docs: add executive summary for team

Co-authored-by: Ishaan Jaff <ishaan-jaff@users.noreply.github.com>

Author: cursoragent

SUMMARY_FOR_TEAM.md

@@ -0,0 +1,213 @@
+# Quick Summary: Opus 4.7 Haiku 4.5 Bedrock Regression
+
+**For:** Ishaan, Krrish, Yuneng, LiteLLM Team  
+**Date:** April 17, 2026  
+**Issue:** Cost spike on Haiku 4.5 Bedrock after Opus 4.7 upgrade
+
+---
+
+## TL;DR
+
+The Opus 4.7 day 0 launch (PRs #25867, #25876 merged April 16) extended hardcoded model pattern checks that are now incorrectly sending thinking/reasoning metadata to Haiku 4.5 models on Bedrock. Bedrock rejects these requests, causing retry storms and cost spikes.
+
+---
+
+## What Happened
+
+```
+User upgrades: v1.82.3-stable.patch.4 → v1.82.3-stable.opus-4.7
+           ↓
+Haiku 4.5 requests now include thinking metadata
+           ↓
+Bedrock API error: "unsupported thinking metadata"
+           ↓
+Retries/fallback logic kicks in
+           ↓
+Cost spike observed
+```
+
+---
+
+## Root Cause
+
+### The PRs
+- **PR #25867** - "Litellm day 0 opus 4.7 support"
+- **PR #25876** - "Litellm hotfix opus 4.7" (merged April 16, 2026 at 7:19pm)
+
+### What Changed
+
+1. **Extended `is_claude_4_5_on_bedrock()` function**
+   - File: `litellm/llms/bedrock/common_utils.py`
+   - Added Opus 4.7 patterns: `"opus-4.7"`, `"opus_4_7"`, `"opus-4-7"`, `"opus_4_7"`
+   - **Problem:** Function name says "4.5" but now covers 4.5, 4.6, AND 4.7
+   - **Impact:** Any code using this function to gate behavior got modified behavior
+
+2. **Added Opus 4.7 to `_supports_extended_thinking_on_bedrock()`**
+   - File: `litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py`
+   - Uses broad pattern matching: `"claude-haiku-4"` in model
+   - **Problem:** Matches Haiku 4.5 too!
+
+3. **Extended adaptive thinking checks**
+   - File: `litellm/llms/anthropic/common_utils.py`
+   - `_is_adaptive_thinking_model()` now includes Opus 4.7
+
+### Why This Breaks Haiku 4.5
+
+Haiku 4.5 **does not support** the same thinking metadata as Opus models. The broad pattern matching added for Opus 4.7 catches Haiku 4.5 models:
+
+```python
+# This pattern now matches BOTH Opus 4.7 AND Haiku 4.5!
+if "claude-haiku-4" in model:
+    # Send thinking metadata
+    pass
+```
+
+When thinking metadata is sent to Haiku 4.5:
+- ❌ Bedrock API rejects the request
+- 🔁 LiteLLM retries
+- 💰 Cost multiplies
+
+---
+
+## Evidence
+
+### From Greptile Review (PR #25876)
+> "`is_claude_4_5_on_bedrock` now covers 4.5, 4.6, *and* 4.7 model families. The name and docstring are misleading to any caller."
+
+### From GitHub Discussion #22555 (Day 0 Model Release Gaps)
+> **Gap G13 - Bedrock:** "Parallel tool use on Bedrock gated on hardcoded Claude 4.5/4.6 patterns"
+> 
+> `is_claude_4_5_on_bedrock()` controls feature flags. Adding new patterns affects all models that match.
+
+### From PR #24053 (Open, addresses same issue)
+> "The real bug is that `supports_reasoning()` fails for Bedrock because `_get_model_info_helper` / `_supports_factory` does not normalize the `'bedrock'` vs `'bedrock_converse'` provider name difference."
+
+---
+
+## Why CI Didn't Catch This
+
+From Slack thread:
+> **Yuneng:** "no not for opus 4.7"  
+> **Krrish:** "oh, why? i assume that would have prevented this"
+
+The Opus 4.7 patch **did not go through full CI/CD**, which would have caught the Haiku 4.5 regression with integration tests.
+
+---
+
+## Immediate Fix Needed
+
+### Option 1: Explicit Exclusion (Fastest)
+Add explicit checks to exclude Haiku 4.5 before broad Claude 4.x matching:
+
+```python
+def _supports_extended_thinking_on_bedrock(model: str) -> bool:
+    # Explicitly exclude Haiku 4.5
+    if "haiku-4-5" in model or "haiku_4_5" in model:
+        return False
+    
+    # Then check for Opus/Sonnet 4.x
+    if "claude-opus-4" in model or "claude-sonnet-4" in model:
+        return True
+    
+    return False
+```
+
+### Option 2: Narrow Pattern Matching (Better)
+Replace broad patterns with specific version checks:
+
+```python
+# Instead of: if "claude-haiku-4" in model
+# Use:
+if any(p in model for p in ["haiku-4-6", "haiku-4-7", "haiku-4-8"]):
+    # Only specific versions, not 4.5
+```
+
+### Option 3: Revert & Rework (Safest)
+1. Revert PRs #25867 and #25876
+2. Fix the root cause (provider name mismatch in `supports_reasoning()`)
+3. Re-implement Opus 4.7 support using JSON capability flags
+4. Run full CI/CD
+
+---
+
+## Long-term Fix
+
+This is the **3rd time** this pattern has caused issues. Need architectural fix:
+
+1. **Fix `supports_reasoning()` for Bedrock**
+   - Normalize `"bedrock"` vs `"bedrock_converse"` provider names
+   - Make `model_prices_and_context_window.json` work correctly
+
+2. **Remove all hardcoded model checks**
+   - Delete: `_is_claude_4_6_model()`, `_is_opus_4_7_model()`, etc.
+   - Replace with: `supports_reasoning(model, provider)` JSON lookups
+
+3. **Add proper JSON flags**
+   - `supports_extended_thinking`
+   - `supports_adaptive_thinking`
+   - `supports_xhigh_reasoning_effort`
+
+This is already documented in **Discussion #22555** - just needs implementation.
+
+---
+
+## Action Items
+
+- [ ] **@Yuneng** - Pull logs from affected customers to confirm "unsupported thinking metadata" errors
+- [ ] **@Sameer** - Test Haiku 4.5 on Bedrock with current main branch
+  - Does it send thinking params?
+  - Does Bedrock reject them?
+- [ ] **@Ishaan** - Review `is_claude_4_5_on_bedrock()` call sites
+  - What behavior is gated by this function?
+  - Did Opus 4.7 changes affect Haiku 4.5 code paths?
+- [ ] **@Krrish** - Implement hotfix (Option 1 or 2 above)
+- [ ] **Team** - Schedule architectural fix (Discussion #22555)
+- [ ] **Team** - Require full CI/CD for all model launches (no exceptions)
+
+---
+
+## Files to Check
+
+Priority order for investigation:
+
+1. `litellm/llms/bedrock/common_utils.py`
+   - Line 565-597: `is_claude_4_5_on_bedrock()`
+   
+2. `litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py`
+   - `_supports_extended_thinking_on_bedrock()`
+   
+3. `litellm/llms/anthropic/chat/transformation.py`
+   - Lines 172-210, 237-248: Thinking/reasoning logic
+   
+4. `litellm/llms/bedrock/chat/converse_transformation.py`
+   - Line 565-580: `get_supported_openai_params()`
+
+---
+
+## Test Commands
+
+```bash
+# See the exact changes
+git diff v1.82.3-stable.patch.4..v1.82.3-stable.opus-4.7 -- \
+  litellm/llms/bedrock/common_utils.py \
+  litellm/llms/anthropic/chat/transformation.py
+
+# Test Haiku 4.5 with thinking (should fail)
+python test_haiku_bedrock_thinking.py  # Create this test
+
+# Run full test suite
+make test-unit
+```
+
+---
+
+## Timeline
+
+- **April 16, 2026 5:19pm** - PR #25867 merged into main
+- **April 16, 2026 7:19pm** - PR #25876 (hotfix) merged
+- **April 17, 2026 7:32pm** - User reports cost spike on Slack
+- **April 17, 2026 7:47pm** - Ishaan asks Cursor to investigate
+
+---
+
+**Bottom line:** Opus 4.7 launch used hardcoded patterns that broke Haiku 4.5. Need immediate hotfix + long-term architectural fix.