Consortio
A pre-user-launch CLI prototype for comparing several coding-model approaches to the same repository task under human review. It uses bounded task intent, isolated candidate work, provider harnesses, run artifacts, verification ledgers and proof plans, comparable diffs, disagreement review, and human-controlled export and apply decisions so model-assisted changes can be inspected before any repository change is applied.
Problem
AI coding agents are useful, but serious repository work quickly becomes coordination work: preparing context, running different tools, comparing diffs, checking verification, tracking failures, and deciding what is safe to merge. A chat transcript is not enough evidence for that workflow.
Consortio compares multiple coding-model approaches to the same repository task. It separates task intent, isolated candidate work, verification output, disagreement review, and reviewed export/apply artifacts so a developer can inspect the path to a decision instead of merging whichever model finished first.
What works today
Consortio is a pre-user-launch CLI prototype. A run can take a task definition, prepare isolated worktrees, route work through configured model/provider harnesses, preserve phase artifacts, run checks, and produce reports for operator review.
It keeps model output as review material, not a delivery decision. Export/apply artifacts remain review material for a human operator.
The useful part is the artifact trail around each candidate's changes: task intent, execution briefs, candidate status, plans, critiques, diffs, verification results, run reports, disagreement points, and export/apply artifacts.
Product decisions
Consortio treats disagreement as information, not a defect to hide. When candidates diverge, the system should preserve alternatives, mark unresolved findings, and make the operator's decision easier rather than flattening the run into artificial consensus.
The delivery boundary is deliberately explicit. The prototype can prepare reviewed export/apply artifacts for selected work, but final delivery stays under human control because repository changes need judgement, not just automation.
The design is organized around comparable artifacts rather than a single privileged transcript. A structured task contract captures the requested work; verification ledgers record check results separately from model self-report; run reports make the remaining risk visible enough for a human operator to decide what happens next.
Known limitations
The CLI engine and artifact model exist; a polished hosted product and broad self-serve UI do not. The broader product surface remains private design work.
Provider automation also has ordinary operational constraints: authentication, quotas, context limits, sandbox setup, and provider-specific behaviour. Those limits are part of what Consortio is meant to expose rather than hide.
Proof
The public proof on this page is a sanitized run shape:
- Task intent: one repository task expressed as durable intent.
- Isolated worktrees: separate candidate workspaces for different model approaches.
- Run reports: preserved plans, critiques, diffs, and status.
- Verification ledger: checks and proof results kept separate from model self-report.
- Disagreement points: unresolved differences kept visible for review.
- Export/apply decision: selected changes prepared for human-controlled delivery.