Parallax
Using it
What arrives, what it costs, and how to read it.
When a session ends
Five quiet minutes (or a Stop hook, or an explicit end) and the door sweeps the session, one at a time: it proves the agent’s own edits in a sandbox, looks up the other sessions on the same repository that touched the same files and re-runs their proven tests, then asks the router-picked reviewers to fork the implementation turn, hunt with each lens, and prove what they claim. Free steps always run; paid steps run in value order under a budget of twice what the session itself cost.
curl -s -H "authorization: Bearer dev" http://127.0.0.1:8787/v1/sweepsThe lenses
| lens | asks | proof |
|---|---|---|
bugs | what did the test not check? | a new test fails without the fix, passes with it, suite green |
security | what could an attacker use? | same three legs, with a severity; a mechanical pass first, free |
perf | a faster, simpler or smaller way? | benchmark 3x each side, at least 1.1x and above the noise; or fewer branches; suite green |
hygiene | what was left behind? | suite green with the cleanup; may only remove or reword |
dead code | what does nothing reference? | each removal proven alone and together; no model call |
On a pull request
One check, Parallax: failure if a bug or a high-severity security finding was proven, neutral if only cleanup, success otherwise. Proven findings are line annotations with the proof steps. The summary names the reviewers, the router version that picked them, and the settled cost. A section titled “Said, not proven” lists what a reviewer claimed and the sandbox refuted, so nothing is hidden.
Merge or close the fix PRs Parallax opens. It reads which, and weighs that reviewer accordingly; a merge counts double.
Memory
A proven finding with a lesson in it (“in this file, guard an empty input before looping”) is kept per repository and given to the next session that touches that file, through the proxy or the MCP parallax_recall tool. Lessons expire in 30 days unless re-proven; list and delete them at GET /v1/recall.
What it costs
Every reviewer call carries a settled receipt. A typical sweep of a small session is $0.10 to $0.60; the budget caps it at twice the session’s own cost with a $0.25 floor. Reviewers are bounded to the Pareto frontier (nothing cheaper ranked above it), asked to reason briefly (reasoning tokens are most of the bill), and one sandbox run serves every copy of the same finding.