Skip to content

docs: conversational custom-order replies + per-turn credits #5

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
mintlify wants to merge 1 commit into
base: main
Choose a base branch
from
Open

docs: conversational custom-order replies + per-turn credits #5

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 24 additions & 4 deletions custom-playlist-order.mdx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,9 +71,22 @@ a **feedback** turn:
- "Use Tokyo time, not UTC."
- "Pick the bumpers in shuffled order, not in sequence."

Each refinement produces a **new immutable version** (**v2**, **v3**, …) with its own dry-run
and diagnostics. Past versions are kept so you can see how the program evolved, and the
conversation transcript shows every intent, feedback turn, and the assistant's reply.
The assistant **always replies with a text message**. Depending on what you asked, that turn
can:

- **Produce a new version** — the assistant generates a fresh program (**v2**, **v3**, …) with
its own dry-run and diagnostics, and the message explains what it changed.
- **Ask a clarifying question** — when the intent is ambiguous (which item is the "intro"?
which timezone?), the assistant replies with a single question and waits for your next turn
before producing anything.
- **Decline** — if the rule isn't something a custom order can safely express (network calls,
external data, behavior the sandbox can't support), the assistant explains why in the reply
and no new version is created.

Only turns that produced a **Ready** version show a **Preview** and **Apply** in the
transcript — clarifying questions and declines are conversation-only and don't add a version
to the program's history. Past versions are kept so you can see how the program evolved, and
the transcript shows every intent, feedback turn, and the assistant's reply.

<Note>
Versions are immutable: you can't edit an existing version's code, only create a new one by
Expand Down Expand Up @@ -195,7 +208,14 @@ them.
- Each pick has a small CPU budget (a few tens of milliseconds). A program that loops or stalls
hits the budget, faults, and falls back.
- Programs are **per-user** — you can't share, fork, or import another user's program.
- The conversation budget is per user per day; very heavy refinement bursts may be rate-limited.
- Each conversation turn — the first intent or any refinement — charges a **fixed amount of
credits** from your wallet up front (the exact price is set by Streamloop and shown in the
composer before you submit). The charge stands whether the assistant produced a new version,
asked a clarifying question, or declined — a real reply is a real reply. If the turn faults
on Streamloop's side (an LLM or system error, not a decline), the credits are **refunded
automatically**, exactly once.
- If your wallet doesn't have enough credits to cover the next turn, the request fails with
`INSUFFICIENT_CREDITS` and nothing is charged. Top up your wallet and try the turn again.

## Next steps

Expand Down