flowproof 0.6.1__tar.gz → 0.7.0__tar.gz

{flowproof-0.6.1 → flowproof-0.7.0}/Cargo.lock

@@ -760,7 +760,7 @@ dependencies = [
 
 [[package]]
 name = "flowproof-adapters"
-version = "0.6.1"
+version = "0.7.0"
 dependencies = [
  "anyhow",
  "flowproof-driver",
@@ -779,7 +779,7 @@ dependencies = [
 
 [[package]]
 name = "flowproof-agent"
-version = "0.6.1"
+version = "0.7.0"
 dependencies = [
  "chrono",
  "flowproof-driver",
@@ -795,7 +795,7 @@ dependencies = [
 
 [[package]]
 name = "flowproof-cli"
-version = "0.6.1"
+version = "0.7.0"
 dependencies = [
  "ab_glyph",
  "chrono",
@@ -816,7 +816,7 @@ dependencies = [
 
 [[package]]
 name = "flowproof-driver"
-version = "0.6.1"
+version = "0.7.0"
 dependencies = [
  "image",
  "postgres",
@@ -830,7 +830,7 @@ dependencies = [
 
 [[package]]
 name = "flowproof-python"
-version = "0.6.1"
+version = "0.7.0"
 dependencies = [
  "flowproof-agent",
  "flowproof-cli",
@@ -842,7 +842,7 @@ dependencies = [
 
 [[package]]
 name = "flowproof-replay"
-version = "0.6.1"
+version = "0.7.0"
 dependencies = [
  "chrono",
  "flowproof-agent",
@@ -856,7 +856,7 @@ dependencies = [
 
 [[package]]
 name = "flowproof-trace"
-version = "0.6.1"
+version = "0.7.0"
 dependencies = [
  "jsonschema",
  "regex",

{flowproof-0.6.1 → flowproof-0.7.0}/Cargo.toml

@@ -3,7 +3,7 @@ resolver = "2"
 members = ["crates/flowproof-driver", "crates/flowproof-trace", "crates/flowproof-replay", "crates/flowproof-agent", "crates/flowproof-adapters", "crates/flowproof-cli", "crates/flowproof-python"]
 
 [workspace.package]
-version = "0.6.1"
+version = "0.7.0"
 edition = "2021"
 license = "Apache-2.0"
 repository = "https://github.com/automators-com/flowproof"

{flowproof-0.6.1 → flowproof-0.7.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: flowproof
-Version: 0.6.1
+Version: 0.7.0
 Classifier: Development Status :: 2 - Pre-Alpha
 Classifier: Intended Audience :: Developers
 Classifier: Programming Language :: Python :: 3

{flowproof-0.6.1 → flowproof-0.7.0}/crates/flowproof-adapters/src/agent_runner.rs

@@ -277,12 +277,54 @@ fn configure(
     child.env("FLOWPROOF_LLM_PROXY", base);
     // The spec's own env goes on LAST so a flow can override any of the
     // above; it knows its client better than this module does.
+    //
+    // Values may reference runtime handles the spec cannot know when it is
+    // written, because the proxy binds an ephemeral port. Without this, an
+    // agent whose client reads a non-standard variable (an AI gateway URL,
+    // say) cannot be pointed at the proxy from the spec at all, and every
+    // such adopter writes the same wrapper script.
     for (key, value) in env {
-        child.env(key, value);
+        child.env(key, substitute_runtime_handles(value, base, env));
     }
     Ok(child)
 }
 
+/// Replace the `${flowproof.*}` handles with values only known at spawn time.
+///
+/// - `${flowproof.proxy_url}` - the model proxy, `/v1` included, the form
+///   OpenAI-compatible clients expect.
+/// - `${flowproof.proxy_url_no_v1}` - the same origin without `/v1`, for a
+///   client that appends its own path (the Anthropic SDK does).
+/// - `${flowproof.mcp_url.<name>}` - the stand-in URL for that MCP server,
+///   for a client that builds its endpoints from a base rather than taking
+///   one URL per server.
+///
+/// Anything else is left untouched: an unknown handle is far more likely to
+/// be a value that happens to look like one than a typo worth failing on,
+/// and `${VAR}` secret refs have already been resolved by this point.
+fn substitute_runtime_handles(value: &str, base: &str, env: &BTreeMap<String, String>) -> String {
+    if !value.contains("${flowproof.") {
+        return value.to_string();
+    }
+    let mut out = value
+        .replace("${flowproof.proxy_url_no_v1}", &anthropic_base(base))
+        .replace("${flowproof.proxy_url}", base);
+    // The MCP stand-in URLs are already in this same map, injected as
+    // `FLOWPROOF_MCP_URL_<NAME>`; this is a friendlier spelling of the same
+    // value, so the two can never disagree.
+    for (key, url) in env {
+        if let Some(name) = key.strip_prefix("FLOWPROOF_MCP_URL_") {
+            out = out.replace(&format!("${{flowproof.mcp_url.{name}}}"), url);
+            // Server names are lowercase in the spec; the env var is upper.
+            out = out.replace(
+                &format!("${{flowproof.mcp_url.{}}}", name.to_ascii_lowercase()),
+                url,
+            );
+        }
+    }
+    out
+}
+
 /// Wait for `child` to the deadline, killing it at the timeout. Returns the
 /// exit status (`None` if killed or unwaitable) and whether it timed out.
 fn wait_to_deadline(
@@ -434,6 +476,77 @@ pub fn run_against_contained(
 
 #[cfg(test)]
 mod tests {
+
+    /// The gap a real adopter hit: their client reads AI_GATEWAY_URL, and
+    /// the proxy's port is not known when the spec is written, so a static
+    /// `agent.env` value could not reach it.
+    #[test]
+    fn the_proxy_url_is_addressable_from_a_spec_env_value() {
+        let env = BTreeMap::new();
+        let got =
+            substitute_runtime_handles("${flowproof.proxy_url}", "http://127.0.0.1:51234/v1", &env);
+        assert_eq!(got, "http://127.0.0.1:51234/v1");
+    }
+
+    /// Some clients append their own `/v1`, so handing them the suffixed
+    /// form produces `/v1/v1`. The no-suffix handle is for those.
+    #[test]
+    fn the_no_v1_handle_strips_the_suffix() {
+        let env = BTreeMap::new();
+        let got = substitute_runtime_handles(
+            "${flowproof.proxy_url_no_v1}",
+            "http://127.0.0.1:51234/v1",
+            &env,
+        );
+        assert_eq!(got, "http://127.0.0.1:51234");
+    }
+
+    /// A client that builds several endpoints off one base (`<base>/mcp`,
+    /// `<base>/mcp-exec/sap`) has no per-server variable to override, so it
+    /// needs the stand-in's base by name.
+    #[test]
+    fn an_mcp_stand_in_is_addressable_by_server_name() {
+        let mut env = BTreeMap::new();
+        env.insert(
+            "FLOWPROOF_MCP_URL_DATAMAKER_EXEC".to_string(),
+            "http://127.0.0.1:44100/mcp".to_string(),
+        );
+        // Upper (as the var is spelled) and lower (as the spec names it).
+        for handle in [
+            "${flowproof.mcp_url.DATAMAKER_EXEC}",
+            "${flowproof.mcp_url.datamaker_exec}",
+        ] {
+            let got = substitute_runtime_handles(handle, "http://127.0.0.1:1/v1", &env);
+            assert_eq!(got, "http://127.0.0.1:44100/mcp", "{handle}");
+        }
+    }
+
+    /// Handles interpolate INTO a larger value, because that is how a base
+    /// URL is actually used.
+    #[test]
+    fn a_handle_substitutes_inside_a_longer_value() {
+        let env = BTreeMap::new();
+        let got = substitute_runtime_handles(
+            "${flowproof.proxy_url_no_v1}/custom/path",
+            "http://127.0.0.1:9/v1",
+            &env,
+        );
+        assert_eq!(got, "http://127.0.0.1:9/custom/path");
+    }
+
+    /// A value that is not a handle must survive untouched - including one
+    /// that merely looks shell-ish. Failing on an unknown handle would be
+    /// worse than passing it through.
+    #[test]
+    fn ordinary_values_and_unknown_handles_pass_through() {
+        let env = BTreeMap::new();
+        for value in ["plain", "$HOME/x", "${flowproof.not_a_handle}"] {
+            assert_eq!(
+                substitute_runtime_handles(value, "http://127.0.0.1:1/v1", &env),
+                value
+            );
+        }
+    }
     use super::*;
     use flowproof_trace::cassette::{Message, ToolCall, Turn, TurnRequest, TurnResponse};
 

{flowproof-0.6.1 → flowproof-0.7.0}/crates/flowproof-cli/src/agent_flow.rs

@@ -1007,7 +1007,8 @@ fn unprotected_tool_warning(spec: &FlowSpec, plan: &Plan, phase: &str) -> Option
          intercept tool execution - {consequence}.\n         \
          For a tool with real side effects, declare it under `mcp:` with a `result:` \
          (flowproof answers it and the real server never runs it), or stub/sandbox it \
-         author-side. See docs/agent-testing.md."
+         author-side. See \
+         https://github.com/automators-com/flowproof/blob/main/docs/agent-testing.md."
     ))
 }
 

{flowproof-0.6.1 → flowproof-0.7.0}/crates/flowproof-cli/tests/examples_resolve.rs

@@ -12,6 +12,7 @@ const CONN_TEST_SPEC: &str = include_str!("../../../examples/api/connection-test
 /// from `npx flowproof` runs, so it has to keep parsing.
 const AGENT_NODE_SPEC: &str = include_str!("../../../examples/agent-demo/weather-node.flow.yaml");
 const AGENT_PY_SPEC: &str = include_str!("../../../examples/agent-demo/weather.flow.yaml");
+const DEMO_SPEC: &str = include_str!("../../../scripts/demo/order-status.flow.yaml");
 
 #[test]
 fn connection_test_example_resolves_with_body_and_headers() {
@@ -119,23 +120,17 @@ fn both_agent_demos_resolve_and_assert_the_same_thing() {
     );
 }
 
-/// The quickstart is the most-read page we have and it is now the front
-/// door for the npm audience, so its YAML must not drift from the file it
-/// claims to quote. A reader who copies a block that no longer parses is
-/// the worst possible first experience.
+/// The two quickstarts are the most-read prose we have and they are now the
+/// front door for the npm audience, so their YAML must not drift from the file
+/// they claim to quote. A reader who copies a block that no longer parses is
+/// the worst possible first experience. Both the README and
+/// docs/getting-started.md open on the same shipped example, so both are held
+/// to it here.
 #[test]
 fn the_quickstart_quotes_the_shipped_agent_example_verbatim() {
+    const README: &str = include_str!("../../../README.md");
     const DOC: &str = include_str!("../../../docs/getting-started.md");
 
-    // Pull the fenced block that names the example file.
-    let marker = "```yaml\n# examples/agent-demo/weather-node.flow.yaml\n";
-    let start = DOC
-        .find(marker)
-        .expect("quickstart quotes the node example")
-        + marker.len();
-    let block = &DOC[start..];
-    let block = &block[..block.find("```").expect("fence closes")];
-
     // The shipped file, minus its comment header.
     let shipped: String = AGENT_NODE_SPEC
         .lines()
@@ -143,12 +138,54 @@ fn the_quickstart_quotes_the_shipped_agent_example_verbatim() {
         .collect::<Vec<_>>()
         .join("\n");
 
-    assert_eq!(
-        block.trim(),
-        shipped.trim(),
-        "docs/getting-started.md has drifted from examples/agent-demo/weather-node.flow.yaml"
+    for (name, prose) in [("README.md", README), ("docs/getting-started.md", DOC)] {
+        // Pull the fenced block that names the example file.
+        let marker = "```yaml\n# examples/agent-demo/weather-node.flow.yaml\n";
+        let start = prose
+            .find(marker)
+            .unwrap_or_else(|| panic!("{name} quotes the node example"))
+            + marker.len();
+        let block = &prose[start..];
+        let block = &block[..block.find("```").expect("fence closes")];
+
+        assert_eq!(
+            block.trim(),
+            shipped.trim(),
+            "{name} has drifted from examples/agent-demo/weather-node.flow.yaml"
+        );
+
+        // And it must still parse, so the block a reader copies actually runs.
+        FlowSpec::parse(block).expect("the quoted quickstart block parses");
+    }
+}
+
+/// The flow behind the README's demo GIF. The GIF is a capture of this spec
+/// running, so a grammar change that breaks it would silently make the README's
+/// hero image a picture of something that no longer works.
+#[test]
+fn readme_demo_spec_resolves() {
+    let spec = FlowSpec::parse(DEMO_SPEC).expect("demo spec parses");
+    assert_eq!(spec.app.id(), "agent");
+
+    // Deliberately unmocked: the demo's tool returns a deterministic result, so
+    // replay needs no substitution and the run earns no unprotected-tool
+    // warning. That is what keeps the captured output to five lines.
+    let tool = spec
+        .tools
+        .iter()
+        .find(|t| t.name == "lookup_order")
+        .expect("lookup_order declared");
+    assert!(
+        tool.result.is_null(),
+        "the demo tool must stay declaration-only"
     );
 
-    // And it must still parse, so the block a reader copies actually runs.
-    FlowSpec::parse(block).expect("the quoted quickstart block parses");
+    // Paths are relative to the repository root, because that is where the GIF
+    // runs its commands and therefore what a reader copies.
+    let command = spec
+        .agent
+        .as_ref()
+        .and_then(|a| a.command.clone())
+        .expect("demo has a command");
+    assert_eq!(command, "python3 scripts/demo/support_agent.py");
 }

{flowproof-0.6.1 → flowproof-0.7.0}/flowproof/__init__.py

@@ -24,7 +24,7 @@ from flowproof.flow import (
     run,
 )
 
-__version__ = "0.6.1"
+__version__ = "0.7.0"
 
 __all__ = [
     "ClarificationNeeded",

{flowproof-0.6.1 → flowproof-0.7.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "flowproof"
-version = "0.6.1"
+version = "0.7.0"
 description = "Deterministic tests for AI agents: record a run once, replay it with zero LLM calls. Assert which tools an agent called, with which arguments, in which order."
 readme = "README.md"
 license = "Apache-2.0"