Pr 12#18
Closed
danielhertz1999-bit wants to merge 15 commits into
Closed
Conversation
The Rust hf-hub client uses a different TLS stack than Python and fails to reach huggingface.co in containers with custom CA certificates (UnknownIssuer). Add _auto_set_embed_offline() which runs at daemon startup: if the bge-small-en-v1.5 snapshot directory already exists in the HF cache, it sets IAI_MCP_EMBED_OFFLINE=1 automatically so the embedder skips the network download entirely. To seed the cache in a restricted environment, download the three model files (model.safetensors, tokenizer.json, config.json) via Python's SSL stack and place them under: ~/.cache/huggingface/hub/models--BAAI--bge-small-en-v1.5/snapshots/<rev>/ Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mf3VFyVtczcK2WxKKCyBS4
…tdfl63 Auto-detect HF model cache to bypass Rust TLS in restricted environments
Introduces src/iai_mcp/_ipc.py as the central abstraction for all daemon socket communication. On POSIX it delegates to the existing Unix-domain socket; on Windows it uses TCP loopback with the ephemeral port persisted in ~/.iai-mcp/.daemon.port. Replaces all nine raw asyncio.open_unix_connection / asyncio.start_unix_server / socket.AF_UNIX call-sites with the new open_ipc_connection / start_ipc_server / make_sync_ipc_socket helpers. The POSIX code-paths are structurally unchanged. This is step 1 of the Windows port. Remaining blockers (fcntl, resource module, POSIX signals, shell hooks, daemon installer) are tracked in the audit at the top of _ipc.py. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add _filelock.py: on POSIX delegates to fcntl.flock; on Windows uses msvcrt.locking with EWOULDBLOCK normalisation so all non-blocking callers work unchanged. Also changes doctor check_c to open the lock file with O_RDWR (msvcrt.locking requires write access; harmless on POSIX). Updated 7 files: capture_queue, lifecycle_event_log, lifecycle, lock_protocol, hippo/_db, hippo/__init__ (dead import), doctor/_lifecycle_checks. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- resource module: lazy-import behind platform check (Windows has no RLIMIT_NOFILE). - Signals: build shutdown-signal list dynamically using hasattr; replace bare signal.SIGKILL with hasattr guard + sys.exit(1) fallback in _self_kill; use getattr for SIGTERM/SIGKILL in CLI stop and doctor orphan killer. - os.getuid(): guarded with hasattr fallback to 0 (4 sites in _daemon.py). - Log path: add _get_daemon_log_path() returning platform-appropriate path (%APPDATA%/iai-mcp/logs on Windows); add Windows branch to cmd_daemon_logs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- crypto.py: skip POSIX mode/uid checks on Windows (chmod is no-op); add _secure_key_file() using icacls for Windows ACL lockdown; guard os.fchmod with hasattr. - cli/_crypto.py: guard st.st_uid and os.geteuid() for status report. - memory_bank.py: guard os.fchmod call with hasattr. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds _is_windows() and SCHTASKS_TASK_NAME to cli/__init__.py and implements cmd_daemon_install/uninstall/start/stop branches for Windows in cli/_daemon.py using schtasks.exe and _render_schtasks_xml(). Also adds the Windows PowerShell turn-capture hook stub and the WINDOWS_PORT_HANDOFF.md guide for continuing the port in a new session (covers the remaining Steps 6-10 with file paths, code examples, and test commands). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Creates three PowerShell scripts (.ps1) to replace shell hooks on Windows: - iai-mcp-turn-capture.ps1: per-turn ambient capture (UserPromptSubmit) - iai-mcp-session-capture.ps1: batch-capture at session end (Stop) - iai-mcp-session-recall.ps1: session-start recall injection (SessionStart) Updates src/iai_mcp/cli/_capture.py: - _hook_ext() detects platform and returns '.ps1' on Windows, '.sh' on POSIX - _capture_hook_paths(), _turn_hook_paths(), _session_recall_hook_paths() now use _hook_ext() for dynamic extension - cmd_capture_hooks_install() uses 'powershell -ExecutionPolicy Bypass -File' for Windows hooks instead of 'bash' on POSIX - Markers (_CAPTURE_HOOK_MARKER, etc.) changed to base names (no extension) so substring matching works for both .sh and .ps1 in settings.json Updates pyproject.toml: - Adds "_deploy/hooks/*.ps1" to package.data so PowerShell hooks are bundled Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
…ps and verification checklist
The bench scripts measured peak RSS via resource.getrusage(RUSAGE_SELF).ru_maxrss, which is POSIX-only and crashes on import on Windows. Switch the Windows branch to psutil.Process().memory_info().peak_wset; POSIX paths are unchanged. - bench/memory_footprint.py: _rss_mb() guards on sys.platform == "win32" - bench/memorygraph_memory.py: rss_mb() same pattern - bench/consolidation_rss_peak.py: defer `import resource` into _ru_maxrss_bytes() - bench/embed_warm_cost.py: rewrite _PAYLOAD_RSS subprocess template to detect platform and pick peak_wset / ru_maxrss; emit rss_platform key; measure_rss decodes the new field for the unit label, returns rss_platform in result dict psutil is already a declared dependency (pyproject.toml), so no new package. Verified end-to-end on Windows: all three helpers and the rewritten payload produce sane values; AST-parse clean. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The fcntl→_filelock rewrite in commit 8154b9b swept timedelta and timezone into the new _filelock import line, but those belong to datetime. The module crashed with ImportError on any platform — masked until now because the Windows porting work hadn't tried to import the chain end-to-end. Move timedelta and timezone back to the `from datetime import` line. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Step 5 added Task Scheduler / schtasks backends to install/uninstall/start/ stop/logs, but the argparse help strings still mentioned only launchd and systemd. Update them to list all three platforms. The `Windows %APPDATA%\iai-mcp\logs` reference in the logs help is escaped as `%%APPDATA%%` so argparse's own %-formatter doesn't choke on it. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add a new "Verified on Windows in-situ" section recording what was actually exercised on a Windows machine in this session: AST parse over all 23 touched files, import smoke test of 10 runtime modules, CLI help, daemon install --dry-run producing a valid Task Scheduler XML, and capture-hooks status detecting the .ps1 templates. Also record the lifecycle_event_log fix and daemon help-text update commits alongside the existing Step-7 entry. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…m shims 5 test modules imported fcntl or resource at top level, aborting pytest collection before any test ran on Windows. Switch to the existing _filelock shim (LOCK_EX/SH/UN/NB + flock) and guard the resource import with a sys.platform check, skipping TestRaiseFdLimitClampsToHard on Windows (resource.RLIM_INFINITY is unavailable there). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two bugs from @warplayer's Windows 11 test report: 1. _filelock.flock() was calling os.lseek(fd, 0, SEEK_SET) before msvcrt.locking(), which silently moved the file offset. fcntl.flock on POSIX never touches the offset; callers that need a specific byte already seek themselves. Removed the lseek to match POSIX behaviour. 2. _patch_claude_desktop_config() called _build_iai_mcp_server_entry() without a FileNotFoundError guard in two places (new-file creation and existing-file update). If the mcp-wrapper wasn't built yet and Claude Desktop was installed, the function would crash after already writing hooks and settings.json, leaving a partial install. Both call sites now catch FileNotFoundError and return an actionable skip message (consistent with the guard already on the Claude Code path). Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Mf3VFyVtczcK2WxKKCyBS4
Owner
|
Huge thanks for taking this on — a Windows port is a big, cross-cutting piece of work and it means a lot that you're pushing on it. I'm following your commits and I want to land it once the port is complete (the PowerShell hooks wired up, the install/uninstall scripts, and ideally a run on a real Windows box to confirm it works end to end). No rush at all — take the time it needs, and I'll merge it as soon as it's ready. Genuinely appreciate the effort here. |
Author
|
Closing in favor of #12, which is the canonical, complete Windows port (head = 🤖 Addressed by Claude Code |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Type of change
Affected areas
Testing
pytestpasses locallyruff check src/ tests/cleanBenchmarks
If this PR touches retrieval, capture, or consolidation, include before/after numbers from the relevant
bench.*harness.python -m bench.___Notes for reviewers