Merge pull request #919 from fabro-sh/codex/cli-model-overrides

Apply CLI model and provider overrides to agent execution
This commit is contained in:
Scott Werner 2026-10-02 13:43:14 -04:00 • committed by GitHub
commit 824b5a2238
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
10 changed files with 349 additions and 34 deletions

View file

@ -44,6 +44,7 @@ mod repo_init;
mod resume;
mod rm;
mod run;
mod run_model;
mod runner;
mod sandbox_cp;
mod sandbox_preview;

View file

@ -0,0 +1,192 @@
use std::fmt::Write as _;
use fabro_test::{TestContext, test_context};
use httpmock::MockServer;
use serde_json::json;
/// Exercise admission and the real agent against two local providers. Checking
/// the request path and wire model catches overrides that only reach run state.
fn assert_agent_route(
mut context: TestContext,
flags: &[&str],
graph_defaults: &str,
node_settings: &str,
expected_provider: &str,
expected_model: &str,
) {
let endpoint = MockServer::start();
let mut settings = "_version = 1\n[server.auth]\nmethods = [\"dev-token\"]\n".to_string();
for provider in ["primary", "alternate"] {
write!(
settings,
r#"
[llm.providers.{provider}]
display_name = "{provider}"
base_url = "{base}/{provider}/v1"
auth = {{ type = "none" }}
default_model = "base"
[llm.providers.{provider}.metadata.agent]
profile = "openai"
"#,
base = endpoint.base_url(),
)
.expect("provider settings should format");
for model in ["base", "override", "node"] {
write!(
settings,
r#"
[llm.providers.{provider}.models.{model}]
display_name = "{model}"
api_model = "{provider}-{model}-wire"
limits = {{ context_tokens = 32000, max_output_tokens = 1000 }}
capabilities = {{ text = true, tools = true }}
"#,
)
.expect("model settings should format");
}
}
context.write_home(".fabro/settings.toml", settings);
context.isolated_server();
context.write_temp(
"workflow.toml",
r#"_version = 1
[workflow]
graph = "workflow.fabro"
[run.model]
provider = "primary"
name = "base"
[run.pull_request]
enabled = false
"#,
);
context.write_temp(
"workflow.fabro",
format!(
r#"digraph ModelSelection {{
{graph_defaults}
start [shape=Mdiamond];
work [shape=box, prompt="Say hello."];
{node_settings}
exit [shape=Msquare];
start -> work -> exit;
}}"#,
),
);
let wire_model = format!("{expected_provider}-{expected_model}-wire");
let chunk = |delta, finish_reason| {
json!({
"id": "scripted-response", "object": "chat.completion.chunk",
"created": 1, "model": wire_model,
"choices": [{"index": 0, "delta": delta, "finish_reason": finish_reason}]
})
};
let response = format!(
"data: {}\n\ndata: {}\n\ndata: [DONE]\n\n",
chunk(
json!({"role": "assistant", "content": "Hello."}),
json!(null)
),
chunk(json!({}), json!("stop")),
);
let expected = endpoint.mock(|when, then| {
when.method("POST")
.path(format!("/{expected_provider}/v1/chat/completions"))
.json_body_includes(json!({"model": wire_model, "stream": true}).to_string());
then.status(200)
.header("Content-Type", "text/event-stream")
.body(&response);
});
// Respond immediately to a wrong route too, so the regression fails with
// the mismatched request instead of waiting through model retry backoff.
let unexpected = endpoint.mock(|when, then| {
when.method("POST");
then.status(200)
.header("Content-Type", "text/event-stream")
.body(&response);
});
let output = context
.run_cmd()
.args(["--auto-approve", "--environment", "local"])
.args(flags)
.arg(context.temp_dir.join("workflow.toml"))
.output()
.expect("run should execute");
assert!(
output.status.success(),
"run failed:\n{}\n{}",
String::from_utf8_lossy(&output.stdout),
String::from_utf8_lossy(&output.stderr),
);
unexpected.assert_calls(0);
assert!(
expected.calls() > 0,
"the agent must call the requested route"
);
}
#[test]
fn workflow_model_controls_agent_requests() {
assert_agent_route(test_context!(), &[], "", "", "primary", "base");
}
#[test]
fn model_flag_overrides_workflow_model_at_execution() {
assert_agent_route(
test_context!(),
&["--model", "override"],
"",
"",
"primary",
"override",
);
}
#[test]
fn provider_flag_overrides_workflow_provider_at_execution() {
assert_agent_route(
test_context!(),
&["--provider", "alternate"],
"",
"",
"alternate",
"base",
);
}
#[test]
fn model_flags_override_workflow_and_graph_defaults_at_execution() {
assert_agent_route(
test_context!(),
&["--model", "override", "--provider", "alternate"],
r#"graph [default_model="base", default_provider="primary"];"#,
"",
"alternate",
"override",
);
}
#[test]
fn explicit_node_model_takes_precedence_over_flags() {
assert_agent_route(
test_context!(),
&["--model", "override", "--provider", "alternate"],
"",
r#"work [model="node", provider="primary"];"#,
"primary",
"node",
);
}
#[test]
fn node_stylesheet_model_takes_precedence_over_flags() {
assert_agent_route(
test_context!(),
&["--model", "override", "--provider", "alternate"],
"",
r##"graph [model_stylesheet="#work { model: node; provider: primary; }"];"##,
"primary",
"node",
);
}

View file

@ -104,6 +104,7 @@ pub fn validate_collected_workflow(
&settings,
&HashMap::new(),
petri_check::launch_without_catalog(&settings),
run_overrides.and_then(|run| run.model.as_ref()),
offline_runtime(run_overrides),
false,
)

View file

@ -12,6 +12,7 @@
use std::collections::{BTreeMap, HashMap, HashSet};
use std::path::PathBuf;
use fabro_config::RunModelLayer;
use fabro_llm::lithos_catalog::Catalog;
use fabro_llm::selection;
use fabro_petri::check::{self, Admitted, Bundle, CheckError, CheckRequest, Diagnostic, Launch};
@ -27,11 +28,12 @@ use lithos_llm::catalog::ProviderId;
pub(crate) const NO_READY_PROVIDER_RULE: &str = "fabro.model.no_ready_provider";
/// The launch Fabro binds around the settings: the run's model and provider
/// below them, and the environment the run selected and the goal the run
/// resolved above them. When the settings name neither model nor provider, the
/// default offering of the eligible providers is bound as the launch model
/// alone: a node that names no model runs on it, and a node that names a model
/// the catalog lacks stays unqualified, so Petri's admission refuses it.
/// as the default below them, and the environment the run selected and the
/// goal the run resolved above them. When the settings name neither model nor
/// provider, the default offering of the eligible providers is bound as the
/// default model alone: a node that names no model runs on it, and a node that
/// names a model the catalog lacks stays unqualified, so Petri's admission
/// refuses it.
pub(crate) fn launch(
catalog: &Catalog,
settings: &WorkflowSettings,
@ -49,11 +51,12 @@ pub(crate) fn launch(
.map(|offering| offering.model.id().to_string())
});
Launch {
model,
provider: settings.run.model.provider.clone(),
default_model: model,
default_provider: settings.run.model.provider.clone(),
environment: environment.map(str::to_owned),
goal: launch_goal(settings),
repository,
..Launch::default()
}
}
@ -81,28 +84,39 @@ fn launch_goal(settings: &WorkflowSettings) -> Option<String> {
/// name, for a check away from the server.
pub(crate) fn launch_without_catalog(settings: &WorkflowSettings) -> Launch {
Launch {
model: settings.run.model.name.clone(),
provider: settings.run.model.provider.clone(),
default_model: settings.run.model.name.clone(),
default_provider: settings.run.model.provider.clone(),
environment: None,
goal: launch_goal(settings),
repository: None,
goal: launch_goal(settings),
repository: None,
..Launch::default()
}
}
/// The check request for `bundle`'s `entrypoint`: every file of every
/// workflow in the bundle at its bundle-relative path, the run's inputs and
/// variables, the launch and the runtime. `unbound_is_warning` makes a
/// template that reads an input nothing binds a warning, for a validation
/// before the run's inputs exist; a run's admission never sets it.
/// variables, the launch and the runtime. `model_overrides` are the run's
/// explicit model flags (`--model`, `--provider`), bound as the launch's
/// model apart from the default the settings resolved, so they outrank the
/// file layers and the graph's defaults during admission. Every check binds
/// them here, so the create, validate and preflight paths judge the same
/// model. `unbound_is_warning` makes a template that reads an input nothing
/// binds a warning, for a validation before the run's inputs exist; a run's
/// admission never sets it.
pub(crate) fn check_request(
bundle: &WorkflowBundle,
entrypoint: &ManifestPath,
settings: &WorkflowSettings,
vars: &HashMap<String, String>,
launch: Launch,
mut launch: Launch,
model_overrides: Option<&RunModelLayer>,
runtime: RuntimeSpec,
unbound_is_warning: bool,
) -> Result<CheckRequest, WorkflowError> {
if let Some(overrides) = model_overrides {
launch.model.clone_from(&overrides.name);
launch.provider.clone_from(&overrides.provider);
}
let mut files = BTreeMap::new();
for workflow in bundle.workflows().values() {
for (path, text) in &workflow.files {
@ -219,3 +233,71 @@ fn fabro_diagnostic(diagnostic: &Diagnostic) -> FabroDiagnostic {
..FabroDiagnostic::default()
}
}
#[cfg(test)]
mod tests {
use super::*;
fn request(launch: Launch, overrides: Option<&RunModelLayer>) -> CheckRequest {
check_request(
&WorkflowBundle::default(),
&ManifestPath::from_wire("workflow.fabro").expect("the path is valid"),
&WorkflowSettings::default(),
&HashMap::new(),
launch,
overrides,
RuntimeSpec::default(),
false,
)
.expect("the request builds")
}
fn settings_default() -> Launch {
Launch {
default_model: Some("settings-model".to_string()),
default_provider: Some("settings-provider".to_string()),
..Launch::default()
}
}
#[test]
fn model_flags_bind_as_the_launch_model_beside_the_default() {
let overrides = RunModelLayer {
name: Some("flag-model".to_string()),
provider: Some("flag-provider".to_string()),
..RunModelLayer::default()
};
let launch = request(settings_default(), Some(&overrides)).launch;
assert_eq!(launch.model.as_deref(), Some("flag-model"));
assert_eq!(launch.provider.as_deref(), Some("flag-provider"));
assert_eq!(launch.default_model.as_deref(), Some("settings-model"));
assert_eq!(
launch.default_provider.as_deref(),
Some("settings-provider")
);
}
#[test]
fn a_provider_flag_alone_leaves_the_launch_model_unset() {
let overrides = RunModelLayer {
provider: Some("flag-provider".to_string()),
..RunModelLayer::default()
};
let launch = request(settings_default(), Some(&overrides)).launch;
assert_eq!(launch.model, None);
assert_eq!(launch.provider.as_deref(), Some("flag-provider"));
}
#[test]
fn no_model_flags_bind_only_the_default() {
let launch = request(settings_default(), None).launch;
assert_eq!(launch.model, None);
assert_eq!(launch.provider, None);
assert_eq!(launch.default_model.as_deref(), Some("settings-model"));
}
}

View file

@ -30,7 +30,7 @@ use std::path::PathBuf;
use fabro_config::parse::{self, ParseError, SettingsSource};
use fabro_config::{
EnvironmentDockerfileLayer, EnvironmentImageLayer, EnvironmentLayer, MergeMap, RunLayer,
SettingsLayer, WorkflowSettingsBuilder,
RunModelLayer, SettingsLayer, WorkflowSettingsBuilder,
};
use fabro_types::settings::interp::{InterpString, ResolveError};
use fabro_types::settings::run::{McpServerSettings, RunGoal};
@ -100,6 +100,7 @@ struct RunMetadata {
/// The environment the run overrides selected, for Petri's settings
/// layer.
environment_id: Option<String>,
model_overrides: Option<RunModelLayer>,
storage_root: PathBuf,
workflow_slug: Option<String>,
workflow_version_id: Option<WorkflowVersionId>,
@ -174,6 +175,10 @@ impl PreparedRun {
self.layered.metadata.environment_id.as_deref()
}
pub(crate) fn model_overrides(&self) -> Option<&RunModelLayer> {
self.layered.metadata.model_overrides.as_ref()
}
pub(crate) fn resolve_run_id(mut self) -> (Self, RunId) {
let run_id = self.layered.metadata.run_id.unwrap_or_default();
self.layered.metadata.run_id = Some(run_id);
@ -316,6 +321,7 @@ pub(crate) fn normalize_source(input: RawRunCompilerInput) -> Result<NormalizedR
.as_ref()
.and_then(|run| run.environment.as_ref())
.and_then(|environment| environment.id.clone());
let model_overrides = run_overrides.as_ref().and_then(|run| run.model.clone());
Ok(NormalizedRun {
workflow_bundle,
@ -332,6 +338,7 @@ pub(crate) fn normalize_source(input: RawRunCompilerInput) -> Result<NormalizedR
metadata: RunMetadata {
run_id,
environment_id,
model_overrides,
storage_root,
workflow_slug,
workflow_version_id,
@ -460,6 +467,7 @@ pub(crate) fn assemble_run(pinned: PinnedRun) -> CreateRunPersistenceInput {
// Consumed at admission, as the launch's environment; the resolved
// settings carry the environment the run persists.
environment_id: _,
model_overrides: _,
storage_root,
workflow_slug,
workflow_version_id,

View file

@ -9,7 +9,7 @@ use fabro_api::types;
use fabro_config::parse::SettingsSource;
use fabro_config::run::resolve_run_goal_from_namespace;
use fabro_config::{
CliLayer, CliOutputLayer, EnvironmentLayer, MergeMap, RunLayer, SettingsLayer,
CliLayer, CliOutputLayer, EnvironmentLayer, MergeMap, RunLayer, RunModelLayer, SettingsLayer,
WorkflowSettingsBuilder, parse_input_overrides, parse_labels, project,
};
use fabro_dot::WorkflowGraph;
@ -53,6 +53,7 @@ pub(crate) struct PreparedManifest {
/// The entrypoint's DOT as written: what the render endpoint draws.
pub root_source: String,
pub settings: WorkflowSettings,
pub model_overrides: Option<RunModelLayer>,
pub target_path: ManifestPath,
pub workflow_bundle: WorkflowBundle,
pub source_directory: PathBuf,
@ -92,6 +93,10 @@ pub(crate) fn prepare_manifest_with_environment_defaults(
let args_overrides =
manifest_args_overrides(manifest.args.as_ref()).context("failed to parse manifest args")?;
let model_overrides = args_overrides
.run
.as_ref()
.and_then(|run| run.model.clone());
let mut workflow_settings_builder = WorkflowSettingsBuilder::new()
.server_manifest_defaults(
manifest_run_defaults.clone(),
@ -166,6 +171,7 @@ pub(crate) fn prepare_manifest_with_environment_defaults(
git: manifest.git.clone(),
root_source,
settings,
model_overrides,
target_path,
workflow_bundle,
source_directory,
@ -206,6 +212,7 @@ pub(crate) fn validate_prepared_manifest(
&prepared.settings,
vars,
launch,
prepared.model_overrides.as_ref(),
runtime,
unbound_is_warning,
)?;

View file

@ -281,6 +281,7 @@ pub(crate) async fn admit(
settings,
prepared.vars(),
launch,
prepared.model_overrides(),
runtime_spec(
state,
eligible,

View file

@ -11,8 +11,11 @@
//! path the map holds the file under.
//!
//! The launch binds the compile variables the Fabro frontend reads:
//! `petri.launch_model` and `petri.launch_provider` as the model default
//! below every file layer, `petri.launch_environment` as the environment
//! `petri.launch_model` and `petri.launch_provider` as the model the run's
//! flags ask for, over every file layer and the graph's defaults but below a
//! model a node names itself, `petri.default_model` and
//! `petri.default_provider` as the run's default model below every file
//! layer, `petri.launch_environment` as the environment
//! the run selected over every file layer, `petri.launch_goal` as the goal
//! the run resolved over every file layer and the graph's own, and
//! `petri.repository` as the repository the root `start` stage checks out. A
@ -26,8 +29,8 @@ use std::path::PathBuf;
use petri_frontend_attractor::kinds::{AGENT_KIND, PROMPT_KIND};
use petri_runtime::LoadError;
use petri_runtime::frontend::{
self, CompileInputs, LAUNCH_ENVIRONMENT_VAR, LAUNCH_GOAL_VAR, LAUNCH_MODEL_VAR,
LAUNCH_PROVIDER_VAR, MapFiles, REPOSITORY_VAR, Severity,
self, CompileInputs, DEFAULT_MODEL_VAR, DEFAULT_PROVIDER_VAR, LAUNCH_ENVIRONMENT_VAR,
LAUNCH_GOAL_VAR, LAUNCH_MODEL_VAR, LAUNCH_PROVIDER_VAR, MapFiles, REPOSITORY_VAR, Severity,
};
use petri_runtime::ir::Graph;
use serde::{Deserialize, Serialize};
@ -65,25 +68,34 @@ impl Bundle {
}
}
/// What the launch binds around the file layers: the model default below
/// them, the environment selection above them, and the repository.
/// What the launch binds around the file layers: the model the run's flags
/// ask for and the environment selection above them, the run's default model
/// below them, and the repository.
#[derive(Clone, Debug, Default)]
pub struct Launch {
pub model: Option<String>,
pub provider: Option<String>,
/// The model and provider the run's flags ask for (`--model`,
/// `--provider`), over every file layer and the graph's defaults but
/// below a model a node names itself.
pub model: Option<String>,
pub provider: Option<String>,
/// The run's default model and provider (the settings', else the
/// catalog's default), below every file layer: they fill what nothing
/// else names.
pub default_model: Option<String>,
pub default_provider: Option<String>,
/// The environment the run selected, by its id in the server's
/// catalog, over every layer's `[run.environment]`, as the intent's
/// selection overrides the bundle in Fabro's own resolution; `None`
/// leaves the layers to select.
pub environment: Option<String>,
pub environment: Option<String>,
/// The goal the run resolved (the intent's override, else the settings'
/// `[run] goal` from any layer), over the bundle's `[run] goal` and the
/// graph's own `goal`, so the stages execute with the goal the run
/// shows; `None` leaves the bundle's layers and the graph to state it.
pub goal: Option<String>,
pub goal: Option<String>,
/// The local repository the root `start` stage checks out into the
/// workspace; `None` starts the run from an empty workspace.
pub repository: Option<PathBuf>,
pub repository: Option<PathBuf>,
}
/// One check: the bundle, the run's inputs and variables, the launch and
@ -219,6 +231,12 @@ fn compile_inputs(
compile
.vars
.insert(LAUNCH_PROVIDER_VAR.into(), text(&launch.provider));
compile
.vars
.insert(DEFAULT_MODEL_VAR.into(), text(&launch.default_model));
compile
.vars
.insert(DEFAULT_PROVIDER_VAR.into(), text(&launch.default_provider));
if let Some(environment) = &launch.environment {
compile.vars.insert(
LAUNCH_ENVIRONMENT_VAR.into(),

View file

@ -140,11 +140,12 @@ async fn a_launch_binds_the_repository_and_the_model_default() {
inputs: BTreeMap::new(),
vars: BTreeMap::new(),
launch: Launch {
model: Some("gpt-5.4".to_string()),
provider: None,
default_model: Some("gpt-5.4".to_string()),
default_provider: None,
environment: None,
goal: None,
repository: Some(repository.path().to_path_buf()),
goal: None,
repository: Some(repository.path().to_path_buf()),
..Launch::default()
},
runtime: RuntimeSpec::default(),
unbound_is_warning: false,
@ -153,7 +154,11 @@ async fn a_launch_binds_the_repository_and_the_model_default() {
let admitted = check::check(&request).expect("the command bundle is admitted");
let launch = &admitted.graph.params["fabro.launch"];
assert_eq!(launch["model"], "gpt-5.4");
assert_eq!(launch["default_model"], "gpt-5.4");
assert!(
launch["model"].is_null(),
"the default is not a launch flag: {launch}"
);
assert_eq!(
launch["clone"]["repository"],
repository.path().to_string_lossy().as_ref()

View file

@ -67,7 +67,7 @@ async fn a_model_call_authenticates_through_the_vault_and_skills_read_the_home()
let graphs = support::admit(
&[("workflow.fabro", &workflow), ("workflow.toml", &settings)],
Launch {
model: Some(OPENAI_MODEL.to_string()),
default_model: Some(OPENAI_MODEL.to_string()),
..Launch::default()
},
&runtime,