New Server Submission: PickySteve

[KernelLord]

Repository: https://github.com/KernelLord/pickysteve
Version: 0.1.0
MCP Protocol: 2025-06-18 (stdio transport via the mcp Python SDK)
Language: Python 3.11
Category: Community Servers (submitted to Under Review per process)

Description:
Skill router and context picker for coding agents. A cheap local model routes each prompt to the right skill via hybrid BM25 + embedding retrieval (RRF-fused), cross-encoder rerank with a calibrated floor, and an LLM judge, then assembles one minimal context bundle. Exposes pick_context(request) and list_skills() over stdio.

Security Features:

• Dual-scan prompt-injection gate (stackone-defender, ONNX): scans the raw request AND every retrieved document before assembly, fail-closed
• Per-call random-nonce boundaries around untrusted content (static delimiters proved forgeable — finding documented in SECURITY_AUDIT.md)
• Red-team results: 100% detection / 0 bypasses on a 180-payload corpus (14 evasion families), 0/43 false positives on the real skill registry
• CI: SHA-pinned actions, minimal token permissions, OSV-Scanner, Trivy, CodeQL, Bandit, OSSF Scorecard, Dependabot

Required Permissions:

• Local filesystem read of its own skill registry (no network required in the security path; local Ollama by default)
• Optional outbound HTTPS only if configured with a cloud model endpoint

Deployment:

• Install: git clone + uv (Python 3.11 venv); no PyPI package yet
• Config via PS_* environment variables (see pickysteve/config.py)
• Connector installer wires 18 coding agents (MCP stdio + OpenAI-compatible proxy on :8077)

Known security considerations (documented in the repo):

• Latin-script non-English injection (e.g. Spanish) can bypass the bundled English-only classifier in some cases
• Thresholds are calibrated per judge model; swapping models requires re-running eval/calibrate.py

Disclosure: I'm the project author; this PR was prepared by Claude (agent) on my behalf. Happy to have the scanners tear into it — that's the point.