scriptoriumadmin/fabro
1
0
Fork 0
mirror of https://github.com/fabro-sh/fabro.git synced 2026-09-11 22:53:00 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

Read catalog policy from lithos core fields and metadata.agent

Browse source 
Fabro's policy layer restated the lithos built-ins under `metadata.fabro`:
enabled flags, credentials, display facts, probe and small-default roles,
and agent profiles. lithos-llm now carries every one of those as a core
field or under the shared `metadata.agent` namespace, so the layer and its
typed view go:

- Delete `fabro-policy.toml` and `FABRO_POLICY_TOML`. The catalog is the
  lithos built-ins plus the operator's `[llm]` overlay, nothing between.
- Delete `fabro_types::catalog_policy`. `enabled`, `stands_in_for`,
  `api_key_url`, `family`, the cutoffs, `estimated_output_tps`,
  `small_default`, and `probe` are read from lithos accessors; the agent
  profile and `reasoning_by_default` come from `metadata.agent`, which
  Pebble reads too.
- `catalog::provider`, `enabled_providers`, and `listed_providers` return
  the lithos `CatalogProvider` directly; `ModelEntry` loses its policy
  field and gains `agent_profile()`.
- Test fixtures move `[providers.x.metadata.fabro] enabled = true` onto
  the provider table, drop `credentials` lists in favor of the secret name
  lithos derives from the provider id, and spell `agent_profile` as
  `metadata.agent.profile`.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
Bryan Helmkamp 2026-09-09 23:20:12 -06:00
parent 6dfc96d3fd
commit 2a2fc41807
No known key found for this signature in database
36 changed files with 314 additions and 1445 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
lib/apps/fabro-cli/src/commands/install.rs
View file

@ -83,22 +83,22 @@ fn supports_install_api_key(provider: &CatalogProvider) -> bool {
fn install_llm_provider_ids(catalog: &Catalog) -> Vec<ProviderId> {
catalog::listed_providers(catalog)
.iter()
.filter(|entry| supports_install_api_key(entry.provider))
.map(|entry| entry.provider.id().clone())
.into_iter()
.filter(|provider| supports_install_api_key(provider))
.map(|provider| provider.id().clone())
.collect()
}
fn provider_env_var_label(provider: &ProviderId, catalog: &Catalog) -> String {
catalog::provider(catalog, provider.as_str())
.map(|entry| fabro_auth::secret_names(entry.provider).join(" / "))
.map(|provider| fabro_auth::secret_names(provider).join(" / "))
.filter(|label| !label.is_empty())
.unwrap_or_else(|| "API_KEY".to_string())
}
fn provider_vault_secret_name(provider: &ProviderId, catalog: &Catalog) -> String {
catalog::provider(catalog, provider.as_str())
.and_then(|entry| fabro_auth::expected_secret_name(entry.provider))
.and_then(fabro_auth::expected_secret_name)
.unwrap_or_else(|| format!("{}_API_KEY", provider.to_string().to_uppercase()))
}

15
lib/apps/fabro-cli/src/shared/provider_auth.rs
View file

@ -61,7 +61,7 @@ fn default_catalog_for_provider_auth() -> Arc<Catalog> {
pub(crate) fn provider_display_name(provider: &ProviderId, catalog: &Catalog) -> String {
catalog::provider(catalog, provider.as_str()).map_or_else(
|| provider.to_string(),
|entry| entry.provider.display_name().to_string(),
|provider| provider.display_name().to_string(),
)
}
@ -69,14 +69,14 @@ fn api_key_catalog_provider<'a>(
provider: &ProviderId,
catalog: &'a Catalog,
) -> Result<&'a CatalogProvider> {
let entry = catalog::provider(catalog, provider.as_str())
let provider = catalog::provider(catalog, provider.as_str())
.with_context(|| format!("provider '{provider}' is not configured in the model catalog"))?;
anyhow::ensure!(
fabro_auth::accepts_api_key(entry.provider),
fabro_auth::accepts_api_key(provider),
"provider '{}' does not define an API-key credential path",
entry.provider.id()
provider.id()
);
Ok(entry.provider)
Ok(provider)
}
pub(crate) async fn validate_api_key(
@ -368,8 +368,6 @@ async fn await_user_response_from_source(
#[cfg(test)]
mod tests {
use fabro_types::catalog_policy;
use super::*;
#[test]
@ -385,8 +383,7 @@ mod tests {
ProviderId::new("inception"),
] {
let provider = api_key_catalog_provider(&provider, &catalog).unwrap();
let policy = catalog_policy::provider_policy(provider);
let url = policy.api_key_url.as_deref().unwrap_or_default();
let url = provider.api_key_url().unwrap_or_default();
assert!(!url.is_empty(), "{} has empty URL", provider.id());
assert!(url.starts_with("https://"), "{} URL: {url}", provider.id());
}

4
lib/apps/fabro-cli/tests/it/cmd/exec.rs
View file

@ -333,7 +333,7 @@ fn exec_accepts_configured_custom_provider_from_settings() {
let context = test_context!();
context.write_home(
".fabro/settings.toml",
"_version = 1\n\n[llm.providers.acme-aws]\ndisplay_name = \"Acme AWS\"\nadapter = \"openai-compatible\"\ncodec = \"openai-chat\"\nbase_url = \"https://bedrock.example.invalid/v1\"\nauth = { type = \"bearer\" }\nallow_passthrough = true\n\n[llm.providers.acme-aws.metadata.fabro]\nagent_profile = \"openai\"\ncredentials = [\"env:ACME_AWS_API_KEY\"]\n\n[cli.exec.model]\nprovider = \"acme-aws\"\nname = \"acme-claude-sonnet-4-6\"\n",
"_version = 1\n\n[llm.providers.acme-aws]\ndisplay_name = \"Acme AWS\"\nadapter = \"openai-compatible\"\ncodec = \"openai-chat\"\nbase_url = \"https://bedrock.example.invalid/v1\"\nauth = { type = \"bearer\" }\nallow_passthrough = true\n\n[llm.providers.acme-aws.metadata.agent]\nprofile = \"openai\"\n\n[cli.exec.model]\nprovider = \"acme-aws\"\nname = \"acme-claude-sonnet-4-6\"\n",
);
let mut cmd = context.exec_cmd();
@ -398,7 +398,7 @@ fn exec_server_target_accepts_configured_custom_provider_from_settings() {
let context = test_context!();
context.write_home(
".fabro/settings.toml",
"_version = 1\n\n[llm.providers.acme-aws]\ndisplay_name = \"Acme AWS\"\nadapter = \"openai-compatible\"\ncodec = \"openai-chat\"\nbase_url = \"https://bedrock.example.invalid/v1\"\nauth = { type = \"bearer\" }\nallow_passthrough = true\n\n[llm.providers.acme-aws.metadata.fabro]\nagent_profile = \"openai\"\ncredentials = [\"env:ACME_AWS_API_KEY\"]\n\n[cli.exec.model]\nprovider = \"acme-aws\"\nname = \"acme-claude-sonnet-4-6\"\n",
"_version = 1\n\n[llm.providers.acme-aws]\ndisplay_name = \"Acme AWS\"\nadapter = \"openai-compatible\"\ncodec = \"openai-chat\"\nbase_url = \"https://bedrock.example.invalid/v1\"\nauth = { type = \"bearer\" }\nallow_passthrough = true\n\n[llm.providers.acme-aws.metadata.agent]\nprofile = \"openai\"\n\n[cli.exec.model]\nprovider = \"acme-aws\"\nname = \"acme-claude-sonnet-4-6\"\n",
);
let server = MockServer::start();
server.mock(|when, then| {

8
lib/apps/fabro-server/src/install.rs
View file

@ -846,14 +846,14 @@ async fn put_install_llm(
}
fn install_catalog_provider(provider: &ProviderId) -> Result<&'static CatalogProvider, String> {
let entry = llm_catalog::provider(&INSTALL_CATALOG, provider.as_str())
let catalog_provider = llm_catalog::provider(&INSTALL_CATALOG, provider.as_str())
.ok_or_else(|| format!("provider '{provider}' is not configured in the model catalog"))?;
if fabro_auth::accepts_api_key(entry.provider) {
Ok(entry.provider)
if fabro_auth::accepts_api_key(catalog_provider) {
Ok(catalog_provider)
} else {
Err(format!(
"provider '{}' does not define an API-key credential path",
entry.provider.id()
catalog_provider.id()
))
}
}

10
lib/apps/fabro-server/src/run_manifest.rs
View file

@ -1763,7 +1763,7 @@ mod tests {
fn openrouter_catalog() -> Catalog {
fabro_llm::test_support::test_catalog_with_overlay(
"[providers.openrouter.metadata.fabro]\nenabled = true\n",
"[providers.openrouter]\nenabled = true\n",
)
}
@ -1883,9 +1883,8 @@ base_url = "{moonshot_url}"
[providers.openrouter]
base_url = "{openrouter_url}"
[providers.openrouter.metadata.fabro]
enabled = true
"#
))
.vault_entries([
@ -3029,9 +3028,8 @@ base_url = "https://api.acme.test/v1"
auth = { type = "bearer" }
default_model = "acme-large"
[providers.acme.metadata.fabro]
agent_profile = "openai"
credentials = ["env:ACME_API_KEY"]
[providers.acme.metadata.agent]
profile = "openai"
[providers.acme.models."acme-large"]
display_name = "Acme Large"

13
lib/apps/fabro-server/src/server/handler/sessions.rs
View file

@ -909,13 +909,13 @@ fn api_model_on_eligible(
eligible: &std::collections::HashSet<ProviderId>,
) -> Option<(ProviderId, String)> {
catalog::enabled_providers(catalog)
.iter()
.filter(|entry| eligible.contains(entry.provider.id()))
.find_map(|entry| {
catalog::provider_models(entry.provider)
.into_iter()
.filter(|provider| eligible.contains(provider.id()))
.find_map(|provider| {
catalog::provider_models(provider)
.into_iter()
.find(|model| model.model.api_model() == api_model)
.map(|model| (entry.provider.id().clone(), model.model.id().to_string()))
.map(|model| (provider.id().clone(), model.model.id().to_string()))
})
}
@ -1554,9 +1554,8 @@ default_model = "gpt-5.6-sol"
[providers.openrouter]
default_model = "gpt-5.6-sol"
[providers.openrouter.metadata.fabro]
enabled = true
"#,
)
}

59
lib/apps/fabro-server/src/server/tests.rs
View file

@ -323,7 +323,10 @@ fn openai_responses_payload(text: &str) -> serde_json::Value {
/// An operator-defined OpenAI-compatible provider `acme` offering one model,
/// `acme-large`, with `credential` (`env:NAME` or `vault:NAME`).
fn acme_overlay(base_url: &str, credential: &str) -> String {
/// An operator-defined provider. Its API key is `ACME_API_KEY`, the name
/// lithos derives from the provider id, whether it lives in the vault or the
/// environment.
fn acme_overlay(base_url: &str) -> String {
format!(
r#"
[providers.acme]
@ -335,21 +338,18 @@ auth = {{ type = "bearer" }}
priority = 120
default_model = "acme-large"
[providers.acme.metadata.fabro]
agent_profile = "openai"
credentials = [{credential}]
[providers.acme.metadata.agent]
profile = "openai"
[providers.acme.models."acme-large"]
display_name = "Acme Large"
api_model = "acme-large"
limits = {{ context_tokens = 128000, max_output_tokens = 8192 }}
capabilities = {{ text = true, tools = true }}
[providers.acme.models."acme-large".metadata.fabro]
probe = true
"#,
base_url = toml::Value::String(base_url.to_string()),
credential = toml::Value::String(credential.to_string()),
)
}
@ -5591,10 +5591,7 @@ async fn validate_endpoint_returns_workflow_summary_without_preflight_checks() {
#[tokio::test]
async fn validate_endpoint_uses_app_state_catalog_for_model_diagnostics() {
let state = TestAppStateBuilder::new()
.llm_overlay_toml(&acme_overlay(
"https://api.acme.test/v1",
"env:ACME_API_KEY",
))
.llm_overlay_toml(&acme_overlay("https://api.acme.test/v1"))
.build();
let app = crate::test_support::build_test_router(state);
let dot = r#"digraph Test {
@ -8826,9 +8823,8 @@ auth = {{ type = "bearer" }}
priority = 120
default_model = "portable-model"
[providers.direct.metadata.fabro]
agent_profile = "openai"
credentials = ["vault:DIRECT_API_KEY"]
[providers.direct.metadata.agent]
profile = "openai"
[providers.direct.models.portable-model]
display_name = "Portable (direct)"
@ -8846,9 +8842,8 @@ auth = {{ type = "bearer" }}
priority = 110
default_model = "portable-model"
[providers.aggregator.metadata.fabro]
agent_profile = "openai"
credentials = ["vault:AGGREGATOR_API_KEY"]
[providers.aggregator.metadata.agent]
profile = "openai"
[providers.aggregator.models.portable-model]
display_name = "Portable (aggregator)"
@ -9011,9 +9006,8 @@ auth = {{ type = "bearer" }}
priority = 120
default_model = "acme-reasoner"
[providers.acme.metadata.fabro]
agent_profile = "openai"
credentials = ["vault:ACME_API_KEY"]
[providers.acme.metadata.agent]
profile = "openai"
[providers.acme.models.acme-reasoner]
display_name = "Acme Reasoner"
@ -9080,7 +9074,7 @@ async fn test_provider_credentials_uses_app_state_catalog() {
"usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}
}));
});
let overlay = acme_overlay(&upstream.base_url(), "vault:ACME_API_KEY");
let overlay = acme_overlay(&upstream.base_url());
let state = TestAppStateBuilder::new()
.runtime_settings(default_test_server_settings(), RunLayer::default())
.max_concurrent_runs(5)
@ -9200,7 +9194,7 @@ async fn list_models_marks_configured_true_when_provider_has_credential_material
#[tokio::test]
async fn list_models_marks_configured_false_when_provider_cannot_register() {
let overlay = acme_overlay("https://api.acme.test/v1", "env:ACME_API_KEY");
let overlay = acme_overlay("https://api.acme.test/v1");
let state = TestAppStateBuilder::new()
.runtime_settings(default_test_server_settings(), RunLayer::default())
.max_concurrent_runs(5)
@ -9270,7 +9264,7 @@ async fn list_models_unknown_provider_returns_empty_page() {
#[tokio::test]
async fn list_models_uses_app_state_catalog_overrides() {
let overlay = acme_overlay("https://api.acme.test/v1", "env:ACME_API_KEY");
let overlay = acme_overlay("https://api.acme.test/v1");
let state = TestAppStateBuilder::new()
.llm_overlay_toml(&overlay)
.build();
@ -9338,9 +9332,7 @@ async fn list_providers_marks_configured_per_provider_and_omits_secrets() {
// this exact provider, not merely be populated.
let catalog = state_test_catalog();
let expected_model_count = fabro_llm::catalog::provider_models(
fabro_llm::catalog::provider(&catalog, "anthropic")
.expect("anthropic should be listed")
.provider,
fabro_llm::catalog::provider(&catalog, "anthropic").expect("anthropic should be listed"),
)
.len();
assert_eq!(
@ -9557,7 +9549,7 @@ async fn test_providers_auth_issue_returns_error_without_upstream_call() {
async fn test_providers_registration_issue_returns_error_without_probe() {
// An adapter lithos does not ship cannot be built, so the provider is
// configured (it has a vault key) yet unavailable.
let overlay = acme_overlay("https://api.acme.test/v1", "vault:ACME_API_KEY").replace(
let overlay = acme_overlay("https://api.acme.test/v1").replace(
"adapter = \"openai-compatible\"",
"adapter = \"not-an-adapter\"",
);
@ -9639,16 +9631,11 @@ auth = {{ type = "bearer" }}
priority = 50
default_model = "zeta-probe"
[providers.zeta.metadata.fabro]
credentials = ["vault:ZETA_API_KEY"]
[providers.zeta.models.zeta-probe]
display_name = "Zeta Probe"
api_model = "zeta-probe"
limits = {{ context_tokens = 128000, max_output_tokens = 8192 }}
capabilities = {{ text = true, tools = true }}
[providers.zeta.models.zeta-probe.metadata.fabro]
probe = true
[providers.alpha]
@ -9660,17 +9647,13 @@ auth = {{ type = "bearer" }}
priority = 40
default_model = "alpha-probe"
[providers.alpha.metadata.fabro]
credentials = ["vault:ALPHA_API_KEY"]
[providers.alpha.models.alpha-probe]
display_name = "Alpha Probe"
api_model = "alpha-probe"
limits = {{ context_tokens = 128000, max_output_tokens = 8192 }}
capabilities = {{ text = true, tools = true }}
[providers.alpha.models.alpha-probe.metadata.fabro]
probe = true
"#,
base_url = toml::Value::String(server.base_url()),
);
@ -18678,7 +18661,7 @@ async fn create_completion_default_model_uses_app_state_catalog() {
.header("content-type", "application/json")
.json_body(json!({"error": {"message": "expected test failure"}}));
});
let overlay = acme_overlay(&upstream.base_url(), "vault:ACME_API_KEY");
let overlay = acme_overlay(&upstream.base_url());
let state = TestAppStateBuilder::new()
.llm_overlay_toml(&overlay)
.vault_entries([("ACME_API_KEY", "acme-test-key")])

2
lib/components/fabro-agent/src/agent_profile.rs
View file

@ -40,7 +40,7 @@ pub trait AgentProfile: Send + Sync {
fn knowledge_cutoff(&self) -> Option<String> {
self.catalog_model()
.and_then(|entry| entry.policy.knowledge_cutoff)
.and_then(|entry| entry.model.knowledge_cutoff().map(str::to_string))
}
/// The catalog row for this profile's route, when the catalog knows it.

20
lib/components/fabro-agent/src/cli.rs
View file

@ -425,8 +425,8 @@ fn cli_client_options(args: &AgentArgs, styles: &'static Styles) -> ClientOption
}
}
/// The catalog the standalone agent runs against: lithos built-ins, Fabro
/// policy, and the operator's `[llm]` overlay from the active settings file.
/// The catalog the standalone agent runs against: the lithos built-ins and
/// the operator's `[llm]` overlay from the active settings file.
#[expect(
clippy::disallowed_methods,
reason = "Standalone agent honors OPENAI_BASE_URL from the process environment."
@ -958,19 +958,18 @@ base_url = "https://example.invalid/v1"
auth = { type = "bearer" }
default_model = "acme-aws-claude"
[providers.acme-aws.metadata.fabro]
agent_profile = "openai"
credentials = ["env:ACME_API_KEY"]
[providers.acme-aws.metadata.agent]
profile = "openai"
[providers.acme-aws.models.acme-aws-claude]
display_name = "Acme AWS Claude"
api_model = "acme-aws-claude"
limits = { context_tokens = 1000, max_output_tokens = 500 }
capabilities = { text = true, tools = true }
[providers.acme-aws.models.acme-aws-claude.metadata.fabro]
family = "claude"
agent_profile = "anthropic"
[providers.acme-aws.models.acme-aws-claude.metadata.agent]
profile = "anthropic"
"#;
/// The same provider with no models, so its default comes from the
@ -984,9 +983,8 @@ base_url = "https://example.invalid/v1"
auth = { type = "bearer" }
allow_passthrough = true
[providers.acme-aws.metadata.fabro]
agent_profile = "openai"
credentials = ["env:ACME_API_KEY"]
[providers.acme-aws.metadata.agent]
profile = "openai"
"#;
fn acme_catalog() -> Catalog {

2
lib/components/fabro-agent/src/profiles/gpt56.rs
View file

@ -237,7 +237,7 @@ mod tests {
/// OpenRouter ships disabled in the built-in catalog.
fn catalog_with_openrouter() -> Arc<Catalog> {
Arc::new(test_catalog_with_overlay(
"[providers.openrouter.metadata.fabro]
"[providers.openrouter]
enabled = true
",
))

2
lib/components/fabro-agent/src/profiles/kimi.rs
View file

@ -172,7 +172,7 @@ mod tests {
/// selectable. Enable it the way they would, to observe gateway routing.
fn catalog_with_openrouter() -> Arc<Catalog> {
Arc::new(test_catalog_with_overlay(
"[providers.openrouter.metadata.fabro]
"[providers.openrouter]
enabled = true
",
))

2
lib/components/fabro-agent/src/profiles/mod.rs
View file

@ -467,7 +467,7 @@ mod tests {
/// OpenRouter ships disabled, so an operator opts in before its models are
/// selectable.
const OPENROUTER_ENABLED: &str = "[providers.openrouter.metadata.fabro]\nenabled = true\n";
const OPENROUTER_ENABLED: &str = "[providers.openrouter]\nenabled = true\n";
fn native_tool_options(
profile_kind: AgentProfileKind,

2
lib/components/fabro-agent/tests/it/guardrails.rs
View file

@ -8,7 +8,7 @@ use fabro_llm::test_support::test_catalog;
fn profile_context_window_matches_catalog_for_default_models() {
let catalog = Arc::new(test_catalog());
for provider in catalog::listed_providers(&catalog) {
let provider_id = provider.provider.id().clone();
let provider_id = provider.id().clone();
let Some(default) = catalog::default_model(&catalog, provider_id.as_str()) else {
// Deployment-defined providers (LiteLLM, Modal, Ollama) carry no
// built-in default model.

2
lib/components/fabro-agent/tests/it/parity_matrix.rs
View file

@ -168,7 +168,7 @@ async fn make_twin_client(twin: &OpenAiTwinOptions) -> Client {
/// resolves the OpenAI-compatible codec the twin speaks.
fn litellm_twin_overlay(base_url: &str) -> String {
format!(
"[providers.litellm]\nbase_url = {}\n\n[providers.litellm.metadata.fabro]\nenabled = true\n",
"[providers.litellm]\nbase_url = {}\nenabled = true\n",
toml::Value::String(base_url.to_string())
)
}

883
lib/components/fabro-llm/catalog/fabro-policy.toml
View file

@ -1,883 +0,0 @@
# Fabro's policy layer over the lithos-llm built-in catalog.
#
# Applied after the built-ins and before operator `[llm]` overlays. Everything
# under `metadata.fabro` is Fabro policy that lithos carries verbatim:
#
# - `enabled`: whether Fabro offers the provider or model.
# - `credentials`: ordered credential references (`env:NAME`, `vault:NAME`,
# `aws_sigv4`); the first that resolves wins.
# - `extra_headers`: request headers, literal or `{{ secrets.NAME }}`.
# - `agent_profile`, `family`, `small_default`, `probe`: agent and selection
# policy read by the Fabro model resolver.
#
# Modal's default model comes from a deployment template, not this layer.
# GPT-OSS is excluded at user request. This layer cannot restore that family.
schema_version = 1
[providers."anthropic"]
priority = 100
default_model = "claude-sonnet-5"
[providers."anthropic".metadata.fabro]
api_key_url = "https://console.anthropic.com/settings/keys"
credentials = ["env:ANTHROPIC_API_KEY", "vault:ANTHROPIC_API_KEY"]
enabled = true
[providers."anthropic".models."claude-fable-5".metadata.fabro]
family = "claude-5"
agent_profile = "claude-5"
small_default = false
probe = false
[providers."anthropic".models."claude-opus-5".metadata.fabro]
family = "claude-5"
agent_profile = "claude-5"
small_default = false
probe = false
[providers."anthropic".models."claude-sonnet-5".metadata.fabro]
family = "claude-5"
agent_profile = "claude-5"
small_default = false
probe = false
[providers."anthropic".models."claude-opus-4.8".metadata.fabro]
family = "claude-4"
agent_profile = "anthropic"
small_default = false
probe = false
[providers."anthropic".models."claude-opus-4.7".metadata.fabro]
family = "claude-4"
agent_profile = "anthropic"
small_default = false
probe = false
[providers."anthropic".models."claude-opus-4.6".metadata.fabro]
family = "claude-4"
agent_profile = "anthropic"
small_default = false
probe = false
[providers."anthropic".models."claude-sonnet-4.5".metadata.fabro]
family = "claude-4"
agent_profile = "anthropic"
small_default = false
probe = false
[providers."anthropic".models."claude-sonnet-4.6".metadata.fabro]
family = "claude-4"
agent_profile = "anthropic"
small_default = false
probe = false
[providers."anthropic".models."claude-haiku-4.5".metadata.fabro]
family = "claude-4"
agent_profile = "anthropic"
small_default = true
probe = true
[providers."bedrock-openai"]
priority = 19
default_model = "gpt-5.5"
[providers."bedrock-openai".metadata.fabro]
api_key_url = "https://docs.aws.amazon.com/bedrock/latest/userguide/api-keys.html"
credentials = ["env:AWS_BEARER_TOKEN_BEDROCK", "env:BEDROCK_API_KEY", "vault:AWS_BEARER_TOKEN_BEDROCK", "vault:BEDROCK_API_KEY"]
enabled = false
[providers."bedrock-openai".models."gpt-5.5".metadata.fabro]
family = "gpt-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock-openai".models."gpt-5.4".metadata.fabro]
family = "gpt-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock"]
priority = 20
default_model = "claude-sonnet-5"
[providers."bedrock".metadata.fabro]
api_key_url = "https://docs.aws.amazon.com/bedrock/latest/userguide/api-keys.html"
credentials = ["env:AWS_BEARER_TOKEN_BEDROCK", "env:BEDROCK_API_KEY", "vault:AWS_BEARER_TOKEN_BEDROCK", "vault:BEDROCK_API_KEY", "aws_sigv4"]
enabled = false
[providers."bedrock".models."anthropic.claude-sonnet-4-6".metadata.fabro]
family = "claude-4"
agent_profile = "anthropic"
small_default = false
probe = false
[providers."bedrock".models."claude-opus-4-8".metadata.fabro]
family = "claude-4"
agent_profile = "anthropic"
small_default = false
probe = false
[providers."bedrock".models."claude-haiku-4-5".metadata.fabro]
family = "claude-4"
agent_profile = "anthropic"
small_default = true
probe = false
[providers."bedrock".models."nova-2-lite".metadata.fabro]
family = "nova-2"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock".models."llama-4-maverick".metadata.fabro]
family = "llama-4"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock".models."mistral-large-3".metadata.fabro]
family = "mistral-large"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock".models."devstral-2".metadata.fabro]
family = "devstral"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock".models."deepseek-v3.2".metadata.fabro]
family = "deepseek-v3"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock".models."kimi-k2.5".metadata.fabro]
family = "kimi-k2"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock".models."glm-5".metadata.fabro]
family = "glm"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock".models."minimax-m2.5".metadata.fabro]
family = "minimax-m2"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock".models."nemotron-3-super".metadata.fabro]
family = "nemotron-3"
agent_profile = "openai"
small_default = false
probe = false
[providers."bedrock".models."claude-fable-5".metadata.fabro]
family = "claude-5"
agent_profile = "claude-5"
small_default = false
probe = false
[providers."bedrock".models."claude-sonnet-5".metadata.fabro]
family = "claude-5"
agent_profile = "claude-5"
small_default = false
probe = false
[providers."deepseek"]
priority = 75
default_model = "deepseek-v4-flash"
[providers."deepseek".metadata.fabro]
api_key_url = "https://platform.deepseek.com/api_keys"
credentials = ["env:DEEPSEEK_API_KEY", "vault:DEEPSEEK_API_KEY"]
enabled = true
[providers."deepseek".models."deepseek-v4-flash".metadata.fabro]
reasoning_by_default = true
family = "deepseek-v4"
agent_profile = "openai"
small_default = true
probe = true
[providers."deepseek".models."deepseek-v4-pro".metadata.fabro]
reasoning_by_default = true
family = "deepseek-v4"
agent_profile = "openai"
small_default = false
probe = false
[providers."fireworks"]
priority = 30
default_model = "kimi-k2.7-code"
[providers."fireworks".metadata.fabro]
api_key_url = "https://app.fireworks.ai/settings/users/api-keys"
credentials = ["env:FIREWORKS_API_KEY", "vault:FIREWORKS_API_KEY"]
enabled = false
# Moonshot documents K3 as always reasoning with low, high, and max effort;
# lithos leaves the levels unverified, so Fabro records them here.
[providers."fireworks".models."kimi-k3"]
capabilities = { reasoning_effort = { minimal = false, low = true, medium = false, high = true, xhigh = false, max = true } }
[providers."fireworks".models."kimi-k3".metadata.fabro]
family = "kimi-k3"
agent_profile = "kimi"
small_default = false
probe = false
[providers."fireworks".models."kimi-k3-fast".metadata.fabro]
family = "kimi-k3"
agent_profile = "kimi"
small_default = false
probe = false
[providers."fireworks".models."kimi-k2.7-code".metadata.fabro]
family = "kimi-k2"
agent_profile = "openai"
small_default = false
probe = false
[providers."fireworks".models."kimi-k2.6".metadata.fabro]
family = "kimi-k2"
agent_profile = "openai"
small_default = false
probe = false
[providers."fireworks".models."deepseek-v4-pro".metadata.fabro]
reasoning_by_default = true
family = "deepseek-v4"
agent_profile = "openai"
small_default = false
probe = false
[providers."fireworks".models."deepseek-v4-flash".metadata.fabro]
reasoning_by_default = true
family = "deepseek-v4"
agent_profile = "openai"
small_default = false
probe = false
[providers."fireworks".models."glm-5.2".metadata.fabro]
family = "glm-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."fireworks".models."minimax-m2.7".metadata.fabro]
family = "minimax-m2"
agent_profile = "openai"
small_default = false
probe = false
[providers."fireworks".models."qwen3.7-plus".metadata.fabro]
family = "qwen3"
agent_profile = "openai"
small_default = false
probe = false
[providers."gemini"]
priority = 80
default_model = "gemini-3.5-flash"
[providers."gemini".metadata.fabro]
api_key_url = "https://aistudio.google.com/apikey"
credentials = ["env:GEMINI_API_KEY", "env:GOOGLE_API_KEY", "vault:GEMINI_API_KEY"]
enabled = true
[providers."gemini".models."gemini-3.1-pro-preview".metadata.fabro]
family = "gemini-3"
agent_profile = "gemini"
small_default = false
probe = false
[providers."gemini".models."gemini-3.1-pro-preview-customtools".metadata.fabro]
family = "gemini-3"
agent_profile = "gemini"
small_default = false
probe = false
[providers."gemini".models."gemini-3.5-flash".metadata.fabro]
family = "gemini-3"
agent_profile = "gemini"
small_default = false
probe = false
[providers."gemini".models."gemini-3-flash-preview".metadata.fabro]
family = "gemini-3"
agent_profile = "gemini"
small_default = false
probe = false
[providers."gemini".models."gemini-3.1-flash-lite".metadata.fabro]
family = "gemini-3"
agent_profile = "gemini"
small_default = true
probe = false
[providers."inception"]
priority = 40
default_model = "mercury-2"
[providers."inception".metadata.fabro]
api_key_url = "https://console.inceptionlabs.ai/api-keys"
credentials = ["env:INCEPTION_API_KEY", "vault:INCEPTION_API_KEY"]
enabled = true
[providers."inception".models."mercury-2".metadata.fabro]
family = "mercury"
agent_profile = "openai"
small_default = false
probe = false
[providers."litellm"]
priority = 50
[providers."litellm".metadata.fabro]
credentials = ["env:LITELLM_API_KEY", "vault:LITELLM_API_KEY"]
enabled = false
[providers."minimax"]
priority = 50
default_model = "minimax-m2.5"
[providers."minimax".metadata.fabro]
api_key_url = "https://platform.minimaxi.com/user-center/basic-information/interface-key"
credentials = ["env:MINIMAX_API_KEY", "vault:MINIMAX_API_KEY"]
enabled = true
[providers."minimax".models."minimax-m2.5".metadata.fabro]
family = "minimax-m2"
agent_profile = "openai"
small_default = false
probe = false
[providers."modal"]
priority = 75
[providers."modal".metadata.fabro]
api_key_url = "https://modal.com/docs/guide/endpoints#proxy-tokens"
enabled = false
[providers."modal".metadata.fabro.extra_headers]
"Modal-Key" = "{{ secrets.MODAL_TOKEN_ID }}"
"Modal-Secret" = "{{ secrets.MODAL_TOKEN_SECRET }}"
[providers."moonshot"]
priority = 70
default_model = "kimi-k3"
[providers."moonshot".metadata.fabro]
api_key_url = "https://platform.kimi.ai/console/api-keys"
credentials = ["env:MOONSHOT_API_KEY", "env:KIMI_API_KEY", "vault:MOONSHOT_API_KEY", "vault:KIMI_API_KEY"]
enabled = true
[providers."moonshot".models."kimi-k2.5".metadata.fabro]
family = "kimi-k2"
agent_profile = "kimi"
small_default = false
probe = false
# Moonshot documents K3 as always reasoning with low, high, and max effort;
# lithos leaves the levels unverified, so Fabro records them here.
[providers."moonshot".models."kimi-k3"]
capabilities = { reasoning_effort = { minimal = false, low = true, medium = false, high = true, xhigh = false, max = true } }
[providers."moonshot".models."kimi-k3".metadata.fabro]
family = "kimi-k3"
agent_profile = "kimi"
small_default = false
probe = false
[providers."ollama"]
priority = 30
[providers."ollama".metadata.fabro]
enabled = false
[providers."openai"]
priority = 90
default_model = "gpt-5.6-sol"
[providers."openai".metadata.fabro]
api_key_url = "https://platform.openai.com/api-keys"
credentials = ["env:OPENAI_API_KEY", "vault:OPENAI_API_KEY"]
enabled = true
[providers."openai".models."gpt-5.6-sol".metadata.fabro]
family = "gpt-5"
agent_profile = "gpt56"
small_default = false
probe = false
[providers."openai".models."gpt-5.6-terra".metadata.fabro]
family = "gpt-5"
agent_profile = "gpt56"
small_default = false
probe = false
[providers."openai".models."gpt-5.6-luna".metadata.fabro]
family = "gpt-5"
agent_profile = "gpt56"
small_default = false
probe = false
[providers."openai".models."gpt-5.4".metadata.fabro]
family = "gpt-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."openai".models."gpt-5.5".metadata.fabro]
family = "gpt-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."openai".models."gpt-5.5-pro".metadata.fabro]
family = "gpt-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."openai".models."gpt-5.4-pro".metadata.fabro]
family = "gpt-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."openai".models."gpt-5.4-mini".metadata.fabro]
family = "gpt-5"
agent_profile = "openai"
small_default = true
probe = true
[providers."openrouter"]
priority = 25
default_model = "claude-sonnet-5"
[providers."openrouter".metadata.fabro]
api_key_url = "https://openrouter.ai/keys"
credentials = ["env:OPENROUTER_API_KEY", "vault:OPENROUTER_API_KEY"]
enabled = false
[providers."openrouter".models."claude-fable-5".metadata.fabro]
family = "claude-5"
agent_profile = "claude-5"
small_default = false
probe = false
[providers."openrouter".models."claude-opus-5".metadata.fabro]
training = "2026-05-01"
knowledge_cutoff = "May 2026"
family = "claude-5"
agent_profile = "claude-5"
small_default = false
probe = false
[providers."openrouter".models."claude-sonnet-5".metadata.fabro]
training = "2026-01-01"
knowledge_cutoff = "Jan 2026"
family = "claude-5"
agent_profile = "claude-5"
small_default = false
probe = false
[providers."openrouter".models."claude-opus-4.8".metadata.fabro]
training = "2026-01-01"
knowledge_cutoff = "Jan 2026"
family = "claude-4"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."claude-opus-4.7".metadata.fabro]
family = "claude-4"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."claude-sonnet-4.6".metadata.fabro]
family = "claude-4"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."claude-haiku-4.5".metadata.fabro]
family = "claude-4"
agent_profile = "openai"
small_default = true
probe = false
[providers."openrouter".models."gpt-5.6-sol".metadata.fabro]
training = "2026-02-16"
knowledge_cutoff = "February 16, 2026"
family = "gpt-5"
agent_profile = "gpt56"
small_default = false
probe = false
[providers."openrouter".models."gpt-5.6-terra".metadata.fabro]
training = "2026-02-16"
knowledge_cutoff = "February 16, 2026"
family = "gpt-5"
agent_profile = "gpt56"
small_default = false
probe = false
[providers."openrouter".models."gpt-5.6-luna".metadata.fabro]
training = "2026-02-16"
knowledge_cutoff = "February 16, 2026"
family = "gpt-5"
agent_profile = "gpt56"
small_default = false
probe = false
[providers."openrouter".models."gpt-5.4".metadata.fabro]
family = "gpt-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."gpt-5.5".metadata.fabro]
family = "gpt-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."gemini-3.1-pro-preview".metadata.fabro]
family = "gemini-3"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."gemini-3.5-flash".metadata.fabro]
family = "gemini-3"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."mimo-v2.5-pro".metadata.fabro]
family = "mimo-v2"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."minimax-m2.7".metadata.fabro]
family = "minimax-m2"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."deepseek-v4-pro".metadata.fabro]
reasoning_by_default = true
family = "deepseek-v4"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."deepseek-v4-flash".metadata.fabro]
reasoning_by_default = true
family = "deepseek-v4"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."kimi-k2.6".metadata.fabro]
family = "kimi-k2"
agent_profile = "kimi"
small_default = false
probe = false
# Moonshot documents K3 as always reasoning with low, high, and max effort;
# lithos leaves the levels unverified, so Fabro records them here.
[providers."openrouter".models."kimi-k3"]
capabilities = { reasoning_effort = { minimal = false, low = true, medium = false, high = true, xhigh = false, max = true } }
[providers."openrouter".models."kimi-k3".metadata.fabro]
family = "kimi-k3"
agent_profile = "kimi"
small_default = false
probe = false
[providers."openrouter".models."laguna-s-2.1".metadata.fabro]
family = "laguna-2"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."laguna-xs-2.1".metadata.fabro]
family = "laguna-2"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."qwen3-coder".metadata.fabro]
family = "qwen3"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."qwen3.6-flash".metadata.fabro]
family = "qwen3"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."qwen3.8-max".metadata.fabro]
family = "qwen3"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."glm-5.2".metadata.fabro]
family = "glm-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."glm-4.6".metadata.fabro]
family = "glm-4"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."nemotron-3-super-120b-a12b".metadata.fabro]
family = "nemotron-3"
agent_profile = "openai"
small_default = false
probe = false
[providers."openrouter".models."devstral-2512".metadata.fabro]
family = "devstral"
agent_profile = "openai"
small_default = false
probe = false
[providers."poolside"]
priority = 65
default_model = "laguna-s-2.1"
[providers."poolside".metadata.fabro]
api_key_url = "https://platform.poolside.ai"
credentials = ["env:POOLSIDE_API_KEY", "vault:POOLSIDE_API_KEY"]
enabled = true
[providers."poolside".models."laguna-s-2.1".metadata.fabro]
family = "laguna-2"
agent_profile = "openai"
small_default = false
probe = false
[providers."poolside".models."laguna-xs-2.1".metadata.fabro]
family = "laguna-2"
agent_profile = "openai"
small_default = true
probe = true
[providers."venice"]
priority = 35
default_model = "deepseek-v4-flash"
[providers."venice".metadata.fabro]
credentials = ["env:VENICE_API_KEY", "vault:VENICE_API_KEY"]
enabled = true
[providers."venice".models."kimi-k3".metadata.fabro]
reasoning_by_default = true
family = "kimi-k3"
agent_profile = "kimi"
small_default = false
probe = false
[providers."venice".models."kimi-k3-fast".metadata.fabro]
reasoning_by_default = true
family = "kimi-k3"
agent_profile = "kimi"
small_default = false
probe = false
[providers."venice".models."grok-4.6".metadata.fabro]
reasoning_by_default = true
family = "grok-4"
agent_profile = "openai"
small_default = false
probe = false
[providers."venice".models."glm-5.3".metadata.fabro]
reasoning_by_default = true
family = "glm-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."venice".models."deepseek-v4-flash".metadata.fabro]
reasoning_by_default = true
family = "deepseek-v4"
agent_profile = "openai"
small_default = false
probe = false
[providers."venice".models."deepseek-v4-pro".metadata.fabro]
reasoning_by_default = true
family = "deepseek-v4"
agent_profile = "openai"
small_default = false
probe = false
[providers."venice".models."qwen3.8-max".metadata.fabro]
reasoning_by_default = true
family = "qwen3"
agent_profile = "openai"
small_default = false
probe = false
[providers."venice".models."qwen3.8-27b".metadata.fabro]
reasoning_by_default = true
family = "qwen3.8"
agent_profile = "openai"
small_default = false
probe = false
[providers."zai"]
priority = 60
default_model = "glm-5.2"
[providers."zai".metadata.fabro]
api_key_url = "https://open.bigmodel.cn/usercenter/apikeys"
credentials = ["env:ZAI_API_KEY", "vault:ZAI_API_KEY"]
enabled = true
[providers."zai".models."glm-5.2".metadata.fabro]
family = "glm-5"
agent_profile = "openai"
small_default = false
probe = false
[providers."zai".models."glm-4.7".metadata.fabro]
family = "glm-4"
agent_profile = "openai"
small_default = false
probe = false
# ChatGPT OAuth (Codex) access to the OpenAI roster. lithos speaks the Codex
# deployment through the `openai` adapter in codex mode; this provider exists
# so a `fabro provider login --provider openai` device-flow credential routes
# here while an API key keeps routing to `openai`. It is hidden from listings
# and stands in for `openai` when `openai` itself has no credentials.
[providers."openai-codex"]
display_name = "OpenAI (ChatGPT)"
adapter = "openai"
codec = "openai-responses"
base_url = "https://chatgpt.com/backend-api/codex"
priority = 89
allow_passthrough = true
default_model = "gpt-5.6-sol"
adapter_options = { mode = "codex" }
default_headers = { originator = "fabro" }
[providers."openai-codex".auth]
type = "bearer"
[providers."openai-codex".metadata.fabro]
enabled = true
stands_in_for = "openai"
credentials = ["vault:OPENAI_CODEX"]
[providers."openai-codex".models."gpt-5.6-sol"]
display_name = "GPT-5.6 Sol"
aliases = ["sol", "gpt-sol", "gpt56-sol", "gpt-56-sol", "gpt-5.6", "gpt56", "gpt-56"]
api_model = "gpt-5.6-sol"
limits = { context_tokens = 1050000, max_output_tokens = 128000 }
capabilities = { text = true, images = true, documents = true, tools = true, response_format = { json_object = true, json_schema = true }, reasoning = true, caching = true, cache_routing = true, tool_choice = { required = true, named = true }, speed = { fast = true, economical = true } }
protocol_options = { reasoning_effort_levels = true }
pricing = { input_usd_micros_per_million = 4000000, output_usd_micros_per_million = 20000000, cached_input_usd_micros_per_million = 400000, cache_write_usd_micros_per_million = 5000000, long_context = { above_input_tokens = 272000, input_usd_micros_per_million = 8000000, output_usd_micros_per_million = 30000000, cached_input_usd_micros_per_million = 800000, cache_write_usd_micros_per_million = 10000000 }, speed = { fast = { input_usd_micros_per_million = 8000000, output_usd_micros_per_million = 40000000, cached_input_usd_micros_per_million = 800000, cache_write_usd_micros_per_million = 10000000 }, economical = { input_usd_micros_per_million = 2000000, output_usd_micros_per_million = 10000000, cached_input_usd_micros_per_million = 200000, cache_write_usd_micros_per_million = 2500000 } } }
[providers."openai-codex".models."gpt-5.6-sol".metadata.fabro]
agent_profile = "gpt56"
family = "gpt-5"
small_default = false
probe = false
[providers."openai-codex".models."gpt-5.6-terra"]
display_name = "GPT-5.6 Terra"
aliases = ["terra", "gpt-terra", "gpt56-terra", "gpt-56-terra"]
api_model = "gpt-5.6-terra"
limits = { context_tokens = 1050000, max_output_tokens = 128000 }
capabilities = { text = true, images = true, documents = true, tools = true, response_format = { json_object = true, json_schema = true }, reasoning = true, caching = true, cache_routing = true, tool_choice = { required = true, named = true }, speed = { fast = true, economical = true } }
protocol_options = { reasoning_effort_levels = true }
pricing = { input_usd_micros_per_million = 2000000, output_usd_micros_per_million = 12000000, cached_input_usd_micros_per_million = 200000, cache_write_usd_micros_per_million = 2500000, long_context = { above_input_tokens = 272000, input_usd_micros_per_million = 4000000, output_usd_micros_per_million = 18000000, cached_input_usd_micros_per_million = 400000, cache_write_usd_micros_per_million = 5000000 }, speed = { fast = { input_usd_micros_per_million = 4000000, output_usd_micros_per_million = 24000000, cached_input_usd_micros_per_million = 400000, cache_write_usd_micros_per_million = 5000000 }, economical = { input_usd_micros_per_million = 1000000, output_usd_micros_per_million = 6000000, cached_input_usd_micros_per_million = 100000, cache_write_usd_micros_per_million = 1250000 } } }
[providers."openai-codex".models."gpt-5.6-terra".metadata.fabro]
agent_profile = "gpt56"
family = "gpt-5"
small_default = false
probe = false
[providers."openai-codex".models."gpt-5.6-luna"]
display_name = "GPT-5.6 Luna"
aliases = ["luna", "gpt-luna", "gpt56-luna", "gpt-56-luna"]
api_model = "gpt-5.6-luna"
limits = { context_tokens = 1050000, max_output_tokens = 128000 }
capabilities = { text = true, images = true, documents = true, tools = true, response_format = { json_object = true, json_schema = true }, reasoning = true, caching = true, cache_routing = true, tool_choice = { required = true, named = true }, speed = { fast = true, economical = true } }
protocol_options = { reasoning_effort_levels = true }
pricing = { input_usd_micros_per_million = 200000, output_usd_micros_per_million = 1200000, cached_input_usd_micros_per_million = 20000, cache_write_usd_micros_per_million = 250000, long_context = { above_input_tokens = 272000, input_usd_micros_per_million = 400000, output_usd_micros_per_million = 1800000, cached_input_usd_micros_per_million = 40000, cache_write_usd_micros_per_million = 500000 }, speed = { fast = { input_usd_micros_per_million = 400000, output_usd_micros_per_million = 2400000, cached_input_usd_micros_per_million = 40000, cache_write_usd_micros_per_million = 500000 }, economical = { input_usd_micros_per_million = 100000, output_usd_micros_per_million = 600000, cached_input_usd_micros_per_million = 10000, cache_write_usd_micros_per_million = 125000 } } }
[providers."openai-codex".models."gpt-5.6-luna".metadata.fabro]
agent_profile = "gpt56"
family = "gpt-5"
small_default = false
probe = false
[providers."openai-codex".models."gpt-5.4"]
display_name = "GPT-5.4"
aliases = ["gpt54", "gpt-54", "gpt-5.2", "gpt5", "gpt-5.3-codex", "codex"]
api_model = "gpt-5.4"
limits = { context_tokens = 1050000, max_output_tokens = 128000 }
capabilities = { text = true, images = true, documents = true, tools = true, response_format = { json_object = true, json_schema = true }, reasoning = true, caching = true, cache_routing = true, sampling = true, tool_choice = { required = true, named = true }, speed = { fast = true, economical = true } }
protocol_options = { reasoning_effort_levels = true }
pricing = { input_usd_micros_per_million = 2500000, output_usd_micros_per_million = 15000000, cached_input_usd_micros_per_million = 250000, long_context = { above_input_tokens = 272000, input_usd_micros_per_million = 5000000, output_usd_micros_per_million = 22500000, cached_input_usd_micros_per_million = 500000 }, speed = { fast = { input_usd_micros_per_million = 5000000, output_usd_micros_per_million = 30000000, cached_input_usd_micros_per_million = 500000 }, economical = { input_usd_micros_per_million = 1250000, output_usd_micros_per_million = 7500000, cached_input_usd_micros_per_million = 125000 } } }
[providers."openai-codex".models."gpt-5.4".metadata.fabro]
agent_profile = "openai"
family = "gpt-5"
small_default = false
probe = false
[providers."openai-codex".models."gpt-5.5"]
display_name = "GPT-5.5"
aliases = ["gpt55", "gpt-55"]
api_model = "gpt-5.5"
limits = { context_tokens = 1050000, max_output_tokens = 128000 }
capabilities = { text = true, images = true, documents = true, tools = true, response_format = { json_object = true, json_schema = true }, reasoning = true, caching = true, cache_routing = true, tool_choice = { required = true, named = true }, speed = { fast = true, economical = true } }
protocol_options = { reasoning_effort_levels = true }
pricing = { input_usd_micros_per_million = 5000000, output_usd_micros_per_million = 30000000, cached_input_usd_micros_per_million = 500000, long_context = { above_input_tokens = 272000, input_usd_micros_per_million = 10000000, output_usd_micros_per_million = 45000000, cached_input_usd_micros_per_million = 1000000 }, speed = { fast = { input_usd_micros_per_million = 12500000, output_usd_micros_per_million = 75000000, cached_input_usd_micros_per_million = 1250000 }, economical = { input_usd_micros_per_million = 2500000, output_usd_micros_per_million = 15000000, cached_input_usd_micros_per_million = 250000 } } }
[providers."openai-codex".models."gpt-5.5".metadata.fabro]
agent_profile = "openai"
family = "gpt-5"
small_default = false
probe = false
[providers."openai-codex".models."gpt-5.4-mini"]
display_name = "GPT-5.4 Mini"
aliases = ["gpt54-mini", "gpt-54-mini", "gpt-5.3-codex-spark", "codex-spark"]
api_model = "gpt-5.4-mini"
limits = { context_tokens = 400000, max_output_tokens = 128000 }
capabilities = { text = true, images = true, documents = true, tools = true, response_format = { json_object = true, json_schema = true }, reasoning = true, caching = true, cache_routing = true, sampling = true, tool_choice = { required = true, named = true }, speed = { fast = true, economical = true } }
protocol_options = { reasoning_effort_levels = true }
pricing = { input_usd_micros_per_million = 750000, output_usd_micros_per_million = 4500000, cached_input_usd_micros_per_million = 75000, speed = { fast = { input_usd_micros_per_million = 1500000, output_usd_micros_per_million = 9000000, cached_input_usd_micros_per_million = 150000 }, economical = { input_usd_micros_per_million = 375000, output_usd_micros_per_million = 2250000, cached_input_usd_micros_per_million = 37500 } } }
[providers."openai-codex".models."gpt-5.4-mini".metadata.fabro]
agent_profile = "openai"
family = "gpt-5"
small_default = true
probe = true

31
lib/components/fabro-llm/src/api.rs
View file

@ -1,7 +1,7 @@
//! API projections of the catalog for `GET /models` and `GET /providers`.
//!
//! Every row is a lithos catalog entry plus its Fabro policy, stamped with
//! whether the caller holds credential material for the provider.
//! Every row is a lithos catalog entry stamped with whether the caller holds
//! credential material for the provider.
use std::collections::HashSet;
@ -9,9 +9,9 @@ use fabro_types::controls::REASONING_EFFORTS;
use fabro_types::{
Model, ModelControls, ModelCosts, ModelFeatures, ModelLimits, Provider, ProviderId,
};
use lithos_llm::catalog::Catalog;
use lithos_llm::catalog::{Catalog, CatalogProvider};
use crate::catalog::{self, ModelEntry, ProviderEntry};
use crate::catalog::{self, ModelEntry};
const USD_MICROS_PER_USD: f64 = 1_000_000.0;
@ -29,7 +29,7 @@ pub fn models(catalog: &Catalog, configured: &HashSet<ProviderId>) -> Vec<Model>
pub fn providers(catalog: &Catalog, configured: &HashSet<ProviderId>) -> Vec<Provider> {
catalog::listed_providers(catalog)
.iter()
.map(|entry| provider_view(entry, configured.contains(entry.provider.id())))
.map(|provider| provider_view(provider, configured.contains(provider.id())))
.collect()
}
@ -41,11 +41,9 @@ fn model_view(entry: &ModelEntry<'_>, configured: bool) -> Model {
Model {
id: model.id().clone(),
provider: entry.provider.id().clone(),
family: entry
.policy
.family
.clone()
.unwrap_or_else(|| model.id().to_string()),
family: model
.family()
.map_or_else(|| model.id().to_string(), str::to_string),
display_name: model.display_name().to_string(),
limits: ModelLimits {
context_window: limits.map_or(0, |limits| saturating_i64(limits.context_tokens)),
@ -54,8 +52,8 @@ fn model_view(entry: &ModelEntry<'_>, configured: bool) -> Model {
.filter(|tokens| *tokens > 0)
.map(saturating_i64),
},
training: entry.policy.training.clone(),
knowledge_cutoff: entry.policy.knowledge_cutoff.clone(),
training: model.training_cutoff().map(str::to_string),
knowledge_cutoff: model.knowledge_cutoff().map(str::to_string),
features: ModelFeatures {
tools: capabilities.tools().is_supported(),
vision: capabilities.images().is_supported(),
@ -81,22 +79,21 @@ fn model_view(entry: &ModelEntry<'_>, configured: bool) -> Model {
.and_then(|pricing| pricing.cached_input_usd_micros_per_million)
.map(usd_per_million),
},
estimated_output_tps: entry.policy.estimated_output_tps,
estimated_output_tps: model.estimated_output_tps(),
aliases: model.aliases().to_vec(),
default: entry.provider.default_model() == Some(model.id().as_str()),
small_default: entry.policy.small_default,
small_default: model.is_small_default(),
configured,
}
}
fn provider_view(entry: &ProviderEntry<'_>, configured: bool) -> Provider {
let provider = entry.provider;
fn provider_view(provider: &CatalogProvider, configured: bool) -> Provider {
Provider {
id: provider.id().clone(),
display_name: provider.display_name().to_string(),
adapter: provider.adapter().as_str().to_string(),
base_url: provider.base_url().to_string(),
api_key_url: entry.policy.api_key_url.clone(),
api_key_url: provider.api_key_url().map(str::to_string),
priority: provider.priority(),
aliases: provider.aliases().to_vec(),
model_count: u32::try_from(catalog::provider_models(provider).len()).unwrap_or(u32::MAX),

286
lib/components/fabro-llm/src/catalog.rs
View file

@ -1,20 +1,22 @@
//! Catalog construction and Fabro-policy queries.
//! Catalog construction and the queries Fabro's dispatch boundaries share.
//!
//! Layer order is fixed: lithos built-ins, then Fabro's policy layer, then the
//! operator's `[llm]` overlay. Every query here reads Fabro policy from the
//! `metadata.fabro` namespace and never bypasses `enabled`.
//! Layer order is fixed: lithos built-ins, then the operator's `[llm]`
//! overlay. Provider and model facts, `enabled`, `stands_in_for`,
//! `small_default`, and `probe` are lithos core fields. The agent harness a
//! model expects lives in the shared `metadata.agent` namespace, which Pebble
//! reads too. Every query here skips disabled providers.
use std::collections::{BTreeMap, HashSet};
use fabro_config::LlmLayer;
use fabro_static::EnvVars;
use fabro_types::catalog_policy::{self, ModelPolicy, ProviderPolicy};
use fabro_types::{AgentProfileKind, Cost, ModelId, ModelRef, ProviderId, TokenCounts};
use lithos_llm::catalog::{Catalog, CatalogError, CatalogModel, CatalogProvider};
use lithos_llm::catalog::{Catalog, CatalogError, CatalogModel, CatalogProvider, Metadata};
use lithos_llm::resolver::ResolvedRoute;
use serde::Deserialize;
/// Fabro's policy layer, applied above the lithos built-ins.
pub const FABRO_POLICY_TOML: &str = include_str!("../catalog/fabro-policy.toml");
/// The metadata namespace agent harnesses read.
const AGENT_METADATA_NAMESPACE: &str = "agent";
/// Builds the effective catalog.
///
@ -25,9 +27,7 @@ pub fn build_catalog(
overlay: &LlmLayer,
env_lookup: &dyn Fn(&str) -> Option<String>,
) -> Result<Catalog, CatalogError> {
let mut builder = Catalog::builder()
.with_builtin()
.toml_layer("fabro-policy.toml", FABRO_POLICY_TOML)?;
let mut builder = Catalog::builder().with_builtin();
if !overlay.is_empty() {
let mut document = overlay.to_overlay_toml();
document.insert_str(0, "schema_version = 1\n");
@ -43,50 +43,78 @@ pub fn build_catalog(
builder.build()
}
/// A provider with its Fabro policy attached.
#[derive(Debug, Clone)]
pub struct ProviderEntry<'a> {
pub provider: &'a CatalogProvider,
pub policy: ProviderPolicy,
}
/// A model with its Fabro policy attached.
#[derive(Debug, Clone)]
pub struct ModelEntry<'a> {
pub provider: &'a CatalogProvider,
pub model: &'a CatalogModel,
pub policy: ModelPolicy,
}
impl ModelEntry<'_> {
/// Whether requests to this model reason when no effort is requested.
///
/// Fabro policy can state it outright. Otherwise a model that supports
/// reasoning and takes named effort levels reasons by default, while one
/// that needs an explicit thinking budget does not.
#[must_use]
pub fn reasons_by_default(&self) -> bool {
self.policy.reasoning_by_default.unwrap_or_else(|| {
self.model.capabilities().reasoning().is_supported()
&& self.model.protocol_options().reasoning_effort_levels
})
}
#[must_use]
pub fn agent_profile(&self) -> AgentProfileKind {
catalog_policy::effective_agent_profile(self.provider, self.model)
}
}
/// The catalog with no operator overlay: lithos built-ins plus Fabro policy.
/// The catalog with no operator overlay: the lithos built-ins.
///
/// Used where no settings file is in play, such as the standalone hook
/// runner. Servers and the CLI build from the operator's `[llm]` overlay
/// with [`build_catalog`] instead.
#[must_use]
pub fn default_catalog() -> Catalog {
build_catalog(&LlmLayer::default(), &|_| None)
.expect("the built-in catalog and Fabro policy layer always build")
build_catalog(&LlmLayer::default(), &|_| None).expect("the built-in catalog always builds")
}
/// A model on the provider that offers it.
#[derive(Debug, Clone)]
pub struct ModelEntry<'a> {
pub provider: &'a CatalogProvider,
pub model: &'a CatalogModel,
}
impl ModelEntry<'_> {
/// Whether requests to this model reason when no effort is requested.
///
/// The catalog can state it outright under `metadata.agent`. Otherwise a
/// model that supports reasoning and takes named effort levels reasons by
/// default, while one that needs an explicit thinking budget does not.
#[must_use]
pub fn reasons_by_default(&self) -> bool {
agent_metadata(self.model.metadata())
.reasoning_by_default
.or(agent_metadata(self.provider.metadata()).reasoning_by_default)
.unwrap_or_else(|| {
self.model.capabilities().reasoning().is_supported()
&& self.model.protocol_options().reasoning_effort_levels
})
}
/// The agent harness this model runs under: the model's own answer, then
/// the provider's, then the profile implied by the provider's adapter.
#[must_use]
pub fn agent_profile(&self) -> AgentProfileKind {
agent_metadata(self.model.metadata())
.profile
.unwrap_or_else(|| provider_agent_profile(self.provider))
}
}
/// The `metadata.agent` namespace on a catalog entry. Malformed metadata
/// falls back to the defaults; the lithos built-ins are validated in lithos.
#[derive(Debug, Default, Deserialize)]
#[serde(default)]
struct AgentMetadata {
profile: Option<AgentProfileKind>,
reasoning_by_default: Option<bool>,
}
fn agent_metadata(metadata: &Metadata) -> AgentMetadata {
metadata
.namespace::<AgentMetadata>(AGENT_METADATA_NAMESPACE)
.ok()
.flatten()
.unwrap_or_default()
}
/// The agent profile a provider's models run under unless a model row says
/// otherwise: the provider's `metadata.agent.profile`, else the profile
/// implied by its wire protocol.
fn provider_agent_profile(provider: &CatalogProvider) -> AgentProfileKind {
agent_metadata(provider.metadata())
.profile
.unwrap_or_else(|| match provider.adapter().as_str() {
"anthropic" | "bedrock" => AgentProfileKind::Anthropic,
"gemini" => AgentProfileKind::Gemini,
_ => AgentProfileKind::OpenAi,
})
}
/// Estimates the catalog cost of `usage` on `model`, when the catalog prices
@ -101,21 +129,16 @@ pub fn estimate_cost(catalog: &Catalog, model: &ModelRef, usage: TokenCounts) ->
/// Enabled providers, highest priority first, ties broken by id.
#[must_use]
pub fn enabled_providers(catalog: &Catalog) -> Vec<ProviderEntry<'_>> {
pub fn enabled_providers(catalog: &Catalog) -> Vec<&CatalogProvider> {
let mut providers: Vec<_> = catalog
.providers()
.map(|provider| ProviderEntry {
provider,
policy: catalog_policy::provider_policy(provider),
})
.filter(|entry| entry.policy.is_enabled())
.filter(|provider| provider.is_enabled())
.collect();
providers.sort_by(|left, right| {
right
.provider
.priority()
.cmp(&left.provider.priority())
.then_with(|| left.provider.id().cmp(right.provider.id()))
.cmp(&left.priority())
.then_with(|| left.id().cmp(right.id()))
});
providers
}
@ -123,10 +146,10 @@ pub fn enabled_providers(catalog: &Catalog) -> Vec<ProviderEntry<'_>> {
/// Enabled providers that Fabro lists to operators. Stand-in providers such
/// as `openai-codex` route requests but are not offerings of their own.
#[must_use]
pub fn listed_providers(catalog: &Catalog) -> Vec<ProviderEntry<'_>> {
pub fn listed_providers(catalog: &Catalog) -> Vec<&CatalogProvider> {
enabled_providers(catalog)
.into_iter()
.filter(|entry| entry.policy.stands_in_for.is_none())
.filter(|provider| provider.stands_in_for().is_none())
.collect()
}
@ -135,82 +158,70 @@ pub fn listed_providers(catalog: &Catalog) -> Vec<ProviderEntry<'_>> {
pub fn enabled_provider_ids(catalog: &Catalog) -> HashSet<ProviderId> {
enabled_providers(catalog)
.into_iter()
.map(|entry| entry.provider.id().clone())
.map(|provider| provider.id().clone())
.collect()
}
/// Looks up an enabled provider by id or alias.
#[must_use]
pub fn provider<'a>(catalog: &'a Catalog, selector: &str) -> Option<ProviderEntry<'a>> {
let provider = catalog.provider(selector).ok()?;
let policy = catalog_policy::provider_policy(provider);
policy
.is_enabled()
.then_some(ProviderEntry { provider, policy })
pub fn provider<'a>(catalog: &'a Catalog, selector: &str) -> Option<&'a CatalogProvider> {
catalog
.provider(selector)
.ok()
.filter(|provider| provider.is_enabled())
}
/// Canonicalizes a provider id or alias to its catalog id, when enabled.
#[must_use]
pub fn canonical_provider_id(catalog: &Catalog, selector: &str) -> Option<ProviderId> {
provider(catalog, selector).map(|entry| entry.provider.id().clone())
provider(catalog, selector).map(|provider| provider.id().clone())
}
/// Enabled models of an enabled provider, in catalog order.
/// The models of a provider, in catalog order.
#[must_use]
pub fn provider_models(provider: &CatalogProvider) -> Vec<ModelEntry<'_>> {
provider
.models()
.map(|model| ModelEntry {
provider,
model,
policy: catalog_policy::model_policy(model),
})
.filter(|entry| entry.policy.is_enabled())
.map(|model| ModelEntry { provider, model })
.collect()
}
/// Every enabled model across listed providers, provider priority order.
/// Every model across listed providers, provider priority order.
#[must_use]
pub fn models(catalog: &Catalog) -> Vec<ModelEntry<'_>> {
listed_providers(catalog)
.into_iter()
.flat_map(|entry| provider_models(entry.provider))
.flat_map(provider_models)
.collect()
}
/// Finds an enabled model on an enabled provider by id, alias, or wire id.
/// Finds a model on an enabled provider by id, alias, or wire id.
#[must_use]
pub fn model_on_provider<'a>(
catalog: &'a Catalog,
provider_selector: &str,
model_selector: &str,
) -> Option<ModelEntry<'a>> {
let entry = provider(catalog, provider_selector)?;
let provider = provider(catalog, provider_selector)?;
// lithos matches ids and aliases. The provider's wire id (an aggregator's
// `vendor/model`) is accepted too, so a selector copied from the
// provider's own listing lands on the catalog row instead of passing
// through unknown.
let model = entry.provider.model(model_selector).or_else(|| {
entry
.provider
let model = provider.model(model_selector).or_else(|| {
provider
.models()
.find(|model| model.api_model() == model_selector)
})?;
let policy = catalog_policy::model_policy(model);
policy.is_enabled().then_some(ModelEntry {
provider: entry.provider,
model,
policy,
})
Some(ModelEntry { provider, model })
}
/// Enabled models matching `selector` by id or alias, ordered like lithos
/// selection: exact ids before aliases, then provider priority.
/// Models matching `selector` by id or alias, ordered like lithos selection:
/// exact ids before aliases, then provider priority.
#[must_use]
pub fn models_matching<'a>(catalog: &'a Catalog, selector: &str) -> Vec<ModelEntry<'a>> {
let mut matches: Vec<_> = enabled_providers(catalog)
.into_iter()
.flat_map(|entry| provider_models(entry.provider))
.flat_map(provider_models)
.filter(|entry| {
entry.model.id().as_str() == selector
|| entry.model.aliases().iter().any(|alias| alias == selector)
@ -220,7 +231,7 @@ pub fn models_matching<'a>(catalog: &'a Catalog, selector: &str) -> Vec<ModelEnt
matches
}
/// Whether `selector` names an enabled model on any enabled provider.
/// Whether `selector` names a model on any enabled provider.
#[must_use]
pub fn is_model_selector(catalog: &Catalog, selector: &str) -> bool {
!models_matching(catalog, selector).is_empty()
@ -232,22 +243,22 @@ pub fn is_provider_selector(catalog: &Catalog, selector: &str) -> bool {
provider(catalog, selector).is_some()
}
/// The enabled default model of an enabled provider.
/// The default model of an enabled provider.
#[must_use]
pub fn default_model<'a>(catalog: &'a Catalog, provider_selector: &str) -> Option<ModelEntry<'a>> {
let entry = provider(catalog, provider_selector)?;
let default = entry.provider.default_model()?;
model_on_provider(catalog, entry.provider.id().as_str(), default)
let provider = provider(catalog, provider_selector)?;
let default = provider.default_model()?;
model_on_provider(catalog, provider.id().as_str(), default)
}
/// The model Fabro probes a provider with: the `probe` model, else the
/// provider default.
#[must_use]
pub fn probe_model<'a>(catalog: &'a Catalog, provider_selector: &str) -> Option<ModelEntry<'a>> {
let entry = provider(catalog, provider_selector)?;
provider_models(entry.provider)
let provider = provider(catalog, provider_selector)?;
provider_models(provider)
.into_iter()
.find(|model| model.policy.probe)
.find(|entry| entry.model.is_probe())
.or_else(|| default_model(catalog, provider_selector))
}
@ -262,9 +273,9 @@ pub fn default_for_ready<'a>(
let providers = enabled_providers(catalog);
providers
.iter()
.filter(|entry| ready.contains(entry.provider.id()))
.filter(|provider| ready.contains(provider.id()))
.chain(providers.iter())
.find_map(|entry| default_model(catalog, entry.provider.id().as_str()))
.find_map(|provider| default_model(catalog, provider.id().as_str()))
}
/// The small utility model across `ready` providers: the first
@ -274,12 +285,11 @@ pub fn small_default_for_ready<'a>(
catalog: &'a Catalog,
ready: &HashSet<ProviderId>,
) -> Option<ModelEntry<'a>> {
let providers = enabled_providers(catalog);
providers
.iter()
.filter(|entry| ready.contains(entry.provider.id()))
.flat_map(|entry| provider_models(entry.provider))
.find(|model| model.policy.small_default)
enabled_providers(catalog)
.into_iter()
.filter(|provider| ready.contains(provider.id()))
.flat_map(provider_models)
.find(|entry| entry.model.is_small_default())
.or_else(|| default_for_ready(catalog, ready))
}
@ -306,14 +316,11 @@ pub fn agent_profile(
provider_selector: &str,
model_selector: Option<&str>,
) -> Option<AgentProfileKind> {
let entry = provider(catalog, provider_selector)?;
let model = model_selector.and_then(|selector| entry.provider.model(selector));
let provider = provider(catalog, provider_selector)?;
let model = model_selector.and_then(|selector| provider.model(selector));
Some(match model {
Some(model) => catalog_policy::effective_agent_profile(entry.provider, model),
None => entry
.policy
.agent_profile
.unwrap_or_else(|| catalog_policy::default_agent_profile(entry.provider)),
Some(model) => ModelEntry { provider, model }.agent_profile(),
None => provider_agent_profile(provider),
})
}
@ -331,7 +338,7 @@ pub fn closest_model<'a>(
.pricing()
.and_then(|pricing| pricing.input_usd_micros_per_million)
.unwrap_or(0);
provider_models(target.provider)
provider_models(target)
.into_iter()
.filter(|entry| {
let caps = entry.model.capabilities();
@ -354,12 +361,12 @@ pub fn closest_model<'a>(
pub fn model_ids_by_provider(catalog: &Catalog) -> BTreeMap<ProviderId, Vec<ModelId>> {
listed_providers(catalog)
.into_iter()
.map(|entry| {
.map(|provider| {
(
entry.provider.id().clone(),
provider_models(entry.provider)
provider.id().clone(),
provider_models(provider)
.into_iter()
.map(|model| model.model.id().clone())
.map(|entry| entry.model.id().clone())
.collect(),
)
})
@ -372,11 +379,11 @@ mod tests {
use crate::test_support::test_catalog;
#[test]
fn policy_layer_builds_over_the_builtins() {
fn builtins_ship_fabro_defaults() {
let catalog = test_catalog();
let ids: Vec<_> = enabled_providers(&catalog)
.iter()
.map(|entry| entry.provider.id().to_string())
.map(|provider| provider.id().to_string())
.collect();
assert_eq!(ids[0], "anthropic");
assert!(ids.contains(&"openai".to_string()));
@ -387,7 +394,7 @@ mod tests {
assert!(
!listed_providers(&catalog)
.iter()
.any(|entry| entry.provider.id().as_str() == "openai-codex"),
.any(|provider| provider.id().as_str() == "openai-codex"),
"stand-in providers are not listed"
);
}
@ -399,7 +406,6 @@ mod tests {
r"
[providers.openai]
priority = 500
[providers.openai.metadata.fabro]
enabled = false
",
)
@ -410,7 +416,7 @@ enabled = false
assert_eq!(
catalog.provider("openai").unwrap().priority(),
500,
"overlay values win over the policy layer"
"overlay values win over the built-ins"
);
}
@ -427,7 +433,7 @@ enabled = false
}
#[test]
fn probe_and_small_default_follow_policy() {
fn probe_and_small_default_follow_the_catalog() {
let catalog = test_catalog();
assert_eq!(
probe_model(&catalog, "openai").unwrap().model.id().as_str(),
@ -488,11 +494,41 @@ enabled = false
);
assert_eq!(
agent_profile(&catalog, "moonshot", None),
Some(AgentProfileKind::Kimi),
"a passthrough model on Moonshot takes the provider's Kimi profile"
);
assert_eq!(
agent_profile(&catalog, "deepseek", None),
Some(AgentProfileKind::OpenAi)
);
assert_eq!(
agent_profile(&catalog, "openrouter", None),
None,
"disabled providers have no profile to offer"
);
assert_eq!(
agent_profile(&catalog, "moonshot", Some("kimi-k3")),
Some(AgentProfileKind::Kimi)
);
assert_eq!(
agent_profile(&catalog, "openai", Some("gpt-6-astra")),
Some(AgentProfileKind::Gpt6)
);
assert_eq!(
agent_profile(&catalog, "anthropic", Some("claude-sonnet-4.5")),
Some(AgentProfileKind::Anthropic)
);
}
#[test]
fn reasoning_by_default_reads_agent_metadata_then_capabilities() {
let catalog = test_catalog();
let kimi = model_on_provider(&catalog, "moonshot", "kimi-k2.5").unwrap();
assert!(kimi.reasons_by_default(), "the catalog row says so");
let sonnet = model_on_provider(&catalog, "anthropic", "claude-sonnet-4.5").unwrap();
assert!(
!sonnet.reasons_by_default(),
"a thinking-budget model reasons only when asked"
);
}
}

7
lib/components/fabro-llm/src/lib.rs
View file

@ -3,8 +3,9 @@
//! lithos owns the LLM vocabulary, the provider catalog, the wire codecs, and
//! the client. This crate adds what is specific to Fabro:
//!
//! - building the catalog from lithos built-ins, Fabro's policy layer, and the
//! operator `[llm]` overlay ([`catalog`]);
//! - building the catalog from the lithos built-ins and the operator `[llm]`
//! overlay, and the catalog queries Fabro's dispatch boundaries share
//! ([`catalog`]);
//! - Fabro's passthrough policy for selections made before a request exists
//! ([`selection`]); at request time the lithos resolver enforces `enabled`
//! and `stands_in_for` itself;
@ -33,7 +34,7 @@ pub mod structured;
#[cfg(any(test, feature = "test-support"))]
pub mod test_support;
pub use catalog::{FABRO_POLICY_TOML, build_catalog, default_catalog};
pub use catalog::{build_catalog, default_catalog};
pub use client::{
ClientOptions, FabroClient, LlmSetupError, RetryListener, RetryNotice, build_client,
build_offline_client,

6
lib/components/fabro-llm/src/probe.rs
View file

@ -104,10 +104,10 @@ pub async fn probe_provider_with_api_key(
api_key: String,
timeout: Duration,
) -> Result<ModelTestOutcome, ApiKeyProbeError> {
let entry = catalog::provider(&catalog, provider.as_str())
let catalog_provider = catalog::provider(&catalog, provider.as_str())
.ok_or_else(|| ApiKeyProbeError::UnknownProvider(provider.to_string()))?;
let provider_id = entry.provider.id().clone();
if !fabro_auth::accepts_api_key(entry.provider) {
let provider_id = catalog_provider.id().clone();
if !fabro_auth::accepts_api_key(catalog_provider) {
return Err(ApiKeyProbeError::NoApiKeyPath(provider_id));
}
let model = catalog::probe_model(&catalog, provider_id.as_str())

18
lib/components/fabro-llm/src/selection.rs
View file

@ -13,8 +13,8 @@
//! - No selector picks the default offering (of the pinned provider, when one
//! is given).
//!
//! Only enabled providers and models take part. Disabled ones are invisible
//! here, exactly as they are to the client's resolver.
//! Only enabled providers take part. Disabled ones are invisible here,
//! exactly as they are to the lithos resolver at request time.
use std::collections::HashSet;
use std::fmt;
@ -115,7 +115,7 @@ pub fn ready_provider(
}
}
/// Finds `selector` as an enabled model on an enabled provider.
/// Finds `selector` as a model on an enabled provider.
pub fn resolve_on_provider<'a>(
catalog: &'a Catalog,
provider: &ProviderId,
@ -182,9 +182,9 @@ pub fn select_default<'a>(
let eligible = canonical_eligible(catalog, eligible);
let providers_with_defaults: Vec<_> = catalog::enabled_providers(catalog)
.into_iter()
.filter_map(|entry| {
catalog::default_model(catalog, entry.provider.id().as_str())
.map(|model| (entry.provider.id().clone(), model))
.filter_map(|provider| {
catalog::default_model(catalog, provider.id().as_str())
.map(|model| (provider.id().clone(), model))
})
.collect();
providers_with_defaults
@ -350,8 +350,7 @@ mod tests {
#[test]
fn slash_selector_with_a_non_provider_prefix_matches_api_ids_on_a_pinned_provider() {
let catalog =
test_catalog_with_overlay("[providers.openrouter.metadata.fabro]\nenabled = true\n");
let catalog = test_catalog_with_overlay("[providers.openrouter]\nenabled = true\n");
let selected = resolve_selection(
&catalog,
Some("openai/gpt-5.6-sol"),
@ -409,8 +408,7 @@ mod tests {
select(&catalog, "gpt-5.4", None, &eligible(&["openrouter"])),
Err(ModelSelectionError::NoEligibleOffering { .. })
));
let enabled =
test_catalog_with_overlay("[providers.openrouter.metadata.fabro]\nenabled = true\n");
let enabled = test_catalog_with_overlay("[providers.openrouter]\nenabled = true\n");
let entry = select(&enabled, "gpt-5.4", None, &eligible(&["openrouter"])).unwrap();
assert_eq!(entry.provider.id(), &ProviderId::new("openrouter"));
}

2
lib/components/fabro-llm/src/test_support.rs
View file

@ -20,7 +20,7 @@ use lithos_llm::types::{
use crate::client::{ClientOptions, build_client, build_offline_client};
/// The lithos built-in catalog with Fabro's policy layer applied.
/// The lithos built-in catalog, as Fabro ships it.
#[must_use]
pub fn test_catalog() -> Catalog {
crate::build_catalog(&LlmLayer::default(), &|_| None).expect("test catalog should build")

5
lib/components/fabro-validate/src/lib.rs
View file

@ -221,9 +221,8 @@ base_url = "https://api.venice.ai/api/v1"
auth = { type = "bearer" }
default_model = "venice-large"
[providers.acme-venice.metadata.fabro]
agent_profile = "openai"
credentials = ["env:VENICE_API_KEY"]
[providers.acme-venice.metadata.agent]
profile = "openai"
[providers.acme-venice.models.venice-large]
display_name = "Venice Large"

2
lib/components/fabro-validate/src/rules/model_support.rs
View file

@ -39,7 +39,7 @@ pub(super) fn check_provider_known(
}
let valid: Vec<String> = catalog::listed_providers(catalog)
.iter()
.map(|entry| entry.provider.id().to_string())
.map(|provider| provider.id().to_string())
.collect();
let valid_str = valid.join(", ");
Some(Diagnostic {

39
lib/components/fabro-workflow/src/handler/llm/api.rs
View file

@ -2065,7 +2065,9 @@ mod tests {
}
/// An OpenAI-compatible mock provider served by `server`, with one model.
fn mock_provider_overlay(provider: &str, model: &str, base_url: &str, env_var: &str) -> String {
/// Its API key is the name lithos derives from the provider id, such as
/// `MOCK_API_KEY`.
fn mock_provider_overlay(provider: &str, model: &str, base_url: &str) -> String {
format!(
r#"
[providers.{provider}]
@ -2076,9 +2078,8 @@ base_url = {base_url}
auth = {{ type = "bearer" }}
default_model = "{model}"
[providers.{provider}.metadata.fabro]
agent_profile = "openai"
credentials = ["env:{env_var}"]
[providers.{provider}.metadata.agent]
profile = "openai"
[providers.{provider}.models.{model}]
display_name = "{model}"
@ -2095,7 +2096,6 @@ capabilities = {{ text = true, tools = true, response_format = {{ json_object =
"mock",
"mock-model",
&server.base_url(),
"MOCK_API_KEY",
)))
}
@ -2103,7 +2103,7 @@ capabilities = {{ text = true, tools = true, response_format = {{ json_object =
/// would so their models become fallback targets.
fn enabled_fallback_catalog() -> Arc<Catalog> {
Arc::new(test_catalog_with_overlay(
"[providers.modal.metadata.fabro]\nenabled = true\n\n[providers.openrouter.metadata.fabro]\nenabled = true\n",
"[providers.modal]\nenabled = true\n\n[providers.openrouter]\nenabled = true\n",
))
}
@ -2132,13 +2132,11 @@ capabilities = {{ text = true, tools = true, response_format = {{ json_object =
"primary",
"test-model",
&format!("{}/primary", server.base_url()),
"PRIMARY_API_KEY",
),
mock_provider_overlay(
"fallback",
"test-model",
&format!("{}/fallback", server.base_url()),
"FALLBACK_API_KEY",
),
);
let catalog = Arc::new(test_catalog_with_overlay(&overlay));
@ -4073,7 +4071,7 @@ capabilities = {{ text = true, tools = true, response_format = {{ json_object =
}))
}
const OPENROUTER_ENABLED: &str = "[providers.openrouter.metadata.fabro]\nenabled = true\n";
const OPENROUTER_ENABLED: &str = "[providers.openrouter]\nenabled = true\n";
/// An operator-defined OpenAI-compatible provider whose models take the
/// provider's `openai` agent profile.
@ -4086,19 +4084,17 @@ base_url = "https://api.acme.test/v1"
auth = { type = "bearer" }
default_model = "acme-llama"
[providers.acme.metadata.fabro]
agent_profile = "openai"
credentials = ["env:ACME_API_KEY"]
[providers.acme.metadata.agent]
profile = "openai"
[providers.acme.models.acme-llama]
display_name = "Acme Llama"
api_model = "acme-llama"
limits = { context_tokens = 131072, max_output_tokens = 8192 }
capabilities = { text = true, tools = true }
[providers.acme.models.acme-llama.metadata.fabro]
family = "llama"
training = "2026-01"
training_cutoff = "2026-01"
"#;
/// The same provider serving a Claude model that overrides the profile.
@ -4111,9 +4107,8 @@ base_url = "https://api.acme.test/v1"
auth = { type = "bearer" }
default_model = "acme-claude"
[providers.acme.metadata.fabro]
agent_profile = "openai"
credentials = ["env:ACME_API_KEY"]
[providers.acme.metadata.agent]
profile = "openai"
[providers.acme.models.acme-claude]
display_name = "Acme Claude"
@ -4121,11 +4116,11 @@ aliases = ["ac"]
api_model = "acme-claude"
limits = { context_tokens = 131072, max_output_tokens = 8192 }
capabilities = { text = true, tools = true }
[providers.acme.models.acme-claude.metadata.fabro]
family = "claude"
training = "2026-01"
agent_profile = "anthropic"
training_cutoff = "2026-01"
[providers.acme.models.acme-claude.metadata.agent]
profile = "anthropic"
"#;
#[tokio::test]

22
lib/components/fabro-workflow/src/handler/prompt.rs
View file

@ -699,9 +699,8 @@ mod tests {
auth = { type = "bearer" }
default_model = "acme-claude"
[providers.acme.metadata.fabro]
agent_profile = "openai"
credentials = ["env:ACME_API_KEY"]
[providers.acme.metadata.agent]
profile = "openai"
[providers.acme.models.acme-claude]
display_name = "Acme Claude"
@ -709,10 +708,10 @@ mod tests {
api_model = "acme-claude"
limits = { context_tokens = 1000, max_output_tokens = 500 }
capabilities = { text = true, tools = true }
[providers.acme.models.acme-claude.metadata.fabro]
family = "claude"
agent_profile = "anthropic"
[providers.acme.models.acme-claude.metadata.agent]
profile = "anthropic"
"#,
));
let mut services = make_services();
@ -774,9 +773,8 @@ mod tests {
auth = { type = "bearer" }
default_model = "acme-claude"
[providers.acme.metadata.fabro]
agent_profile = "openai"
credentials = ["env:ACME_API_KEY"]
[providers.acme.metadata.agent]
profile = "openai"
[providers.acme.models.acme-claude]
display_name = "Acme Claude"
@ -784,10 +782,10 @@ mod tests {
api_model = "acme-claude"
limits = { context_tokens = 1000, max_output_tokens = 500 }
capabilities = { text = true, tools = true }
[providers.acme.models.acme-claude.metadata.fabro]
family = "claude"
agent_profile = "anthropic"
[providers.acme.models.acme-claude.metadata.agent]
profile = "anthropic"
"#,
));
let mut services = make_services();

4
lib/components/fabro-workflow/src/model_fallback.rs
View file

@ -405,7 +405,7 @@ mod tests {
}
fn openrouter_catalog() -> Catalog {
test_catalog_with_overlay("[providers.openrouter.metadata.fabro]\nenabled = true\n")
test_catalog_with_overlay("[providers.openrouter]\nenabled = true\n")
}
#[test]
@ -508,7 +508,7 @@ mod tests {
#[test]
fn resolves_the_requested_production_policy_as_independent_chains() {
let catalog = test_catalog_with_overlay(
"[providers.modal.metadata.fabro]\nenabled = true\n\n[providers.openrouter.metadata.fabro]\nenabled = true\n",
"[providers.modal]\nenabled = true\n\n[providers.openrouter]\nenabled = true\n",
);
let eligible = [
ProviderId::new("modal"),

3
lib/components/fabro-workflow/src/operations/create.rs
View file

@ -746,9 +746,8 @@ mod tests {
[providers.openrouter]
priority = 25
default_model = "gpt-5.6-sol"
[providers.openrouter.metadata.fabro]
enabled = true
"#,
))
}

14
lib/components/fabro-workflow/src/operations/start.rs
View file

@ -1454,9 +1454,8 @@ mod tests {
[providers.openrouter]
priority = 25
default_model = "gpt-5.6-sol"
[providers.openrouter.metadata.fabro]
enabled = true
"#,
)
}
@ -1493,9 +1492,8 @@ mod tests {
auth = { type = "bearer" }
default_model = "acme-claude"
[providers.acme.metadata.fabro]
agent_profile = "openai"
credentials = ["env:ACME_API_KEY"]
[providers.acme.metadata.agent]
profile = "openai"
[providers.acme.models.acme-claude]
display_name = "Acme Claude"
@ -1503,10 +1501,10 @@ mod tests {
api_model = "acme-claude"
limits = { context_tokens = 1000, max_output_tokens = 500 }
capabilities = { text = true, tools = true }
[providers.acme.models.acme-claude.metadata.fabro]
family = "claude"
agent_profile = "anthropic"
[providers.acme.models.acme-claude.metadata.agent]
profile = "anthropic"
"#,
);
let mut settings = ResolvedRunSettings::default();

5
lib/components/fabro-workflow/src/pipeline/pull_request.rs
View file

@ -782,9 +782,8 @@ base_url = "http://mock.invalid/v1"
auth = { type = "bearer" }
allow_passthrough = true
[providers.mock.metadata.fabro]
agent_profile = "openai"
credentials = ["env:MOCK_API_KEY"]
[providers.mock.metadata.agent]
profile = "openai"
[providers.mock.models.mock-model]
display_name = "Mock Model"

7
lib/components/fabro-workflow/src/transforms/model_resolution.rs
View file

@ -174,9 +174,8 @@ auth = { type = "bearer" }
priority = 200
default_model = "venice-large"
[providers.acme-venice.metadata.fabro]
agent_profile = "openai"
credentials = ["env:VENICE_API_KEY"]
[providers.acme-venice.metadata.agent]
profile = "openai"
[providers.acme-venice.models.venice-large]
display_name = "Venice Large"
@ -355,7 +354,7 @@ capabilities = { text = true, tools = true }
#[test]
fn fallback_resolution_keeps_ready_preference_for_unpinned_nodes() {
let catalog = Arc::new(test_catalog_with_overlay(
"[providers.openrouter.metadata.fabro]\nenabled = true\n",
"[providers.openrouter]\nenabled = true\n",
));
let mut graph = Graph::new("test");
let mut portable = Node::new("portable");

7
lib/components/fabro-workflow/tests/it/integration.rs
View file

@ -2672,9 +2672,8 @@ base_url = {base_url}
auth = {{ type = "bearer" }}
default_model = "compact-model"
[providers.compact.metadata.fabro]
agent_profile = "openai"
credentials = ["env:COMPACT_API_KEY"]
[providers.compact.metadata.agent]
profile = "openai"
[providers.compact.models.compact-model]
display_name = "Compact Model"
@ -2834,8 +2833,6 @@ async fn workflow_persists_authoritative_openrouter_cost_for_agent_stage() {
&format!(
"[providers.openrouter]
base_url = {}
[providers.openrouter.metadata.fabro]
enabled = true
",
toml::Value::String(server.base_url()),

5
lib/foundation/fabro-config/src/builders.rs
View file

@ -693,10 +693,7 @@ adapter = "openai-compatible"
codec = "openai-chat"
base_url = "https://api.acme.test/v1"
auth = { type = "bearer" }
[llm.providers.acme.metadata.fabro]
enabled = true
credentials = ["env:ACME_API_KEY"]
[llm.providers.acme.models."acme-large"]
display_name = "Acme Large"
@ -710,7 +707,7 @@ api_model = "acme-large"
let overlay = settings.llm_overlay.0;
let acme = &overlay["providers"]["acme"];
assert_eq!(acme["display_name"].as_str(), Some("Acme"));
assert_eq!(acme["metadata"]["fabro"]["enabled"].as_bool(), Some(true));
assert_eq!(acme["enabled"].as_bool(), Some(true));
assert_eq!(
acme["models"]["acme-large"]["api_model"].as_str(),
Some("acme-large")

20
lib/foundation/fabro-config/src/layers/llm.rs
View file

@ -2,17 +2,14 @@
//!
//! Operator model catalog overrides. The table uses the lithos-llm catalog
//! schema verbatim, minus `schema_version`, and is applied as an overlay layer
//! on top of the lithos built-in catalog and Fabro's policy layer:
//! on top of the lithos built-in catalog:
//!
//! ```toml
//! [llm.providers.moonshot]
//! [llm.providers.openrouter]
//! priority = 60
//!
//! [llm.providers.moonshot.metadata.fabro]
//! enabled = true
//! credentials = ["env:MOONSHOT_API_KEY", "vault:MOONSHOT_API_KEY"]
//!
//! [llm.providers.moonshot.models."kimi-k2.5".metadata.fabro]
//! [llm.providers.openrouter.models."kimi-k2.5"]
//! small_default = true
//! ```
//!
@ -83,7 +80,6 @@ mod tests {
r"
[providers.acme]
priority = 10
[providers.acme.metadata.fabro]
enabled = false
",
);
@ -92,19 +88,19 @@ enabled = false
[providers.acme]
priority = 5
base_url = "https://acme.test"
[providers.acme.metadata.fabro]
enabled = true
small_default = true
[providers.acme.metadata.agent]
profile = "openai"
"#,
);
let merged = higher.combine(lower).0;
let acme = &merged["providers"]["acme"];
assert_eq!(acme["priority"].as_integer(), Some(10));
assert_eq!(acme["base_url"].as_str(), Some("https://acme.test"));
assert_eq!(acme["metadata"]["fabro"]["enabled"].as_bool(), Some(false));
assert_eq!(acme["enabled"].as_bool(), Some(false));
assert_eq!(
acme["metadata"]["fabro"]["small_default"].as_bool(),
Some(true)
acme["metadata"]["agent"]["profile"].as_str(),
Some("openai")
);
}

11
lib/foundation/fabro-config/src/parse.rs
View file

@ -295,22 +295,21 @@ mod tests {
#[test]
fn llm_overlay_merges_across_layers() {
let higher = r#"
[llm.providers.openai.models."gpt-5.4".metadata.fabro]
[llm.providers.openai.models."gpt-5.4"]
small_default = true
"#
.parse::<SettingsLayer>()
.unwrap();
let lower = r#"
[llm.providers.openai.models."gpt-5.4".metadata.fabro]
[llm.providers.openai.models."gpt-5.4"]
probe = true
"#
.parse::<SettingsLayer>()
.unwrap();
let merged = crate::Combine::combine(higher, lower);
let fabro =
&merged.llm.unwrap().0["providers"]["openai"]["models"]["gpt-5.4"]["metadata"]["fabro"];
assert_eq!(fabro["small_default"].as_bool(), Some(true));
assert_eq!(fabro["probe"].as_bool(), Some(true));
let model = &merged.llm.unwrap().0["providers"]["openai"]["models"]["gpt-5.4"];
assert_eq!(model["small_default"].as_bool(), Some(true));
assert_eq!(model["probe"].as_bool(), Some(true));
}
#[test]

229
lib/foundation/fabro-types/src/catalog_policy.rs
View file

@ -1,229 +0,0 @@
//! Fabro's `metadata.fabro` catalog namespace.
//!
//! lithos-llm owns provider and model facts. Fabro attaches its own policy to
//! each entry under `metadata.fabro`, which lithos carries verbatim and never
//! interprets. These types are the typed view of that namespace. Every field
//! is optional in the TOML; the accessors here apply Fabro's defaults.
use lithos_llm::catalog::{CatalogModel, CatalogProvider};
use serde::{Deserialize, Serialize};
use crate::AgentProfileKind;
/// Name of the metadata namespace Fabro owns on catalog entries.
pub const FABRO_METADATA_NAMESPACE: &str = "fabro";
/// Provider-level Fabro policy.
#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
#[serde(default)]
pub struct ProviderPolicy {
/// Whether Fabro offers this provider at all. Missing means enabled.
pub enabled: Option<bool>,
/// Default agent profile for models on this provider.
pub agent_profile: Option<AgentProfileKind>,
/// Where an operator obtains an API key.
pub api_key_url: Option<String>,
/// Ordered credential references (`env:NAME`, `vault:NAME`, `aws_sigv4`).
/// The first that resolves wins.
pub credentials: Vec<String>,
/// Extra request headers. Values are literal text or `{{ secrets.NAME }}`
/// interpolation strings resolved against the vault.
#[serde(skip_serializing_if = "std::collections::BTreeMap::is_empty")]
pub extra_headers: std::collections::BTreeMap<String, String>,
/// Another provider this one serves requests for when that provider has
/// no credentials of its own. Used by `openai-codex`, which answers
/// `openai` requests with a ChatGPT OAuth credential.
pub stands_in_for: Option<String>,
}
impl ProviderPolicy {
#[must_use]
pub fn is_enabled(&self) -> bool {
self.enabled.unwrap_or(true)
}
}
/// Model-level Fabro policy.
#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
#[serde(default)]
pub struct ModelPolicy {
/// Whether Fabro offers this model. Missing means enabled.
pub enabled: Option<bool>,
/// Agent profile override for this model.
pub agent_profile: Option<AgentProfileKind>,
/// Model family label for display and grouping.
pub family: Option<String>,
/// Training data cutoff label.
pub training: Option<String>,
/// Public knowledge cutoff label.
pub knowledge_cutoff: Option<String>,
/// Estimated output tokens per second.
pub estimated_output_tps: Option<f64>,
/// Preferred for small utility calls such as title generation.
pub small_default: bool,
/// Preferred for provider connectivity probes.
pub probe: bool,
/// Whether requests reason when no effort is requested. Missing means
/// "reasons when the model supports reasoning".
pub reasoning_by_default: Option<bool>,
}
impl ModelPolicy {
#[must_use]
pub fn is_enabled(&self) -> bool {
self.enabled.unwrap_or(true)
}
}
/// Reads a provider's Fabro policy. Malformed metadata falls back to the
/// defaults; the catalog build is the place to validate shape, and Fabro's
/// own policy file is checked in tests.
#[must_use]
pub fn provider_policy(provider: &CatalogProvider) -> ProviderPolicy {
provider
.metadata()
.namespace::<ProviderPolicy>(FABRO_METADATA_NAMESPACE)
.ok()
.flatten()
.unwrap_or_default()
}
/// Reads a model's Fabro policy.
#[must_use]
pub fn model_policy(model: &CatalogModel) -> ModelPolicy {
model
.metadata()
.namespace::<ModelPolicy>(FABRO_METADATA_NAMESPACE)
.ok()
.flatten()
.unwrap_or_default()
}
/// The agent profile a model runs under: the model override, then the
/// provider default, then the profile implied by the provider's adapter.
#[must_use]
pub fn effective_agent_profile(
provider: &CatalogProvider,
model: &CatalogModel,
) -> AgentProfileKind {
model_policy(model)
.agent_profile
.or(provider_policy(provider).agent_profile)
.unwrap_or_else(|| default_agent_profile(provider))
}
/// The agent profile implied by a provider's wire protocol.
#[must_use]
pub fn default_agent_profile(provider: &CatalogProvider) -> AgentProfileKind {
match provider.adapter().as_str() {
"anthropic" | "bedrock" => AgentProfileKind::Anthropic,
"gemini" => AgentProfileKind::Gemini,
_ => AgentProfileKind::OpenAi,
}
}
#[cfg(test)]
mod tests {
use lithos_llm::catalog::Catalog;
use super::*;
fn catalog() -> Catalog {
Catalog::builder()
.toml_layer(
"test",
r#"
schema_version = 1
[providers.acme]
display_name = "Acme"
adapter = "openai-compatible"
codec = "openai-chat"
base_url = "https://acme.test/v1"
auth = { type = "bearer" }
default_model = "large"
[providers.acme.metadata.fabro]
enabled = false
credentials = ["env:ACME_API_KEY"]
agent_profile = "kimi"
[providers.acme.models.large]
display_name = "Large"
api_model = "large"
[providers.acme.models.large.metadata.fabro]
small_default = true
probe = true
family = "acme"
[providers.acme.models.small]
display_name = "Small"
api_model = "small"
[providers.acme.models.small.metadata.fabro]
agent_profile = "openai"
enabled = false
"#,
)
.unwrap()
.build()
.unwrap()
}
#[test]
fn reads_provider_and_model_policy() {
let catalog = catalog();
let provider = catalog.provider("acme").unwrap();
let policy = provider_policy(provider);
assert!(!policy.is_enabled());
assert_eq!(policy.credentials, vec!["env:ACME_API_KEY"]);
assert_eq!(policy.agent_profile, Some(AgentProfileKind::Kimi));
let large = provider.model("large").unwrap();
let policy = model_policy(large);
assert!(policy.small_default && policy.probe && policy.is_enabled());
assert_eq!(policy.family.as_deref(), Some("acme"));
assert_eq!(
effective_agent_profile(provider, large),
AgentProfileKind::Kimi
);
let small = provider.model("small").unwrap();
assert!(!model_policy(small).is_enabled());
assert_eq!(
effective_agent_profile(provider, small),
AgentProfileKind::OpenAi
);
}
#[test]
fn missing_namespace_yields_defaults() {
let catalog = Catalog::builder()
.toml_layer(
"test",
r#"
schema_version = 1
[providers.bare]
display_name = "Bare"
adapter = "anthropic"
codec = "anthropic-messages"
base_url = "https://bare.test"
auth = { type = "none" }
[providers.bare.models.m]
display_name = "M"
api_model = "m"
"#,
)
.unwrap()
.build()
.unwrap();
let provider = catalog.provider("bare").unwrap();
assert!(provider_policy(provider).is_enabled());
let model = provider.model("m").unwrap();
assert!(model_policy(model).is_enabled());
assert_eq!(
effective_agent_profile(provider, model),
AgentProfileKind::Anthropic
);
}
}

2
lib/foundation/fabro-types/src/lib.rs
View file

@ -8,7 +8,6 @@ pub mod billing_rollup;
pub mod blob_hash;
pub mod blob_ref;
pub mod catalog_api;
pub mod catalog_policy;
pub mod checkpoint;
pub mod command_output;
pub mod conclusion;
@ -75,7 +74,6 @@ pub use billing::{
pub use blob_hash::BlobHash;
pub use blob_ref::{format_blob_ref, parse_blob_ref, parse_managed_blob_file_ref};
pub use catalog_api::{Model, ModelControls, ModelCosts, ModelFeatures, ModelLimits, Provider};
pub use catalog_policy::{ModelPolicy, ProviderPolicy};
pub use checkpoint::Checkpoint;
pub use command_output::{CommandOutputStream, CommandTermination};
pub use conclusion::{Conclusion, StageSummary};