RSIBench
Public train registryv4 · 84

Task Registry

Explore all 84 held-in training tasks used across the six formal domains. The matching 84 test tasks are reserved for final A0/A_last reporting and their trajectories are never exposed during evolution.

168total tasks
84public train
0validation
84sealed test

Tasks by domain

Difficulty across six domains.

Compare the public training distribution by domain and effort tier. Hover or focus an axis to inspect its mix; click a label to filter the task registry below.

84 public train tasks · 6 domainsEasyMediumHardUnratedlabels are clickable filters

Held-in tasks

Search the public training set

generated from official splits

Loading the public task registry…

Test trajectories are never returned to this page.

Editable capabilities

Six shared modules plus twelve Pi source modules.

Prompt, Rules, Skills, Hooks, MCP, and Workflow are comparable across Pi, Claude Code, and Codex. Pi additionally exposes twelve physical source-code owners with clean rebuild and runtime-load verification.

S1

Prompt

Native Agent guidance.

S2

Rules

Scoped behavioral constraints.

S3

Skills

Reusable procedures and references.

S4

Hooks

Lifecycle-triggered scripts.

S5

MCP

Typed external tool adapters.

S6

Workflow

Commands and subagent profiles.

P1

Agent Loop

Turns, tools, retry, cancel, and stop.

P2

Harness Runtime

Runtime assembly and lifecycle.

P3

Coding Session

Session orchestration and resume.

P4

Observation

Tool results, errors, and truncation.

P5

Context

Provider-call context construction.

P6

Compaction

Summaries and overflow recovery.

P7

Session Storage

Persistence and working-directory state.

P8

Prompt Loader

Prompt discovery and assembly.

P9

Skill Loader

Skill discovery and precedence.

P10

Tool Runtime

Read, edit, write, search, and shell.

P11

Extension Runtime

Extensions, events, and dispatch.

P12

Provider Runtime

Requests, streams, tools, usage, and errors.