Developer: write code until the tests pass

examples/73-code-writer.halo policies |<> loops

Reads a test file (the spec), has an agent write the implementation, runs the tests, and while they fail sends the agent the failing output to fix (|<> loop). The tests decide when it's done, not the model. This runs model-written code on your machine: point workdir at a sandbox (a scratch folder or a container), not at anything you care about. Block(DestructiveOps) rejects code with things like rm -rf.

Open in the playground →

halo run examples/73-code-writer.halo --arg workdir=/tmp/calc \
--arg test_file=calc.test.ts --arg target_file=calc.ts \
--arg task="Evaluate arithmetic expressions"

// ============================================================
// Developer: write code until the tests pass
// ============================================================
// Reads a test file (the spec), has an agent write the implementation,
// runs the tests, and while they fail sends the agent the failing
// output to fix (|<> loop). The tests decide when it's done, not the model.
//
// This runs model-written code on your machine: point `workdir` at a
// sandbox (a scratch folder or a container), not at anything you care
// about. Block(DestructiveOps) rejects code with things like `rm -rf`.
//
// RUN:  halo run examples/73-code-writer.halo --arg workdir=/tmp/calc \
//         --arg test_file=calc.test.ts --arg target_file=calc.ts \
//         --arg task="Evaluate arithmetic expressions"
// ============================================================

policy SafeCode : Block(DestructiveOps) & Block(APIKey) & MaxSpend(0.50)

tool Shell : Bash()

agent Coder [anthropic/claude-sonnet-4-5, temp: 0.0]
  system """
    You are a careful TypeScript engineer working in a Bun project.
    Write the complete contents of the requested file: no Markdown fences,
    no commentary, no placeholders. Make the tests pass without special-
    casing test inputs. When test output is shown, fix the root cause.
  """
  requires SafeCode

pipeline WriteCode(task: String, workdir: String, test_file: String, target_file: String) -> Result:
  task
    // 1. The tests are the spec
    |> Shell.run(command: "cat {{workdir}}/{{test_file}}") as spec
    |? (context.data.spec.exitCode == 0)

    // 2. Write the implementation
    |> Coder.implement(
         "Task: {{task}}. Write {{target_file}} so that {{test_file}} passes.",
         context.data.spec.stdout
       ) as code

    // 3. Save it and run the tests (pipefail: the exit code is bun test's)
    |> Shell.run(command: """
         set -o pipefail
         cd {{workdir}}
         printf '%s\n' {{code}} | sed '/^```/d' > {{target_file}}
         bun test {{test_file}} 2>&1 | tail -60
       """) as result

    // 4. Failing? The coder sees the test output and tries again
    |<> loop(Coder, "context.data.result.exitCode != 0", max: 4) as fixes

    |> System.Value(context.data.result.stdout)