oneagentgraph

Compose agents into a graph over oneharness and onejudge, and merge their outputs into one NDJSON event stream.

One config file and one CLI call: the graph names an oneharness config per role and side, and oneagentgraph prepares each member's launch, supervises liveness, and emits every turn, tool call, fallback, and settle as an event you can pipe into anything.

Install

pip install oneagentgraph-cli      # prebuilt binary, no Rust toolchain
npm install -g oneagentgraph-cli   # the same binary, via npm
cargo install oneagentgraph        # from crates.io, compiled locally

To install a revision that has not been released yet, build it from the repository:

cargo install --git https://github.com/nickderobertis/oneagentgraph --locked

Prebuilt archives for Linux (x86-64, arm64), macOS (Intel, Apple silicon), and Windows (x86-64) are attached to every release, with sha256 checksums.

onejudge is a library dependency, linked into this binary — there is nothing to install for it. run, smoke, and interrupt drive the oneharness CLI, so that has to be on PATH, at the release the justfile pins as oneharness-version or newer — smoke and interrupt ask it for JSON by name (--format json), which an older release refuses as an unknown argument. health, validate, history, persona, trigger, reset-timer, cancel, and sweep need nothing at all — health reads oneharness's own identity sweep through its library, in this process. ONEAGENTGRAPH_ONEHARNESS_BIN names a pinned install instead.

A kind: oneharness member's turn runs on the linked oneharness-core in this process too, so what that engine reads from the environment it reads from the graph's env: block: with ONEHARNESS_HISTORY=1 and ONEHARNESS_HISTORY_POINTER_FILE=<file> there, every such turn appends one pointer line to that file saying where its session went — oneharness's own contract, read back through the core's io::history::read_pointers. Nothing here writes or sets it; the requirement in Cargo.toml is what decides the linked core has it, and tests/inventory.rs holds it there.

What it does

oneagentgraph run graph.yaml --task "add the retry" --output json

A graph is YAML — members, the oneharness config each side uses, personas, schedules, and dependencies:

version: 10
name: node-scope
members:
  worker:
    kind: onejudge
    base_config: ./onejudge.base.yaml
    agent: { oneharness_config: ./oneharness.toml }
    judge:                            # one side, or a list of sides judged as one panel
      - oneharness_config: ./oneharness.judge.toml
        label: reviewer
      - kind: llmlint
        config: ./llmlint.yml
        diff_base: origin/main
      - command: [./scripts/checks]
    mode: bypass

run streams one envelope per line. Exit 0 means every member settled, 1 that one failed or died, 2 that the config is invalid. --output text renders one line per envelope as it arrives — the stream at the top of this page, filling in real time — and is a rendering of those same events rather than a second report.

A member may name a persona — the role delta it layers over its base config, by built-in name (persona: engineer), by a name in the graph's own catalog, or by path. A persona is a onejudge config fragment: it is written in onejudge's own field names, and onejudge's schema decides what it may say. The persona format documents it, including the agent: block earlier versions defined, which is now refused with no compatibility path. oneagentgraph persona validate PATH takes a file or a whole catalog:

The images on this page are captures of the real CLI, rendered from its actual output by just screenshots and hash-gated by screencomp — so they cannot drift away from what the binary prints.

A two-party member's judge: is a list of sides judged as one panel — a harness reviewer, an llmlint run over the worker's tree, and a repository's own command, on one worker at once — and a single side is the one-element shorthand for it. A harness side names an oneharness_config, an llmlint side says kind: llmlint with its config resolved from the graph's own directory, and a command side names a command; each may carry a label. A list of one harness side composes the same onejudge provider it always did; any other list composes onejudge's split provider with the judges in list order, and each judge's verdict on each worker turn is published as a judge-decided event naming the judge, its kind, its decision and its reason.

A single-sided member may declare pre_turn commands — run immediately before each of its turns, with what they printed prepended to what that turn is asked. A supervisory member's first act becomes reading a prepared view rather than spending tool calls rediscovering state that already exists. A view that cannot be started, fails, prints nothing, or outruns its own bound leaves the turn happening without it and publishes a pre-turn-context saying which.

When a member's turn goes the wrong way, oneagentgraph interrupt RUN MEMBER --input "do this instead" redirects it in place instead of discarding it the way cancel does. Exit 3 means there was no controllable turn in flight, and says which — a fact, not an error.

oneagentgraph owns no harness, model, or fallback logic. oneharness keeps owning identity chains, fallback, model pins, and quota classification; onejudge keeps owning the two-party conversation. This composes them.

Check a graph before you run it

oneagentgraph validate GRAPH reads the document, resolves every config and persona it names, and builds each member's invocation, without launching anything. A refusal names the member, the entry inside it, and the value it will not take:

What past runs did

A run's record is written when it starts and rewritten as it settles, so oneagentgraph history lists the run that is still going alongside the ones that finished — each with its exit code:

history show ID prints that run's whole record — which members it declared and what became of each, the digest of every config the run resolved, and where the merged stream was written:

Identities and quota

oneagentgraph health forwards oneharness's own per-identity sweep: which identities this host has, how each was selected, its auth mode, and the headroom left in its subscription window. Every probe is free — no harness takes a model turn — so it is a pre-flight check rather than a thing that costs what it measures. Here it is swept on a machine with no harness installed at all, which is why every identity answers with the program it could not find or the reason it has no headroom to report:

Reclaiming scratch

Under disk pressure, oneagentgraph sweep --dry-run says what scratch exists, what is reclaimable, and what it could not examine; without --dry-run it reclaims what it proves is dead. --format json writes the same report as one JSON object a script reads by field name, and --min-age-hours takes a non-negative decimal — 0.5 is thirty minutes, the same floor onevcs sweep takes. What counts as proof, and the JSON's shape, are stated in the contract.

The command surface

Eleven verbs beside clap's own help, in the three groups they fall into: run, smoke and interrupt drive the oneharness CLI; trigger, reset-timer and cancel signal a run that is already going; and validate, history, health, sweep and persona need nothing installed at all. oneagentgraph <VERB> --help has the flags for each, and the contract is the approved statement of the whole surface.

Develop

just bootstrap   # from a clean clone; installs the toolchain and oneharness CLI
just check       # the deterministic gate: format, clippy, tests, coverage, docs
just gate        # check + the LLM-judge tier; the pre-push bar

Licence

MIT. See LICENSE.