arbor-agent-executor

Arbor Executor

Use this when a pending Idea Tree leaf is selected for implementation.

Dispatch Contract

The executor receives:

• Isolated worktree path and experiment branch.
• Node id and exact hypothesis.
• Evaluation info from tree metadata with {cwd} and {node_id} substituted.
• Ancestor insights.
• Additional context from the coordinator.

The executor must implement the assigned idea faithfully. It may choose how to implement the idea, but it must not silently switch to a different direction.

Worktree Lifecycle

Native RunExecutor does this automatically:

1. Validate node exists and is pending or running.
2. Enforce leaf-only dispatch when max_tree_depth is set.
3. Mark node running.
4. Create a git worktree from current trunk.
5. Run the executor agent in that worktree.
6. Finalize useful code changes with a commit.
7. Snapshot plugin outputs such as submission.csv if configured.
8. Remove the worktree but preserve the experiment branch.
9. Parse the executor report.
10. Update the node as done with score, insight, result, and code_ref.
11. Save experiment artifacts under experiments/<node_id>/.
12. Propagate insights upward.

When emulating manually, do the same sequence. Use arbor-agent-tools for prompt generation and tree updates.

Executor Workflow

The executor follows this loop:

1. Understand: read relevant files before editing.
2. Implement: make the idea active in code.
3. Validate implementation: run 2-3 small representative checks.
4. Iterate until solid: fix bugs and integration misses before judging the idea.
5. Evaluate: run the full B_dev eval once implementation is credible.
6. Report: include changes, baseline vs result, absolute score, and insight.

A bad score is useful only if the implementation was correct. Do not conclude an idea failed from broken code.

Evaluation Rules

• Use B_dev only. Never run B_test during routine executor work.
• Use the eval command injected from metadata. It should already target the executor worktree after {cwd} substitution.
• Save results to results/<node_id>-<brief-description>/.
• Report absolute score, not delta.
• If full eval is expensive, run smoke/subset checks first, then full eval when code is correct.