feat: session-configurable reasoning effort + media provider aliases

os-apps/paw-agent/wasm/provider_caller/src/lib.rs

@@ -2967,6 +2967,21 @@ fn build_openai_responses_input(messages: &[Value]) -> OpenAiResponsesInput {
 ///
 /// Uses the Responses API format (not Chat Completions): instructions, input, stream=true.
 /// The WASM http_call buffers the full SSE stream — we parse the response.completed event.
+fn reasoning_effort_from_options(provider_options_json: &str) -> &'static str {
+    let parsed: Option<Value> = serde_json::from_str(provider_options_json).ok();
+    match parsed
+        .as_ref()
+        .and_then(|v| v.get("reasoning_effort"))
+        .and_then(Value::as_str)
+        .unwrap_or("")
+    {
+        "minimal" => "minimal",
+        "low" => "low",
+        "high" => "high",
+        _ => "medium",
+    }
+}
+
 fn call_openai(
     ctx: &Context,
     temper_api_url: &str,
@@ -2980,6 +2995,7 @@ fn call_openai(
     tools: &[Value],
     temperature: f64,
     provider: &str,
+    provider_options_json: &str,
 ) -> Result<LlmResponse, String> {
     // Convert Anthropic-format messages to Responses API input format
     let pre_convert_types: Vec<String> = messages
@@ -3046,7 +3062,9 @@ fn call_openai(
         "stream": true,
         "store": false,
         "reasoning": {
-            "effort": "medium",
+            // Session-configurable via provider_options_json
+            // {"reasoning_effort": ...}; "medium" preserves prior behavior.
+            "effort": reasoning_effort_from_options(provider_options_json),
             "summary": "auto",
         },
     });
@@ -3954,6 +3972,7 @@ pub fn run_provider_caller() -> Result<(), String> {
                 &prepared.tools,
                 temperature,
                 &provider,
+                &provider_options_json,
             ),
             other => Err(format!("unsupported LLM provider: {other}")),
         },
@@ -4235,6 +4254,18 @@ fn check_phase_budget(
 
 #[cfg(test)]
 mod tests {
+    #[test]
+    fn reasoning_effort_defaults_to_medium_and_accepts_valid_levels() {
+        use super::reasoning_effort_from_options as effort;
+        assert_eq!(effort(""), "medium");
+        assert_eq!(effort("not json"), "medium");
+        assert_eq!(effort("{}"), "medium");
+        assert_eq!(effort(r#"{"reasoning_effort":"high"}"#), "high");
+        assert_eq!(effort(r#"{"reasoning_effort":"low"}"#), "low");
+        assert_eq!(effort(r#"{"reasoning_effort":"minimal"}"#), "minimal");
+        assert_eq!(effort(r#"{"reasoning_effort":"xhigh"}"#), "medium");
+    }
+
     use super::*;
 
     #[test]

os-apps/paw-media/specs/media_generation.ioa.toml

@@ -158,7 +158,7 @@ kind = "input"
 from = ["Created", "Failed"]
 to = "Authorizing"
 params = ["prompt", "media_type", "operation", "provider", "model", "size", "quality", "output_format", "background", "workspace_id", "output_path"]
-hint = "Start a governed media generation request."
+hint = "Start a governed media generation request. provider: omit it or pass 'openai_codex' (aliases: openai, codex) — the only v1 provider; media_type 'image', operation 'generate'."
 effect = [{ type = "trigger", name = "ensure_provider_auth" }]
 
 [[action.triggers]]

os-apps/paw-media/wasm/openai_codex_image_generate/src/lib.rs

@@ -187,7 +187,10 @@ fn validate_request(fields: &Value) -> Result<(), String> {
 fn normalize_provider(provider: &str) -> String {
     match provider.trim().to_ascii_lowercase().as_str() {
         "" => DEFAULT_PROVIDER.to_string(),
-        "codex" | "openai-codex" => DEFAULT_PROVIDER.to_string(),
+        // "openai" is the guess models make first; both observed hero-image
+        // failures (Mullion, Madori 2026-07) burned an attempt on it. There
+        // is exactly one v1 provider, so alias every obvious spelling.
+        "openai" | "codex" | "openai-codex" | "openai codex" => DEFAULT_PROVIDER.to_string(),
         other => other.to_string(),
     }
 }