fix: speed up multi-file apply-patch

2026-06-06 23:12:51 +00:00
parent 27acd2a98f
commit b81585bcab
4 changed files with 319 additions and 3 deletions
@@ -202,7 +202,7 @@ export function sshHelp(): unknown {
      "When a one-line shell command is easier to type through the script path, `script -- '<command && command>'` runs that single string through the remote shell without waiting for stdin. When `script --` is followed by multiple tokens, it stays a direct argv form for commands such as `trans D601:/work script -- sed -n '1,20p' file`.",
      "script and shell helper modes inject a tiny POSIX-compatible printf wrapper before user shell text, so portable printf headings such as `printf \"--- section ---\\n\"` work consistently under dash/sh and bash. Direct argv commands are unchanged.",
      "For arbitrary stdin streams into a workload command, use a workload route plus `exec --stdin -- <command> ...`; this keeps the route as location-only and avoids heredoc/base64/tar shell wrapping.",
-      "`apply-patch` is the default remote text patch entry and uses the v2 local line-based patch engine with remote read/write operations, including Windows routes such as `D601:win/c/test`, so long Unicode/Chinese lines and pure insertion hunks avoid the legacy remote shell hunk parser. Its stdout follows Codex apply_patch text output rather than UniDesk JSON output; stderr keeps Codex-style failure text and appends one `UNIDESK_APPLY_PATCH_TIMING` JSON summary with durationMs, patchBytes, fileCount, hunkCount, changedCount, remoteOperationCount, remoteOperationCounts and remoteElapsedMs so slow patch runs can be attributed without changing success stdout.",
+      "`apply-patch` is the default remote text patch entry and uses the v2 local line-based patch engine with remote read/write operations, including Windows routes such as `D601:win/c/test`, so long Unicode/Chinese lines and pure insertion hunks avoid the legacy remote shell hunk parser. Plain multi-file Update File patches on POSIX host/k3s routes use bulk read/write operations to avoid per-file SSH round trips. Its stdout follows Codex apply_patch text output rather than UniDesk JSON output; stderr keeps Codex-style failure text and appends one `UNIDESK_APPLY_PATCH_TIMING` JSON summary with durationMs, patchBytes, fileCount, hunkCount, changedCount, remoteOperationCount, remoteOperationCounts and remoteElapsedMs so slow patch runs can be attributed without changing success stdout.",
      "`upload` and `download` are the default whole-file transfer entries for non-text and generated files. They write through remote temp files, verify byte count and SHA-256 on both sides, and return `verification.automatic=true`, `verification.verified=true`, and `verification.match.{bytes,sha256}=true`; this JSON is the transfer integrity proof, so callers do not need a separate manual `sha256sum` check. The client falls back from a single stdin payload to bounded chunks before treating provider-gateway limits as a server-side problem.",
      "`apply-patch-v1` is the only legacy fallback entry: it rejects low-context update hunks by default, reports the matched file:line for each hunk on stderr, and only accepts --allow-loose when the caller has manually reviewed an intentionally ambiguous insertion.",
      "script defaults to target /bin/sh and inherits provider proxy variables such as HTTP_PROXY/HTTPS_PROXY/ALL_PROXY/NO_PROXY; it is for host/k3s POSIX shell only. Use --shell bash only for bash syntax such as pipefail, arrays, or [[ ... ]], not as a proxy workaround.",