feat(parser): add Windsurf workspace chat support (#696)

[rodboev]

Agentsview currently has no Windsurf parser, so Windsurf workspace chat history is invisible to sync and import. Windsurf stores workspace chat in workspaceStorage/*/state.vscdb, and the payloads are close enough to the existing VS Code family that the right implementation is an adapter over the shared VS Code parser, not a separate parser stack.

This adds Windsurf as a first-class parser provider. The provider discovers workspaceStorage/*/state.vscdb, reads the known chat rows from ItemTable, translates the tab-container shape into VS Code session JSON when needed, and then reuses the existing VS Code chat parser while relabeling the normalized session and usage events as windsurf. Each parseable session is exposed as a state.vscdb#sessionID virtual source so discovery, fingerprinting, watch updates, and restart-time lookup all stay scoped to one session instead of one shared database file.

The rest of the change is the minimum consumer routing needed for a new DB-backed virtual source family: generic source-path resolution, per-session parse-diff scoping, sync-by-path lookup, session export, and narrow workspaceStorage remote targeting. The SSH resolver change stays because the longer resolver script hit the Windows command-length limit in CI, and sending the script over stdin fixes that at the transport layer without changing parser ownership.

This PR intentionally keeps Cascade trajectory protobufs, encrypted .pb cache files, and other non-workspace Windsurf storage out of scope. Those are a separate format surface and can be handled later without blocking workspace chat support. The remaining proof gap is format ownership: validation here is based on synthetic SQLite fixtures and public non-owner references, not a captured real Windsurf session database, so the claim stays scoped to the handled workspace chat shapes rather than all Windsurf transcript variants.

Closes #696

[roborev-ci]

roborev: Combined Review (c8366f4)

Medium-risk issues remain in the Windsurf parser around deleted DB handling and numeric chat bubble type decoding.

Medium

• internal/parser/windsurf_provider.go:294
  Problem: Fingerprint returns an error when state.vscdb has been deleted. SourcesForChangedPath can still return stored virtual sources for that DB, and Parse has deletion tombstone handling, but the engine fingerprints before parsing, so deleted Windsurf chats can remain archived.
  Fix: Return an empty keyed fingerprint on os.IsNotExist so Parse can emit ForceReplace/SkipNoSession; add a deletion changed-path test.

• internal/parser/windsurf_provider.go:637
  Problem: windsurfChatBubble.Type is a string, but isAssistant tries to support "2" numeric-style assistant types. If actual chat data uses JSON numbers like "type": 2, unmarshalling the whole tab container fails and no sessions are discovered.
  Fix: Decode type with a custom string-or-number helper or json.RawMessage, and add coverage for numeric user/assistant bubble types.

---

Reviewers: 2 done | Synthesis: codex, 9s | Total: 7m50s

[roborev-ci]

roborev: Combined Review (6b93c2a)

Medium findings:

• internal/parser/windsurf_provider.go:566 - Windsurf persists file_path as a virtual <state.vscdb>#<session> path, but existing source-path resolution only understands Visual Studio Copilot and Aider virtual paths. Sync-by-ID, SyncPaths, source export, and mtime lookup may treat it as a literal path and fail or return empty data. Add Windsurf virtual-path parsing and route stored Windsurf paths through provider-backed session resolution/export handling.

• internal/parser/types.go:293 - The Windsurf provider registers the whole Windsurf/User profile as a file-based default root. Once provider-authoritative, remote archive sync can tar and stream unrelated editor state such as settings, extension/global storage, or other secret-bearing files. Narrow Windsurf archive targets to the required workspaceStorage DB/manifest files, or add provider-specific target selection that transfers only session data while preserving import expectations.

---

Reviewers: 2 done | Synthesis: codex, 9s | Total: 10m51s

[roborev-ci]

roborev: Combined Review (1602c4c)

High-level verdict: changes need fixes before merge due to one high-risk Windsurf archive deletion path and one medium virtual-path parsing bug.

High

• Location: internal/parser/windsurf_provider.go:101
• Problem: A missing state.vscdb is treated as ForceReplace + SkipNoSession, so a remove event for the whole Windsurf workspace DB can cause archived Windsurf sessions to be parser-excluded and deleted. Whole source-file disappearance should preserve the archive.
• Fix: Do not set ForceReplace when the physical DB is missing; only force-replace when the DB still exists and the selected session row is absent. Also avoid emitting stored tombstones for a removed whole DB.

Medium

• Location: internal/parser/windsurf_provider.go:753
• Problem: SplitWindsurfVirtualPath accepts any path containing #, so ordinary source files with # in their names can be misclassified as Windsurf DB virtual paths by session export, ResolveSourceFilePath, and sync paths.
• Fix: Scope the parser to Windsurf containers, e.g. split via ParseVirtualSourcePathForBase(path, windsurfStateDBName), and add coverage for non-Windsurf # paths.

---

Reviewers: 2 done | Synthesis: codex, 9s | Total: 8m25s

[roborev-ci]

roborev: Combined Review (a403234)

Summary: One medium-severity correctness issue needs attention before merge.

Medium

• Location: internal/parser/windsurf_provider.go:618
  Problem: Empty raw sessionId values fall back to the constant "session" while the JSON payload is left unchanged. parseVSCodeCopilotData then falls back to filepath.Base(virtualPath), so multiple workspaces without embedded session IDs parse to the same windsurf:state.vscdb#session ID and overwrite each other.
  Fix: Generate a stable unique fallback from the workspace/DB identity and make the parsed session ID use that value, with a regression test covering two workspaces missing sessionId.

---

Reviewers: 2 done | Synthesis: codex, 7s | Total: 12m36s

[roborev-ci]

roborev: Combined Review (090c8ab)

Summary verdict: One medium issue needs attention; no high or critical findings were reported.

Medium

• internal/service/direct.go:428 - session sync <path-to-state.vscdb> syncs Windsurf sessions but then fails to resolve the synced session because resolveSessionIDByPath only expands physical container paths to virtual path#... rows for Visual Studio Copilot.
  • Fix: Extend the virtual-container lookup to Windsurf state.vscdb paths and add single-match plus ambiguous-match tests.

---

Reviewers: 2 done | Synthesis: codex, 7s | Total: 9m55s

[roborev-ci]

roborev: Combined Review (da36a2e)

Review verdict: changes need fixes before merge due to two Medium Windsurf data-loss/reporting risks.

Medium

• internal/sync/parsediff.go:409 - Windsurf emits one virtual source per chat as state.vscdb#<sessionID>, but state.vscdb is not included in perSessionDBVirtualSourceBases. Parse-diff can collapse all Windsurf chats in the same DB to one source key, so --limit may falsely report unparsed sibling chats as missing, and a per-session parse error can fan out to every sibling row. Add state.vscdb/windsurfStateDBName to the per-session virtual DB base list and cover a multi-tab Windsurf parse-diff limit case.

• internal/parser/windsurf_provider.go:640 - Malformed Windsurf chat JSON is treated as “no records” instead of a parse error. On changed-path sync, stored source paths are appended, Parse then sees sql.ErrNoRows and force-replaces the archived session, so transient or corrupt chatdata can delete previously synced sessions. Return an error when the chatdata value cannot unmarshal into any supported shape, so sync fails safely instead of interpreting it as deletion.

---

Reviewers: 2 done | Synthesis: codex, 9s | Total: 13m7s

[roborev-ci]

roborev: Combined Review (0875538)

No issues found.

---

Reviewers: 2 done | Synthesis: codex | Total: 12m45s

[roborev-ci]

roborev: Combined Review (439a9a4)

Summary verdict: one medium issue remains; no high or critical findings were reported.

Medium

• internal/parser/windsurf_provider.go:330 - Windsurf computes and stores a source hash, but the sync freshness path does not require that hash for Windsurf. A same-size/same-mtime change to state.vscdb, WAL, or workspace.json can be skipped by DB freshness or skip cache, leaving stale messages.
  • Fix: Make the Windsurf fingerprint hash stable/content-based and include AgentWindsurf in the provider hash freshness checks, with a same-size/same-mtime regression test.

---

Reviewers: 2 done | Synthesis: codex, 6s | Total: 13m58s

[roborev-ci]

roborev: Combined Review (cd4edb2)

Verdict: One medium issue should be addressed before merge.

Medium

• internal/remotesync/resolve.go:89: resolveWindsurfTarget allows the entire <Windsurf User>/workspaceStorage directory as a remote-sync target. A client with access to /api/v1/remote-sync/archive can request that target and receive a recursive tar archive, exposing unrelated Windsurf and extension workspace state beyond the session data the parser needs.

  Fix: Avoid archiving workspaceStorage wholesale. Enumerate only the required files, or generate a sanitized Windsurf export containing only the needed chat-data records and workspace metadata.

---

Reviewers: 2 done | Synthesis: codex, 8s | Total: 15m6s

[roborev-ci]

roborev: Combined Review (08a4011)

Summary verdict: one medium severity issue remains; no high or critical findings were reported.

Medium

• Location: internal/ssh/resolve.go:80
• Problem: SSH remote sync still resolves Windsurf to the whole workspaceStorage directory, and buildTarCommand archives directory targets recursively. This bypasses the new file-scoped TargetSet.Files behavior used by HTTP remote sync, so the default SSH transport can still transfer unrelated workspace files such as extension secrets.
• Fix: Extend SSH resolution/transfer to carry file-scoped Windsurf targets (state.vscdb, state.vscdb-wal, state.vscdb-shm, workspace.json) and pass them through import as TargetSet.Files, or have the SSH tar command enumerate only those files for Windsurf.

---

Reviewers: 2 done | Synthesis: codex, 7s | Total: 13m43s

[roborev-ci]

roborev: Combined Review (95fec13)

Remote sync has one medium-risk scalability issue; no high or critical findings were reported.

Medium

• internal/ssh/transfer.go:36 - Windsurf remote sync expands every workspace DB, WAL/SHM, and workspace.json into a single tar ... <paths> shell command. With many Windsurf workspaces, this can exceed the remote shell/exec argument length and fail before transfer starts. Feed paths to tar via stdin or a temporary list file, or batch transfers into bounded chunks.

---

Reviewers: 2 done | Synthesis: codex, 6s | Total: 14m11s

[roborev-ci]

roborev: Combined Review (360fc2e)

Summary verdict: Two medium-risk Windsurf remote-sync issues should be addressed before merge.

Medium

• internal/remotesync/resolve.go:128
  Windsurf remote sync exposes the raw state.vscdb workspace database, plus WAL/SHM files, through the HTTP archive path. A bearer-token holder authorized for agentsview API access could download the full workspace state DB, including extension state or recently deleted WAL content that may contain non-session secrets.
  Fix: Do not expose raw state.vscdb via remote sync. Query server-side and archive a sanitized artifact containing only Windsurf chat keys and required metadata, or generate a filtered temporary SQLite/JSON source for import.

• internal/parser/windsurf_provider.go:323 and internal/remotesync/resolve.go:130
  Windsurf fingerprints and remote-sync targets include state.vscdb-shm. SQLite SHM files are transient WAL-index/lock files and may be rewritten or removed during normal SQLite activity. Including them can make unchanged sessions appear stale, and exact remote archive selection can fail if the file disappears between resolution and archive creation.
  Fix: Drop -shm from the Windsurf fingerprint and remote file list. If -wal remains, treat it as optional or re-resolve at archive time so checkpointed WAL removal does not abort sync.

---

Reviewers: 2 done | Synthesis: codex, 10s | Total: 14m18s

[roborev-ci]

roborev: Combined Review (19f29b4)

Medium issue found: tar path-list mode needs option-shaped path protection.

Medium

• internal/ssh/transfer.go:55
  The new tar -T - path-list mode no longer protects option-shaped paths. Entries read via -T can be interpreted by GNU tar as options unless verbatim mode is used, so a resolved path beginning with - after trimming the leading slash can alter tar behavior or fail. The previous argv form used --, which avoided this issue.

  Fix: Emit file-list entries in a verbatim-safe form, such as prefixing relative paths with ./ after the existence check, or use tar’s verbatim/null file-list mode where supported.

---

Reviewers: 2 done | Synthesis: codex, 8s | Total: 18m17s

[roborev-ci]

roborev: Combined Review (e650e31)

Summary verdict: One medium-severity issue remains; no high or critical findings were reported.

Medium

• internal/ssh/transfer.go:53 — SSH tar paths are emitted as a newline-delimited tar -T - list. A resolved path containing an embedded newline is shell-quoted as one argument to av_emit_tar_path, passes the existence check, then printf '%s\n' splits it into multiple tar entries. This can archive unintended readable paths or fail on crafted filenames under a scanned agent directory.

  Suggested fix: Emit a NUL-delimited file list and use tar’s null-file-list mode where supported, or reject/skip resolved paths containing \n/\r before building the tar script.

---

Reviewers: 2 done | Synthesis: codex, 7s | Total: 13m59s

[roborev-ci]

roborev: Combined Review (234f411)

No issues found.

---

Reviewers: 2 done | Synthesis: codex | Total: 16m31s