fix(providers): recover Groq tool_use_failed as a plain assistant turn (v0.7.6) (#97)

Groq rejects a generation with HTTP 400 code=tool_use_failed when the model
emits a malformed tool call — but preserves the text the model was producing
in error.failed_generation. Previously this failed the whole run; now the
adapter recovers that text as a plain assistant turn (no intents — a call
that cannot parse never reaches the runtime), so the chat gets the model's
words instead of an error. A humanized fallback covers variants without
recoverable text. With Auto mode advertising the full tool surface to
llama-class models this failure is common, so recovery matters.

Versions bumped to 0.7.6.

Verified: cognition + runtime crates build and test green. Not verified
here: a live Groq tool_use_failed reproduction (depends on model behavior).

Co-authored-by: grysz <117801274+grysz@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: 3 people

thymos/Cargo.lock

@@ -16,7 +16,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.7.5"
+version = "0.7.6"
 edition = "2021"
 license = "Apache-2.0"
 rust-version = "1.80"

thymos/Cargo.toml

@@ -1,6 +1,6 @@
 {
   "name": "thymos-desktop",
-  "version": "0.7.5",
+  "version": "0.7.6",
   "private": true,
   "description": "OpenThymos Desktop — governed-cognition runtime, install once.",
   "scripts": {

thymos/clients/desktop/package-lock.json

@@ -5,7 +5,7 @@
 
 [package]
 name = "thymos-desktop"
-version = "0.7.5"
+version = "0.7.6"
 edition = "2021"
 license = "Apache-2.0"
 description = "OpenThymos Desktop — a governed-cognition client that supervises a local runtime."

thymos/clients/desktop/package.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "OpenThymos",
-  "version": "0.7.5",
+  "version": "0.7.6",
   "identifier": "com.openthymos.desktop",
   "build": {
     "frontendDist": "../src"

thymos/clients/desktop/src-tauri/Cargo.lock

@@ -230,6 +230,28 @@ impl Cognition for OpenAiCognition {
             .json()
             .map_err(|e| Error::Other(format!("openai response parse: {e}")))?;
         if !status.is_success() {
+            // Groq's `tool_use_failed`: the model emitted a malformed tool
+            // call and the API rejected the whole generation with a 400 —
+            // but the text the model was trying to produce is preserved in
+            // `error.failed_generation`. Recover it as a plain assistant
+            // turn instead of failing the run: no tool executes (a call that
+            // can't parse never reaches the runtime), the model just "spoke".
+            if status.as_u16() == 400 {
+                if let Some(failed) = resp_json
+                    .pointer("/error/failed_generation")
+                    .and_then(|v| v.as_str())
+                    .filter(|s| !s.trim().is_empty())
+                {
+                    let text = failed.to_string();
+                    self.messages
+                        .push(json!({ "role": "assistant", "content": text }));
+                    return Ok(CognitionStep {
+                        intents: vec![],
+                        final_answer: Some(text),
+                        usage: crate::CognitionUsage::default(),
+                    });
+                }
+            }
             // Prefer the standard header; fall back to providers that put the
             // precise wait only in the JSON body ("Please try again in 9.62s").
             let hint_ms = retry_after_header

thymos/clients/desktop/src-tauri/Cargo.toml

@@ -42,7 +42,9 @@ fn parse_retry_after_hint(msg: &str) -> Option<u64> {
 /// some providers, embed request context). Unknown errors get a generic line.
 pub fn humanize_provider_error(msg: &str) -> String {
     let m = msg.to_lowercase();
-    if m.contains("429") || m.contains("rate limit") {
+    if m.contains("tool_use_failed") {
+        "The model fumbled a tool call (a provider-side formatting failure). Just ask again — smaller steps help, or switch models in Providers.".to_string()
+    } else if m.contains("429") || m.contains("rate limit") {
         "The provider is rate-limiting requests right now (token budget).".to_string()
     } else if m.contains("401") || m.contains("invalid api key") || m.contains("unauthorized") {
         "The provider rejected the API key. Check it in Settings.".to_string()