Skip to main content
Pullfrog carries useful context forward between agent runs so each run does not have to rediscover the same repo and PR facts from scratch. There are three kinds of learnings:
  • Repo-level learnings apply to every run in a repository.
  • PR-level learnings apply only to a specific pull request while it is being reviewed or updated.
  • Cross-repo learnings apply to cross-repo (--xrepo) runs across an organization.

​
Repo-level learnings

Repo-level learnings are durable notes about how a repository works. They are meant for facts that future runs can reuse across many tasks. Good repo-level learnings include:
  • Setup and verification commands, like the right package manager, lint command, or targeted test command.
  • Project conventions, like where shared utilities live or how components are usually styled.
  • Repo-specific gotchas, like a flaky integration test, a generated file that should not be edited, or a deployment detail that changes how code should be validated.
  • Stable architectural facts, like which package owns shared runtime logic or where auth decisions are enforced.
Repo learnings should be high-confidence and broadly useful. Pullfrog avoids turning them into a full run log: one-off discoveries, speculative guesses, and stale notes should be skipped or pruned. You can view repo-level learnings in the repo console under Learnings.

​
PR-level learnings

PR-level learnings are scoped to one pull request. Internally, Pullfrog stores these as a rolling PR summary snapshot that helps review and re-review runs understand what has already happened on that PR. Good PR-level learnings include:
  • What the PR is trying to change.
  • Which subsystems, files, or behaviors are materially affected.
  • Important risks, unresolved questions, or contracts the next review run should preserve.
  • Which parts of the PR have already been reviewed, and what changed since the last review.
  • Prior feedback that has been addressed or is no longer relevant.
PR-level learnings are not a public changelog and not a replacement for the review body. The visible GitHub review still gets its own human-readable summary. The PR-level snapshot is mainly agent context for the next run.

​
Cross-repo learnings

Cross-repo learnings are org-level structural knowledge for cross-repo (--xrepo) runs — the map of how your repositories relate, so a cross-repo run doesn’t have to rediscover the topology every time. There are two org-level inputs:
  • Xrepo brief is operator-authored prose. It is authoritative and never edited by the agent. Use it to state relationships directly: “api consumes types from shared”, “web and mobile both depend on design-system”, “run pnpm -r build from platform before touching downstream repos”.
  • Xrepo learnings are agent-curated, like repo-level learnings but org-scoped. When you don’t spell out the structure in the brief, the agent learns it once — from READMEs, dependency manifests, and imports — and writes it down so later cross-repo runs inherit it instead of paying the discovery cost again.
Good cross-repo learnings stay structural and org-level: dependency direction, where shared code and types live, and per-repo build and test entrypoints. Single-repo facts belong in that repo’s own learnings, not here. You can view and edit the xrepo brief and xrepo learnings in the organization console.