Crate vela_scientist 

vela-scientist — the agent layer

Sits on top of vela-protocol. Reads a researcher’s working directory and emits StateProposals tagged with an AgentRun for reviewer-facing provenance. Never signs canonical state.

v0.22 ships Literature Scout only: PDF folder → finding.add proposals. Other agents (Notes Compiler, Code Analyst, Contradiction Finder, Experiment Planner, Reviewer Agent) land one at a time in v0.23+.

Modules

agent

Shared agent infrastructure

code_analyst

Code & Notebook Analyst (v0.24)

datasets

Datasets agent (v0.25)

experiments

Experiment Planner (v0.28)

extract

Literature Scout extractor — claude -p backend

llm_cli

claude -p shared wrapper

notebook

Jupyter notebook parser

notes

Notes Compiler (v0.23)

reviewer

Reviewer Agent (v0.28)

scout

Literature Scout

tensions

Contradiction Finder (v0.28)

Constants

AGENT_ACTOR_ID_LITERATURE_SCOUT

AGENT_LITERATURE_SCOUT

Stable agent name + actor id for Literature Scout. Pairs with StateProposal::actor.id == AGENT_ACTOR_ID_LITERATURE_SCOUT so the Workbench can group its proposals.

Functions

new_run_id

Generate a fresh run id for an agent invocation. Format: vrun_<16 hex chars> derived from the agent name + a UTC timestamp. Not content-addressed (two identical inputs at different wall-clock instants produce different ids); the run id’s job is to group proposals from the same invocation in the reviewer UI, not to act as a substrate primitive.