Search returns 0 results for Claude Code checkpoints despite successful capture (CLI 0.6.1 + 0.6.2-nightly)

Summary

In a real production monorepo with Entire installed via Path B (Entire wrapping Lefthook), capture works end-to-end for Claude Code sessions — trailers are attached, checkpoints land in entire/checkpoints/v1, the branch pushes to a separate private checkpoint remote, entire activity / entire dispatch / entire checkpoint explain all see the data — but entire search and entire checkpoint search return 0 results for every query, including queries that match exact content from the captured transcripts.

This is the Claude Code variant of #1171 (Codex variant, same author, brand-new repo repro). Filing separately because:

• Different agent (Claude Code, not Codex)
• Different repo profile (real multi-package monorepo, not a clean canary)
• Same end-state regardless: search is empty while every other surface sees the data

Environment

• Entire CLI: pre-flight done on 0.6.2-nightly.202605090651.30e58ea3, production rollout on stable 0.6.1
• macOS arm64
• Agent: Claude Code (claude.ai / Claude Code CLI)
• Lefthook 1.13.6 (project) / 2.1.6 (system PATH)
• Hook architecture: Path B — Entire wraps Lefthook via entire configure --force in prepare script in package.json
• Separate private checkpoint remote: agiterra/tankloop-checkpoints-private
• GitHub App: Entire installed on agiterra org with repository_selection=all

What works

• entire status — Enabled, Claude Code + Codex agents active
• Capture: every commit gets an Entire-Checkpoint: <id> trailer
• Push: entire/checkpoints/v1 syncs to the checkpoint remote on every git push
• entire checkpoint list — shows local checkpoints (2 on main)
• entire checkpoint explain <id> — shows the full transcript, prompts, file list, tokens
• entire activity — shows 11 checkpoints across 3 repos, including all of today's agiterra/tankloop work, with proper agent attribution (Claude Code 64% / Codex 36%)
• entire dispatch — generates digests citing real repo activity
• PII redaction working in captured transcripts (verified: real emails replaced with REDACTED_EMAIL)

What fails

All search variants return empty:

$ entire search "lefthook" --json
{ "results": [], "total": 0, "page": 1, "total_pages": 1, "limit": 25 }

$ entire search "redaction" --json
{ "results": [], "total": 0, "page": 1, "total_pages": 1, "limit": 25 }

$ entire search "Path B" --json
{ "results": [], "total": 0, "page": 1, "total_pages": 1, "limit": 25 }

$ entire checkpoint search "lefthook" --json
{ "results": [], "total": 0, "page": 1, "total_pages": 1, "limit": 25 }

$ entire checkpoint search "" --branch main --json
{ "results": [], "total": 0, "page": 1, "total_pages": 1, "limit": 25 }

These tokens are all in the captured transcripts and commit messages — verified via entire checkpoint explain <id> which prints them inline.

Sample checkpoints (real)

From entire checkpoint list on agiterra/tankloop main:

• 632b58b88c75 — commit 9f71259 ("chore(agents): notify on pull when entire CLI is missing")
• b5bbb47b6e39 — commit 5465328 ("feat(agents): adopt entire for session provenance (#25)")

Both contain extensive Claude Code transcripts; neither is discoverable via search.

Expected behavior

Once a checkpoint is visible in entire activity and entire checkpoint explain returns the transcript, an exact-token search from the captured prompts or transcript should be discoverable via entire search / entire checkpoint search after indexing completes.

Actual behavior

Search service returns empty for every query, no matter how recent or how stale the checkpoint, no matter which --repo, --branch, or --author filter is applied.

Impact

search is a load-bearing skill for cross-agent context recall. Without it, what-happened, search, and the Entire skills wrapping these commands are effectively non-functional for our team. The marketing value of "find prior work before changing X" disappears.

Related

• entireio/cli#1171 — Codex variant, clean repro, also from us (agiterra / WrangleMeThis). Same symptom, different agent.

🤖 Peace , B.Sweet

Reproduce with the documented entire search and entire checkpoint search commands against the sample checkpoints, then trace the search and indexing entry points. Compare their behavior with entire checkpoint explain and entire activity, using exact transcript tokens such as lefthook and redaction. Done means indexed checkpoint content is returned by both search commands with the relevant filters.