What is UmadEv
UmadEv is a local Rust binary released on GitHub that wraps existing logged-in sessions of Claude Code, Codex, or OpenCode. It imposes a fixed nine-stage delivery pipeline with explicit role assignments and quality gates. The project reached GitHub Trending in the last week after its maintainers published the full specification and example run logs. It treats the base model as the execution engine while the binary itself handles orchestration, session forking for reviews, and audit logging.
Pipeline Stages and Governance
The core of
A separate director role sits at each gate. It reads structured outputs from the blackboard files, compares them against the host spec, and either passes the artifact forward or injects a single corrective instruction back into the main session. Review roles execute in read-only forked sessions so they cannot alter the primary context. Termination is driven by measurable signals such as contract compliance, coverage thresholds, and security scan results rather than model self-assessment.
This design removes free-form multi-agent chat. All communication occurs through versioned files and JSON verdicts, which the Rust layer validates before allowing the next stage.
Integration with Coding Models
UmadEv does not call model endpoints itself. It requires an already authenticated Claude Code, Codex, or OpenCode session and reuses that session across the entire pipeline. The binary monitors the session state, writes role-specific prompts at each stage, and parses the resulting artifacts.
Because the base session stays warm, earlier decisions remain in context. The maintainers note that the older single-prompt-per-stage approach often produced disconnected outputs; the continuous session reduces that drift. Users can still fall back to one-shot prompts when the persistent session fails to start, but the default path is the long-lived one.
The tool ships with a UMADEV_HOST_SPEC_V1 file that defines the exact artifacts expected at each gate. Any deviation triggers an immediate rollback instruction rather than continued generation.
Implementation Details and Trade-offs
The project is written in Rust and produces a single static binary with minimal dependencies. It reads configuration from a local .env file and writes all intermediate state to a workspace directory that can be inspected or replayed. Cargo.lock is committed, so builds are reproducible across machines.
One practical constraint is that the base model must already be capable of the underlying tasks. UmadEv only sequences and gates; it cannot add capabilities the model lacks. Another limit is session length. Long pipelines can exceed the base model's context window even with continuous reuse, forcing manual compaction or stage restarts.
On the positive side, the explicit gates produce an audit trail of every decision and artifact. Teams that need traceable delivery can review the blackboard files and director verdicts after the run completes. The Rust implementation also starts quickly and runs with low overhead on developer laptops.
FAQs
Does UmadEv replace the base coding model? No. It loads an existing authenticated session of Claude Code, Codex, or OpenCode and only adds the pipeline logic around it.
Can I use models other than the three supported bases? The binary itself does not route to other providers. Any additional model support must come from the base tool.
How are quality failures handled? The director role converts gate failures into one targeted instruction that is injected back into the main session, then the affected stage restarts with the new context.
---
📖 Related articles
- Agentic Coding: Una Trappola per lo Sviluppo Software?
- Claude Code: da prompt a agente con skills e subagent
- File agents.md: utili per gli agenti di coding?
Need a consultation?
I help companies and startups build software, automate workflows, and integrate AI. Let's talk.
Get in touch