Consortio

Consortio is a private, pre-user-launch CLI prototype. It is not a hosted product or public demo.

Problem

Serious coding-agent work quickly becomes coordination work: preparing repository context, isolating changes, comparing approaches, running checks, tracking failures, and deciding what is safe to deliver. A chat transcript cannot carry that decision on its own.

What exists now

A run can take a task definition, prepare isolated Git worktrees, route work through configured provider harnesses, preserve phase artifacts, run checks, and produce reports for operator review. Provider CLIs can be driven headlessly, with a PTY fallback where needed.

The engine supports configurable Codex, Claude, and Gemini workflows. Model output remains proposed work; verification is recorded separately; export and apply remain explicit operator decisions.

A run that produced no export

The best public artifact is not a polished success case. It is a sanitized trace from a real run against a product repository I own.

The run carried one candidate family through several implementation rounds. Static analysis passed, but the candidate-level test status remained false; the final end-to-end command failed; and required dark- and light-theme visual evidence was never proven. The recorded disposition was incomplete and the export list was empty. I treat that as the correct boundary: keep the work as evidence, but do not export or apply it from that run.

The trace supports a narrow claim: this run preserved task intent, candidate states, separate verification results, proof gaps, and a no-export outcome. It does not prove a multi-model comparison or final implementation quality. The trace used one candidate family and says so plainly.

Decisions, alternatives, and costs

Preserve artifacts before ranking

Decision: write task intent, candidate state, diffs, and verification results as durable artifacts.

Alternative: treat the conversation transcript and final model message as the run record.

Cost: artifacts need schemas, storage, lifecycle rules, and interfaces that keep detail from becoming noise. The benefit is a record that can be inspected after the session ends.

Keep disagreement visible

Decision: preserve competing approaches, failed proof, and unresolved review findings instead of flattening them into artificial consensus.

Alternative: select whichever candidate scores highest and hide the losing evidence.

Cost: the operator has more judgement work to do, and some runs correctly end without a winner.

Make delivery a separate boundary

Decision: candidate generation and verification do not themselves authorize repository delivery.

Alternative: merge or apply the top-ranked result automatically.

Cost: there is no hands-off finish. That is intentional: a repository change can pass some checks and still be poorly scoped, insufficiently proven, or wrong for the product.

Current limits

There is no public hosted product, self-serve UI, or public demo. The sanitized trace demonstrates one single-family run, not the product's broader multi-harness ambition. Authentication, quotas, context limits, sandbox setup, and provider-specific behaviour remain ordinary constraints that the prototype has to expose and handle.

For the product rationale, read What I want from agent tooling. To see the current prototype at its most concrete, inspect the no-export run trace.