Establish native-Windows CI and compatibility harness

[KenTaniguchi-R]

Why

AgentCall has Linux and macOS coverage, but no Windows compatibility evidence. The published CLI explicitly declares "os": ["darwin"], the main CI job runs on Ubuntu, the packed-consumer job runs on macOS, and README still names non-macOS platforms as a v1 non-goal.

Before changing supervision or packaging, establish a native-Windows feedback loop and turn the current failures into an explicit implementation inventory. A Linux Docker container on a Mac is not Windows evidence: it cannot exercise NT paths, ACLs, PowerShell quoting, Windows process trees, Task Scheduler, or Windows agent binaries.

This is the Windows successor to #14, which was closed by #195 after delivering Linux/systemd/container support.

Scope

• Add a windows-2025 GitHub Actions job that runs on a real GitHub-hosted Windows VM.
• Exercise Node 20, 22, and 24 where the published engines contract matters; avoid multiplying the entire slow suite unless the results justify it.
• Run build, typecheck, and platform-neutral tests under PowerShell.
• Pack the shared and CLI workspaces and install the tarballs into a clean prefix on Windows.
• Run agentcall --version and agentcall --help; record the expected setup/doctor/listener failures until the implementation issue lands.
• Audit and classify Windows blockers in at least these areas:
  • npm os metadata
  • absolute/managed-policy paths
  • Unix-only fixture shebangs and executable bits
  • chmod(0600/0700) versus Windows ACLs
  • launchd/systemd listener supervision
  • detached process groups, negative PIDs, and SIGTERM/SIGKILL teardown
  • binary discovery and .cmd/.exe resolution
  • shell/PowerShell quoting and paths containing spaces

Acceptance criteria

• A manually dispatchable native-Windows CI job exists and produces useful results even while full Windows support is incomplete.
• Platform-neutral tests pass on Windows; genuinely Unix-specific tests are named and gated explicitly rather than failing accidentally.
• The packed CLI smoke test runs from a clean Windows install location.
• Every remaining failure is linked to the functional Windows-support issue, with no undocumented red tests.
• A short manual probe is documented for a real Windows VM (Daytona is acceptable) covering setup, foreground listen, one inbound call, cancellation/timeout, status, and uninstall.

Testing environment decision

Use GitHub-hosted Windows VMs as the required regression gate. Use a Daytona Windows VM for interactive debugging and the end-to-end probe. Docker on macOS may remain a Linux portability check, but is not part of the Windows support claim.

[KenTaniguchi-R]

Implementation spec

#250 is startable and claimed exclusively by this account. The implementation will remain an evidence harness, not functional Windows support (tracked by #251).

• Add workflow_dispatch plus one windows-2025 matrix job. Node 20/22/24 each consume the Linux-produced packed tarballs from a clean Windows prefix and run the npm-generated .cmd shim; only Node 24 runs workspace build/typecheck and platform-neutral source tests so the slow suite is not tripled.
• Preserve the published os restriction. The harness first proves ordinary npm installation fails with EBADPLATFORM, then uses --force only inside CI to inspect artifact startup. --version and --help must succeed from paths containing spaces.
• Run packed setup, doctor, and listen; until Support Windows as a native AgentCall callee #251 lands they must all stop at the one documented Managed policy is not supported on win32 boundary. Any different failure is red.
• Add an explicit CLI test:windows allowlist for suites that do not construct unsupported machine paths, rely on POSIX modes/shebangs, launch Unix process groups, or inspect launchd/systemd. Run all shared and relay tests natively on Windows.
• Add a living blocker inventory covering npm metadata, managed paths, fixtures/modes/ACLs, supervision, process-tree teardown, executable discovery, and PowerShell/path semantics, plus the Daytona/manual setup/listen/inbound-call/cancel/timeout/status/uninstall probe for Support Windows as a native AgentCall callee #251.
• Pin the workflow/docs contract in a deterministic local test, then validate the real PowerShell behavior through the PR checks before merging.