fix(windows): native ConPTY + kernel32 metrics + agent loop cap (v0.7.6)
All checks were successful
PR Check / check (pull_request) Successful in 1m0s

Three issues reported on Windows + one user-requested limit bump:

1. Dashboard CPU/RAM/Network all at 0
   handleSystemMetrics read /proc/* exclusively. Replaced with a
   platform-split:
   - metrics_unix.go (!windows): existing /proc reading code.
   - metrics_windows.go: kernel32!GetSystemTimes for CPU
     (delta of idle vs kernel+user FILETIMEs) and
     kernel32!GlobalMemoryStatusEx for memory. Network left at zero
     for now — MIB_IF_ROW2 is too version-sensitive to parse by hand.
   handlers_info.go::handleSystemMetrics reduced to one delegating
   call.

2. Terminal black screen on Windows
   creack/pty/v2 returns "unsupported" on Windows; the v0.7.1 pipe
   fallback works but pipes don't carry TTY signals, so cmd/pwsh/wsl
   go silent. Implemented native ConPTY:
   - terminal_conpty_windows.go: CreatePseudoConsole + STARTUPINFOEX
     + PROC_THREAD_ATTRIBUTE_PSEUDOCONSOLE wiring via
     windows.NewProcThreadAttributeList. CreateProcessW launches
     child with the PC attached, full ANSI / line discipline /
     resize.
   - canUseConPTY() probes once at startup (Win10 1809+ check).
   - Restructure: terminal_session.go now holds just the interface
     + ptySession + pipeSession structs. terminal_session_unix.go
     wires creack/pty. terminal_session_windows.go tries ConPTY
     first, falls back to pipeSession.

3. Agent stops after 15 tool calls
   MaxToolIterations bumped 15 → 500. Doc comment explains why the
   cap exists at all (infinite-loop safety) and that 500 is well
   above realistic usage.

- internal/version/version.go: 0.7.5 → 0.7.6
- CHANGELOG.md: v0.7.6 entry covers the three fixes
This commit is contained in:
Muyue
2026-04-27 14:04:41 +02:00
parent 79e467c32a
commit d557b8e74c
10 changed files with 596 additions and 119 deletions

View File

@@ -10,9 +10,13 @@ import (
"github.com/muyue/muyue/internal/orchestrator"
)
const (
MaxToolIterations = 15
)
// MaxToolIterations bounds the inner tool-call loop in RunWithTools /
// RunNonStream. The cap exists only to avoid an infinite loop when a model
// keeps calling tools forever; the value is intentionally generous so a
// realistic agent run (multi-file refactor, exploratory debugging…) never
// hits it. If you find yourself raising this to absurd values, look for a
// loop bug in the model output instead.
const MaxToolIterations = 500
// ToolLimiter checks if a tool call is allowed and returns a release function.
type ToolLimiter func(toolName string) (release func(), err error)