Arbor agent setup intake

Arbor Setup And Intake

Use this before the coordinator starts. The output is a concrete research contract plus a clean workspace/session ready for the Arbor cycle.

Fast Path

1. Confirm the target project directory. Treat the launch cwd as the default unless evidence says it is wrong.
2. Inspect README/config/eval files yourself. Do not ask the user to recite data you can read.
3. Identify the primary metric, direction, and real evaluation command.
4. Determine B_dev and B_test. If only one split exists, call it B_dev and record that no separate B_test is available.
5. Run or locate a cheap baseline when feasible. If not feasible, state baseline unknown - measure during INIT.
6. Propose one complete contract and ask for a single yes/edit confirmation.
7. Initialize or select .arbor/sessions/<run_name>/ and hand the contract to the coordinator.
8. If real merges are allowed, define a non-protected trunk_branch such as arbor/trunk/<run_name>. Treat main/master as the base branch, not the merge target.

For smoke/forward tests, never run expensive setup, data prep, training, GPU jobs, or the discovered full eval command. Locate an existing score in cached metadata/logs or use a clearly labelled mocked score, and include smoke-only in the contract.

Research Contract

The instruction passed to the coordinator must contain all five components:

• Metric: exact score name, command that prints it, and maximize/minimize.
• Baseline anchor: current value if known, otherwise say it will be measured in INIT.
• Ambition: beat baseline, reach a target, or push as high as possible within the cycle budget.
• Scope preference: novelty-leaning, effect-leaning, or mixed. Infer it from the repo/task when possible.
• Hard constraints: at minimum, B_test is not for iteration, data/eval harness must not be modified to game the metric, and project-specific protected paths must be respected.

Do not prescribe a specific approach in the contract. The coordinator owns idea generation.

Scaffold When Not Eval-Ready

Some targets arrive as code only — no runnable eval, no dev/held-out split, or no clean git repo. Do not dead-end: scaffold the measurement plumbing (never the solution) so the coordinator has a metric to optimize.

• Prefer the keyless MCP tool scaffold_benchmark. style="light" produces a runnable target (eval entrypoint printing score:, a dev/test split, an editable solution.py); style="zoo" additionally writes the README front-matter contract + PROVENANCE.md and runs the structural verifier. Pass git_init=true to initialize and commit a baseline.
• Fallback when the MCP server is absent: arbor benchmark scaffold <dir> [--style zoo] [--git-init].
• Only scaffold after the user confirms what counts as success. The tool is idempotent and non-destructive — report created vs skipped, never overwrite.

Persist The One-Screen Contract

Once the contract is confirmed, persist it at the target root so it outlives the chat (the durable upgrade over an ephemeral on-screen contract):

• ARBOR_CONTRACT.md — one screen: target dir, metric (name / command / direction), baseline anchor, ambition, scope, dev/test discipline, edit surface, and budget (suggested max cycles).
• research_config.yaml — machine-readable, auto-detected by arbor: task (the contract paragraph), coordinator.max_cycles, coordinator.ui.interaction_mode: review (safest default; change on request). Follow examples/research_config.example.yaml.

These are written during setup, before handing the contract to the coordinator.

Preflight Checks

Mirror arbor run preflight:

• LLM credentials or provider config exist if using the real CLI.
• cwd exists and is non-empty.
• git exists. A git repo should be clean before a fresh run.
• An eval entry point exists or can be discovered (eval.sh, run_eval.py, evaluate.py, README instructions, Makefile targets, etc.).
• If the repo is on a non-base branch, either switch to the base branch or explicitly allow starting from that branch.

For skill-only smoke tests, do not require external API credentials unless an LLM call will actually be made.

Session Layout

Use the open-source layout:

<project>/.arbor/sessions/<run_name>/
  REPORT.md
  COORDINATOR_FINAL_REPORT.txt
  events.jsonl
  run_stats.json
  conversation.md
  .coordinator/
    idea_tree.json
    idea_tree.md
    checkpoint.json
    messages.jsonl
    baseline_cache.json
  experiments/<node_id>/
    report.md
    metrics.json
    diff.patch
  submissions/

When no native Arbor runtime exists, initialize the same layout with:

TOOLS="<skill-dir>/arbor-agent-tools/scripts/arbor_state.py"
python "$TOOLS" init \
  --cwd <project> --run-name <run_name> --task "<contract>"

Config And Plugins

Honor this precedence:

pydantic defaults < plugin.config_overrides < profile < project YAML < CLI

Recognize auto config names in the target project:

• research_config.yaml
• arbor.yaml
• autoresearch.yaml

If plugin: mle_kaggle is selected, load arbor-agent-plugins-hitl-budget before INIT because the plugin pre-fills eval contract, protected paths, required outputs, time budget, and skill behavior.

Baseline Metadata

The setup phase can discover metadata, but the coordinator must persist it:

• baseline_score
• trunk_score
• test_baseline_score
• test_trunk_score
• eval_cmd
• eval_cmd_test
• eval_timeout
• eval_retries
• dataset_info
• metric_direction
• trunk_branch
• submission_path
• sample_submission_path

Use command templates:

cd {cwd} && <eval command> --run-name {node_id}

Never bake the main checkout path into executor commands.

When using arbor_state.py meta, quote any --set value that contains spaces, braces, shell metacharacters, or JSON:

python "$TOOLS" meta --cwd <project> --run-name <run_name> \
  --set "eval_cmd=cd {cwd} && python eval.py --split dev" \
  --set "trunk_branch=arbor/trunk/<run_name>"

In smoke mode, if a cached baseline log exists, parse only metric lines with arbor_state.py parse-log or an equivalent parser that normalizes carriage returns. Do not cat, raw rg, raw grep, or tail long training logs; if debugging requires context, cap output to 20 lines.

Launch Commands

Native runtime examples:

arbor setup
arbor doctor
arbor run "<contract>" --yes --yes-cwd <project> --config <config> --max-cycles 3
arbor report <project>/.arbor/sessions/<run_name>

Skill-only smoke tests should use arbor-agent-tools and a copied project directory unless the user explicitly wants to modify the live repo.