[zwanski-start]

Closes #44

Problem

On Windows, claude is a .cmd shim, so the whole command goes through cmd.exe and its ~8191 character limit. Two separate problems came from how the chat sessions passed their system prompts:

• Expand Project passed its system prompt (expand-project.md, about 6.6KB, plus the project path) inline as --system-prompt. With the other arguments (--mcp-config JSON, tool lists, settings path), that goes over the limit and fails with "The command line is too long."
• Project Assistant / Spec creation worked around the limit (07c2010) by overwriting the project's CLAUDE.md with the chat prompt. That deletes whatever the user had in CLAUDE.md. It also leaks the chat-only instructions into the coding agents, which load CLAUDE.md via setting_sources=["project"] (for example, the assistant's "You must NEVER implement code").

Fix

• New helpers in server/services/chat_constants.py:
  • write_system_prompt_file() writes the prompt to a per-session file, .autoforge/.system_prompt.<kind>.<uuid>.md.
  • remove_system_prompt_file() deletes it.
• Assistant, spec and expand sessions now pass system_prompt={"type": "file", "path": ...}, which the SDK sends as --system-prompt-file <path>. Each session deletes its file in close().
• None of the sessions write to CLAUDE.md any more. The assistant and spec sessions still use setting_sources, so the project's real CLAUDE.md is now read as context instead of being overwritten.
• Added get_system_prompt_path() to autoforge_paths.py, and added .system_prompt.*.md to the .autoforge/.gitignore template.
• Raised the minimum claude-agent-sdk to 0.1.51, the first release that supports the file system prompt option. I checked this by comparing the 0.1.50 and 0.1.51 wheels.

Verification

• Captured the real CLI command each session builds, with SDK 0.1.81 and a long project path:
  • Assistant: about 1.9k characters
  • Expand: about 1.5k characters (before this change it carried the 6.6KB prompt inline)
  • Spec: about 0.8k characters
• Confirmed the prompt files are created on start and deleted on close.
• Confirmed an existing CLAUDE.md is left untouched.
• End-to-end on Linux with the real claude CLI: started an AssistantChatSession and asked "what is your role / which project". It answered "I'm the project assistant and backlog manager for the "demo" project…", so the CLI reads the prompt from the file.
• New test_chat_system_prompt.py (4 tests) passes. ruff check ., test_security.py, test_client.py and test_ws_security.py also pass.
• mypy . shows no new errors. The diff against master is only line-number shifts of the existing unused-ignore warnings.
• Not tested on an actual Windows machine. The fix relies on the argument length being well under 8191, which the measurements above show.

Assumptions / notes

• I can't tell which CLAUDE.md files older versions already overwrote, so this PR doesn't try to restore them. Users who see an assistant or create-spec prompt in their project's CLAUDE.md should delete it or restore it from git.
• Raising the SDK minimum changes the requirements hash, so npm installs will rebuild their venv once on the next start.

🤖 Generated with Claude Code

[zwanski-start]

🔍 QA report: ✅ Passed

Tester: Ken (cubefarm agent, in a fresh session) · Round: 1 · Author: Guido

Assistant, spec and expand chat sessions now hand their system prompt to the Claude CLI via --system-prompt-file (ephemeral .autoforge/.system_prompt.<kind>.<uuid>.md) instead of passing it inline or overwriting the project's CLAUDE.md. With a stub claude CLI recording argv, all three commands contain no inline prompt and stay ~1.3k–2.4k chars (expand was ~10k inline on master, above cmd.exe's 8191 limit), a user CLAUDE.md is left untouched, and prompt files are removed on session close and server shutdown. ruff/mypy show no new problems, new tests and security tests pass, and the 21 pytest failures are identical on master. Not tested on a real Windows machine; the fix is proven by argv length measurement.

	Check	Details
✅	Code review of diff	Small, focused change: write/remove helpers in chat_constants.py, a path helper plus a .gitignore pattern in autoforge_paths.py, and three sessions switched to system_prompt={'type':'file'}, with cleanup in close(). No debug leftovers, security.py and the localhost middleware are untouched, no secrets logged, and no new hardcoded OS paths (pathlib, UTF-8 writes).
✅	SDK support for file system prompt	Installed SDK 0.1.81 maps {'type':'file'} to --system-prompt-file <path>, and the installed claude CLI help lists --system-prompt[-file]. Downloaded the 0.1.50 and 0.1.51 wheels: only 0.1.51 contains system-prompt-file, so the new >=0.1.51 floor is correct.
✅	Issue #44 repro: command length (stub claude CLI capturing argv)	Long unicode project path. Assistant argv 2431 chars (prompt file 3813 B), Spec 1260 chars (24470 B file), Expand 2089 chars (7988 B file). No --system-prompt inline flag in any of them. On master, expand would carry the ~8KB prompt inline, about 10k chars total, which is over the 8191 limit.
✅	CLAUDE.md no longer overwritten	Pre-existing CLAUDE.md content was byte-identical after running assistant, spec and expand sessions, both in a direct probe and through the live server's Assistant panel.
✅	Prompt file lifecycle / cleanup	The file exists while the CLI runs (verified from inside the stub) and is deleted by close(), including on the client-init-failure path. In the live server the remaining assistant file persisted while the session was kept for resume (existing design) and was removed on graceful shutdown. No orphan processes left.
✅	Graceful degradation with no API key / failing CLI	Assistant panel shows a 'Failed to initialize Claude' error message without crashing. The only console errors are the expected assistant error logs.
✅	UI smoke (project list, dashboard, kanban, assistant panel, phone width)	Served SPA on :5345 loads, project selection works, Kanban renders, assistant panel works at 390px width. The PR does not touch the UI.
✅	Backend test suite	148 passed, 21 failed. All 21 failures are in test_devserver_security.py and fail identically on origin/master (diffed failure lists), so they predate this PR.
⏭️	UI e2e tests	PR touches no UI code.
⏭️	Real Windows run	No Windows machine available. Fix verified by argv length measurement on Linux.

🧪 Commands run

Command	Result
venv/bin/ruff check .	All checks passed
venv/bin/mypy .	24 errors in 9 files, identical set to origin/master (no new errors)
venv/bin/python -m pytest -q	148 passed, 21 failed (same 21 test_devserver_security failures on master)
venv/bin/python -m pytest -q test_chat_system_prompt.py	4 passed
venv/bin/python test_security.py	ALL TESTS PASSED
venv/bin/python test_security_integration.py	ALL INTEGRATION TESTS PASSED
npm --prefix ui run build	success
npm --prefix ui run lint	0 errors, 6 warnings
pip download claude-agent-sdk==0.1.50 / 0.1.51 + grep system-prompt-file	0.1.50: absent, 0.1.51: present
probe.py with stub claude on PATH (assistant/spec/expand start+close)	no inline prompt; argv 1260–2431 chars; files cleaned; CLAUDE.md intact
uvicorn server.main:app --port 5345 (isolated HOME, stub CLI) + Playwright	assistant argv 1905 chars, graceful error, file removed on shutdown, no orphans

No browser screenshots were taken in this round.

_{Posted by cubefarm · merges automatically once GitHub's checks pass}