# Aura v1.8.52 Release Notes

This release is a big internal smoothness pass for Aura’s Planner/Worker loop. The headline... Aura now carries less prompt noise, learns from repeated failure patterns more selectively, and presents Planner/Worker handoffs more cleanly in the UI.

Highlights

Learned Hazard Guards

Aura now has a learned-hazard system that records objective failure signals from Worker runs and graduates repeated failure patterns into prompt guards.

This means Aura can recognize terrain that has burned it before, such as repeated validation failures or tool/edit patterns, and warn itself before making the same mistake again.

The system is built in three layers:

• A hazard substrate that logs raw failure facts to .aura/hazards.db
• A graduation reader that clusters repeated failures by readable fingerprints
• Prompt guard injection that turns graduated hazards into compact learned warnings

Hazards are guards, not fixes. They do not mutate the workspace, rewrite history, or make silent decisions. They simply tell Aura, “this ground has teeth, verify the important behavior before calling it done.”

Terrain-Scoped Hazard Injection

Hazard guards are no longer just “top 5 recent hazards everywhere.”

Worker runs now pass real terrain into the hazard guard system:

• Worker model
• Inferred task kind
• Target files

Aura uses that terrain to select relevant learned hazards instead of injecting unrelated warnings into every prompt. This keeps the loop sharper and cuts down on prompt static.

Cleaner Planner/Worker Prompt Context

A major prompt cleanup pass removed repeated Worker rules and made context routing more selective.

The result is a lighter, less noisy Worker packet. Aura now spends more of its context budget on the task and less on repeated instruction boilerplate.

Smoother Planner/Worker UI Flow

Fixed a long-running UI boundary bug where Planner continuation text after a Worker dispatch could get fused into the same assistant card as the dispatch flow.

Now:

• Planner pre-dispatch text stays in its original Aura card
• Dispatch/spec UI stays visible
• Worker completion remains attached to the correct flow
• Planner’s final completion summary starts in a fresh Aura card

This makes the conversation timeline much easier to read and stops separate phases from getting stapled together visually.

Final Response Scroll Fix

Fixed a UI issue where the last bit of Aura’s response could appear visually cut off near the input box.

When the user is already following the bottom of the chat, Aura now performs delayed bottom-sync passes after final markdown/layout reflow. If the user has scrolled upward to read earlier output, Aura preserves their position and does not yank them back down.

Internal Improvements

• Added readable hazard fingerprints instead of opaque hashes
• Added distinct-dispatch graduation thresholds for repeated failures
• Preserved hazard dormancy naturally by only reading recent windows
• Kept hazard guard injection fail-safe so context failures never break a run
• Routed Worker hazard terrain through the normalized Worker task spec
• Preserved message persistence and dispatch semantics while fixing visual card boundaries
• Preserved all spec/tool/summary artifacts when closing an assistant card for Planner continuation