| name | autoresearch |
| description | Standalone AIDE-style autoresearch agent: tree-search over candidate solutions (draft/debug/improve operators) for any task directory with a task.md and score.py — ML engineering, heuristic optimization, or harness engineering. Runs the current best generation evolved by the rsi-loop outer loop (falling back to baseline/gen-000). Use when the user wants to solve an optimization task by iterative search, or invokes /rsi:autoresearch. |
autoresearch
Runs the rsi-loop inner agent standalone: point it at a task directory containing task.md,
score.py, and public/, and it executes tree-search autoresearch (parallel drafts →
greedy debug/improve loop) under a token budget, returning the best-scoring solution.
Follow the procedure in commands/rsi-autoresearch.md:
sandbox from public materials only (scripts/rsi-sandbox.sh), launch the generation's
inner-agent.workflow.mjs with {sandbox, genDir, taskName, seed, policy}, report the best
node's public score. Private scoring is outer-loop-only (RSI_OUTER_LOOP=1) and never enters
the inner agent's context.
The generation is resolved through the run's best pointer (falling back to
baseline/gen-000), so improvements the outer loop discovers (/rsi:step) flow to standalone
users automatically. First verified run: see
docs/experiments/m1-smoke-bin-packing.md.
Design details: docs/PLAN.md, §3 "autoresearch as a standalone skill".
Naming note: this is distinct from the third-party
uditgoenka/autoresearch skill (a single-level
metric hill-climb loop, evaluated in PLAN.md §5.1 as a candidate outer-loop chassis). The §5.2
experiment resolved in favor of the native chassis (Arm B, /rsi:step / /rsi:run); the
third-party chassis was not adopted, so no rename is needed and this skill keeps its name. See
docs/experiments/chassis-ab/.
