commons 0.2.0 introduces three big features in addition to a number of minor improvements and big fixes.
The commons() function was split into two,
enabling users to use commons with other coding agents and to
reuse trusted code across multiple projects. Calls to
commons() previously looked like this:
commons <- commons(client, data_sources, semantic_layer, context_layer)The construction of the data structure that holds the data sources, semantic layer, and context layer and the construction of the agent harness that makes that information available to an LLM are now split into two pieces.
commons <- commons(data_sources = ..., semantic_layer = ..., context_layer = ...)
agent <- agent(client, commons = commons)We call that first data structure “the commons.” The commons can be
used to invoke trusted code and search trusted context directly from R
and is intended for use with other coding agents. See
?commons() and the new sharing vignette at
vignette("sharing", package = "commons") to learn more.
agent() returns what commons() used to
return—a commons agent.
This transition was implemented backward-compatibly; existing calls
to commons() will raise a deprecation warning with
migration instructions.
commons agents can now write self-contained, sharable reports as
Quarto documents. When users mention writing reports or creating
something sharable, the agent can load an agent skill and then stream
document code directly into a chat drawer. Documents get their data by
calling commons() methods in their code, and the agent
revises them with a new edit_artifact tool.
On Posit Connect, users can now share conversations, documents, plots, or tables as static content they own. Existing apps need no changes to enable this feature. Shared content includes the code that made it, the date it was created, and a link back to the original agent.
agent() (formerly commons()) gains a
mode argument. With mode = "trusted only", the
agent answers only with trusted calculations. It never writes its own
SQL or R, and it tells the user when no trusted calculation answers a
question. These answers carry no provenance markers.
Citation and provenance markers keep their icons in conversations restored after commons is reinstalled or the app is redeployed. Conversations saved before this change still lose them.
A number of bug fixes in data_source():
data_source() errors when a dictionary
file has no top-level tables field, rather than silently
loading an empty dictionary (#473).data_source() imports Databricks metric views from
Unity Catalog through ODBC. It previously listed them as plain views and
could not read their definitions (#496).data_source() given a DBI connection and no
tables selects the tables and views in the connection’s
current schema. When that schema has too many tables to list in the
system prompt, the agent finds them with search_catalog
instead, as for Snowflake and Databricks connections. A
DBI::Id naming only a schema, like
DBI::Id(schema = "marts"), selects every table and view in
that schema. A data dictionary can name those tables relatively, like
orders for marts.orders, when the name matches
only one selected table.A number of fixes for Snowflake/Databricks semantic models (#439, #441, #446, #447, #486).
A number of bug fixes in describle_table() the tool
that the agent uses to grab table metadata and documentation:
describe_table works on SQL Server, Oracle, and other
databases whose SQL has no LIMIT clause (#468).describe_table shows at most 50 columns and reports how
many additional columns it omits, keeping wide-table descriptions
compact. Columns documented in the data dictionary are shown first. The
system prompt no longer limits the agent to columns that
describe_table or search_context confirmed
(#475).describe_table describes any table the connection can
query, not only the data source’s tables, so it agrees with
run_sql. The result notes when a table is outside the data
source. Agents in mode = "trusted only" still describe only
the data source’s tables (#469).describe_table reports table schemas and documentation
without a sampled-row summary. DuckDB descriptions use database-native
types. Snowflake and Databricks continue to use native metadata, while
other DBI connections infer R column classes from five sampled rows, or
zero rows in trusted-only mode (#319).A number of security hardenings in run_r():
run_r worker reads no R startup files, so variables
set in a .Renviron file, like API keys, stay out of reach
of model-written code.run_r can no longer send a signal to the
application process’s ID or to its process group. On Linux, the programs
it starts can also no longer leave its process group, so they all stop
when the R session ends.run_r worker has its own temporary folder, so
model-written code can no longer read or write files in the
application’s temporary directory (#463).run_r returns plots as image data rather than file
paths, so model-written code can no longer pass other files on the host
to the model or the viewer (#515).trajectory_read() now returns a user
field with the originating Shiny session username supplied by Connect.
With log = TRUE, authenticated session users are recorded
with conversation trajectories; older logs and anonymous sessions return
NA_character_ (#316).
trajectory_review() has been removed. The
coding-agent skill provides guidance for building interactive reviews
and trust-over-time plots from trajectory_read()
output.
Fixes an issue with the R code sandbox where the generated policy would be too long when there were thousands of R packages installed.
A number of improvements to documentation.