Prefer compatible package binaries by default - #156
Conversation
Keep source candidates in pak's dependency solve while preferring compatible repository binaries, including older releases. Preserve selected artifacts through renv and re-solve after missing binary downloads. Add IR_PREFER_BINARIES=0, version policy and artifact cache identities, and cover release lag, constraints, fallback, and cache transitions with local repository CLI fixtures. Document the default and opt-out.
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Resolve ordinary requests through pak's lazy dependency policy and restore selected artifact URLs through renv's lockfile interface. Re-solve once with source requests for missing binaries, preserving constraints and separate policy cache identities.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d6d9e1c794
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Allow the documented driver entry point without private policy flags and report conflicting flags explicitly. Exercise its binary default and version-freshness mode against the local repository fixture. Verify that ambient user-library installations cannot supply the isolated environment's artifact. pak already resolves against an empty library; document that boundary without changing its public API usage.
The direct-driver regression is cross-platform, but stderr() was gated to macOS. Remove that restriction so the Linux and Windows test targets compile with the same diagnostic assertion.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 56e87090c5
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Read a fresh inventory before verifying restored packages and versions. Qualify namespace calls so the resolver works with reduced default packages. Cover stale inventories and reduced-default binary installation and source fallback through the public CLI.
Place renv's mutable archive downloads under the existing resolver lock so concurrent resolutions with separate IR caches cannot overwrite each other. Keep explicit download-cache overrides and the shared installed-package cache.
Independent IR caches have separate resolver locks. Sharing the same renv artifact entry lets one restore temporarily remove a package still linked by another. Include the normalized IR cache root in artifact and library identities, and update the resolution cache format. Reproduce cache replacement with local packages and socket checkpoints, and cover the cache transition through the CLI.
Drop inherited R_LIBS so ambient packages cannot override the fixture's fake installer through R_LIBS_USER.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f96dbcd3e0
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Use the temporary directory name to keep the synthetic root distinct from repository packages. Cover direct and transitive requests for irrequirements through the public CLI.
Use a nextest group for real Quarto renders so concurrent tests cannot contend for the shared Sass database. Keep other tests parallel.
1 participant
A new CRAN release can be available as source for days, sometimes up to a week, before its pre-built binaries are ready. During that gap, IR currently picks the newest version. I recently ran into this with DuckDB: IR started a slow source build even though an older compatible binary was available.
This PR makes IR prefer compatible binaries, including older versions, for ordinary package requests and their dependencies. Source-only packages can still install alongside binaries.
Set
IR_PREFER_BINARIES=0to keep prioritizing newer versions. This does not force source installation. Exact pins, minimum versions, and explicit Git, local, URL, or source requests still control what IR installs. MCP Console gets the new default automatically, without configuration changes.IR still uses pak to resolve dependencies and renv to install them:
DESCRIPTIONfile with a unique package name so pak applies its binary preference to requested packages too. This uses pak's public API and adds no pkgdepends dependency or vendored solver code.renv::restore(), preserving the chosen binary through installation.1share cached resolutions;0is separate. Old plans and incomplete installations are bypassed.Downloaded archives stay under IR's resolver lock. renv's installed-package cache remains shared, with artifact keys scoped to the IR cache root so concurrent runs using separate roots cannot replace each other's entries.
Local repository tests reproduce the release-lag case: a 1.0.0 binary, a 2.0.0 source release, and a source-only dependency. They check the installed artifact, constraints, fallback, and cache reuse. A synchronized regression also replaces a cached binary during installation and checks that IR still produces a complete library. Documentation includes opt-out examples and a release note about the default and cache transition.
With renv 1.3.0 in an isolated library,
NEXTEST_TEST_THREADS=2 scripts/check.shpassed on macOS: formatting, Clippy, and all 222 tests, with none skipped. Quarto render tests share a nextest group to avoid concurrent writes to Quarto’s Sass cache.