rlm

rlm is an opt-in research/REPL mode. Use it when the task is exploration, source-backed research, or iterative analysis rather than direct mutation in the normal coding-agent loop.

gjc rlm

What it does

rlm runs a notebook-style research session over the agent loop. It is backed by the shared persistent Python kernel and a hard-gated research toolset:

• python
• read
• web_search
• search_tool_bm25
• read-only bash for inspection commands (grep, rg, tree, ls, pwd, wc, du, file, stat)
• goal
• complete_research

Runs aggregate under the current GJC session root:

.gjc/_session-<gjcSessionId>/rlm/<rlmSessionId>/notebook.ipynb

On exit, the session synthesizes:

.gjc/_session-<gjcSessionId>/rlm/<rlmSessionId>/report.md

When to use it

Reach for rlm when you need a durable research scratchpad: compare sources, run small Python probes, preserve intermediate cells, and leave a report. Stay in the normal gjc loop when the work is implementation, review, or workflow execution through deep-interview, ralplan, team, and ultragoal.

Related