OpenSpace/agentpilot/configs

Configuration Reference

For installation and the reading order, start at Quick Start. This page describes the core TOML fields. Unknown framework keys are rejected; relative paths resolve from the TOML file's directory. environment.settings is validated by its adapter.

Templates: Textual Policy Learning, UCB-based Policy Learning, and a runnable synthetic example. The Textual and UCB templates require an adapter and replacement of every REPLACE_ME value.

Policy

Both methods learn the policy for improving agents while keeping language-model weights fixed. UCB-based Policy Learning updates outcome statistics and uses KL-UCB to balance exploration and exploitation. Textual Policy Learning updates reusable lessons stored in Policy Memory.

Table / field	Purpose
policy.method	Select "memory" for Textual Policy Learning or "ucb" for UCB-based Policy Learning.
policy.direction_slate_size	Number of directions in a proposed slate.
policy.max_planning_attempts	Maximum whole-slate replacements per logical round.
policy.proposer	Model and investigation limits for proposing changes.
policy.memory.ranker	Model that ranks directions using Policy Memory; Textual selects the first ranked item.
policy.memory.summarizer	Model that distills and revises Policy Memory from completed execution and evaluation feedback.
policy.memory.max_chars	Maximum Policy Memory size.
policy.semantic	UCB embedding model, calibration identity, and matching thresholds.
policy.cooldown, policy.effectiveness, policy.neighborhood	UCB retry, improvement, and exploration rules.

In the core batch runner, Textual uses one frozen Policy Memory snapshot throughout a batch. After the batch's actions settle, MemorySummarizer combines the previous lessons with proposals, code changes, and evaluation outcomes into a validated summary; SingleFileMemoryCarrier saves it to the run's memory.md. The updated lessons guide the next batch's parent, direction-allocation, and proposal-selection decisions.

The asynchronous Arena research workflow reuses MemorySummarizer as build results and their B comparisons settle. It saves revised lessons in the research-policy/memory checkpoint for subsequent decisions, without waiting for a whole search batch to finish.

Each completion profile explicitly sets provider, model, max_output_tokens, temperature, and disable_reasoning; profiles do not inherit one another's values. Provider names must match the adapter's ProviderRegistry. For UCB, use the policy tables in the UCB template; embedding dimensions, identities, and thresholds must match an accepted calibration bundle. Keep your existing environment and worker settings when replacing the policy tables, and finish this configuration before collecting baseline evidence.

For Arena research, edit each policy profile's model field separately. The task solver uses environment.settings.action_runtime.program.model_name, in the form openrouter/vendor/model-id. The coding worker uses its own model configuration.

For proposer investigation, max_proposer_turns bounds turns; max_trace_reads and max_repository_reads bound reads; optional max_read_calls adds a shared read limit. max_read_bytes bounds returned evidence and max_context_bytes bounds serialized request history, not the model's token window. See the research reference for tool behavior.

Runtime

Field under runtime	Purpose
database_path	SQLite database for durable run records.
runs_root	Candidate workspaces and generated artifacts.
worker_slots	Maximum concurrent actions.
sqlite_timeout_seconds	SQLite lock timeout.
lease_ttl_seconds	How long an action ownership claim remains valid.
lease_renew_interval_seconds	Claim renewal interval; must be shorter than the TTL.
require_provider_reported_cost	Require authoritative provider cost data; missing data cannot be treated as zero.

Create the configured output directories before running, and keep them under an ignored directory such as runs/.

Environment Adapter

Field under environment	Purpose
factory	Import path, for example my_adapter:create_dependencies. Called with AppConfig; returns AssemblyDependencies.
identity	Stable name for the task environment.
settings	Adapter-specific program, task, worker, and evaluation configuration.

AssemblyDependencies supplies a provider registry and a per-batch environment factory, plus optional UCB calibration. The environment connects durable search inputs, an action runner, a frozen evaluation plan, and an effectiveness rule. See the core environment interface reference for implementation contracts and worker configuration for Arena research workers.

For benchmark-specific task splits and evaluation options, run agentpilot-arena-aime-build --help or agentpilot-arena-tb2-build --help through uv run --project agentpilot/integrations/arena --no-sync from the repository root. Both builders accept --worker-config and --model; role-specific model flags override --model.

Campaigns and Resume

The optional campaign table is used by agentpilot campaign:

Field	Purpose
total_processes, actions_per_process	Number of search processes and actions per process.
max_planning_batches_per_process	Bound on replacement planning batches.
max_stalled_resumes	Stop after this many consecutive resumes without durable progress.
poll_interval_seconds	Supervisor polling interval.
reconciliation_deadline_seconds	Maximum wait for provider accounting.

Resume requires the original configuration and batch ID. AgentPilot rejects configuration or adapter changes that alter an active batch. In the batch runner, exhausted Policy Memory summarization attempts are a terminal policy-update failure rather than a resumable interruption.

Credentials

Keep credentials in the environment or a local .env based on the repository .env.example; store only environment-variable names in adapter settings. --env-file loads the file for execution, with existing host variables taking precedence. validate-config checks framework TOML structure without loading credentials, importing the adapter, or contacting providers.