(PR#77 Feature Add) Claude Code Desktop Improvements

[austinc3030]

Builds on / supersedes #77 (@sjalife's transcript-conflict + desktop-record sync work). This branch is a strict superset of #77 — its head commit is an ancestor of this branch — plus:

• Cross-platform CCD session-store detection fixes (Linux/macOS paths, MSIX fallback on Windows)
• desktop list / desktop forget CLI commands
• cwd/originCwd path translation across devices for Desktop session records
• WebDAV: reuse pooled connections to avoid a concurrent-handshake TLS race under the sync worker pool
• Desktop pointer sync now skips records whose cliSessionId has no locally-synced transcript (avoids empty sidebar entries), and pull leaves a pointer alone if it was touched in the last 5 minutes (avoids racing a live Desktop write)

Note for reviewers: since #77 isn't merged yet, this diff includes its commits. Once #77 merges into main, this PR should shrink to just the additional commits above. Happy to rebase once that happens.

[Copilot]

🟡 Changes recommended

There are correctness issues in changed code paths (notably stale state timestamps on no-op pushes with CCD uploads, and incomplete ErrNotFound wrapping in WebDAV Head) that should be fixed before approval.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR expands claude-sync’s multi-device correctness by adding safer conflict handling/merge semantics for append-only JSONL files (session transcripts + history.jsonl), plus first-class syncing and management of Claude Code Desktop “session pointer” records, and a WebDAV transport tweak to improve connection reuse under concurrency.

Changes:

• Add JSONL fast-forward/merge logic to avoid spurious .conflict artifacts and prevent truncation clobbers; union-merge history.jsonl by appending missing lines.
• Add cross-platform Claude Code Desktop session-record sync under _ccd-sessions/, plus desktop list / desktop forget CLI commands.
• Introduce a typed storage.ErrNotFound signal and update adapters to wrap provider-specific “not found” errors; increase WebDAV idle connections per host.

File summaries

File	Description
internal/sync/sync.go	Core push/pull changes: debris exclusion, history merge, JSONL conflict resolution, CCD session sync hooks
internal/sync/sync_push_pull_test.go	Retarget conflict tests from history.jsonl to a normal file (settings.json)
internal/sync/state.go	Add HashBytes; prevent CCD session state entries from being treated as deletions in DetectChanges
internal/sync/session_merge.go	Implement append-only merge strategy for diverged session transcripts
internal/sync/session_merge_test.go	Tests for transcript merge behavior (DAG integrity, idempotence, state resolution)
internal/sync/paths.go	Add JSON-safe path token substitution modes for .json/.jsonl content
internal/sync/paths_test.go	Tests for JSON-mode path substitution across OSes and for non-JSON text behavior
internal/sync/history_merge.go	Implement union-merge for history.jsonl and append-only application helper
internal/sync/history_merge_test.go	Extensive tests for history merge correctness and convergence across “machines”
internal/sync/fastforward.go	Add byte-prefix classifier and session JSONL eligibility check
internal/sync/fastforward_test.go	Tests for prefix classification, fast-forward, merge vs conflict behavior, and push shrink guard
internal/sync/debris_test.go	Tests ensuring .lock and tool-generated .conflict.<ts> artifacts never sync
internal/sync/ccdsessions.go	Implement Desktop session record push/pull, store detection, normalization, safety checks
internal/sync/ccdsessions_test.go	Tests for CCD store detection, LWW merge, traversal safety, transcript-linked filtering, freshness guard
internal/sync/ccdlist.go	Add remote listing API for CCD records with display-friendly cwd resolution
internal/sync/ccdlist_test.go	Tests for CCD listing ordering, cwd display resolution, and forget/list interoperability
internal/sync/ccdforget.go	Add “forget” operation to remove CCD records locally and remotely by exact id match
internal/sync/ccdforget_test.go	Tests for forget behavior and non-destructive matching semantics
internal/storage/webdav/webdav.go	Clone transport and increase idle conns per host; wrap not-found for Head
internal/storage/storage.go	Add ErrNotFound + IsNotFound helper
internal/storage/s3/s3.go	Wrap S3 NotFound errors from Head with storage.ErrNotFound
internal/storage/r2/r2.go	Wrap R2 NotFound errors from Head with storage.ErrNotFound
internal/storage/gcs/gcs.go	Wrap GCS not-exist errors from Head with storage.ErrNotFound
cmd/claude-sync/main.go	Add quiet-mode stderr reporting; add desktop CLI command group (list, forget)

Review details

• Files reviewed: 24/24 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🔵 Needs a closer look

It introduces a potential runtime panic in the WebDAV transport setup (unchecked http.DefaultTransport type assertion) and has misleading documentation about “exact-byte” line matching in the new merge code.

Review details

Suppressed comments (3)

Previously missed (3) — in code that hasn't changed since the last review.

internal/sync/history_merge.go:24

• MergeHistoryPayloads claims it dedupes by an "exact-byte multiset match", but the implementation uses forEachLine, which TrimSpaces each line before counting. That means the match is not truly byte-exact, and the comment is misleading about the dedupe semantics.

// Remote lines are dropped as duplicates when either:
//   - an identical raw line exists locally (exact-byte multiset match — this
//     is what dedupes blank submissions and unknown-shape records, which have
//     no prompt signature and would otherwise multiply on every cycle), or

internal/sync/session_merge.go:50

• This comment says uuid-less records use an "exact-raw-line multiset union", but the merge counts lines via forEachLine which TrimSpaces them first. If the whitespace normalization is intentional, the comment should reflect that; if not, the merge should avoid TrimSpace to truly preserve raw bytes.

//   - other uuid-less records (queue-operation, ...): exact-raw-line multiset
//     union, the same rule that makes the history merge idempotent.

internal/storage/webdav/webdav.go:56

• http.DefaultTransport.(*http.Transport) can panic if http.DefaultTransport has been overridden with a non-*http.Transport RoundTripper (common in tests or embedding). Using a checked type assertion avoids introducing a runtime panic in New().

	transport := http.DefaultTransport.(*http.Transport).Clone()
	transport.MaxIdleConnsPerHost = 32

• Files reviewed: 24/24 changed files
• Comments generated: 0 new
• Review effort level: Lite

[Copilot]

🟡 Changes recommended

appendHistoryLines currently reads entire existing files into memory just to check a trailing newline, which will be prohibitively expensive for the very large files this PR aims to support.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

• Files reviewed: 24/24 changed files
• Comments generated: 1
• Review effort level: Lite

[Copilot]

🔵 Needs a closer look

A cross-platform bug in the new exclusion logic can fail to recognize debris basenames on Windows due to slash-normalized relPaths, risking unintended uploads.

Review details

Suppressed comments (2)

Previously missed (2) — in code that hasn't changed since the last review.

internal/sync/sync.go:195

• GetLocalFiles normalizes relPath to forward slashes (filepath.ToSlash). Using filepath.Base here is OS-separator dependent and won’t strip directories on Windows when relPath contains '/', so .lock / .conflict artifacts may be uploaded instead of excluded.
  internal/sync/sync.go:593
• The comment above localForm is inaccurate: localForm is captured before NormalizeContent, so it’s the on-disk bytes used for state hashing, not necessarily the exact bytes uploaded after path-token normalization.

• Files reviewed: 24/24 changed files
• Comments generated: 0 new
• Review effort level: Lite

[Copilot]

🔵 Needs a closer look

The new desktop commands don’t fully honor the global --quiet flag, and the JSONL fast-forward path does an unnecessary full reread that can significantly increase peak memory/IO for large transcripts.

Review details

Suppressed comments (4)

Previously missed (2) — in code that hasn't changed since the last review.

cmd/claude-sync/main.go:2790

• The global --quiet flag is documented as "Suppress output", but this command still prints to stdout (both the "no records" message and the record listing). In quiet mode it should perform the operation but return without printing anything.

This issue also appears in the following locations of the same file:

• line 2884
• line 2896
  internal/sync/sync.go:859

• In the fast-forward case this rereads the entire file to confirm equality and compute a hash. For large transcripts this can double peak memory (local+remote+final) and add significant IO. Since PrefixRemoteAhead implies we only appended, you can detect concurrent appends cheaply via file size and update state using the already-downloaded remote bytes when sizes match.

cmd/claude-sync/main.go:2887

• In quiet mode this branch still prints "No matching record found" to stdout, which contradicts the global --quiet behavior used elsewhere (push/pull). Consider suppressing this message when quiet is set (still return success).

			if len(result.RemovedRemoteKeys) == 0 && len(result.RemovedLocalPaths) == 0 {
				fmt.Printf("%s⋯%s No matching record found for %q\n", colorDim, colorReset, id)
				return nil
			}

cmd/claude-sync/main.go:2898

• This always prints a success summary line even when --quiet is set. That makes scripting harder and is inconsistent with the rest of the CLI's quiet behavior; please guard this output behind !quiet.

			fmt.Printf("%s✓%s Removed %d remote record(s), %d local file(s)\n",
				colorGreen, colorReset, len(result.RemovedRemoteKeys), len(result.RemovedLocalPaths))
			return nil

• Files reviewed: 24/24 changed files
• Comments generated: 0 new
• Review effort level: Lite

[Copilot]

🔵 Needs a closer look

CCD session sync has correctness issues on pull (state hash form mismatch and Windows pointer-file replacement via os.Rename) that can cause redundant re-uploads and failed record updates.

Review details

Suppressed comments (2)

Previously missed (2) — in code that hasn't changed since the last review.

internal/sync/ccdsessions.go:252

• ccdWriteAtomic uses os.Rename to replace an existing record. On Windows, os.Rename fails when the destination already exists, so pulls that need to update an existing pointer file can error and leave stale records behind.
  internal/sync/ccdsessions.go:496
• pullCCDSessions updates SyncState.Hash using the resolved (local) record bytes, but pushCCDSessions compares against a tokenized form (NormalizeContent). This mismatch will make the same record appear changed after every pull and cause redundant re-uploads/ping-ponging.

• Files reviewed: 24/24 changed files
• Comments generated: 0 new
• Review effort level: Lite

[tawanorg]

Thanks for picking up #77 and extending it — the extra ground this covers (Desktop store detection on Linux and macOS, cwd/originCwd translation across devices, desktop list / desktop forget, the WebDAV connection-pool and MKCOL race fixes) is genuinely useful work.

Holding it for now, and the reasons are mostly mechanical rather than about the code:

1. It conflicts with main. #82 merged earlier today and substantially reworked content path mapping in internal/sync/paths.go.

2. It duplicates part of what just landed. Your ad31e936 sync: JSON-escape substituted paths in .json/.jsonl content fixes the same bug #82 fixed — a Windows device writing a raw backslash path into .jsonl producing an invalid escape. You diagnosed it independently and correctly. #82 took a different route: it derives a content kind from the file extension, makes separator handling depend on whether the mapping root is a Windows path, and canonicalizes matched tails to / in the remote form so both platforms produce byte-identical remote content. On rebase I'd suggest dropping ad31e936 in favour of what's on main, rather than merging the two — and if you think #82's approach misses something yours caught, say so and I'll reopen that specifically. One known gap is already tracked in #90.

3. It contains #77 verbatim. Your first seven commits are #77's by SHA. The plan is to land #77 first and have this rebase onto it, which keeps attribution with @sjalife and reduces this PR to its own 16 commits — a much more reviewable diff than the current 3,683 lines across 24 files. Sequencing is tracked in #91.

So the path forward: #77 lands, you rebase onto main, drop ad31e936, and this becomes a focused review of your own additions. I'd rather do that than review a 24-file diff where two thirds of it is already approved work from another PR.

Two of your commits look independently valuable and low-risk, and I'd merge either as its own PR without waiting on any of the above:

• f0fe4035 webdav: reuse pooled connections to avoid concurrent-handshake TLS — the syncer runs a 10-worker pool, so this is a real bug for every WebDAV user
• a4daf886 webdav: fix directory-creation race and silent MKCOL failures

Splitting those two out would get them in quickly and shrink this PR further. Nothing else needed from you until #77 lands.