Cora helps humans and AI code better together.
Cora is a reflection and coaching layer for AI-assisted software development. It analyzes real coding sessions, finds friction in prompting and execution, and helps developers get to working outcomes in fewer turns.
Coding agents are already useful. Most teams still use them inefficiently.
The bottleneck is not only model quality. It is interaction quality:
- unclear prompts
- repeated exploration
- too many correction turns
- wasted reasoning and tool loops
- weak feedback loops between developer and agent
Cora exists to close that gap.
The long-term vision is to build a reflection system for human-AI software development:
- observe how developer-agent sessions actually unfold
- measure friction in prompting, execution, and correction
- identify repeated patterns that waste time or tokens
- coach developers toward better prompting and tighter workflows
- reduce turns-to-completion for real feature work
Today, Cora focuses on session reflection.
It can:
- analyze session history after work completes
- measure prompting time, execution time, correction time, and token usage
- detect repeated tool and visible reasoning patterns
- generate private developer-facing summaries
- surface actionable coaching for getting to completion faster
- run on historical sessions to learn from past work
Implemented:
- standalone dream / reflection analysis
- session-level metrics for user turns, assistant turns, tool calls, and token usage
- time split across prompting, execution, dormancy, and correction lag
- repeated pattern detection for tools and visible reasoning traces
- deterministic summary generation
- optional LLM-written summary layer
- historical session analysis against an existing local session database
- persistence model for dream records and dream state
- idle-trigger integration kept separate from the main session loop
- focused test coverage for deterministic dream analysis
Not done yet:
- polished UI for browsing reflections
- ranking and trend analysis across sessions
- developer progress tracking over time
- repo or project dashboards
- automatic prompt improvement before the next task
- team-level analytics
- stronger feature-completion scoring
- production validation that Cora reduces turns-to-completion
- Use real session evidence, not vague advice
- Focus on measurable developer outcomes
- Keep reflection separate from the main agent execution loop
- Respect boundaries around hidden reasoning
- Prefer coaching that is concrete, short, and operational
- Optimize for fewer turns to completion
This repository is derived from work originally developed in the anomalyco/opencode codebase and remains subject to the included MIT license.
The Cora-specific reflection work in this repo builds on that foundation and keeps the original license and attribution intact.
See ATTRIBUTION.md and LICENSE.