How it works
What happens when you ask it to fix something.
You describe what you want. It works out what is actually there. Codex and Claude each write a plan and check the other's. You approve, and it runs small steps with a record of each.
Consensus Logic is a development controller: the layer around Codex and Claude, not another coding assistant. This page explains what it does, not the prompts.
The everyday workflow
Six things you do. Everything else is the controller's job.
- 01
Select a project
Register a folder as a project. Its record is its own: nothing from another project leaks in, and that boundary is enforced in trusted code, not by a filter on the screen.
- 02
Describe the outcome
A Task is the outcome you ask for, in your own words. Not a ticket, not a prompt, not an agent run.
- 03
Start review
Authorises discovery, two plans written separately, challenge and reconciliation. Nothing is built. If the two planners disagree on something that matters, the decision comes to you with the trade-off stated.
- 04
Inspect the programme in plain English
Steps, dependencies, limits and checks, summarised for the person in charge. The Technical tab shows the same evidence at full depth.
- 05
Choose creator and policies
Pick which AI creates; the other independently reviews. Set supervision to Manual or Automatic, and repair to Decision required or Auto-bounded. Once per task, not per step.
- 06
Start development
A separate authorisation, available only after the summary for that exact plan revision has been shown. Small, checked steps run inside the budget you set; the app calls that budget your allowance. Stop is always one click away.
Two authorities, two records
Review authority
Start review lets the AIs look, plan and argue. It cannot change your project.
Execution authority
Start development lets the approved programme run, one small, checked step at a time. It is never implied by review.
Planning never changes your project. Running it needs your separate go-ahead.The rule the product is built around
A disagreement, resolved
What “the decision comes to you” looks like.
The task: bring section 3 of a customer guide back into agreement with the product. Both planners read the same source. They disagreed on one thing that mattered.
Illustrative. The first real disagreement story will be posted to the build log from the first live planning run, with the record visible.
Rewrite section 3 from source
“Section 3 describes a setup flow that no longer exists. Rewrite it from the current source in two steps, then run the check that no section contradicts another.”
Fix section 1 first
“Section 1 defines the term section 3 contradicts. Correct the definition first or the rewrite inherits the contradiction. Three steps.”
Challenge
A accepts B's dependency but objects to size: three steps for a definition change. B accepts A's reading of the source. Neither withdraws the point that matters.
Order of work
Trade-off stated: fix the definition first (one extra step, lower risk of a second pass) or rewrite directly (faster, with the risk that the contradiction survives). You chose the definition first.
Agreed programme
Three steps, one dependency, the contradiction check at the end. Plain-English summary shown. Start development is now available.
The short version
Not just a better prompt. A better development process.
One AI can run a large task in a single pass. Consensus Logic breaks it into stages, has both AIs check each stage before the next one starts, and validates the finished work before calling it done.
Narrated, with captions. Under a minute. It plays only when you press play; nothing on this site moves on its own.
Narrated, under a minute. Captions on screen.
The control system, in order
What happens between Start review and the report.
- 01
It reads your project before anyone plans
Before anyone plans, the controller works out what is actually in the project: what exists, what changed recently, what the evidence says works and what it does not. Recovery starts from the real state, not from the last chat's memory of it.
- 02
It remembers what it found
What was examined, what was decided and what remains open is kept as a durable record of the project. The next task starts from that record. Nothing from another project is visible to it.
- 03
Every plan is checked against the real code
Before a plan is trusted, the AI that did not write it reads the source the plan depends on and reports what it actually found. A plan built on a misreading is caught here, not after it runs.
- 04
Two plans, written without seeing each other
Both planners receive the same evidence and write their plans without seeing each other's. Two models repeating a claim is not corroboration, and the product says so.
- 05
Each plan is challenged by the other
Each plan is cross-examined by the other for omissions, dependencies, size and proof. A finding can be handed back for a response. A clean pass is allowed, manufactured objections are not, and the exchange is limited and recorded.
- 06
Real disagreements come to you
A real disagreement becomes your decision, with the trade-off stated, never a dead end. Consensus is not manufactured by weakening one side until it agrees.
- 07
Work runs in small steps, each with a limit
The agreed programme is a list of small steps, each with a size limit, a list of what it depends on, and a check. One executing task at a time. No unattended scheduler.
- 08
A model saying “done” proves nothing; the output does
A model's “done” and a green test count cannot unlock the next step. The controller reads the actual output first. Evidence is labelled product or mechanism and never mixed; only the installed application can produce a product pass.
- 09
Every step leaves a record and stays inside the budget
Every AI call is reserved against the budget before it starts and settled after. Splits and retries never reset the budget. Each step closes with what changed, what passed, what failed, what is unknown and what to do next.
- 10
Stop works, and you can pick up tomorrow
Stop starts nothing new immediately and tells you whether the work actually stopped. If it cannot confirm that, nothing else runs. Close the app after a step and come back to Where we left off: completed work, open findings, what to do next and the budget you have spent, all intact.
Why the small steps matter
Keep AI focused. Keep development in scope.
Given a vague task, an AI still tries to help: it refactors what it was not asked to touch, tries another approach, widens the job. That costs hours and usage without getting closer to what you asked for. Consensus Logic scopes the work first, then checks after each small step that it is still solving your actual request.
Narrated, with captions. Under a minute. It plays only when you press play; nothing on this site moves on its own.
Narrated, under a minute. Captions on screen.
The merge motif
Two plans. One path.
Two plans that cross, challenge each other, and resolve into one direction. Six seconds. It plays only when you ask; nothing on this site moves on its own.
Silent, six seconds.
Authority stays with you
Five controls. Each one names exactly what will happen.
| Control | Options | What it means |
|---|---|---|
| Creator and Reviewer | Codex or Claude | You choose which AI creates; the other independently reviews. Once per task, not per step. |
| Supervision | Manual · Automatic | Approve each prepared step yourself, or let the next steps continue within the authority and budget you already gave. Automatic is not unattended: stopping conditions and the budget still apply. |
| Repair | Decision required · Auto-bounded | A failure comes to you first, or a limited number of corrections run within the recorded budget. A failure is never converted into a success because a model says it is fixed. |
| Allowance | Account · Project · Task · Step | The finite budget of calls, time and repairs at every level. Reserved before each call, settled after. No model can grant itself scope. |
| Stop | Always available | Starts nothing new immediately. Reports whether the work actually stopped, or says plainly that it cannot tell. If it cannot tell, nothing else runs. |
What it is not
Said plainly, so nobody buys the wrong thing.
Not a coding agent
Codex and Claude write the code. Consensus Logic decides what they are allowed to do, checks what they actually did, and keeps the record.
Not an always-on agent
No unattended scheduler. One executing task at a time. Automatic supervision runs inside the permissions and budget you set, and its stopping conditions, then stops.
Not a pay-per-use AI wrapper
It sells no tokens and needs no API key. It uses the command-line tools of the subscriptions you already have, and their limits still apply.
Not cloud, not a team product yet
Local-first, single user, Windows desktop. The record lives on your machine. Teams are a stated horizon, not a hidden tier.
Next