parallax-review
Skill2 starsChanged 9 days agoApache-2.0View on GitHub
Run a Parallax review on a PR β two independent model lenses (π Claude correctness + 𧨠Codex adversarial) posted as two branded GitHub reviews. Use when the user says "parallax review <url>", "parallax-review", "run the two-lens review on this PR", or points at a PR (usually authored by someone else) and asks for the deeper cross-model review. For PRs our own agents just created, the orchestrator already auto-dispatches parallax β this skill is the manual entry point, mainly for external PRs.
In this repository16 other files it gives its agents.
- AGENTS.mdAGENTS.md
- adopt-evalSkill
- contract-validationSkill
- create-prSkill
- designSkill
- graphifySkill
- helm-upgradeSkill
- ingest-readingSkill
- investigate-alertSkill
- k8s-debugSkill
- shiny-engineerSkill
- stakeholder-commsSkill
- stop-slop-codeSkill
- systematic-debuggingSkill
- upstream-triageSkill
- verify-claimsSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
β0
Common Β· estimate
Many files like this have a rule of this kind.
β0
Distinctive Β· estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 61
Read-only by construction: the adversarial lens works in an isolated worktree at the merge-base and never mutates.
β βCommonβ is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. βVery specificβ is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Parallax Review
Dispatches the pr-reviewer sub-agent in
Parallax mode: it runs its own π correctness lens (Claude) and dispatches an
independent 𧨠adversarial lens (Codex, via
agent-knowledge/scripts/codex-dispatch.sh)
that tries to break head against the PR's claims in a read-only worktree at the
merge-base, then posts two separately-branded GitHub reviews.
This skill is a thin orchestrator β it only picks the action mode and tier and
dispatches. The full procedure lives in
core/agents/pr-reviewer.md (Operating modes,
Step 3.5, Step 7.5) and the dispatch matrix in
core/protocols/pr-review-loop.md.
When to use
- The user asks for a "parallax review" / "two-lens review" / "cross-model review" of a PR.
- A PR (usually authored by someone else) needs the deeper adversarial pass, not just a single-lens read.
- Our own agents' PRs are already auto-dispatched to parallax by the orchestrator β reach for this skill mainly for external PRs, or to re-run parallax on our PR explicitly.
What to do
- Capture the input: a PR URL (or
owner/repo#N). If none was given, ask for one. - Decide the action mode:
- External PR (someone else's branch) β
Mode: review-only(comment-only, never push). This is the common case for this skill. - Our own PR β
Mode: fix(the reviewer may push fixes). Usually the orchestrator already auto-dispatched parallax on our PRs, so only usefixhere if explicitly re-running.
- External PR (someone else's branch) β
- Pick the tier (the reviewer self-triages too):
sensitivefor RBAC/secrets/NetworkPolicy/production-cluster/CRD/auth-ingress-filter config; elsestandard. For asensitivePR, request a stronger model for the correctness lens at dispatch when the runtime supports it. - Dispatch
pr-reviewerwith the dispatch-prompt template fromcore/protocols/pr-review-loop.md, setting:Style: parallaxMode: review-only(external) orfix(ours)Tier: <standard | sensitive>Output: briefβ only when the user asked for a short comment, an agent prompt for the author, or said "brief". The reviewer then posts nothing, writes the draft to$PXD/brief.mdand returns it; you show it to the user and post only after they approve. Omitting this switch on a request for a draft turns it into two published reviews on someone else's PR.- the PR URL, a one-line context, and any constraints the user gave.
- Relay the outcome: confirm both lenses posted (π correctness + 𧨠adversarial), surface the two verdicts, and flag any disagreement between the lenses for the human to adjudicate β never reconcile it silently.
Notes
- The two lenses are independent β the Codex prompt never receives the Claude findings, and the correctness pass runs before its output is read. Divergence is signal; both are posted even when they agree.
- That independence is gated, not assumed: the adversarial prompt is checked for copy-paste of the correctness lens's findings before dispatch, with a planted-finding control proving the check can fail (pr-reviewer Step 3.5, Isolation gate). A lens that quotes a finding back is contaminated and its verdict is discarded.
- Read-only by construction: the adversarial lens works in an isolated worktree at the merge-base and never mutates. In
review-onlymode nothing is pushed at all. - Cost: two model runs per PR. Fine for sensitive/standard; for a trivial PR prefer a single-lens review.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

