tool output and code blocks dropped), a one-line list of earlier user messages so it doesn't repeat them, and the current page with any selected entity (e.g. the task and its status).
Guidance through the prompt, kept in internal/prompts: tell it what Archie can do on that page and ask for follow-ups that act on the reply. No hard allowlist in code.
Output: JSON, at most three short suggestions. An empty list is valid and is preferred over generic filler.
Runs on the cheap default model, and doesn't add latency to the reply.
Suggestions are stored with the turn, so reopening the chat doesn't regenerate them.
The static per-page prompts stay, but only for an empty conversation.
Done when
After a reply, the chips reflect it (e.g. offer to retry or archive what the reply discussed), they never repeat a question already asked, and they show nothing when nothing useful fits.
## Problem
Chat suggestions are a fixed list per page. They don't follow from the reply just given, so they're often irrelevant or repeat what was already asked. The current stopgap only hides exact repeats.
## Wanted
After each assistant turn finishes, a separate model call (its own context, not the chat agent) proposes follow-up suggestions.
- Input: the last user message and the last assistant reply (trimmed, tool output and code blocks dropped), a one-line list of earlier user messages so it doesn't repeat them, and the current page with any selected entity (e.g. the task and its status).
- Guidance through the prompt, kept in `internal/prompts`: tell it what Archie can do on that page and ask for follow-ups that act on the reply. No hard allowlist in code.
- Output: JSON, at most three short suggestions. An empty list is valid and is preferred over generic filler.
- Runs on the cheap default model, and doesn't add latency to the reply.
- Suggestions are stored with the turn, so reopening the chat doesn't regenerate them.
- The static per-page prompts stay, but only for an empty conversation.
## Done when
After a reply, the chips reflect it (e.g. offer to retry or archive what the reply discussed), they never repeat a question already asked, and they show nothing when nothing useful fits.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Problem
Chat suggestions are a fixed list per page. They don't follow from the reply just given, so they're often irrelevant or repeat what was already asked. The current stopgap only hides exact repeats.
Wanted
After each assistant turn finishes, a separate model call (its own context, not the chat agent) proposes follow-up suggestions.
internal/prompts: tell it what Archie can do on that page and ask for follow-ups that act on the reply. No hard allowlist in code.Done when
After a reply, the chips reflect it (e.g. offer to retry or archive what the reply discussed), they never repeat a question already asked, and they show nothing when nothing useful fits.