character v0.10.5

No new functionality for us over v0.10.4. This release is to export some MCP functionality for use elsewhere.

character v0.10.4

We now embed usage instructions into the MCP agent so that AIs can use it more fluidly. The skill becomes less important.

character v0.10.3

v0.10.3 improves the MCP handling to use pagination. Some of our queries could return enough data to overflow context. So we revisit the limits and the amount of data in per-rune items in block overviews and pagination, to let Claude handle the results much more smoothly, without needing to resort to local Python to extract from its own history.

The new pagination replaces, for MCP, some maximum display limits intended as a sanity-check before spewing to console. The pagination is designed to survive AI API mishaps stalling, and re-issue queries if entries have aged out of cache. We're fortunate that we're querying over static data, and thus search results should be the same if re-issued.

v0.10.2 was not released (CI mishap, tag poisoned).

We now explicitly warn that character agent sub-commands can change incompatibly between patch releases, and agents are expected to use live introspection at all times, rather than encoding knowledge about how these end-points behave.

character v0.10.1

v0.10.1 refines the MCP support from v0.10.0 with added safe-guards against hallucinated API calling parameters and a Skill file which can be loaded to reduce hallucinations further.

It's not perfect, given a search term which returns many results and exceeds context token count, but Claude recovers and writes the response to a file and then proceeds to efficiently query the file with Python, so it all works.

Also include a Go dependencies bump which was missed from the previous release.

character v0.10.0

This release adds the agent sub-command tree, including character agent mcp to provide a stdio MCP Unicode oracle for agents to mitigate the risk of hallucinations or poor transformations.