[zwanski-start]

What changed

The Web UI spec creation chat (server/services/spec_chat_session.py) now writes a Markdown transcript to .autoforge/prompts/spec_input_summary.md, next to app_spec.txt.

• render_spec_conversation_markdown() is a pure function. It builds a header (project, start time, status: In progress / Complete), an Initial Prompt section (the user's first message) and a Conversation section with alternating ### Claude / ### You turns. Claude's questions and suggested options appear exactly as Claude wrote them, followed by the user's answer or follow-up. Text chunks streamed from Claude are merged into one turn, and attachment filenames are listed on the user turn that included them.
• SpecChatSession._save_conversation_summary() runs in a finally after start() and after every send_message(). Abandoned or errored conversations are saved too. A failed write only logs a warning and never breaks the chat.
• Starting a new spec session overwrites the file. This matches start(), which already deletes the old app_spec.txt.
• Docs: README project tree and CLAUDE.md project structure.

Verification

• New test_spec_conversation_summary.py (4 tests) covers rendering, chunk merging, status, attachments and the file write: python -m pytest test_spec_conversation_summary.py passes.
• ruff check . is clean and python test_security.py passes.
• mypy shows no new errors. The existing unused type: ignore on the effort= line is unchanged.
• No UI files changed, so I didn't do a browser check. Running a real spec chat needs live Claude API calls, so the persistence path is covered by unit tests instead.

Assumptions

• The file goes in .autoforge/prompts/, the current layout. The issue's prompts/ path is the legacy layout.
• The file is written by the server, not by asking Claude in the skill to write it. This makes it an exact, reliable record and doesn't depend on the model.
• The CLI /create-spec path (start.py → interactive claude) is not covered, because AutoForge doesn't see that conversation. The Claude CLI already keeps that session's own .jsonl history.
• Structured answers sent via the answer WS message are recorded as the flattened text the router sends to Claude (e.g. Quick; React).

Closes #84

🤖 Generated with Claude Code

[zwanski-start]

🔍 QA report: ✅ Passed

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

The Web UI spec creation chat now writes .autoforge/prompts/spec_input_summary.md with the initial prompt, Claude's questions and options, and every user answer, as issue #84 asks. A full session over the real /api/spec/ws WebSocket with a mocked Claude client (no API calls) produced a correct transcript after every exchange, including structured answers, attachment names, UTF-8 text and the Complete status. Errored or abandoned sessions are still saved, and an unwritable file only logs a warning without breaking the chat. ruff is clean; mypy and pytest give the same results as master (pre-existing failures only); the UI build passes and lint has 0 errors.

	Check	Details
✅	Transcript has initial prompt, questions/options and answers (issue AC)	Ran a mocked full session over the real /api/spec/ws/{project} WebSocket: start → message → structured answer → message with PNG attachment → spec write. The file contains a header, an Initial Prompt section ('A todo app with tags') and alternating ### Claude / ### You turns. Option lists appear exactly as Claude wrote them, the structured answer is recorded as 'Quick; React, Tailwind', and the attachment is listed as mock.png.
✅	File location and status updates	Written to .autoforge/prompts/spec_input_summary.md next to app_spec.txt. The file exists right after start() (greeting only, 'In progress') and switches to 'Status: Complete' once app_spec.txt and initializer_prompt.md are both written and spec_complete/complete are sent.
✅	Error / partial conversation persisted	When the fake client raised RuntimeError mid-reply, the WS sent 'error' and the file still held the user message and the partial Claude text. A new session overwrote the earlier transcript, matching how start() resets app_spec.txt.
✅	Write failure degrades gracefully	With spec_input_summary.md replaced by a directory, the server logged 'Failed to write spec conversation summary: [Errno 21]' as a warning and both start and message still finished with response_done.
✅	Encoding / cross-platform	The file is written with an explicit encoding='utf-8' and the path is built from get_prompts_dir(); there are no hardcoded OS paths. Non-ASCII text ('Ünïcödé ✓') round-tripped correctly.
✅	Code review	The change is small and self-contained: a pure render function, a save call in finally blocks with no awaits, and a new 'attachments' key on the message dict that nothing else reads. Security settings, the allowlist and the localhost/origin middleware are untouched, and no secrets are written. Minor cosmetic point: text from separate assistant messages around a tool call is joined without a separator ('now.Done'), but the chat UI (useSpecChat.ts) joins it the same way, so the file matches what the user saw.
✅	Backend boots and spec REST routes respond	uvicorn server.main:app on 127.0.0.1:5820: /api/health 200, /api/spec/sessions 200 [], /api/spec/sessions/nope 404, /api/projects 200, / 200.
✅	UI smoke check in browser	The PR has no UI changes. The served SPA loads with 0 console errors at desktop and 390x844 phone sizes. test:e2e was not run because no UI files changed.
⏭️	CLI /create-spec path	Out of scope as the PR states: AutoForge doesn't see the interactive claude CLI conversation.

🧪 Commands run

Command	Result
ruff check .	All checks passed
mypy .	25 errors in 10 files — same as origin/master (spec_chat_session unused-ignore only moved from line 162 to 222)
python -m pytest -q	148 passed, 21 failed — the 21 test_devserver_security.py failures also happen on origin/master (144 passed, 21 failed)
python -m pytest test_spec_conversation_summary.py -v	4 passed
python test_security.py	ALL TESTS PASSED
python test_security_integration.py	ALL INTEGRATION TESTS PASSED
python probe_ws.py (throwaway mocked-ClaudeSDKClient WebSocket probe)	ALL PROBES OK — happy path, error path, unwritable path
npm --prefix ui run build	success (tsc -b && vite build)
npm --prefix ui run lint	0 errors, 6 warnings (already on master)
uvicorn server.main:app --host 127.0.0.1 --port 5820 + curl key routes	health 200, spec sessions 200/404, projects 200; server stopped afterwards

No browser screenshots were taken in this round.

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