Coding agents - Archal

Archal is designed to be set up and operated by the coding agent already working in your repository. The agent drives the process, but the human still chooses the test boundary and approves destructive state changes.

Pair once

Open Dashboard > Setup, choose the coding-agent target, and paste the generated prompt into Claude Code, Codex, Cursor, or Windsurf. The prompt connects the CLI and installs five repository-local skills:

Skill	Responsibility
`archal-setup`	Pair, install, verify, and stop before application changes
`archal-session`	Discover environments and control sandbox lifecycle
`archal-state`	Inspect, validate, load, diff, and reset explicit state
`archal-ci`	Fit Archal into an existing test or CI workflow
`archal-diagnose`	Isolate auth, fidelity, state, routing, usage, and cleanup failures

Checksums must pass for all five. A successful setup ends with Archal is connected and ready.

Let the agent inspect before it edits

A good first request gives the coding agent a real workflow, not a vague goal:

Inspect our existing Slack and Datadog clients and the integration tests that
exercise them. Show me the matching Archal environments and five starting
states for each. Propose the smallest disposable sandbox that proves one real
workflow. Keep our current test runner and observability provider.

The coding agent should discover:

where provider base URLs and credentials enter the application
which tests need stateful behavior
which environment operations are actually supported
whether a curated sample already matches the test
the smallest safe code and configuration change

Authority boundaries

The protected Archal credential controls sandbox lifecycle. Do not put it in a prompt or expose it to the agent under test. The code being tested receives only the short-lived provider credentials returned for its environments. State reads and validation are safe discovery operations. Loading or resetting state overwrites data. The archal-state skill requires explicit approval for the named sandbox, environment, and artifact unless the user deliberately configured a newly created disposable CI sandbox.

Archal does not own the harness

Keep Braintrust, LangSmith, custom graders, traces, and test runners in place. Archal provides stateful provider environments. Your coding agent connects those environments to the harness you already trust.

Adding CI

The setup skill never edits CI. After one useful manual proof, ask the agent to prepare a CI plan that names the files, triggers, environments, sample hashes, TTL, parallelism, secret boundary, cleanup behavior, and maximum expected cost. Review that plan before allowing the agent to edit a workflow. See CI and test suites for the current safety requirements and product limits.

​Pair once

​Let the agent inspect before it edits

​Authority boundaries

​Archal does not own the harness

​Adding CI

Pair once

Let the agent inspect before it edits

Authority boundaries

Archal does not own the harness

Adding CI