The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.
schema
Defines the Config struct, the single source of truth for config.toml,
and everything needed to load, save, and migrate it. AGENTS.md points
contributors here: "Rust configuration is defined under
crates/openhuman-core/src/config/schema/ and loaded through its config
operations" (the operations live in ../ops/, re-exported as config::rpc).
Config itself is split into submodules: types.rs declares types/config.rs
(the struct definition), types/model_ids.rs (model constants), and
types/defaults.rs / types/output_language.rs / types/resolvers.rs (small
helper types) purely to keep any one file under the repo's ~500-line
guideline; treat them as one unit. load_user_state.rs sits at this level but
is mounted as a submodule of load/dirs.rs via #[path].
Layout: [section] to file to struct
config.toml section |
File | Struct |
|---|---|---|
[agent], [orchestrator], [teams.*] |
agent.rs |
AgentConfig, OrchestratorModelConfig, TeamModelConfig |
[autonomy] |
autonomy.rs |
AutonomyConfig: feeds security::SecurityPolicy |
[[capability_providers]] |
capability_providers.rs |
CapabilityProviderConfig |
[channels_config], [sandbox] |
channels.rs |
ChannelsConfig and per-provider configs re-exported from tinychannels_bus; SandboxConfig, ResourceLimitsConfig, AuditConfig, and SecurityConfig (only DaemonConfig embeds the last one) |
[claude_agent_sdk] |
claude_agent_sdk.rs |
ClaudeAgentSdkConfig |
[[cloud_providers]], primary_cloud, *_provider slugs |
cloud_providers.rs |
CloudProviderCreds, AuthStyle |
[context] |
TinyAgents ContextConfig re-exported by context.rs; OpenHuman environment overlays remain in load/env_overlay/context.rs |
ContextConfig |
[cost] |
identity_cost.rs |
CostConfig, ModelPricing |
[dashboard] |
dashboard.rs |
DashboardConfig, DiagramViewerConfig, EventStreamConfig, ModelHealthConfig |
[channel_voice] |
channel_voice.rs |
ChannelVoiceConfig, ChannelVoiceOverride, VoiceReplyMode |
[ephemeral_route] |
ephemeral_route.rs |
EphemeralRoute |
[cron] |
cron.rs |
CronConfig |
[hooks] |
hooks.rs |
HooksConfig |
[hosting] |
hosting.rs |
HostingConfig |
[learning] |
learning.rs |
LearningConfig, ReflectionSource |
[inbound_debounce] (+ by_channel.<name>) |
inbound_debounce.rs |
InboundDebounceConfig, InboundDebounceOverride: per-conversation batching of inbound channel messages before a turn |
[local_ai] |
local_ai.rs |
LocalAiConfig, LocalAiUsage |
[modules] |
modules.rs |
ModulesConfig, ModuleOverride: controls only whether compiled-in modules load. The loadable set is fixed by crate::modules::registry |
[observability] |
observability.rs |
ObservabilityConfig, AgentTracingConfig |
[privacy] |
privacy.rs |
PrivacyConfig, PrivacyMode |
[proxy] |
proxy.rs |
ProxyConfig, ProxyScope, plus runtime_proxy_config() / set_runtime_proxy_config() process-wide accessors |
[[model_routes]], [[embedding_routes]] |
routes.rs |
ModelRouteConfig, EmbeddingRouteConfig |
[runtime] (+ [runtime.docker]), [shell], [reliability], [scheduler] |
runtime.rs |
RuntimeConfig, DockerRuntimeConfig, ShellConfig, ReliabilityConfig, SchedulerConfig |
[runtime.local_jail] |
runtime_local_jail.rs |
LocalJailConfig: filesystem grants for the local OS jail |
[scheduler_gate] |
scheduler_gate.rs |
SchedulerGateConfig, SchedulerGateMode |
[memory] (engine, engines.<id>, conversations, context, sources, embedding settings) |
memory.rs |
MemoryConfig, MemoryEngineSettings, MemoryConversationsConfig, MemoryContextConfig, MemorySourceConfig (see docs/specs/memory-v2.md) |
[task_sources] |
task_sources.rs |
TaskSourcesConfig |
[tokenjuice] |
tokenjuice.rs |
TokenjuiceConfig |
| tool-related sections (see below) | tools/ |
(multiple structs) |
[update] |
update.rs |
UpdateConfig, UpdateRestartStrategy |
[voice_server] |
voice_server.rs |
VoiceServerConfig, SttEngine |
[[voice_providers]], stt_provider / tts_provider |
voice_providers.rs |
VoiceProviderCreds, BuiltinVoiceProvider, BUILTIN_VOICE_PROVIDERS |
top-level Config and its scalar keys |
types.rs (+ types/config.rs, types/model_ids.rs) |
Config, ModelRegistryEntry, provider-neutral workload roles |
| built-in defaults | defaults.rs |
impl Default for Config and per-field default_* fns |
tools/mod.rs groups the tool-facing sections: browser.rs (BrowserConfig,
BrowserComputerUseConfig, and the unattended_actions allow-list with its
known names in UNATTENDED_BROWSER_ACTIONS), http.rs (HttpRequestConfig, CurlConfig),
integrations.rs (IntegrationsConfig, ComposioConfig, SecretsConfig),
mcp.rs (McpServerConfig, McpClientConfig, GitbooksConfig),
multimodal.rs (MultimodalConfig, MultimodalFileConfig), search.rs
(SearchConfig, WebSearchConfig, SearxngConfig, SeltzConfig).
SearchConfig (schema version 2) holds providers (enabled + managed /
direct route per provider), roles (ordered providers per capability role),
presentation, limits, and direct keys; fresh installs default to managed Exa
and managed Gemini. Files without schema_version carry the single-engine
format and are migrated on load (tools/search_migrate.rs,
load/migrate.rs::migrate_search_settings); Parallel keeps its key as a
direct-only provider, and a keyless managed-Parallel selection is dropped. Seltz and SearXNG keep their options in their own sections, but only
the providers map decides whether they are on.
The parser drops the retired top-level [agents] delegate table before typed
decoding, so existing config files with stale [agents.*] entries continue to
load. Agent definitions now come from the agent registry and live sub-agent
tools.
Most sections have a matching *_tests.rs (some further split into several
*_tests.rs siblings, e.g. types_model_pin_tests.rs); this is the repo's
file-size-splitting convention, not separate modules.
Engine selection keys
This is where a config file names which engine handles a given concern. The keys that select an implementation, rather than tune one:
search.engine(SearchConfig.engine,tools/search.rs): one of the string constants re-exported frommod.rs(SEARCH_ENGINE_MANAGED,SEARCH_ENGINE_PARALLEL,SEARCH_ENGINE_BRAVE,SEARCH_ENGINE_QUERIT,SEARCH_ENGINE_EXA,SEARCH_ENGINE_TAVILY,SEARCH_ENGINE_DISABLED), plus SearXNG through the separate toggle described above.agent.tool_search.ranker(ToolSearchConfig,agent.rs):"jev"by default, with"auto","bm25", and"compare"as the other values. It picks which ranker answers a tool search over tools exposed asToolExposure::Deferred;top_k(default 3) caps how many matches come back.[memory](memory.rs): Memory v2 config, including[[memory.sources]]. Old[subsystems.memory],[memory_tree]and v1[memory]keys still parse and are ignored.
autonomy.rs is not an engine selector; it is the config-side half of the
sandbox policy contract described under Workspace/identity helpers below.
Loading
Config::load_or_init (in load/impl_load.rs) is the entry point used by
config::ops::loader::load_config_with_timeout. For an existing
config.toml it runs, in this order:
- Resolve the config and workspace directories (
load/dirs.rs,resolve_runtime_config_dirs_with). With no config file and the default config dir, it returns a pre-login in-memoryConfigwithout writing anything. - Read and parse
config.toml(parse_config_with_recovery), recovering from non-UTF-8 or unparsable content by renaming the file to.corrupted[.<ts>], trying the.bakcopy, and finally falling back to defaults withrecovered_from_corruption = true. - Fill
config_path/workspace_dir/action_dir(resolve_action_dir), then apply the two pre-schema-version legacy rewrites inload/migrate.rs(migrate_legacy_inference_url,migrate_cloud_provider_slugs). - Apply environment-variable overrides through
Config::apply_env_overrides_frominload/env_overlay.rs, split into submodules underload/env_overlay/(context.rs,learning_memory.rs,observability.rs,proxy.rs,runtime.rs,search.rs);load/env.rsis only theEnvLookupseam that lets tests supply a fake environment. - Run pending schema migrations (
../migrations/, viamigrations::run_pending), which bumpschema_versionand save as they go. - Decrypt at-rest secrets (
load/secrets.rs); a legacyenc:(XOR) value is force-migrated toenc2:(ChaCha20-Poly1305 viasecurity::secret_storage) and the config is saved again immediately.
A brand-new workspace is instead stamped at CURRENT_SCHEMA_VERSION, seeded
via migrations::seed_new_workspace, saved (0600 on Unix), and then given the
same env overlay.
Config::save() restores CLI-overridden inference fields
(cli_overrides::restore_persisted_inference_fields), re-encrypts secrets,
writes and fsyncs a 0600 temp file, then hands off to
load/atomic_commit.rs::commit_replacement, which preserves the previous
config as .bak and renames the temp file into place. The split exists so
callers can tell "nothing written" (an Err before the rename) from "already
committed" (after it) and roll back in-memory state accordingly; the
migrations runner depends on this.
load/mod.rs also exports CONFIG_OWNER_MISMATCH_MARKER: the loader appends
this marker to a config-read failure when the file's owner differs from the
reading process, and core::observability::expected_error_kind keys on it to
keep that failure paging instead of being demoted as ordinary user-environment
state.
Workspace / identity helpers (load/dirs.rs)
Re-exported through config::mod and config::schema::mod:
default_root_openhuman_dir,user_openhuman_dir: the per-user~/.openhuman/users/<user-id>/root.resolve_action_dir,default_action_dir,action_dir_env_override(OPENHUMAN_ACTION_DIR): the agent's sandboxed read/write root.active_workspace_dir/active_workspace_dir_cached: resolve (and synchronously cache, viaload/active_workspace.rs) the workspace the loader last resolved, for callers (like the developer Event Log's SSE stream) that cannot afford a disk read per lookup.PRE_LOGIN_USER_ID("local"),pre_login_user_dir,read_active_user_id/write_active_user_id/clear_active_user: the pre-authentication identity scope and theactive_user.tomlmarker (implemented in../load_user_state.rs).
config only describes these roots. Per AGENTS.md, action_dir is the
agent's permitted read/write root, and workspace_dir stores internal state
and is never an acting-tool target. Enforcement of that boundary lives in
security::SecurityPolicy (security/policy/), not here. autonomy.rs
(AutonomyConfig) is the config-side half of the same contract: it is read
into SecurityPolicy at startup and on every settings change
(config::ops::agent::apply_autonomy_settings calls
security::live_policy::reload_from).
cli_overrides/
Process-local inference overrides supplied by the standalone CLI
(set_cli_inference_overrides, apply_cli_inference_overrides,
restore_persisted_inference_fields, all pub(crate), and the
#[doc(hidden)] AppliedInferenceOverride snapshot stored on
Config::cli_inference_snapshot). This lets a CLI invocation temporarily swap
model/provider without the override ever reaching the persisted config.
Tests
Per-section *_tests.rs files, plus load_tests.rs (split into
load_active_user_and_dirs_tests.rs, load_env_overlay_tests.rs,
load_corruption_recovery_tests.rs, load_migration_tests.rs,
load_backup_tests.rs) for the loader/env-override/migration surface.
Related docs
- ../README.md: the
configmodule overview. - ../ops/README.md: the mutation/RPC surface built on top of this schema.
- ../migrations/README.md: automatic schema-version upgrades run during load.
- ../../security/README.md: enforces the
action_dir/workspace_dirboundary this module only describes. - Settings
- Deep architecture reference
Config sources
load/source/ defines the ConfigSource port: where the TOML text of a config lives. FileConfigSource is config.toml with the atomic replace and .bak of load/atomic_commit.rs (the file stays hand-editable). DocumentConfigSource is the config/{scope} document on a shared (multi-tenant) storage backend, used by Config::save, snapshot reloads (load_from_config_path) and load_or_init. A process boots before its backend is installed, so its first load normally reads the file. A document that does not parse is an error (no file-style recovery, which would touch the bootstrap file). Bootstrap tables ([storage]) are never written to a document and are always re-applied from the file on a document read. SaasConfig is not part of this.