--page-all returns only the first page for cursor-paginated endpoints, with has_more true and exit code 0

[BenHuss360]

Version

elevenlabs 1.1.0 (npm @elevenlabs/cli@1.1.0), Linux x64.

Summary

--page-all is documented as "Auto-paginate and stream all results as NDJSON". On endpoints that paginate by cursor it emits the first page only, leaves has_more: true in the output, and exits 0. Nothing signals that the result set is incomplete.

Reproduction

Workspace with 294 voices:

Invocation	Voices returned	has_more	Exit
voices search --page-all	10	true	0
voices search --page-size 100 --page-all	100	true	0
voices search --page-size 100 --page-all --page-limit 30	100	true	0

--page-limit makes no difference, so it is not being clamped.

Not specific to voices. history list --page-all behaves the same way: 100 items, has_more: true, exit 0.

Underlying pagination works when driven by hand, so this is --page-all not following the cursor rather than a broken endpoint:

$ elevenlabs voices search --page-size 100 --next-page-token "<token from page 1>" --format json | jq -c '{returned:(.voices|length), has_more}'
{"returned":100,"has_more":true}

Scope

This looks like it affects most of the API surface. In reference.md, 36 commands expose --next-page-token or a cursor flag, against 1 exposing --offset. If --page-all only implements offset pagination, then it silently under-reports nearly everywhere it can be used.

Expected

Follow next_page_token while has_more is true, up to --page-limit. Failing that, a non-zero exit or a warning on stderr when the stream ends with has_more: true would at least make the truncation visible.

Impact

Silent truncation with a success exit code is difficult to catch downstream. We reached for this to audit stored voice IDs against the workspace, and taken at face value it presents 194 of 294 live voices as absent.

[tech-aria]

Confirming this on elevenlabs 1.1.0 (npm @elevenlabs/cli@1.1.0), macOS arm64. Two things to add.

1. It is confirmed at the network level, not just in the output. Using --debug to count outgoing requests, --page-all emits exactly one HTTP request in every combination — so the truncation happens before any output formatting, and --query / --format / --no-extract are not involved:

Invocation	HTTP requests emitted
--page-size 100	1
--page-size 100 --page-all	1
--page-size 100 --page-all --page-limit 5	1
--page-size 100 --page-all --no-extract	1

2. It also affects the Conv AI endpoints, in addition to the voices and history cases above. Forcing pagination with --page-size 1 on three different endpoints, all with a cursor available in the response:

Command	Items returned	has_more	Cursor in response	Requests with --page-all
agents conversations list --page-size 1 --page-all	1	true	yes	1
agents list --page-size 1 --page-all	1	true	yes	1
history list --page-size 1 --page-all	1	true	no	1

--page-size is honoured (one item comes back), so there is definitely more to fetch and the cursor is right there in the response.

Note on the manual workaround: driving --cursor by hand does work on agents conversations list (page 2 came back with has_more: false, and the totals matched a hand-rolled curl loop). But history list returns has_more: true with no cursor field in the response, so the workaround does not apply uniformly — worth keeping in mind if the fix is generic.

Agreed that a non-zero exit or a stderr warning when a stream ends on has_more: true would be valuable on its own, independently of the cursor fix. Silent truncation with exit 0 is the part that is hard to defend against downstream.