image-scoring-backend

Bump semver + promote CHANGELOG Unreleased via compiled harness. Use for /release, version bumps, or tagging. Runs scripts/agent_skills/release_bump.py; LLM only when Unreleased is ambiguous. Never commit/push unless the user asks.

Vexlum Scoring MCP — compact search+dispatch on is-be-mcp / is-be-live and gallery is-ui-mcp / is-ui-live.

Choose CLI tools and safe patterns before repo work. Install checklist, bounded search/read commands, and confirmation gates. Use when navigating the codebase, running shell commands, or setting up agent tooling on Windows, WSL2, or macOS.

Inspect JSON, YAML, and API responses with jq, yq, dasel, curl, and httpie. Use for package.json, config.json, and backend health checks. Never output secrets or full credential-bearing config.

Backend build, test, lint, and verify workflows via WSL pytest, ruff, doctor, and docker compose. Use before claiming work complete or opening a PR. Do not invent ad hoc Python commands outside documented venvs.

Choose and run the right search tool: rg vs grep vs ast-grep vs fd vs fzf. Bounded repo navigation with ripgrep, fd, tree, bat, and structural search. Use when finding symbols, files, syntax patterns, or deciding which CLI search to use before reading or editing code.

Use before letting an agent run without step-by-step supervision — unattended or overnight runs, long scoring/backfill jobs, "keep iterating until it passes", ratchet loops, subagent fan-out, or swarm work. Also use to decide how much orchestration a task needs (single call vs loop vs chain vs parallel vs DAG). Produces a written contract with metric, budget, revert rule, and stop conditions before the run starts.

Vexlum Scoring MCP — compact search+dispatch on is-be-mcp / is-be-live and gallery is-ui-mcp / is-ui-live.

Routine read-only debugging for the image-scoring Python backend via MCP—is-be-mcp search+dispatch, scoring/tagging failures, job errors, Postgres questions, DB integrity, and config sanity.

Use at the end of a session, after a correction, or when the user says bake this in, turn this into a skill, update the skill based on this conversation, capture the lesson, or don't make that mistake again. Analyzes the conversation for corrections, mistakes, repeated actions, and workarounds, then enriches an existing skill/command/rule or creates a new one. For mining an external document or repo instead, use /mine.

Compare MCP code-intelligence layers — CLI wrappers, ast-grep, symbol/graph tools, Zoekt, embeddings. Use when choosing search/dispatch vs heavyweight indexes. Backend domain MCP is image-scoring-mcp (is-be-mcp).

Use image-scoring MCP tools to debug scoring failures, check database health, and investigate system issues. Trigger when the user asks to debug scoring, investigate failure, or check database health.

Routine read-only debugging for the image-scoring Python backend via MCP—is-be-mcp search+dispatch, scoring/tagging failures, job errors, Postgres questions, DB integrity, and config sanity.

Cross-repo GitHub Project board is the canonical task queue. Use whenever picking work, claiming an issue, transitioning Stage, or filing/closing a backlog issue across image-scoring-backend or image-scoring-gallery.

Maintain image-scoring-backend docs/ as an OKF-aligned wiki: frontmatter, planning vs features/planned, guides, architecture, indexes, and log. Triggers: wiki maintenance, docs update, documentation audit, wiki ingest/lint/query.

Use when the user asks to commit and push, publish, ship, or prepare a release commit. Guides staging only intended changes, writing a Conventional Commit, committing, and pushing to origin after verifying status and diff.

Use Graphify knowledge-graph MCP (graphify-be) or CLI for architecture / cross-module connectivity. Prefer after rg/fff fail for “how does X connect to Y”; never for pipeline/DB triage (use is-be-mcp).

Apply Karpathy's anti-pattern checklist before/during non-trivial coding — surface assumptions, resist overcomplication, keep diffs surgical, frame work as verifiable goals. Use on ambiguous or multi-step tasks, or when the user asks for a "Karpathy pass".

Use when creating, reviewing, or improving first-party agent skills in this repo. Apply whenever a user asks to add a skill, edit SKILL.md, optimize skill triggering, or adapt patterns from Anthropic's public skills into the canonical .cursor skill tree.

Use when encountering a bug, failing test, build failure, unexpected behavior, flaky behavior, regression, performance anomaly, or integration issue. Apply before proposing fixes so the agent finds root cause, compares working patterns, tests one hypothesis at a time, and avoids guess-and-check patches. Also use when the user asks to diagnose, investigate, root-cause, or fix a flaky failure.

Use when implementing a feature, bug fix, refactor, or behavior change that can be tested. Apply before production code changes to enforce red-green-refactor, prove tests fail for the intended reason, and keep implementation minimal. Also use for risky vertical slices or when the user asks for TDD, red/green/refactor, or tests first.

Use before claiming work is complete, fixed, passing, ready to commit, or ready for PR. Apply to ensure fresh command output supports every success claim and to report warnings or failures honestly.

Verify implementation against spec AC-n criteria via compiled harness. Use after /implement or /test-and-fix, before /pr-ready. Parses ACs and emits the report; LLM assigns Verified/Failed/Unknown when evidence is not a clean command exit.

Safe git and GitHub CLI workflows — status, bounded diffs, gh issues/PRs. Use before commits, when reviewing changes, or for backlog/PR operations. Destructive git commands require user confirmation.

Log agent sessions, consolidate project memory (dream), promote reviewed memory, and load context for new chats. Use when ending a session, improving cross-session recall, or when the user mentions agent memory, dream, log-session, or memory.md.

Choose Windows native vs WSL2 for agent shell work. Backend Python/GPU, pytest, and Unix-path MCP in WSL2; gh and light search on Windows. Use when picking where to run installs, tests, or MCP servers.

PostgreSQL custom-format backup for image-scoring-backend via Backup-Postgres.ps1. Use when the user runs /backup-db, asks for a local pg_dump, database backup, or Postgres dump. Default workflow keeps at most 3 dumps in backups/postgres and mirrors the latest copy to D:\Dropbox\Photos\Scoring (also capped at 3 files).

Capture task quality signals and log them to agent memory to build a feedback loop. Use at the end of each implemented task or merged PR, or whenever a task required more than one agent iteration.

Design and implement an MCP server's tools, resources, and prompts with safe transport and input validation. Use when building or extending a Model Context Protocol server for agent integration.

Runs detect_subagents then run_subagent (codex/gemini) in review mode via subagent-orchestrator MCP, reads .agent-runs outputs, and summarizes findings. Use when the user asks for external CLI review, codex/gemini review, subagent panel, tie-breaker, or invokes /check-subagents, /run-codex-review, /run-gemini-review, /run-subagent-review.

Threat-model agentic tooling — MCP/tool abuse, hook/prompt injection, remote control surfaces, and secret leakage. Use during security reviews or when adding features that expose the system to AI agents or remote callers.

Prevent Windows sleep and display timeout via SetThreadExecutionState. Use when the user asks to keep Windows awake, pause sleep, prevent screen off, or stop sleep during long jobs. Starts a detached hidden PowerShell worker (not a Cursor background terminal).

Set up, run, and maintain the WSL2 environment image-scoring-backend depends on — Ubuntu distro, ~/.venvs/tf (app/scripts/ML) and ~/.venvs/image-scoring-tests (pytest -m wsl), GPU/CUDA passthrough, Docker Desktop Postgres, Firebird libs. Use when provisioning the venvs, launching long-running GPU jobs robustly, or recovering from WSL instability (Wsl/Service/E_UNEXPECTED, distro Stopped, OOM-kill, disk bloat). For which venv a single command needs, see wsl-tf-python-runner; this skill owns the environment lifecycle.

Runs Python, scripts, and tests the way image-scoring-backend expects — WSL, ~/.venvs/tf for app and scripts that import modules/DB/ML, separate ~/.venvs/image-scoring-tests for pytest -m wsl, LD_LIBRARY_PATH for bundled Firebird when DB/Firebird FFI is involved. Use when running launch.py or webui.py, scripts under scripts/, anything importing modules.*, resolving Windows vs WSL Python confusion, or choosing the correct pytest environment and markers.

Read-only scan for files ≥1000 LoC and functions/methods ≥150 LoC across image-scoring-backend and image-scoring-gallery. Produces a markdown or JSON report with refactor priority hints. Use when the user asks for a codebase size audit, large file scan, god-module review, or refactoring hotspot analysis.

Implements or styles the image-scoring-backend React SPA at /ui/, Gradio design-token sync, and frontend/ components using @synthet/image-scoring-design. Use for frontend/ changes, visual components, Tailwind theming, or design:check failures—not Python backend modules unless API contract is in scope.

GitHub backlog housekeeping for image-scoring-backend and image-scoring-gallery: sync Project board Stage, label hygiene, tier-1 closes. Use when the user runs /backlog-housekeeping, asks to close stale or obsolete issues, fix board drift, or clean up issue labels on the synthet Project #1 queue.

Development workflow, running commands, and debugging patterns for the Gradio WebUI.

Runs detect_subagents then run_subagent (codex/gemini) in review mode via subagent-orchestrator MCP, reads .agent-runs outputs, and summarizes findings. Use when the user asks for external CLI review, codex/gemini review, subagent panel, tie-breaker, or invokes /check-subagents, /run-codex-review, /run-gemini-review, /run-subagent-review.

Runs Python, scripts, and tests the way image-scoring-backend expects — WSL, ~/.venvs/tf for app and scripts that import modules/DB/ML, separate ~/.venvs/image-scoring-tests for pytest -m wsl, LD_LIBRARY_PATH for bundled Firebird when DB/Firebird FFI is involved. Use when running launch.py or webui.py, scripts under scripts/, anything importing modules.*, resolving Windows vs WSL Python confusion, or choosing the correct pytest environment and markers.