[AdeRistanto]

Pull Request

Description

Adds a Quick PRD mode: a complete Product Requirements Document is generated in one streaming LLM call and shown in a new web UI. Two templates are available: fullstack_nextjs (the default, "PRD Full Stack untuk AI Coding (Starter Next JS)", written in Indonesian) and general. This PR also hardens reference-website fetching against SSRF and validates every generated PRD automatically.

Type of Change

• 🐛 Bug fix (non-breaking change which fixes an issue)
• ✨ New feature (non-breaking change which adds functionality)
• 💥 Breaking change (fix or feature that would cause existing functionality to not work as expected)
• 📚 Documentation update
• 🔧 Code refactoring (no functional changes)
• ⚡ Performance improvement
• 🧪 Test addition or improvement
• 🔨 Build/CI changes (dependency declaration only)
• 🎨 Code style/formatting changes

Changes Made

Quick PRD generator (gtplanner/agent/prd_generator.py, POST /api/prd/stream in fastapi_main.py, static/index.html)

• Streams the PRD as SSE events: status, reference, chunk, validation, done.
• Fixed template parts are inserted verbatim: the header (usage notes plus a section 0 setup prompt with NGODINGPAKEAI_TOKEN taken from .env) and the footer. The LLM fills prd_templates/fullstack_nextjs/skeleton.md.
• Revisions: the previous PRD is sent back without its fixed parts and the whole document is regenerated.
• The latest framework is in the template:
  • architecture flowchart and sequence diagrams for critical flows
  • "used" markers for give-backs
  • provider deadline and cancel rules
  • follow-up flags, owner columns and configurable limits
  • a 29-item checklist and six cross-checks

SSRF hardening of reference fetching (fetch_reference). URLs found in the request are fetched with these rules:

• Only http/https on ports 80/443. URLs with userinfo are rejected, and so are internal host names (localhost, *.internal, *.local, single-label names).
• Every DNS answer must be public. If any address in the answer is not public, the request is rejected. Explicit IPv4/IPv6 deny lists cover cloud metadata (169.254.169.254, 168.63.129.16, 100.100.100.200, fd00:ec2::254), CGNAT, ULA, link-local, 6to4 and Teredo. IPv4-mapped and NAT64 addresses are unwrapped first. The check does not rely on is_global alone.
• The connection is pinned to the validated IP, with the original Host header and sni_hostname. TLS verification stays on.
• Redirects are handled manually (max 3), and each hop is validated independently.
• trust_env=False, so proxy environment variables are ignored.
• Response limits: only text/html, application/xhtml+xml and text/plain; Content-Length over 1 MB is rejected; the streamed body is capped at 1 MB.
• Timeouts and concurrency: 15 s total per URL, up to 4 concurrent fetches across the server, and references within one request are fetched in parallel.
• The client only sees generic error messages. Details go to the server log.
• Kill switch: PRD_REFERENCE_FETCH=0 disables fetching completely, with no network access.

Post-generation validation (validate_prd, shown per item in the UI)

• finish_reason and max_tokens are reported. length/max_tokens marks the PRD as truncated, and a truncated PRD is never shown as done.
• Checks:
  • no leftover [placeholder] outside code
  • all required headings (0–12, "Hasil pengecekan silang", "Pertanyaan terbuka")
  • an even number of code fences
  • every actor in sequenceDiagram arrows declared as a participant
  • the repo version (section 0 removed) contains no ngpk_ or NGODINGPAKEAI_TOKEN value

Config

• settings.toml: llm.max_tokens = 65536. Without it the provider's default cut long PRDs off.

⚠️ Unrelated changes included in this PR

Whole files were committed, so this PR also carries some earlier local edits to the same files that are not part of the feature:

• pyproject.toml / uv.lock: adds truststore>=0.10. The code that uses it (gtplanner/__init__.py) is not in this PR, so the dependency is currently unused.
• settings.toml: Chinese comments translated to English; timeout = 600 and max_retries = 1 for llm.
• fastapi_main.py: everything changed in this file, not only the /api/prd/stream route and serving static/index.html at /.

I can split these into a separate PR if you prefer.

Testing Performed

• CLI functionality tested
• API endpoints tested (if applicable)
• MCP service tested (if applicable)
• Error handling tested
• Edge cases tested

uv run pytest tests -q
# 226 passed on the full working tree.
# 188 passed on the committed snapshot alone (test_prd_generator, test_prd_reference_security,
# test_prd_validation, test_basic), verified to import the snapshot's modules.

• tests/test_prd_reference_security.py (135 tests) runs fully offline: DNS is faked, HTTP goes through httpx.MockTransport, and httpcore's real connect_tcp is replaced by a recorder, so no socket is ever opened. The tests show that the real connection layer connects to the pinned IP and that start_tls receives the original host name with CERT_REQUIRED and check_hostname=True. Mutation check: disabling each control in memory makes the suite fail (IP checks 69, pinning 18, kill switch 6, error sanitising 68 failures; without the size cap a test never finishes).
• tests/test_prd_validation.py (43 tests) covers a passing and a failing case for each check, plus finish_reason handling. Each check was also disabled in memory to confirm the tests fail.
• Manual end-to-end run through /api/prd/stream (e-commerce PRD with Midtrans and shipping):
  • glm-5.3-flash: finish_reason=length after about 22 min. The PRD was correctly flagged as truncated, probably because reasoning tokens count against max_tokens.
  • deepseek-v4-flash: finish_reason=stop after about 5.5 min, about 121k characters, all 5 checks passed.

Not verified: real-network behaviour of the SSRF hardening (an actual TLS handshake to a pinned IP, OS/container DNS and IPv6 behaviour). Please test from a staging environment before relying on it in production.

Environment Tested

• OS: Windows 11
• Python: 3.11.18 (uv), httpx 0.28.1, httpcore 1.0.9
• LLM: OpenAI-compatible proxy (glm-5.3-flash, deepseek-v4-flash)

Breaking Changes

• No breaking changes
• Yes, breaking changes (describe below)

The existing agent planner endpoints are unchanged. Note that llm.max_tokens now applies to every LLM call.

Documentation

• Code comments added/updated
• README.md updated (if needed)
• CONTRIBUTING.md updated (if needed)
• API documentation updated (if applicable)
• No documentation changes needed

Dependencies

• No new dependencies

• New dependencies added (list below)

• httpx>=0.28.1: declared explicitly. It was already installed through openai. The locked version is unchanged (no other package versions changed in uv.lock).

• truststore>=0.10: see "Unrelated changes" above.

Performance Impact

• No performance impact
• Performance improvement
• Potential performance regression (explain below)

A full-stack PRD takes several minutes to generate and uses a large system prompt (about 12k characters plus the skeleton). Reference fetches add at most 15 s and now run in parallel.

Security Considerations

• No security implications
• Security improvements
• Potential security concerns (explain below)

Out of scope and still open:

• /api/prd/stream has no authentication.
• CORS is *.
• The server binds to 0.0.0.0.
• Other fetchers (URL_to_Markdown.py, search.py via Jina) are unchanged.
• Section 0 of the fullstack template contains the NGODINGPAKEAI_TOKEN value. The UI's "Download untuk repo" strips it, and the validation checks that.

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[AdeRistanto]

Opened by mistake (intended for a personal repository). Closing, sorry for the noise.