fix: handle reasoning/thinking content from models

PR_DESCRIPTION.md

@@ -0,0 +1,56 @@
+# Fix: Handle reasoning/thinking content from models
+
+## Problem
+When using reasoning-capable models (e.g., Claude with extended thinking, Grok 3, OpenAI o3/o4), the application fails with:
+```
+❌ Request failed
+[error-kind: unknown]
+error: assistant stream produced no content
+```
+
+This occurs when the model returns **only** thinking/reasoning blocks without regular text content.
+
+## Root Cause
+The SSE stream parser and event converter were explicitly ignoring `Thinking` and `RedactedThinking` content blocks:
+
+```rust
+// In rust/crates/tools/src/lib.rs
+OutputContentBlock::Thinking { .. } | OutputContentBlock::RedactedThinking { .. } => {}
+
+ContentBlockDelta::ThinkingDelta { .. } | ContentBlockDelta::SignatureDelta { .. } => {}
+```
+
+When a model returned only thinking content, zero `AssistantEvent` content events were produced. The `build_assistant_message` function then correctly rejected this as "no content".
+
+## Solution
+1. **Added `ThinkingDelta` event variant** (`rust/crates/runtime/src/conversation.rs`)
+   - New `AssistantEvent::ThinkingDelta { thinking, signature }` variant
+   - Accumulates thinking content and flushes it as text blocks wrapped in `<thinking>` tags
+   - Updated "no content" check to consider thinking as valid content
+
+2. **Emit thinking events from stream** (`rust/crates/tools/src/lib.rs`)
+   - `push_output_block` now emits `ThinkingDelta` for thinking blocks
+   - `ContentBlockDelta` handler processes `ThinkingDelta` and `SignatureDelta`
+   - Synthetic `MessageStop` check includes thinking as valid content
+
+## Changes Checklist
+
+| File | Change | Why |
+|------|--------|-----|
+| `runtime/src/conversation.rs` | Added `ThinkingDelta` variant to `AssistantEvent` | Allow thinking content to flow through the runtime |
+| `runtime/src/conversation.rs` | Added `flush_thinking_block()` helper | Convert accumulated thinking to displayable text blocks |
+| `runtime/src/conversation.rs` | Updated `build_assistant_message()` | Accept thinking as valid content; prevent false "no content" errors |
+| `runtime/src/conversation.rs` | Added tests for thinking content | Verify fix works for thinking-only and thinking+signature cases |
+| `tools/src/lib.rs` | Updated `push_output_block()` | Emit thinking events instead of ignoring |
+| `tools/src/lib.rs` | Updated `ContentBlockDelta` handler | Process thinking deltas and signatures |
+| `tools/src/lib.rs` | Updated synthetic stop check | Treat thinking as valid content for stream completion |
+
+## Testing
+- Added `build_assistant_message_accepts_thinking_content` test
+- Added `build_assistant_message_accepts_thinking_with_signature` test
+- All 23 conversation tests pass
+
+## Impact
+- Enables use of reasoning models that return thinking content
+- Backward compatible: regular text/tool content flows unchanged
+- Redacted thinking is intentionally skipped (no useful content to display)

README.md

@@ -208,3 +208,14 @@ Claw Code is built in the open alongside the broader UltraWorkers toolchain:
 
 - This repository does **not** claim ownership of the original Claude Code source material.
 - This repository is **not affiliated with, endorsed by, or maintained by Anthropic**.
+
+---
+
+### ☕ Support This Project
+
+Help keep this project going — use a referral link below and both of us get credits!
+
+| Service | Your Bonus | Details | Referral Code |
+|---------|-----------|---------|---------------|
+| [**Neuralwatt**](https://portal.neuralwatt.com/auth/register?ref=NW-ROGER-ET3Y) | $10 in credits | Spend $10+ → you get $10, we get $20 | `NW-ROGER-ET3Y` |
+| [**Synthetic**](https://synthetic.new/?referral=UAWqkKQQLFkzMkY) | $10 in credits | Subscribe → both get $10 credit | `UAWqkKQQLFkzMkY` |

rust/crates/api/src/error.rs

@@ -15,6 +15,7 @@ const CONTEXT_WINDOW_ERROR_MARKERS: &[&str] = &[
     "prompt is too long",
     "input is too long",
     "request is too large",
+    "no parseable body",
 ];
 
 #[derive(Debug)]
@@ -55,6 +56,9 @@ pub enum ApiError {
         retryable: bool,
         /// Suggested user action based on error type (e.g., "Reduce prompt size" for 413)
         suggested_action: Option<String>,
+        /// Parsed Retry-After header value (seconds) for 429 responses.
+        /// When present, overrides the exponential backoff delay.
+        retry_after: Option<Duration>,
     },
     RetriesExhausted {
         attempts: u32,
@@ -123,6 +127,18 @@ impl ApiError {
     }
 
     #[must_use]
+    /// Return the `Retry-After` delay if this error came from a 429 response
+    /// that included a `retry-after` header. Callers should prefer this value
+    /// over the computed backoff delay when it exists.
+    #[must_use]
+    pub fn retry_after(&self) -> Option<Duration> {
+        match self {
+            Self::Api { retry_after, .. } => *retry_after,
+            Self::RetriesExhausted { last_error, .. } => last_error.retry_after(),
+            _ => None,
+        }
+    }
+
     pub fn is_retryable(&self) -> bool {
         match self {
             Self::Http(error) => error.is_connect() || error.is_timeout() || error.is_request(),
@@ -491,6 +507,7 @@ mod tests {
             body: String::new(),
             retryable: true,
             suggested_action: None,
+        retry_after: None,
         };
 
         assert!(error.is_generic_fatal_wrapper());
@@ -514,6 +531,7 @@ mod tests {
                 body: String::new(),
                 retryable: true,
                 suggested_action: None,
+            retry_after: None,
             }),
         };
 
@@ -535,6 +553,7 @@ mod tests {
             body: String::new(),
             retryable: false,
             suggested_action: None,
+        retry_after: None,
         };
 
         assert!(error.is_context_window_failure());