BambooAI 2.1.0 (2026-10-06)

The analyst's workspace, the review of its work, and what it can hand back to you.

Documents in the analysis

Attach up to four documents to a thread (PDF, Word, Markdown, text). They are parsed into units with locators,
mapped in every prompt, and loaded in the kernel as objects the analyst can read.
READ returns verbatim, verified passages; the report cites them as [D1.17] with the passage on hover.
A Reader seat with its own brief; the Documents view in the Dataset cache.

One page of Markdown: the format as a template of one turn, the actions as a table, the note's seven headings.
A reply with more than one action runs the first; the rest is named as not run. The budget line comes on the
last turn itself. The prompts carry no task-specific or model-specific content - a battery check enforces it.
DS holds the dataset as attached in every run: DS.load() restores it; DS.save(frame, "name") saves a dataset
for you - csv by default, json, parquet, xlsx or txt by the name, a string as a text file - listed in the Dataset
cache under Generated with a Download button, and linked under the run's REPLAY row.

RESULT(what, estimate, low, high, unit, direction) prints the line and keeps a record; the ledger in every
prompt is built from the records, never from printed text. A number typed into the call instead of computed is
marked, and a report quoting it is flagged.
test="after turn 16" and corrects=14 mark a test's answer and a correction.

One call per review. The session hands over the code and complete output of the cells the current answer rests
on; the reviewer writes a Checked line per cell, kept for later reviews; a Re-check brings a cell back next time.
Before it judges, five perspectives: identification, alternative explanation, heterogeneity, measurement, and the
question's frame. A TEST comes from one of them.
A REPORT verdict binds only when every cell it cites has been checked. A review that finds a line does not describe
its code marks it in the ledger, and a report quoting it is flagged. The reader's note under a report names what
was checked, what was not, and any caution.
Quick and Deep make no reviewer call.

One kernel in the package and in the executor image (checked byte for byte), executor build v51.
The replay defines what the analysis kernel had - DS, RESULT - and runs to the run's last figure cell, so
figures drawn after the last cited cell reach the Plots tab.
The Dataset cache's details keep the buttons in view while the columns scroll; the run card shows each dataset
pill once; a reply cut at the model's length limit is reported as such.

vLLM and Ollama adapters (2.0.1, 2.0.2) with the INT4 serving notes in the handover.