# Bug Report: Opus Used During Initialization Failures (v2.7.2-beta.10)

[azrlb]

Checklist

• I searched existing issues and this hasn't been reported

Area

Backend

Operating System

Windows

Version

OS: Windows 10 (MSYS_NT-10.0-19045) - Auto-Claude Desktop: v2.7.2-beta.10 - Claude Code CLI: 2.0.75

What happened?

Problem: Auto-Claude uses Opus during "Control request timeout: initialize" failures, even with Balanced/Sonnet profile selected.

Expected Behavior: All API calls should use Sonnet when Balanced profile is selected, including error handling/retry logic.

Actual Behavior: ~40K Opus tokens consumed during failed initialization attempts.

Steps to reproduce

1. Install v2.7.2-beta.10
2. Set Default Agent Profile to "Balanced" (Sonnet)
3. Create and start a task
4. Task fails with "Agent error: Control request timeout: initialize"
5. Check Anthropic console - Opus tokens increased despite Sonnet setting

Expected behavior

All API calls should use Sonnet when Balanced profile is selected, including error handling/retry logic.

Suggested Fix

Check all error handling and retry logic paths for hardcoded Opus fallbacks. The model selection from settings/task_metadata should be used consistently across all code paths, including:

• Agent initialization retries
• Spec writing retries
• Any fallback/recovery logic

Additional Context

• Normal operation uses Sonnet correctly (PR fix(model): respect task_metadata.json model selection #415 fix working)
• Only during failures/retries does Opus get used
• This can be expensive if tasks fail repeatedl

Logs / Screenshots

Agent error: Control request timeout: initialize
Starting phase 4: HISTORICAL CONTEXT
Knowledge graph not configured, skipping
Starting phase 5: CONTEXT DISCOVERY
Starting phase 6: SPEC DOCUMENT CREATION
Agent error: Control request timeout: initialize
Agent error: Control request timeout: initialize
Agent error: Control request timeout: initialize
Phase 'spec_writing' failed: Attempt 1: Agent did not create spec.md; Attempt 2: Agent did not create spec.md; Attempt 3: Agent did not create spec.md
Phase spec_writing failed
# Anthropic Console Evidence

| Time | Opus 4.5 | Sonnet 4.5 |
|------|----------|------------|
| 17:06 | 65,112 | 0 |
| 17:10 | 106,000 | 0 |
| 17:16 | 0 | 16,700 |
~40K Opus tokens consumed during failures

## Settings Files

**settings.json** (`C:\Users\Bob\AppData\Roaming\auto-claude-ui\settings.json`):

{
  "defaultModel": "sonnet",
  "selectedAgentProfile": "balanced",
  "featureModels": {
    "insights": "haiku",
    "ideation": "sonnet",
    "roadmap": "sonnet"
  }
}


## Related Issues

- PR #415 (merged Dec 30, 2025) fixed normal model selection - this works correctly
- This bug appears to be in error handling/retry code path that wasn't covered by PR #415

[azrlb]

Files with Hardcoded claude-opus-4-5-20251101:

File	Line	Context
auto-claude/ideation/config.py	28	model: str = "claude-opus-4-5-20251101"
auto-claude/ideation/generator.py	59	model: str = "claude-opus-4-5-20251101"
auto-claude/ideation/runner.py	44	model: str = "claude-opus-4-5-20251101"
auto-claude/ideation/types.py	34	model: str = "claude-opus-4-5-20251101"
auto-claude/runners/ideation_runner.py	97-98	default="claude-opus-4-5-20251101"
auto-claude/runners/roadmap/models.py	26	model: str = "claude-opus-4-5-20251101"
auto-claude/runners/roadmap/orchestrator.py	30	model: str = "claude-opus-4-5-20251101"
auto-claude/runners/roadmap_runner.py	58-59	default="claude-opus-4-5-20251101"

File with Hardcoded "planning": "opus":

File	Line	Context
auto-claude/phase_config.py	52	DEFAULT_PHASE_MODELS = { ... "planning": "opus" ... }

Suggested Fix

1. Change all hardcoded defaults from Opus to Sonnet - Sonnet should be the safe default
2. Opus should remain a selectable option - Users who want Opus can explicitly choose it
3. All code paths should respect task_metadata.json - The user's model selection should propagate everywhere

Quick Fix Command:

# Replace hardcoded opus model IDs with sonnet
find auto-claude -name "*.py" -exec sed -i 's/claude-opus-4-5-20251101/claude-sonnet-4-5-20250929/g' {} \;

# Fix phase_config.py planning default
sed -i 's/"planning": "opus"/"planning": "sonnet"/' auto-claude/phase_config.py

This ensures users on "Balanced" profile don't accidentally burn Opus tokens during ideation, roadmap, or planning phases.

[joelfuller2016]

@openhands-agent Please fix this critical model selection bug during failures.

Issue Summary: Auto-Claude uses Opus model during "Control request timeout: initialize" failures, even when Balanced/Sonnet profile is selected. ~40K Opus tokens consumed during failed initialization attempts despite settings showing Sonnet.

Platform: Windows 10, v2.7.2-beta.10, Claude Code CLI 2.0.75

Priority: HIGH - This causes unexpected API costs

Error Pattern:

Agent error: Control request timeout: initialize
Phase 'spec_writing' failed: Agent did not create spec.md (after 3 attempts)

Token Evidence from Anthropic Console:

• 17:06: Opus +65K, Sonnet 0
• 17:10: Opus +106K, Sonnet 0
• 17:16: Opus 0, Sonnet +16K (normal operation works)

User Settings (correctly configured):

{
  "defaultModel": "sonnet",
  "selectedAgentProfile": "balanced",
  "featureModels": { "insights": "haiku", "ideation": "sonnet", "roadmap": "sonnet" }
}

Root Cause Analysis:

• PR fix(model): respect task_metadata.json model selection #415 fixed normal model selection (working)
• Bug is in error handling/retry code path NOT covered by PR fix(model): respect task_metadata.json model selection #415
• Likely hardcoded Opus fallback in retry logic

Investigation Areas:

1. apps/backend/core/client.py - Check retry/fallback logic
2. apps/backend/agents/ - Agent initialization retry code
3. apps/backend/spec_agents/ - Spec writing retry logic
4. Search for hardcoded opus or claude-3-opus in error paths

Required Fix:

1. Find all retry/fallback code paths
2. Remove hardcoded Opus references
3. Ensure model from settings/task_metadata is used in ALL code paths
4. Add tests to verify model selection in failure scenarios

Target Branch: develop