docs: align agent instructions and README guidance (#323)

This commit is contained in:
Daniel Volz
2026-02-25 21:35:35 +01:00
committed by GitHub
parent bfc9aaaa6d
commit aa29d1c699
6 changed files with 104 additions and 81 deletions
+5 -5
View File
@@ -13,6 +13,7 @@ You are the release manager for **MedAssist-ng**. Your job is to guide code from
## Critical Safety Rules
- **NEVER release, tag, push, or create PRs without explicit user confirmation at each step.** Always present your plan and wait for approval.
- **This specialist agent is the only agent allowed to perform remote release operations after explicit confirmation.**
- **NEVER push directly to `main`** — GitHub will reject it (`GH013: Repository rule violations`). All changes go through Pull Requests.
- **NEVER skip CI checks.** Wait for all status checks to pass before merging.
- **Testing ownership belongs to `@testing-manager`**. Do not plan or implement tests in this agent; request/hand off to testing-manager when testing work is required.
@@ -48,12 +49,11 @@ This repository intentionally uses only two operational agents for CI/CD handoff
- Never use `gh` commands that can open an interactive pager and block execution (requiring `q`).
- Always run `gh` commands in non-interactive mode using `GH_PAGER=cat` (or `--no-pager` where supported).
- Do not use these commands in agent flows:
- `gh pr view 155 --json statusCheckRollup --jq '.statusCheckRollup[] | {name:.name,conclusion:.conclusion,detailsUrl:.detailsUrl,workflowName:.workflowName}'`
- `SHA=$(gh pr view 155 --json headRefOid --jq .headRefOid) && gh api repos/DanielVolz/medassist-ng/commits/$SHA/check-runs --jq '.check_runs[] | {name,conclusion,details_url,html_url,app:.app.name}'`
- Use safe variants instead:
- Avoid hardcoded PR/repo examples in instructions; always use parameterized placeholders.
- Use safe command patterns:
- `GH_PAGER=cat gh pr view <PR_NUMBER> --json statusCheckRollup --jq '<jq-filter>'`
- `GH_PAGER=cat gh api repos/<owner>/<repo>/commits/<sha>/check-runs --jq '<jq-filter>'`
- `SHA=$(GH_PAGER=cat gh pr view <PR_NUMBER> --json headRefOid --jq .headRefOid)`
- `GH_PAGER=cat gh api repos/<owner>/<repo>/commits/$SHA/check-runs --jq '<jq-filter>'`
---
+5 -2
View File
@@ -15,6 +15,8 @@ You are the testing manager for **MedAssist-ng**. Your job is to ensure every fe
- **Tests are mandatory**: Every new feature and every bug fix MUST have corresponding tests.
- **Fix bugs, don't test around them**: If behavior is incorrect, fix the implementation first, then write tests for correct behavior.
- **Run tests non-interactively**: Use `CI=true` where required to avoid watch-mode hangs.
- **Playwright must disable auto-open reports**: Always prefix Playwright runs with `PLAYWRIGHT_HTML_OPEN=never`.
- **Keep CI E2E stable**: Use `PLAYWRIGHT_WORKERS=1` in CI unless a change is explicitly requested.
- **Never start interactive report servers**: Do not run commands that wait for manual input (for example Playwright HTML report server: `Serving HTML report ... Press Ctrl+C to quit`). Always use finite, non-interactive commands and reporters.
- **No remote git operations**: Do not push, merge, create PRs, tags, or releases. Hand over to `@release-manager` when ready.
- **Keep scope focused**: Do not fix unrelated failures unless explicitly requested.
@@ -66,8 +68,9 @@ cd frontend && npm run build
### Playwright E2E
```bash
cd frontend && npm run test:e2e
cd frontend && npm run test:e2e -- --project=chromium
cd frontend && PLAYWRIGHT_HTML_OPEN=never npm run test:e2e
cd frontend && PLAYWRIGHT_HTML_OPEN=never PLAYWRIGHT_WORKERS=4 npm run test:e2e:local
cd frontend && PLAYWRIGHT_HTML_OPEN=never npm run test:e2e -- --project=chromium
# Never use interactive UI/headed/report-server commands in agent runs.
# Do not use: npm run test:e2e:ui, npm run test:e2e:headed, npx playwright show-report
```
+8 -72
View File
@@ -1,77 +1,13 @@
# MedAssist-ng - AI Coding Instructions
# MedAssist-ng - Copilot Entry Point
## Purpose
Use `AGENTS.md` as the single source of truth for all governance, workflow, and skill rules.
Use `AGENTS.md` as the canonical governance source. Read the referenced skill files before starting any task.
## Required Startup Steps
## Project Orientation (Read First)
1. Read `AGENTS.md` first.
2. Identify triggered skills from `AGENTS.md` and read each referenced `SKILL.md` before making changes.
3. Follow delegation boundaries exactly (`@testing-manager` for testing, `@release-manager` for release orchestration).
- **Product**: MedAssist-ng is a medication planner with stock tracking, reminders (email/push), refill history, and schedule sharing.
- **Tech stack**: React + TypeScript + Vite (`frontend/`), Fastify + TypeScript + Drizzle + SQLite (`backend/`).
- **Request path**: Frontend uses `/api/*` only; backend route handlers live in `backend/src/routes/`.
- **Primary backend modules**:
- Auth/SSO: `backend/src/routes/auth.ts`, `backend/src/routes/oidc.ts`, `backend/src/plugins/auth.ts`
- Medications/data: `backend/src/routes/medications.ts`, `backend/src/db/schema.ts`
- Reminders: `backend/src/services/reminder-scheduler.ts`, `backend/src/routes/planner.ts`, `backend/src/routes/settings.ts`
- **Primary frontend modules**:
- Pages: `frontend/src/pages/`
- Shared app state: `frontend/src/context/AppContext.tsx`
- Domain hooks: `frontend/src/hooks/`
- Translations: `frontend/src/i18n/en.json`, `frontend/src/i18n/de.json`
## Scope
Use this orientation for quick navigation before applying the rules below.
## Always-On Rules
- English only for project artifacts.
- **NEVER run remote git commands** — no `git push`, no `gh pr create/merge`, no `gh release`, no `git tag`. Prepare locally, then hand off to `@release-manager`.
- Testing work belongs to `@testing-manager`.
- PR/release/CI orchestration belongs to `@release-manager`.
- Keep changes local, focused, and consistent with existing UI/API patterns.
- **Hard PR scope + size rule**: one cohesive objective per PR; if scope drifts or diff becomes large (target <= 500 changed lines, hard split at ~800+), split into logical follow-up PRs instead of bundling.
- Remove obsolete code when re-implementing — never leave dead code behind.
- **Document behavioral discoveries**: When you discover or clarify how a feature works (e.g., what triggers notifications, how thresholds interact, which code paths exist), **always** add or update the relevant section in `doku/APP_BEHAVIOR.md`. This is mandatory — do not rely on conversation context alone.
## MedAssist Essentials
- Frontend calls backend through `/api/*`.
- DB changes must stay backward-compatible (schema default + alter migration + null-safe reads).
---
## Skills (MANDATORY — read before every task)
Before starting any task, identify which skills apply and **read their full SKILL.md file** for detailed rules.
| Skill | Trigger | File |
|---|---|---|
| **Architecture Guard** | API endpoints, frontend API calls, routing, code placement | `.github/skills/medassist-architecture-guard/SKILL.md` |
| **DB Compatibility** | Persisted data, schema changes, migrations | `.github/skills/medassist-db-compat-check/SKILL.md` |
| **i18n Enforcer** ⚠️ | Any user-facing text in frontend or backend | `.github/skills/medassist-i18n-enforcer/SKILL.md` |
| **UI Consistency** | UI flows, modals, buttons, forms, settings | `.github/skills/medassist-ui-consistency/SKILL.md` |
| **Frontend Polish** | Visual quality improvements | `.github/skills/medassist-frontend-polish/SKILL.md` |
| **Security Sanity** | Backend routes, auth, file handling, external input | `.github/skills/medassist-security-sanity/SKILL.md` |
| **Observability Guard** | Services, schedulers, startup, failure handling | `.github/skills/medassist-observability-guard/SKILL.md` |
| **Config Change Guard** | `.env`, Docker, Vite proxy, runtime defaults | `.github/skills/medassist-config-change-guard/SKILL.md` |
| **Doc Sync Guard** | Behavior changes, setup, env vars, workflows | `.github/skills/medassist-doc-sync-guard/SKILL.md` |
| **Testing Handoff** | Writing/running tests, CI test failures | `.github/skills/medassist-testing-handoff/SKILL.md` |
| **Release Handoff** | Branch push, PR, merge, tagging, release | `.github/skills/medassist-release-handoff/SKILL.md` |
| **Skill Quality Review** | Creating/modifying skills | `.github/skills/medassist-skill-quality-review/SKILL.md` |
### Non-negotiable parity rules (always apply)
1. **Desktop + Mobile Parity**: Medication edit has two paths — `MedicationsPage.tsx` (desktop) and `MobileEditModal` (mobile). **Always update BOTH**.
2. **Notification Dual Code Paths**: Notifications have two code paths — `backend/src/services/reminder-scheduler.ts` (scheduler) and `backend/src/routes/planner.ts` (manual). **Always update BOTH**.
---
## Delegation
- **Testing handoff → `@testing-manager`**: test planning, writing, execution, CI test triage.
- **Release handoff → `@release-manager`**: PR/release orchestration, merge flow, workflow monitoring.
## Key References
- Canonical governance: `AGENTS.md`
- Skill files: `.github/skills/*/SKILL.md`
- Specialist agents: `.github/agents/testing-manager.agent.md`, `.github/agents/release-manager.agent.md`
This file intentionally stays minimal to prevent duplicated or conflicting instructions.
+1 -1
View File
@@ -13,7 +13,7 @@ Use one governance source to avoid duplicated or conflicting policy text.
## Skills
- `medassist-karpathy-core` — enforce assumption clarity, simplicity, surgical diffs, and verifiable execution.
- `medassist-karpathy-core` — enforce think-before-coding, simplicity-first changes, surgical diffs, and goal-driven verification.
- `medassist-architecture-guard` — enforce frontend/backend boundary and `/api/*` data-flow conventions.
- `medassist-db-compat-check` — enforce backward-compatible SQLite/Drizzle schema changes.
- `medassist-i18n-enforcer` — enforce translation-key-only UI copy with EN/DE parity.
@@ -0,0 +1,69 @@
---
name: medassist-karpathy-core
description: Apply assumption clarity, simplicity-first implementation, surgical diffs, and goal-driven verification for non-trivial coding tasks.
---
# Skill Instructions
Use this skill as an execution style layer for implementation tasks where overengineering, broad refactors, or unclear assumptions are likely.
## Use When
- The request is ambiguous and assumptions must be made explicit.
- The change can easily balloon in scope.
- A bug fix or feature needs explicit success criteria and verification.
- You need to keep diffs minimal and directly tied to the request.
## Do Not Use When
- The task is trivial and can be completed safely without extra process overhead.
- The task is only about ownership routing (use `medassist-testing-handoff` / `medassist-release-handoff`).
- The task is only about domain guardrails already covered by specialized skills (architecture, DB, i18n, UI, security, config, observability).
## Core Principles
### 1. Think Before Coding
- Do not assume silently.
- State assumptions explicitly.
- If multiple interpretations exist, present them instead of picking one invisibly.
- If uncertain or blocked by ambiguity, stop and ask.
- If a simpler approach exists, call it out.
### 2. Simplicity First
- Implement the minimum code required to solve the asked problem.
- Do not add speculative features, abstractions, or configurability.
- Avoid defensive handling for impossible scenarios.
- If the solution feels overcomplicated, simplify before finalizing.
### 3. Surgical Changes
- Touch only lines required for the request.
- Do not refactor unrelated areas.
- Match existing local style and patterns.
- Remove only unused code introduced by your own change.
- If unrelated dead code is discovered, mention it but do not remove it unless requested.
### 4. Goal-Driven Execution
- Translate requests into verifiable outcomes before implementation.
- For multi-step tasks, define short steps with checks.
- Verify the requested behavior explicitly before declaring done.
Example execution frame:
```text
1. [Step] -> verify: [check]
2. [Step] -> verify: [check]
3. [Step] -> verify: [check]
```
## Response Format
When this skill is used, report briefly:
- Assumptions made (or clarifications requested)
- Why the chosen approach is the simplest viable one
- What was changed (and what was intentionally not changed)
- Verification performed and result