cli: add code search with multi-region fan-out (ENT-993) by evisdren · Pull Request #1616 · entireio/cli

evisdren · 2026-07-02T20:43:34Z

Summary

Adds cmd/entire/cli/codesearch/ — client package for peregrine's /search/api/search endpoint
Wires --code and --case-sensitive flags into entire search, gated behind ENTIRE_CODE_SEARCH=1
Multi-region fan-out matching the BFF pattern: list repos from entire-core → group by cell/jurisdiction → parallel search with per-cell timeouts → merge results
Single-cell path when --repo filter is specified (routes to repo's owning cell)
One bad cell never sinks the search — errors are logged and skipped

How it works

Scenario	Auth path	Behavior
`--code query` (no filter)	`ListRepos` → group by cell → `NewEntireAPICellClient` per jurisdiction	Fan-out across all cells
`--code --repo owner/name query`	`NewAuthenticatedEntireAPICellClient`	Single cell (repo's owner)
`--code repo:owner/name query`	Same (inline filter parsed)	Single cell
`--code --all-repos query`	Fan-out (same as no filter)	Fan-out across all cells

Files

cmd/entire/cli/codesearch/codesearch.go — client types and Search() function
cmd/entire/cli/codesearch/codesearch_test.go — client tests (success, API error, non-JSON error, response too large)
cmd/entire/cli/search_cmd.go — command wiring, fan-out orchestration, output formatters
cmd/entire/cli/search_cmd_test.go — command tests (gate, flags, output, groupReposByCell, mergeSearchResults)

Test plan

go test ./cmd/entire/cli/codesearch/... passes
go test ./cmd/entire/cli/ -run "TestSearchCmd|TestWriteCodeSearch|TestCodeSearch|TestGroupRepos|TestMergeSearch" passes
mise run lint clean
CI passes

🤖 Generated with Claude Code

Copilot

Pull request overview

Adds a new cmd/entire/cli/codesearch package to provide a small, typed Go client for peregrine’s POST /search/api/search endpoint using the existing authenticated api.Client, including response-size capping and basic HTTP error parsing.

Changes:

Introduces codesearch.Search() plus request/response types for peregrine code search.
Implements an 8 MiB response read cap with explicit overflow detection.
Adds unit tests covering success, API error responses, non-JSON error bodies, and oversized responses.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
cmd/entire/cli/codesearch/codesearch.go	New code search client wrapper around `api.Client.Post`, including response cap and error parsing.
cmd/entire/cli/codesearch/codesearch_test.go	Unit tests for the new client (success + error/edge cases).

+		if apiErr.Message == "" && len(body) > 0 {
+			apiErr.Message = string(body)
+		}


Add `entire search --code <query>` to search code content across all jurisdictions via the peregrine search service, gated behind ENTIRE_CODE_SEARCH=1. Architecture: - codesearch package: thin client for peregrine's search endpoint through the cell's entire-api gateway (GET /api/v1/search/api/search) - Multi-region fan-out: list repos from core → group by jurisdiction → search each cell in parallel with per-cell timeouts → merge results by score, deduplicate, cap to --limit - Per-cell repo ID scoping: each cell only receives ULIDs belonging to its jurisdiction, avoiding cross-jurisdiction leakage - Partial failure resilience: a down region warns but returns results from healthy cells; only all-cells-fail is an error Repo filter formats: - gh/owner/repo, et/proj/repo, bare owner/repo, or raw ULID - Defaults to current repo (derived from git origin), matching checkpoint-search behavior; use --all-repos or repo:* for all repos - Multiple filters supported via --repo flag or inline repo: syntax Output: - Text: grep-style (repo:path:line: context), truncated to 200 chars - JSON: full response with stats, repo_stats, and results (--json) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Group fan-out by cell (not jurisdiction), matching BFF pattern - Resolve cell apiUrl from cluster catalog for direct routing - Align resolveRepoFilters with BFF: only strip gh/, try unstripped filter as full_name fallback, match id on raw filter - Use sort.SliceStable with tiebreaker for deterministic output - Route single-cell path through mergeSearchResults for uniform sort/dedup/truncate - TrimSpace error body fallback in codesearch client - Strengthen TestSearch_NonJSONError assertions - Add searchAllCells test seams and 5 integration tests covering cell grouping, cluster URL resolution, partial failure, filter resolution, and single-cell merge - Document case_sensitive as optimistic (peregrine proto lacks it) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

The cluster catalog is keyed by Cluster.Slug, which corresponds to RepoIndexEntry.ClusterSlug — not RepoIndexEntry.Cell. Using Cell for the lookup would silently miss baseURL resolution when Cell differs from ClusterSlug, falling back to the home-cell URL. - Add clusterSlug field to cellGroup, populated from r.ClusterSlug - Use cells[i].clusterSlug for slugToCluster lookup - Test ResolvesBaseURLFromClusters now uses distinct Cell/ClusterSlug values to prove the lookup path - All test RepoIndexEntry literals updated with ClusterSlug field Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

The BFF uses r.cell with a URL template to resolve cell URLs. No other CLI code path does a Cluster.Slug lookup — the experts resolver matches by ClusterHost instead. Revert to using cell for the cluster catalog lookup to mirror the BFF pattern. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

resolveExpertsCellTarget was experts-named but nothing in it is experts-specific: it maps any repo (ULID or owner/repo) to the entire-api cell hosting it, exactly what every future repo-scoped or fan-out command needs (code search is next, PR #1616). Rename the file and identifiers — resolveRepoCellTarget, cellCoreClient, newCellCoreClient, cellResolveTimeout — and keep experts as the documented example consumer. Mechanical rename; no behavior change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Entire-Checkpoint: 89b4d13bd2b0

The data plane has no server-side cross-cell aggregator: a query over all of the caller's repos must be fanned out to each cell hosting them and merged client-side (the entire.io BFF's code-search pattern). PR #1616 inlines that orchestration into search_cmd.go; this extracts the generic layer so code search — and any later repo-set command — shares one implementation: - groupReposByCell: repo index → per-cell groups, deterministic order. - resolveCellBaseURLs: cluster-catalog join on ClusterSlug↔Cluster.Slug. The catalog exposes no cell field, so joining the cell name against Slug (as #1616 currently does) only works when the two coincide. - fanOutCells: parallel per-cell calls under a per-cell timeout, one shared auth.CellClientFactory (subject resolved once, one identity token per jurisdiction), partial failures isolated per result slot. Merge semantics stay with the command. Also documents the three cell-routing shapes in CLAUDE.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Entire-Checkpoint: 842e421e74a8

resolveExpertsCellTarget was experts-named but nothing in it is experts-specific: it maps any repo (ULID or owner/repo) to the entire-api cell hosting it, exactly what every future repo-scoped or fan-out command needs (code search is next, PR #1616). Rename the file and identifiers — resolveRepoCellTarget, cellCoreClient, newCellCoreClient, cellResolveTimeout — and keep experts as the documented example consumer. Mechanical rename; no behavior change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Entire-Checkpoint: 89b4d13bd2b0

The data plane has no server-side cross-cell aggregator: a query over all of the caller's repos must be fanned out to each cell hosting them and merged client-side (the entire.io BFF's code-search pattern). PR #1616 inlines that orchestration into search_cmd.go; this extracts the generic layer so code search — and any later repo-set command — shares one implementation: - groupReposByCell: repo index → per-cell groups, deterministic order. - resolveCellBaseURLs: cluster-catalog join on ClusterSlug↔Cluster.Slug. The catalog exposes no cell field, so joining the cell name against Slug (as #1616 currently does) only works when the two coincide. - fanOutCells: parallel per-cell calls under a per-cell timeout, one shared auth.CellClientFactory (subject resolved once, one identity token per jurisdiction), partial failures isolated per result slot. Merge semantics stay with the command. Also documents the three cell-routing shapes in CLAUDE.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Entire-Checkpoint: 842e421e74a8

resolveExpertsCellTarget was experts-named but nothing in it is experts-specific: it maps any repo (ULID or owner/repo) to the entire-api cell hosting it, exactly what every future repo-scoped or fan-out command needs (code search is next, PR #1616). Rename the file and identifiers — resolveRepoCellTarget, cellCoreClient, newCellCoreClient, cellResolveTimeout — and keep experts as the documented example consumer. Mechanical rename; no behavior change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Entire-Checkpoint: 89b4d13bd2b0

The data plane has no server-side cross-cell aggregator: a query over all of the caller's repos must be fanned out to each cell hosting them and merged client-side (the entire.io BFF's code-search pattern). PR #1616 inlines that orchestration into search_cmd.go; this extracts the generic layer so code search — and any later repo-set command — shares one implementation: - groupReposByCell: repo index → per-cell groups, deterministic order. - resolveCellBaseURLs: cluster-catalog join on ClusterSlug↔Cluster.Slug. The catalog exposes no cell field, so joining the cell name against Slug (as #1616 currently does) only works when the two coincide. - fanOutCells: parallel per-cell calls under a per-cell timeout, one shared auth.CellClientFactory (subject resolved once, one identity token per jurisdiction), partial failures isolated per result slot. Merge semantics stay with the command. Also documents the three cell-routing shapes in CLAUDE.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Entire-Checkpoint: 842e421e74a8

Soph · 2026-07-06T11:02:30Z

Heads-up: the generic multi-cell plumbing this PR inlines in search_cmd.go now exists as a shared layer in #1641 (green, awaiting review), extracted so code search and future repo-set commands share one implementation:

groupReposByCell / resolveCellBaseURLs (cell_fanout.go) — repo index → per-cell groups, then a cluster-catalog join keyed on ClusterSlug↔Cluster.Slug. On the join-key back-and-forth in this PR's history: the BFF's r.cell usage works because the BFF templates the URL from the cell name (https://<cell>.api…); this PR instead looks up the catalog, and the catalog exposes no cell field — its key is Slug. So slugToCluster[cells[i].cell] only matches while slug and cell name coincide, and silently degrades to jurisdiction routing (plus a hand-rolled catalog re-fetch inside auth per cell) when they don't. The earlier fix: use ClusterSlug commit had it right; cli: cell-routing foundation — client factory, generic repo→cell resolver, multi-cell fan-out #1641 pins that with a slug≠cell test.
fanOutCells — parallel per-cell dispatch, per-cell timeout, partial failures isolated per result slot. Builds every client from one auth.CellClientFactory, which resolves the login subject once and mints one identity token per jurisdiction (tokens are per-jurisdiction, not per-cell). Calling auth.NewEntireAPICellClient inside the per-cell loop re-runs discovery + login refresh + RFC 8693 exchange once per cell.
resolveRepoCellTarget (cell_target.go, generalized from the experts resolver) — the single-cell path for the --repo case.

Suggested shape for a rebase onto #1641: the codesearch client package, merge/sort/dedup semantics, and text/JSON rendering carry over as-is; the inline searchAllCells/searchCell/groupReposByCell orchestration collapses onto the shared helpers. The three routing shapes are documented in CLAUDE.md under "Entire-API Cell Routing". Happy to help with the rebase once #1641 lands.

Copilot AI review requested due to automatic review settings July 2, 2026 20:43

evisdren requested a review from a team as a code owner July 2, 2026 20:43

This was referenced Jul 2, 2026

cli: wire code search into entire search --code (ENT-994) #1617

Closed

cli: add code search via peregrine (ENT-993, ENT-994) #1615

Closed

Copilot started reviewing on behalf of evisdren July 2, 2026 20:43 View session

Copilot AI reviewed Jul 2, 2026

View reviewed changes

Comment thread cmd/entire/cli/codesearch/codesearch.go

Comment on lines +84 to +86

if apiErr.Message == "" && len(body) > 0 {

apiErr.Message = string(body)

}

Comment thread cmd/entire/cli/codesearch/codesearch_test.go

evisdren changed the title ~~cli: add codesearch client package for peregrine (ENT-993)~~ cli: add code search with multi-region fan-out (ENT-993) Jul 2, 2026

evisdren force-pushed the evis/ent-993-code-search-client branch 7 times, most recently from 1e7a2b3 to 0ef9bc1 Compare July 3, 2026 05:17

evisdren and others added 2 commits July 3, 2026 17:43

evisdren force-pushed the evis/ent-993-code-search-client branch from 1fdef33 to 8ec2ed5 Compare July 4, 2026 00:59

evisdren and others added 2 commits July 3, 2026 20:35

Soph mentioned this pull request Jul 5, 2026

cli: cell-routing foundation — client factory, generic repo→cell resolver, multi-cell fan-out #1641

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

cli: add code search with multi-region fan-out (ENT-993)#1616

cli: add code search with multi-region fan-out (ENT-993)#1616
evisdren wants to merge 4 commits into
mainfrom
evis/ent-993-code-search-client

evisdren commented Jul 2, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Soph commented Jul 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

evisdren commented Jul 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

How it works

Files

Test plan

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Soph commented Jul 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

4 participants

evisdren commented Jul 2, 2026 •

edited

Loading