Repository navigation
Cross-model review: agent PRs reviewed by a different provider than the author #259
Description
Inspired by Greptile's Model Inversion (models catch more bugs in other models' code) and their v5 swarm review (parallel single-hypothesis reviewers, 52% -> 66% addressed-comment rate).
Add an optional review stage to the PR flow: a thread's PR gets reviewed by a different provider than the one that wrote it (claude authors, codex reviews), optionally fanned out as parallel single-hypothesis review passes using the existing best-of-N infrastructure. Review findings route back through #161.
Activity
currentbits commentedon Aug 17, 2026
Research note (Cognition's 'Multi-Agents: What's Actually Working', Apr 2026, measured on Devin Review): the reviewer should share NO context with the coder. Give the reviewing agent the diff only; context rot makes the coder blind to its own bugs, and a reviewer anchored on the same spec misses the same things. Second finding: the coder must be the filter on review findings to prevent fix-loops. So: clean-context review thread type (fork zero context), findings routed back to the author thread for filtering.
currentbits commentedon Aug 18, 2026
Research note (Anthropic multi-agent patterns): the epistemics experiment validates cross-checking — newer models detected lies only because scout reports OVERLAPPED, letting contradictions surface. Cross-provider review works the same way: overlap is what makes contradictions visible.
currentbits commentedon Aug 19, 2026
The steal from Claude Code Desktop's self-review is the rubric, not the button.
In the diff view there is a Review code action, and the docs state its scope explicitly (docs):
That negative list is the whole feature. An unconstrained "review this diff" prompt returns twelve nits and one real bug, and the user learns to skip the panel — the failure mode that kills every review surface. Whatever provider Solenta routes the cross-model review to, the prompt needs those four exclusions in it verbatim, and the results need to land as comments in the diff rather than as a prose turn (which is how #162's comment thread becomes bidirectional: agent comments and user comments in the same gutter).
Codex frames the same idea as severity tiers instead of exclusions — it surfaces P0/P1 by default with the threshold configurable via review guidelines in
AGENTS.md(26.415 writeup). Repo-configurable review scope pairs well with #378 (repo-versioned agent definitions).