Home

Welcome to the GraphAgent wiki.

GraphAgent is a fork of opencode that adds a DAG workflow engine: the agent decomposes a task into a dependency graph of child-agent sessions and drives it to completion. State is durable, crashes are recoverable, and you can watch and control the whole graph from the terminal.

If you are brand new

The whole project in one block:

• What it is. An opencode fork whose agent can decompose a task into a dependency graph of child-agent sessions (a "workflow") and drive it to completion — parallel where safe, gated where it matters, recoverable if it crashes.
• Minimum setup. Install a release binary and drop one file: ~/.config/opencode/dag.jsonc with a single model entry. Everything else has a working default.
• First graph. Type /dag-flow <a task with stages> in the TUI — the agent designs the graph, starts it, and reports the workflow ID. On a repo with a GitHub/GitLab remote, run /dag-init once to verify the platform connection; /dag-auto then drives a full six-block run (explore → design → develop → verify → release → summarize) and stops only at user-owned decisions.
• Watch it. Command palette (ctrl+k) → dag.open: workflow list, wave-ordered node status, enter drops into any node's child session; p/r/s/x pause/resume/step/cancel.
• What to expect. Waves of nodes going queued → running → completed. Review gates can issue REJECT — the workflow then completes at the checkpoint and waits for a decision; a rejection is a verdict, not a crash. Failed nodes stay visible until a replan replaces them.
• The cost knob. Judgment nodes (reviews, arbitration, anything required) run on the advanced model tier, volume work on standard — declared once in dag.jsonc; one tier is fine to start.

Then read the Quickstart end to end.

---

If you were not here when this project was built, read in this order:

1. Quickstart — from install to your first running graph in about five minutes.
2. Why GraphAgent — what the graph buys you, and when to keep using a single loop.
3. Differences from opencode — what the fork adds, and how to choose.
4. Concepts — nodes, waves, gates, revisions, ownership: the five words everything else uses.
5. Harness — the whole machinery in lifecycle order: entry, authoring boundary, scheduler, event sourcing, parent-child contracts, gates, revisions, budgets and recovery.
6. Authoring Workflows — writing and saving your own graph specs.
7. Real Examples — worked examples with transcripts of actual runs.
8. Engine Internals — the implementation detail behind each layer.
9. FAQ — model configuration, cost control, relation to upstream opencode.

中文区：首页-zh（九页完整镜像，从快速上手-zh开始）。

The README covers the same ground in condensed form; this wiki is where the walkthroughs live.