[quickbeard]

Summary

When rg is neither on PATH nor in ~/.cache/codev/bin, CoDev Code downloads it from GitHub releases on first use — which corporate networks commonly block. PR #146 made that failure visible and retryable, and codevhub#230 stages the binary at install time, but users who never ran codevhub (fresh machine, cleared cache) were still stuck. This adds a second source inside the agent itself: GitHub first, then the CoDev landing page's mirror of the official 15.1.0 executables (quickbeard/codev-landing-page#64, rg-<version>-<platform>-<arch>[.exe]).

Changes

• packages/core/src/ripgrep/binary.ts
  • The download tail is split into fromGithub (unchanged archive → extract flow) and fromMirror (raw executable, no extraction — staged in a scoped temp dir, then copied to the target + chmod; the cache check is bare file existence, so a partial target must never appear).
  • Fallback engages via Effect.catchCause, not Effect.catch: a corporate proxy block page returns 200 + HTML, which passes filterStatusOk and then dies inside extract as a defect — that case must reach the mirror too.
  • Each attempt is bounded to 30s (Effect.timeout) so a blackholing proxy can't hang first-time file search on OS TCP timeouts before the fallback is even tried.
  • Empty response bodies are now a typed failure (Effect.fail) rather than a defect, so they trigger the fallback as well.
• packages/core/test/ripgrep-binary.test.ts (new) — first coverage of binary resolution at all, using the mocked-HTTP-client pattern from models.test.ts (AppNodeBuilder + Layer.fresh): fallback order (GitHub URL then exact mirror URL — pinning the hosted-filename contract), bytes land at the target with the exec bit and no leftover staging dir, success is cached (no further requests), failure is not cached (recovers mid-session), and an empty mirror body never creates the target.
• packages/core/test/preload.ts — redirects XDG_CACHE_HOME to a fresh temp dir for the whole suite, so tests that exercise real binary resolution can never read or clobber the developer's actual ~/.cache/codev (the new tests write fake bytes at the rg target path; models.test.ts's cache writes also become hermetic).
• Docs — extended the existing PR-146 DIVERGENCES row (mirror behavior, new files, and the three-way version-pin invariant with codev-landing-page and codevhub) and added a CHANGELOG entry under Unreleased.

Cross-repo invariant

VERSION here, RG_VERSION in codevhub, and the landing page's hosted filenames must move together on a ripgrep bump — host new binaries first. Until codev-landing-page#64 deploys, the mirror GET 404s and behavior degrades to exactly today's GitHub-only failure.

Test plan

• bun typecheck in packages/core and packages/opencode
• Full core suite from packages/core: 1083 pass, 0 fail (caught and fixed a leak where the fake test binary outlived the file and broke the live ripgrep suites)
• New tests pass in isolation and alongside test/ripgrep.test.ts + test/filesystem/
• Manual on a GitHub-blocked network once the landing-page PR deploys: first @ search downloads via the mirror (log shows "github ripgrep download failed, trying the CoDev mirror")

🤖 Generated with Claude Code