v3.14.4: Current models, honest outcomes

What's New

Agentic QE 3.14.4 moves to the current AI model generation and makes results honest about what actually happened. Most of the fixes in this release come from @rudycelekli.

Current models, fewer failures

• Routing now uses Claude Haiku 4.5, Sonnet 5 (cheaper than Sonnet 4.6) and Opus 5.5. OpenAI defaults to GPT-6, and the local default is qwen3-coder:30b.
• Claude requests no longer fail on Opus 4.7 and the Claude 5 models, and GPT-6 requests work too. Budgets now price every routable model instead of counting some as free.
• Security agents stay on Sonnet 4.6 for Opus/Fable-class models until you enable Cyber Verification (AQE_CYBER_VERIFIED=true).

Security

• File paths and git refs passed to MCP tools can no longer run shell commands or smuggle git options.
• All open Dependabot alerts are patched.

Honest outcomes

• aqe fleet run and aqe task submit --wait exit 1 when work fails or never ran. MCP task_submit reports the real task state, and failed MCP tool calls are marked as errors.
• Workflow approval gates now wait for approval before acting; use the new pipeline_approve / pipeline_reject MCP tools. Parallel steps can no longer overwrite each other's outputs.
• Cancelled tasks stay cancelled, and aborting a test run stops the whole Vitest process tree.

Your learning data

• Backups are consistent snapshots, and restore takes a safety copy first.
• aqe brain import no longer aborts on pattern conflicts, and --dry-run shows the real merge result.
• Running aqe from a subfolder no longer creates stray stores in several common cases.

Heads-up

• aqe fleet run test / analyze currently exit 1, because --target is not yet mapped to task inputs. Before this release they reported success without doing the work.
• Pull the new local default with ollama pull qwen3-coder:30b.
• Azure GPT-5/6 deployments should set model explicitly.

Getting Started

npx agentic-qe init --auto

See CHANGELOG and the release notes for full details.