Getting Started
Documentation source v0.5.0-rc.1. Published releases
Install and prove the runtime
With Go 1.26 or newer, install the latest published stable release. Ensure
Go’s binary directory (go env GOBIN, or $(go env GOPATH)/bin when GOBIN
is empty) is on PATH.
go install github.com/justyn-clark/loopexec/cmd/loopexec@latestloopexec --versionloopexec demoThe command resolves the latest public stable release. GitHub Releases
is authoritative; documentation source identity does not establish publication.
For a pinned install, substitute a tag listed there as public for latest. Prebuilt archives and
checksums are available there if Go is unavailable.
After installation, loopexec demo needs no model, API key, Python runner,
or network access. It creates a red local fixture, executes the real bounded
loop, repairs the fixture, observes an external green check, writes a receipt,
and replays the check fingerprint. This proves the execution/check/receipt/replay
contract, not agent intelligence. Its successful process exit is 0.
No clone or hidden fixture is needed. The output gives the retained temporary
workdir and receipt path. For a known location, use loopexec demo --workdir <fresh-dir>;
it refuses to overwrite existing .loopexec or status.txt fixtures.
Run your own bounded loop
In your own project, supply a real work command and a real check:
loopexec run \ --check "go test ./..." \ --exec "make fix" \ --max-iterations 10This example assumes that your project has Go tests and a make fix target
that performs one repair attempt. Replace them with your project’s commands.
LoopExec runs the work command, then the check, until the check passes, a bound
trips, or execution fails. Halt reasons are computed from observed state.
An external check exits 0 when green. A converged loopexec run exits 10,
so shell/CI wrappers must handle that successful convergence class explicitly.
Use the reported receipt and run ID with loopexec replay and loopexec report;
see the CLI guide for exact flags and halt classes.
Go deeper
- Capabilities: shipped cores and planned sub-parts.
- Workflows: deadlines, usage accounting, numeric progress and candidate restoration.
- Integrations: adapter and operator responsibilities, including isolation hooks.
- Shared model-call budgets: call limits across run IDs.
- SMALL: optional state/audit integration.
- Acceptance and evidence: retention, convergence and owner decisions.