CosmosMind-ai/RSI-Harness

English · 简体中文

Turn an agent's harness into something you can version, share, and generate automatically.

Built on the Pi coding agent, with a configuration layer called Genome on top. A Genome is a complete, self-contained, deliverable harness configuration — system prompt, tool set, skills, MCP servers, extensions, runtime policies, memory, keybindings, themes — all in one directory. Switching contexts is switching Genomes.

git clone https://github.com/CosmosMind-ai/RSI-Harness.git && cd RSI-Harness
./install.sh

Requires Node 22.19+. The script checks dependencies, builds, installs into ~/.local/bin, and asks you to pick an install mode: --copy moves the binary and all its assets out of the repo for good; --link symlinks to the build output, for hacking on RSIH itself. With bun installed it compiles a single-file binary; without it, it falls back to a node wrapper.

Once installed, rsih is pi, zero difference: every flag, subcommand, and slash command works unchanged. Launched without a Genome it behaves exactly like pi — a test guards this invariant — except the config directory becomes ~/.rsih:

rsih
rsih --resume
rsih --fork <session>
rsih -p "Review the current workspace"

For scripted use, add --run-id: every call carrying the same id appends to the same conversation, and the Genome is stated once, on the first call, then restored automatically:

rsih :notes -p "Outline this week's lab notes" --run-id week-32 --cwd ~/lab
rsih       -p "Section 3 is too long; split it" --run-id week-32 --cwd ~/lab
rsih       -p "Export as markdown"              --run-id week-32 --cwd ~/lab --json

--cwd sets the working directory and where the session lands, --model swaps the model for one call, and --json emits a structured event stream.

Piped stdin joins the prompt, so a -p call in a non-TTY context waits for stdin to close — when spawning rsih -p from another process, close the stdin pipe or redirect it from /dev/null.

The full CLI operating manual — one-shot calls, --run-id conversations, --json events, and the persistent RPC mode for driving RSIH from another program or agent — is RSIH in the CLI.

Everything that separates RSIH from pi comes from one switch: name a Genome at launch. Four spellings are equivalent:

rsih --genome paperlab   # explicit
rsih :paperlab           # colon shorthand
rsih ::paperlab          # double colon
rsih +paperlab           # plus

Only the first argument counts. Pi treats positional arguments as the message and most options take a value, so a genome marker anywhere else would swallow a message word or an option value. (( is not offered: it is a shell metacharacter — rsih (paperlab is a syntax error in both zsh and bash.)

Two Genomes ship with the distribution:

paperlab is the example Genome: not a blessed setup, but a real one — it distills an actual workflow, running paper experiments. For a concrete sense of what a Genome is, unfold everything rsih :paperlab layers on top of bare pi:

More Genomes live in examples/genomes/ — contributed by the community and installable from a clone; see Community Genomes.

Managing Genomes:

rsih genome list                    # what's installed, what shipped, what's outdated
rsih genome show paperlab           # resolved Genome + the settings patch it would write (read-only)
rsih genome validate ./my-genome
rsih genome install paperlab        # restore the factory version

Drop a Genome directory someone shared with you into ~/.rsih/genomes/ and it runs.

GEE — short for Genome Expression Engine — is a one-word command that launches the harness-rsi Genome:

gee            # == rsih :harness-rsi == rsih --genome harness-rsi

How it works. GEE never asks what system prompt you want — it reads what you actually did. It starts by asking one question in conversation: what is this Genome for? Then it asks which session stores to analyze (RSIH's own is included by default; you can add Pi's ~/.pi/agent/sessions, Claude Code's ~/.claude/projects, Codex's session and archive stores under CODEX_HOME (default ~/.codex), and Copilot CLI's session-state under COPILOT_HOME, defaulting to ~/.copilot). It groups history by working directory, then aggregates tool-call histograms, frequent bash commands, hot files, and the corrections you keep repeating — aggregating first, reading raw text selectively, never dumping a whole JSONL into context. The scenario you described becomes the yardstick for classifying evidence: which recurring pattern should become a skill, which a tool, which an MCP server, and which is mere preference for memory. Before writing anything it walks you through the whole plan with its evidence, in enough detail that you can object to the wording. It writes nothing until you confirm.

What it produces. A Genome directory, ~/.rsih/genomes/<name>/: the genome.json manifest plus configuration for the 12 components (instructions, tools, skills, commands, model, runtime, policies, integrations, …), passing the rsih genome validate gate on completion. Launch it with rsih :<name>. Send the directory to someone else and it runs from their ~/.rsih/genomes/.

→ harness-rsi in depth

Harnesses should be first-class. Tuning an agent today scatters configuration across settings.json, CLI flags, prompts pasted around, and "I remember that prompt worked well last time" — none of it versionable, diffable, reproducible, or handable to someone else. Genome collects it into one object.

No fork of the Core, only its public configuration surface. CLI, TUI, slash commands, keybindings, session tree/fork/resume, the model and settings screens, extension UI — all provided by Pi; this project re-implements none of it.

Pi coding-agent      ← immutable Core, not forked
  ↓
Genome adapter       ← this project
  ↓
harness-rsi          ← a Genome whose output is other Genomes

Two invariants follow, each held by a test:

1. Without --genome, rsih behaves exactly like pi, except the config directory becomes ~/.rsih.
2. Anything Pi can configure, a Genome can configure. test/pi-surface.test.ts extracts every settings key and keybinding id from Pi's own .d.ts; any key left unrouted turns the test red — when a Pi upgrade adds a switch, the test tells you first.

Configuration is a patch, not a replacement. An absent field inherits the Pi default; null resets explicitly to the default; only a present value overrides (objects merge recursively, arrays replace wholesale). A Genome that sets only model still gets Pi's full system prompt, full tool set, and AGENTS.md discovery. You never reimplement a harness to change one field.

Self-referential. Not one line in src/ serves harness-rsi itself: its charter lives in the instructions component, its methodology in a file-based skill, its interactive tools in its own extension — built entirely with means available to any ordinary Genome. That is where "RSI" lands: not the model editing its own weights, but editing its own harness, by exactly the means you would use to hand-write one.

Personalization from evidence. harness-rsi does not ask what system prompt you want; it reads what you have done. Fields without evidence stay empty — empty means inheriting the Pi default, which is always the safe answer.

A directory bundle whose manifest is always named genome.json:

my-genome/
  genome.json                 # base + component list
  components/<id>.json        # each component's configuration
  contracts/<id>.dev.md       # each component's contract (what it may touch)
  skills/**                   # bundled skills
  extension/*.ts              # bundled extensions

Twelve components with mutually exclusive field ownership; an out-of-bounds write fails at load time (a tools component trying to set system_prompt → immediate failure):

→ Component protocol · 12 contracts · harness-rsi in depth

Genomes are meant to be shared, and examples/genomes/ is the community shelf: one self-contained directory per Genome, installable straight from a clone of this repo.

rsih genome install examples/genomes/paperlab
rsih :paperlab

Contributing your own is a PR: one self-contained directory, a unique genome_id, a passing rsih genome validate, and no private residue.

→ How to contribute a Genome

Discovery order (first match wins); a path also works directly:

./.rsih/genomes/<name>.json      or  ./.rsih/genomes/<name>/genome.json
~/.rsih/genomes/<name>.json      or  ~/.rsih/genomes/<name>/genome.json
<dist>/genomes/<name>/genome.json        # seed

The third layer is a seed, not the run location. A built-in Genome is copied into ~/.rsih/genomes/ the first time it is used by name, and loaded from there ever after — the Genome you run always lives in your own directory, and its skills and extensions load from there too. Send someone a Genome folder; it runs from their ~/.rsih/genomes/.

Seeds update, but never overwrite you. At install time a .rsih-seed.json inside the bundle records the seed's content hash:

Settings are compiled. Pi has no API for injecting settings, so keys a Genome declares are rewritten into ~/.rsih/settings.json at every startup; keys it doesn't declare (including theme and defaultModel, which Pi writes itself) pass through untouched. Switching Genomes clears keys the previous one managed and the current one no longer declares. The managed set is recorded in the file as $rsih.managedKeys. A project-level <cwd>/.rsih/settings.json outranks the Genome — an explicit escape hatch.

Semantic fields and the escape hatch. Semantic fields like runtime.steering_mode, policies.compaction, and model.cycle project onto the corresponding Pi settings; the settings component is the raw escape hatch at the bottom layer — values written there override the projections.

Works: building Genomes end to end — 12 components covering Pi's whole configuration surface, inherit-by-default merging, the settings compilation layer, directory bundles, seeding and upgrades, and harness-rsi generating interactively from the RSIH, Pi, Claude Code, Codex, and Copilot CLI session stores. Codex support reads plain JSONL rollouts using explicit user-message events, with a 2 MiB head limit per file. Copilot support reads per-session events.jsonl (not session-store.db) using root user.message events, also with a 2 MiB head limit. Sampling limits and skipped files are reported; compressed rollouts and history stored only in a database are not supported.

Not yet: one-command remote installs. genome install accepts built-in names and local paths only, not git/npm URLs — for someone else's Genome, clone this repo and install from examples/genomes/<name>. There is no pre-publish redaction gate either — a Genome is distilled from private transcripts, and before sharing it must be able to flag absolute paths, intranet domains, and likely secrets.

npm run check          # typecheck + test + build
npm run build:binary
npm run smoke:binary
npm run sync:contracts # after editing docs/genome/components/, sync into the bundle

The documentation index lives at docs/README.md. The repository contains no benchmark, data-generation, training, or evaluation code.