- Call getMultiRepoJSONLPaths() once at sync start instead of 3 times
- Eliminates redundant config lookups and path constructions
- Performance optimization for multi-repo workflows
Optimization to avoid JSON unmarshalling when strings match exactly (common case).
Simple 1-line change instead of complex streaming rewrite for P3 issue.
- Add snapshotMetadata struct with version, timestamp, commit SHA
- Validate snapshots are <1 hour old, from compatible version
- Auto-cleanup stale snapshots from interrupted syncs
- All snapshot functions now write/validate metadata files
- Fixes issue where crash/kill -9 left stale snapshots
Removed old global daemon infrastructure that was replaced by per-workspace
daemon architecture. Package had no imports and all functions were unreachable.
Closes bd-irq6
- Added getMultiRepoJSONLPaths() helper to enumerate all JSONL files
- Updated daemon_sync.go to capture/update snapshots for each repo
- Added TestMultiRepoDeletionTracking to verify fix
- Added TestMultiRepoSnapshotIsolation for isolation guarantees
Previously, deletion tracking only worked for single-repo mode because
snapshot operations were hardcoded to the primary JSONL path. This fix
ensures snapshots are managed per-repository in multi-repo mode.
The previous implementation required hash IDs to contain letters a-f,
but SHA256 hashes can be all digits (probability ~2.4%). This caused
TestMigrateHashIDs to fail when the generated hash ID was all numeric.
Updated isHashID to:
- Strip hierarchical suffixes (.1, .1.2) before checking
- Accept any valid hex string (0-9, a-f)
- Distinguish hash IDs by presence of letters a-f
Fixes bd-6ku3
Root cause: Import and export commands tried to open the database directly
while the daemon already held the lock, causing indefinite blocking.
Solution: Both commands now explicitly close the daemon connection before
opening direct database access, avoiding SQLite lock contention.
Changes:
- import.go: Close daemon connection and open direct SQLite connection
- export.go: Close daemon connection before direct access
- Added debug logging to help diagnose similar issues
Tests: Existing TestImport and TestExport tests pass
Oracle identified a critical race condition in the initial fix:
- Pre-push hook checked for changes but didn't flush first
- Pending 5s-debounced flushes could land after the check
- Result: stale JSONL could still be pushed
Improvements:
1. Pre-push now flushes pending changes FIRST (bd sync --flush-only)
2. Uses git status --porcelain to catch ALL change types:
- Staged, unstaged, untracked, deleted, renamed, conflicts
3. Handles both beads.jsonl and issues.jsonl (backward compat)
4. Works even without bd installed (git-only check)
5. Pre-commit stages both JSONL files (simpler loop)
This completely eliminates the race condition.
The original pre-push hook tried to export DB → JSONL during the push,
then run 'git add', but this doesn't work because:
1. The commit is already created when pre-push runs
2. git add in pre-push stages files for a FUTURE commit
3. The current push sends the old commit with stale JSONL
4. Result: dirty git status after push
Fix:
- Pre-push now CHECKS for uncommitted JSONL changes
- If found, it FAILS the push with clear instructions
- User must commit JSONL before pushing
- This prevents stale JSONL from reaching remote
The pre-commit hook already properly flushes changes, so this
catch-all prevents changes made BETWEEN commit and push.
Amp-Thread-ID: https://ampcode.com/threads/T-39a89553-c301-4d4f-b39f-6df9c403d22b
Co-authored-by: Amp <amp@ampcode.com>
- Changed all test commands to 'go test -short ./...'
- Clarified that full tests run in CI
- Updated: Code Standards, Before Committing, Landing the Plane, Building and Testing, Release Process
- Agents should only run short tests; CI runs full suite
