skip to content
What it is

Parallax

What it is

Your coding agent works all day. When it stops, other models check its work, and each finding tells you what was verified.

A coding-agent session is an artifact: what the user asked, what the agent read, what it changed, what it claimed. Parallax records it (no extra tokens: it reads what the agent already wrote), and when the session ends it asks the Conifer router which models from other labs are best for that exact task. Those models re-do the important part blind, hunt for bugs, security holes, dead code and slower-than-needed code. The sandbox attempts to verify their findings in a throwaway copy of your repo. Every finding reaches you labelled with what the sandbox established: proven, suite green, unverified, could not run, or refuted. A fix PR carries eligible proven or suite-green patches; other findings remain in the review summary and console.

Fig. 01a review

From a finished session to a fix pull request

Every finding reaches you labelled with what the sandbox established: proven, suite green, unverified, could not run or refuted.

  1. your sessionCaptured as it endsWhat was asked, read, changed and claimed. No extra tokens.
  2. the routerSeats other labsNever the lab that did the work, picked per task from what each seat has proven on your code.
  3. the reviewersRe-do it blindHunting bugs, security holes, dead code and slower-than-needed code.
  4. the sandboxTries to verifyIn a throwaway copy of your repo: a test that fails before the fix and passes after.
  5. youOne fix PRProven and suite-green patches. Every other finding stays in the summary, labelled.

What “proven” means

A bug is proven only if the reviewer writes a test that fails on your current code and passes with its fix. The suite then runs with the fix and its result is reported beside the proof (green, not run, or regressed); a patch that regresses the suite is never offered as a one-click suggestion. A faster rewrite counts only if the sandbox measures it faster and the suite still passes. Dead code counts only if removing it leaves the suite green. A claim without a successful reproduction is not labelled proven.

What you get

One fix pull request per review, one commit per admitted mechanism, with the patch and its verification evidence. While the source PR is open in the same repository, fixes target its branch. Eligible follow-ups for merged source PRs or forks target the base branch; a source PR closed without merging receives no follow-up fix PR. GitHub checks annotate proven mechanisms. Other findings remain in the review summary and console, labelled by evidence. A memory: a lesson from every proven finding, injected into the next session that touches that file, so the same defect is not written twice. And a scoreboard: which model is right on which kind of task, on your code, with the cost attached.

Source: packages/github/src/index.ts · packages/deliver/cards/policy.json · packages/lens/cards/delivery.json

Why other labs

A model is bad at catching the kinds of bugs it tends to write. Reviewers are always from labs other than the one that did the work, picked per task by the router from what each seat has proven on your code and what it cost per call once measured. The hosted door runs unrestricted: no spend cap and no wall clock on a review; each seat has an inactivity clock only.