cphu-aipla-app

Keep the AIPLA user-facing how-to guides (docs/guides/) up to date and in sync with the product. Use when the user says "update the guides", "regenerate / republish the guides", "the guides are stale / out of date", "re-capture the guide screenshots", "add a new guide", or mentions "guide-staleness" — and also proactively when a change lands on a UI surface a guide documents (the teacher class/activity/materials/co-pilot flows, the student join/workspace, or the researcher views). Covers the full pipeline: render, screenshot capture on deployed dev, publish into the app, seed the in-product corpus + onboarding tutors, and the staleness check.

In-app help assistant for teachers and researchers learning to use the AIPLA platform. Answers natural-language questions about setting up classes, building activities, adding curriculum materials, using the authoring co-pilot, the student experience, and the researcher views — grounded in the AIPLA how-to guides. Read-only: it explains how, it doesn't change anything. Visible to teachers and researchers (tagged role:teacher).

Teacher-facing co-pilot that helps a teacher author a good activity. The teacher describes, in plain Danish or English, what they want to teach; the assistant PROPOSES a Socratic lesson prompt (and, later, fitting elements) as editable suggestions the teacher accepts or edits — it never publishes on its own. Teacher-only (tagged role:teacher). Dark-flagged until the researcher teaching framework lands (designs 1.1.39 + 1.1.50).

Add a new MCP App artefact (hand-curated HTML/JS rendered as a sandboxed iframe in the chat workspace) to the AIPLA fork. Covers the static-artefact path (one shared mcp-sandbox Cloud Run service, files under infrastructure/mcp-sandbox/artefacts/<name>/v<version>/), the decision tree for "static vs. dynamic MCP server", the ADR-013 security gates (CSP, sandbox flags, 200 KB size limit, library-bypass review), and the frontend wiring (NEXT_PUBLIC_MCP_SANDBOX_URL, MCPAppToolCallRouter, workspace surface mount). Use when the user says "add a new MCP app", "new artefact", "build a sim", "deploy a sandbox iframe", "embed an interactive visualization", or references Boldkast / mcp-sandbox / mcp-ext-apps-map. Do NOT use for the agent-skills authoring workflow (that's the inherited template's skill creator) or for dynamic-MCP-server scaffolding (a separate v1 skill).

Teacher-facing analytics assistant. Ask natural-language questions about your class's session data — message counts, time-on-task, common misconceptions, group engagement. Only visible to teachers (tagged role:teacher). Data queries are scoped to the requesting teacher's own classes; no cross-teacher data exposure.

Teacher-facing hub for running your classes from chat — the conversational alternative to the React /teacher dashboard. Create a class, list your classes, mint group join-codes, look up your activities, and ask about a class's engagement (it consults the analytics assistant for session-data questions). Same backend, one chat. Only visible to teachers (tagged role:teacher).

Wire a new student-facing workbench/activity element so it reaches the tutor on BOTH surfaces — the iframe-context state push AND the visible "shared with the AI" human-tool-use trust card — AND make it proposable by the activity-authoring co-pilot (the add_element tool). Use when adding a workbench element, a new activity element, when a student interaction (table/calculator/checklist/etc.) does not show in the chat, when "the AI didn't see what the student did", when wiring or debugging the trust card, when a new element isn't offered by the authoring co-pilot, or reviewing an element PR for the dropped-card or dropped-co-pilot-coverage bug.

Triage runbook for AIPLA's dependency security pile. Use whenever the CI gate's `security-audit` job fails on a PR, when the Monday weekly rolling issue surfaces a new CVE, or when a developer says "run the security audit", "check dependabot", "vulnerability sweep", "triage the security pile", or "is this gate going to pass". Encodes the reachability rubric (direct prod / transitive / dev-only / deprecated) developed during the 2026-06-05 20-alert sweep, plus the per-ecosystem commands (npm audit, pip-audit via uvx, dependabot via gh api) and the npm-override conflict pattern. The policy of record is docs/design/aipla/v1.1.0-feedback/security-monitoring-pipeline.md — this skill is the human-facing operationalisation of it.

Create design documents for Aitana Labs features in the correct format and location. Use when user asks to create a design doc, plan a feature, write a spec, or document a design. Also use when user says "design", "RFC", "proposal", or wants to think through a feature before implementing. Handles both planned and implemented docs.

English-language Socratic kinematics tutor for NCERT/CBSE Class 11 Physics. Paired with an interactive workbench (seven simulations, motion graphs, a topic-keyed formula reference, notes, and a per-topic quiz). The tutor sees which topic the student picked, which sim they ran with what parameters, and how their quiz answers are going.

Dansk stx fysik-A virtuelt laboratorium — bestem Plancks konstant ved at måle tærskelspændingen for seks LEDs med forskellig farve. Sokratisk tutor guider eleven gennem fire trin: kredsløbssamling, I-U-karakteristik, spektroskopi, og rapport. Tutoren afslører aldrig formlen direkte.

Danish stx physics tutor for ONE specific problem — projectile motion ("Boldkast", v0=15 m/s @ 40°). Scaffolds students through the sub-steps without giving the final answer. v0.1 demo skill; v1 ships problem-set-helper-config so teachers can author their own per-class tutors with different seeded problems.

Standalone Socratic concept-exploration tutor for Danish stx physics — NO simulator, chat-only. The teacher sets the topic and focus via their activity configuration ({teacher_focus}); the tutor draws the student into a dialogue about the concept rather than solving a numbered problem. The base "engine" skill for teacher-authored no-workbench concept activities (v1.1 teacher-activity-authoring, TAA-1).

Vendored protocol specifications for the agent platform's full stack: Agent Skills (SKILL.md format), AG-UI (streaming transport), A2UI (declarative UI), MCP (tools/resources/prompts), and MCP Apps (interactive iframe UIs). Use when writing a design doc that involves any of these protocols, when implementing a tool that emits A2UI or MCP App content, when wiring AG-UI events, or when answering "is this A2UI, MCP App, or AG-UI?" — they get confused often. References are vendored at fetch time so they survive spec-site outages and are quotable from disk. Do NOT use for project-specific ADK wiring (that's adk-cheatsheet) or for deployed-service ops (that's the *-deploy skills).

How to inspect and verify ADK session state — events, artifacts, traces — on the running Aitana v6 backend using the ADK-native HTTP endpoints that ship for free with `get_fast_api_app(web=True, ...)`. Load when the user asks "are sessions actually being saved?", "where do messages live?", "how do I see what the agent saw?", "can I view artifacts the loader produced?", "how do I use `adk web` / `adk api_server` against this backend?", or is debugging a session that looks empty in the UI but should have events. Also load when verifying that `make_document_loader` saved the right `doc:{id}.json` artifact, when reproducing a bug from a known threadId, or when handing a session over to `adk eval`. Covers the app_name / user_id / session_id triple, the agents_dir-vs-APP_NAME quirk that makes the dev UI's app picker misleading, and the relationship between ADK's canonical store and the Firestore `chat_sessions` mirror.

How to drive a real Chrome instance via the chrome-devtools MCP to verify Aitana v6 frontend changes end-to-end — rendered DOM, console errors, /api/proxy network traffic, AG-UI SSE streams, and Firebase auth state. Load whenever a frontend change has been made and you need to confirm it *actually works* before reporting success, or when debugging behavior that static checks (tsc/eslint/vitest) cannot see. Trigger phrases include "verify this in the browser", "did the UI change work", "check the console", "watch the network panel", "is the SSE stream connecting", "click through the chat flow", "screenshot the page", "run lighthouse", "is fetchWithAuth being used", "are we getting 401s", "reproduce the bug in the browser", "test the threadId/doc-injection/session-delete flow". Also load when investigating a UI regression report from the user where the description involves something that only manifests at runtime (streaming, auth, layout, hydration). Do NOT load for pure static-code questions (rename a prop, re