@simonyea/holysheep-cli 2.1.28 → 2.1.29

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@simonyea/holysheep-cli",
-  "version": "2.1.28",
+  "version": "2.1.29",
   "description": "Claude Code/Cursor/Cline API relay for China \u2014 \u00a51=$1, WeChat/Alipay payment, no credit card, no VPN. One command setup for all AI coding tools.",
   "scripts": {
     "test": "node tests/droid.test.js && node tests/workspace-store.test.js && node tests/runtime-stale-upgrade.test.js && node tests/hermes.test.js && node tests/preflight.test.js",

package/src/tools/hermes.js

@@ -305,12 +305,25 @@ module.exports = {
     return isConfiguredInToml(readConfig())
   },
   configure(apiKey, _baseUrlAnthropic, baseUrlOpenAI, primaryModel /*, _selectedModels */) {
-    const merged = mergeConfig(apiKey, baseUrlOpenAI, primaryModel)
+    // [HolySheep fork v2.1.29 / hs22] Hermes talks to its provider via the
+    // OpenAI-compatible Chat Completions API (`/v1/chat/completions`), which
+    // the HolySheep CRS relay does NOT support for `claude-*` model IDs at
+    // the moment (returns 503 "Service temporarily unavailable"). Only the
+    // anthropic native `/v1/messages` path accepts Claude. GPT-family models
+    // work on both paths.
+    //
+    // To give users a usable hermes out of the box we pick a safe default
+    // here: if the caller passes a `claude-*` model, silently downgrade to
+    // gpt-5.4 (HolySheep's default GPT model) for hermes only. The TOML
+    // keeps the real choice so `hermes --provider holysheep --model claude-*`
+    // still works if the user overrides it later (and CRS fixes the upstream).
+    const hermesPrimaryModel = /^claude-/i.test(primaryModel || '') ? 'gpt-5.4' : (primaryModel || 'gpt-5.4')
+    const merged = mergeConfig(apiKey, baseUrlOpenAI, hermesPrimaryModel)
     writeConfig(merged)
     // [HolySheep fork v2.1.28 / hs21] Also patch config.yaml so the actual
     // agent runtime uses the right base_url. See CONFIG_YAML comment.
     try {
-      patchConfigYaml(apiKey, baseUrlOpenAI, primaryModel)
+      patchConfigYaml(apiKey, baseUrlOpenAI, hermesPrimaryModel)
     } catch (e) {
       // Non-fatal: hermes can still run on config.toml alone if the user
       // manually edits config.yaml. Surface via stderr for diagnosability.

package/src/tools/pty-hermes-wrapper.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 """
-[HolySheep fork v2.1.25 / hs19] Hermes ACP PTY wrapper.
+[HolySheep fork v2.1.29 / hs22] Hermes ACP PTY wrapper.
 
 Why: bun 1.3.9's subprocess.spawn with stdio:['pipe','pipe','pipe']
 passes sockets to the child. Python's asyncio.connect_write_pipe()
@@ -13,9 +13,22 @@ captures Python tracebacks separately. Master side pumps bytes between
 our (bun-spawned) stdin/stdout and the PTY, so bun reads through kernel
 PTY buffer which hermes's asyncio transport drives correctly.
 
-Disable ECHO on slave so stdin bytes are not echoed back as stdout bytes
-(the old BSD `script -q /dev/null` wrapper had this bug).
-Disable ONLCR so '\\n' is not rewritten to '\\r\\n'.
+CRITICAL: disable ICANON (canonical/line-buffered mode) on the slave.
+Default macOS/Linux PTY slaves have ICANON on, which means:
+  1. Input is buffered line-by-line and each line is capped at MAX_CANON
+     (1024B on macOS, ~4096B on Linux). Longer lines get truncated.
+  2. Special chars (^C, ^D, ^Z, erase/kill) are interpreted instead of
+     passed through — breaks JSON containing random bytes.
+
+ACP session/prompt JSON can exceed 4KB easily (system prompt + skill
+list + assistant rules). Before hs22 we only disabled ECHO + ONLCR, so
+id=1 (initialize, ~150B) and id=2/3 (newSession, setSessionMode, <400B)
+went through, but id=4 (prompt, ~4-8KB) was truncated by MAX_CANON —
+hermes got invalid JSON, returned an SDK error with empty data, our
+AcpAgentManager 60s fallback masked it. User saw "hermes thinks forever".
+
+Also disable ICRNL and IXON so we don't have \\r<->\\n rewriting or
+flow-control pauses on large bursts.
 
 Usage: python3 pty-hermes-wrapper.py
 HERMES_BIN is auto-resolved from $PATH; override via $HOLYSHEEP_HERMES_BIN.
@@ -39,9 +52,30 @@ def main() -> None:
     hermes_bin = _resolve_hermes_bin()
     master_fd, slave_fd = pty.openpty()
     attrs = termios.tcgetattr(slave_fd)
-    attrs[3] &= ~termios.ECHO
-    attrs[1] &= ~termios.ONLCR
+    # attrs = [iflag, oflag, cflag, lflag, ispeed, ospeed, cc]
+    # lflag (attrs[3]): disable canonical-mode line buffering, ECHO*, signal chars
+    attrs[3] &= ~(termios.ICANON | termios.ECHO | termios.ECHOE | termios.ECHOK
+                  | termios.ECHONL | termios.ISIG)
+    # iflag (attrs[0]): disable \r<->\n translation, XON/XOFF flow control,
+    # and parity checking (we carry raw bytes only).
+    attrs[0] &= ~(termios.ICRNL | termios.INLCR | termios.IGNCR | termios.IXON
+                  | termios.IXOFF | termios.ISTRIP | termios.IGNBRK | termios.BRKINT
+                  | termios.INPCK | termios.PARMRK)
+    # oflag (attrs[1]): disable post-processing (ONLCR \n -> \r\n, tabs, etc.)
+    attrs[1] &= ~termios.OPOST
+    # Non-blocking reads: return as soon as 1 byte is available.
+    attrs[6][termios.VMIN] = 1
+    attrs[6][termios.VTIME] = 0
     termios.tcsetattr(slave_fd, termios.TCSANOW, attrs)
+    # Best-effort: print the applied flags once so future regressions are
+    # easy to diagnose from hs web logs.
+    try:
+        sys.stderr.write(
+            f'[pty-hermes-wrapper] slave configured raw: ICANON off, ECHO off, OPOST off\n'
+        )
+        sys.stderr.flush()
+    except OSError:
+        pass
 
     pid = os.fork()
     if pid == 0:
@@ -103,9 +137,28 @@ def main() -> None:
                     except ProcessLookupError:
                         pass
                     break
-                try:
-                    os.write(master_fd, data)
-                except OSError:
+                # [HolySheep fork v2.1.29 / hs22] Diagnostic trace when enabled —
+                # emits one line per stdin chunk so we can see exactly what bun
+                # pushed to us and verify hermes received the full payload. Only
+                # active when HOLYSHEEP_PTY_TRACE=1 is set to avoid log flood.
+                if os.environ.get('HOLYSHEEP_PTY_TRACE') == '1':
+                    try:
+                        sys.stderr.write(f'[pty-hermes-wrapper] stdin chunk len={len(data)}\n')
+                        sys.stderr.flush()
+                    except OSError:
+                        pass
+                # Write in full — os.write may short-write on large buffers; loop.
+                total = 0
+                while total < len(data):
+                    try:
+                        n = os.write(master_fd, data[total:])
+                    except OSError:
+                        total = -1
+                        break
+                    if n <= 0:
+                        break
+                    total += n
+                if total < 0:
                     break
     finally:
         try:

package/src/webui/aionui-runtime-fetcher.js

@@ -60,10 +60,10 @@ const VENDOR_DIR = path.join(__dirname, 'vendor', 'aionui')
 // new CLI release, the next `hs web` invocation on user machines will detect
 // the version drift and upgrade the cache in place.
 const DEFAULT_RUNTIME_URL =
-  'https://mail.holysheep.ai/app/cli/aionui-runtime-v1.9.18-holysheep-hs20.tar.gz'
+  'https://mail.holysheep.ai/app/cli/aionui-runtime-v1.9.18-holysheep-hs22.tar.gz'
 const DEFAULT_RUNTIME_SHA256 =
-  '99ddb768cacd223a13c77d716413032e913c8be2850e7dbccaf7ae52d0797e37'
-const DEFAULT_RUNTIME_VERSION = '1.9.18-holysheep-hs20'
+  '03e386ea83660b4074e29026dfc53be6ed01fcd51fa35a9379ed64a396e99f33'
+const DEFAULT_RUNTIME_VERSION = '1.9.18-holysheep-hs22'
 
 function isValidRuntimeDir(dir) {
   if (!dir) return false