Unreleased: goal-driven agent improvement

Add a project, detect agents and open a goal. Detected agents become active immediately. Choose a goal, optionally add context and click Go; a durable coordinator prepares measurements, builds a reviewed evaluator, measures a baseline and searches for verified improvements. Missing inputs appear on the goal page. Evaluation details and execution limits remain available as advanced controls.

A run retains its exact goal and measurement revisions, child tasks and shared time/trial allowance. It protects affected sibling behaviors, reuses compatible frozen evidence and preserves failed attempts. Pause, continue and stop control the parent run. Reads never start work, and interrupted runs require explicit continuation after a service restart. Verified changes still require user selection; publication, merge and deployment remain separate actions.

Dashboard and MCP expose the same detection and goal-run operations. Evaluation cases, native evaluator selection, execution profiles and durable drafts remain available. Trace-driven Fix and explicit production monitoring use the shared task runtime.

Obsolete setup components, route aliases, legacy record fallbacks and temporary planning documents have been removed. Application metadata now uses version 5 and rejects earlier versions. Fresh application and project state are required; existing private data is not migrated or deleted. Retired public website routes return the branded 404.

Validation results and environment-specific limitations are recorded in the pull request.