Proposal: confirmation-gated TaskMarket requester toolset
Haystack's Toolset abstraction already supports cohesive groups of tools and
dynamic external services. A TaskMarket toolset would give Haystack agents a
real delegation path when a request is better sent to external workers rather
than handled by repeated inference or local computation.
The proposed opt-in TaskMarketToolset would provide:
- read-only discovery of live tasks and task status;
- submission retrieval for human review;
- a read-only preview containing the exact description, reward, deadline,
Base/USDC network, and maximum spend; and
- a create tool that requires a matching preview token, an explicit boolean,
and an application-provided approval callback before calling the official
TaskMarket CLI once.
The implementation would keep wallet custody in the first-party CLI, never
accept private keys or tokens, verify Base/USDC configuration and balance
before the write, avoid shell interpolation, and report an unknown result after
a timeout without retrying. It would include focused tests, serialization of
the endpoint descriptor, documentation, and a release note.
This is intentionally an optional toolset and would not change existing
Haystack agent behavior. I would appreciate feedback on whether the API belongs
in core haystack.tools or should be maintained as an external integration.
Proposal: confirmation-gated TaskMarket requester toolset
Haystack's
Toolsetabstraction already supports cohesive groups of tools anddynamic external services. A TaskMarket toolset would give Haystack agents a
real delegation path when a request is better sent to external workers rather
than handled by repeated inference or local computation.
The proposed opt-in
TaskMarketToolsetwould provide:Base/USDC network, and maximum spend; and
and an application-provided approval callback before calling the official
TaskMarket CLI once.
The implementation would keep wallet custody in the first-party CLI, never
accept private keys or tokens, verify Base/USDC configuration and balance
before the write, avoid shell interpolation, and report an unknown result after
a timeout without retrying. It would include focused tests, serialization of
the endpoint descriptor, documentation, and a release note.
This is intentionally an optional toolset and would not change existing
Haystack agent behavior. I would appreciate feedback on whether the API belongs
in core
haystack.toolsor should be maintained as an external integration.