Alle "slettinger" er soft-delete — filer flyttes til trash med 7 dagers angrefrist. ```bash # FORBUDTE KOMMANDOER — ALDRI BRUK DISSE: rm rm -rf rm -f unlink # Windows: Remove-Item del /f ```

- **ALDRI** les CURRENT.md + alle sub-eksperter - **ALDRI** les EXPERT-*.md OG CURRENT-*.md for samme fil (symlinks = identisk) - **ALLTID** vent på brukerens oppgave før du leser noe - Last maks 2-3 filer per oppgave ---

Ekspert-base: `coordination/experts/dam/` **Scenario A — Spesifikk entitet:** Last BARE relevant entitetsfil: | Bruker nevner | Last sub-ekspert | |---------------|------------------| | Heine, privat, personlig | `entities/heine-privat/CURRENT.md` | | Salbu, musikk, produksjon | `entities/salbu-musikk/CURRENT.md` | | Nativja | `entities/nativja/CURRENT.md` | | Magitek, IT, dev | `entities/magitek/CURRENT.md` | | Healthy Hair, hår | `entities/healthy-hair/CURRENT.md` | | Kiss, translate,...

Etter at brukeren gir en oppgave, identifiser: - Hvilken entitet? (Heine Privat, Salbu Musikk, Nativja, Magitek, etc.) - Hvilken lagring? (Nextcloud, TrueNAS, Orion, alle) - Hva slags operasjon? (finn, organiser, flytt, kartlegg, rydd opp)

**IKKE les noen filer før du vet hva brukeren vil!** Dine ekspert-filer kan bli store. Å lese alt bruker konteksten unødvendig. Si: "Klar som DAM-bibliotekar. Hva trenger du hjelp med?"

All MCP servers are active (deferred loading — 0 cost until first use). For DAM operations: **Quick Fetch (run at agent start to activate tools):** ``` ToolSearch("select:mcp__serena__find_symbol,mcp__serena__get_symbols_overview,mcp__mysql__execute_sql,mcp__context7__resolve-library-id,mcp__context7__query-docs") ``` **NEVER use keyword search in ToolSearch** — always use `select:` with exact tool names. - **Serena** — Navigate code for file management logic - **MySQL** — Query file...

Du er bibliotekaren for alle digitale eiendeler i Magitek-nettverket. Du organiserer, finner, flytter og forvalter filer på tvers av forretningsenheter og lagringslokasjoner. Du er IKKE en infrastruktur-agent. Du bryr deg om **innholdet** — hva filene er, hvem de tilhører, hvordan de bør organiseres — ikke om serverne de bor på. ---

{git diff --stat output} ```

- Agent 4: Syntax error → Resumed with claude-sonnet-4.6 → ✅

- PHP syntax: ✅ All files pass - PHPStan delta: ✅ No new errors - File conflicts: ✅ None detected - File size: ✅ All under 800 lines

- Total premium requests: 2.33 (of ~X monthly budget) - Free agents: 1 (saved 1.0 requests) - Budget agents: 1 (saved 0.67 requests vs standard)

| Agent | Task | Model | Tier | Status | Duration | Cost | Files Changed | |-------|------|-------|------|--------|----------|------|---------------| | 1 | {task} | claude-sonnet-4.6 | 1x | ✅ | 12s | 1.0 | {files} | | 2 | {task} | gpt-5-mini | 0x | ✅ | 8s | FREE | {files} | | 3 | {task} | claude-haiku-4.5 | 0.33x | ✅ | 6s | 0.33 | {files} | | 4 | {task} | gpt-5.3-codex | 1x | ❌→✅ | 25s | 1.0 | {files} |

After orchestration completes, report: ```

| Capability | Copilot CLI | Codex CLI | Gemini CLI | |------------|-------------|-----------|------------| | Non-interactive flag | `-p "prompt"` | `exec "prompt"` | `-p "prompt"` | | Auto-approve | `--yolo` | `--full-auto` | `--approval-mode yolo` | | Model selection | `--model {any of 17}` | `-m {gpt models only}` | `-m {gemini models only}` | | Multi-provider | ✅ Claude+GPT+Gemini | GPT only | Gemini only | | Reasoning control | Model-based | `-c model_reasoning_effort=` | Model-based | |...

| Error | Cause | Fix | |-------|-------|-----| | "argument is invalid. Allowed choices are..." | Model not available in CLI | Use model from Allowed Choices list | | Exit code 1 with no output | Rate limit or auth issue | Wait 10s, retry; check `gh auth status` | | Agent edited wrong files | Prompt too vague | Add explicit SCOPE + DO NOT EDIT | | Budget model overwrote recent fix | No PROTECTED FILES in prompt | Run Step 0.5, inject PROTECTED_ALL; restore with `git checkout -- {file}` | |...

``` ⚠️ SANDBOX PERSISTENCE: Copilot agents may use sandbox overlays. File changes can be LOST on process exit unless committed to git. Each agent MUST git add + git commit its own changes before exiting. BEFORE parallel spawn: 1. git stash (if uncommitted changes exist) 2. Verify clean working tree 3. Run Step 2.5 trait detection for PHP tasks DURING execution: - Each agent prompt MUST include "git add + git commit" in WHEN DONE - Agents commit to the SAME branch — non-overlapping files...

Before spawning parallel agents, verify no overlap: ```bash # Extract file lists from all prompts, check for duplicates echo "${ALL_FILES[@]}" | sort | uniq -d # If any output → files appear in multiple agents → SERIALIZE those tasks ```

Include in EVERY agent prompt (via the SAFETY block in Standard Prompt Structure). > **⚠️ Do NOT write literal forbidden command names** (e.g., migrate:fresh, db:wipe) in prompt text. > The production-safety hook scans Bash commands — if prompts containing these strings > are passed as heredoc/inline arguments, the hook triggers false positives. > Use the generic SAFETY block from Standard Prompt Structure instead. > Write prompts to temp files via Write tool to bypass the hook entirely.

``` 10 trivial tasks (renames, simple edits): gpt-5-mini × 10 = 0 premium requests (FREE!) 5 moderate tasks (extract methods, add features): claude-haiku-4.5 × 5 = 1.65 premium requests 3 complex tasks (refactoring, multi-file logic): claude-sonnet-4.6 × 3 = 3 premium requests Mixed workload (typical orchestration of 10 tasks): 4 × gpt-5-mini (0) + 3 × haiku (0.99) + 3 × sonnet (3) = 3.99 requests vs. all sonnet: 10 requests — 60% savings! ``` ---

- **Recommended:** 3-5 parallel agents - **Maximum:** 8 (Copilot rate limits per subscription tier) - **If rate limited:** Reduce to 2-3, add 3s stagger between spawns - **Mixed-model strategy:** Spread across providers to avoid per-provider limits

```bash # Fastest spawn (skip all extras) — ONLY for Tier 2+ (sonnet, gpt-5.x) on isolated tasks: copilot -p "..." --yolo --no-ask-user --no-custom-instructions --disable-builtin-mcps 2>&1 # Budget/free models — NEVER skip custom instructions: copilot -p "..." --yolo --no-ask-user 2>&1 # With GitHub context (slower but has repo awareness): copilot -p "..." --yolo --no-ask-user 2>&1 ``` `--no-custom-instructions` skips loading AGENTS.md and copilot-instructions.md. `--disable-builtin-mcps`...

**For GPT models (gpt-5.x):** Be explicit about file paths, step-by-step instructions. **For Claude models:** Can handle more abstract instructions, better at inferring patterns. **For Gemini models:** Good with multi-file context, strong at pattern matching. ---

**Include this block in EVERY Copilot agent prompt to prevent premature stopping:** ``` AUTONOMY RULES (CRITICAL — follow these exactly): - Complete the ENTIRE task in full. Do NOT stop early. - Do NOT suggest breaking the task into stages or batches. - Do NOT ask the user to confirm, clarify, or choose between options. - Do NOT say "too large", "unmanageable", or "please narrow scope". - If a subtask is ambiguous, make the best reasonable decision and proceed. - If you encounter an error,...

**Include this block in EVERY Copilot agent prompt:** ``` CODE QUALITY RULES (ReportMaker enforces these — do not violate): PHP: - Max file size: 400 lines target, 800 lines hard limit - No new PHPStan errors at Level 7 - Follow PSR-12 / Laravel code style (Pint-compatible) - Constructor DI only — no app() / resolve() calls inside classes - No business logic in controllers (<300 lines, delegate to services) - Do NOT run vendor/bin/phpstan yourself — only use php -l JavaScript/Vue: - NO...

When using `gpt-5-mini`, `gpt-4.1`, `gpt-5.1-codex-mini`, or `claude-haiku-4.5`, inject this block **BOTH at the start AND at the end** of the prompt (in addition to the standard structure above): ``` ⚠️ BUDGET MODEL — STRICT SCOPE ENFORCEMENT ⚠️ You MUST NOT edit any file outside of SCOPE. BEFORE touching any file, check: "Is this file in my SCOPE list?" Answer NO → skip it entirely. Do not read it. Do not write to it. PROTECTED FILES (these were recently modified — overwriting them...

``` TASK: {one-line description} ⛔ PROTECTED FILES — DO NOT EDIT UNDER ANY CIRCUMSTANCES: {INJECT $PROTECTED_ALL HERE — one file per line} - Any file not listed in SCOPE below SCOPE (only edit these files): - {file1.php} - {file2.php} TRAIT AWARENESS: - If a method you need to fix is NOT in the class file, it is in a Trait file. - Check for `use SomeTrait;` in the class — the method lives in that trait file. - If the method is in a trait, edit the TRAIT file (which MUST be in your SCOPE). -...

```bash # Resume most recent session with higher-capability model copilot --continue --yolo --no-ask-user \ --model claude-sonnet-4.6 2>&1 ``` For retries: escalate model tier (free→budget→standard→premium). ---

```bash # 1. Syntax check on files committed by agents CHANGED=$(git diff HEAD~N --name-only | grep '\.php$') echo "$CHANGED" | xargs -I{} php -l {} # 2. Check for merge markers or corruption echo "$CHANGED" | xargs grep -l '<<<<<<' 2>/dev/null # 3. PHPStan delta — run CENTRALLY, agents must NOT run PHPStan themselves vendor/bin/phpstan analyse --no-progress -c phpstan-no-baseline.neon $CHANGED 2>/dev/null | tail -20 # 4. File size check (800 line limit) echo "$CHANGED" | xargs awk...

After all agents complete: 1. **Verify commits exist**: `git log --oneline -N` (N = number of agents) - Each agent should have one commit with `[copilot-agent-{N}]` tag - If a commit is MISSING → that agent's work was lost → re-run it 2. `git diff --name-only` should be EMPTY if all agents committed 3. Read each agent's result file 4. Check exit codes (0 = success) 5. Parse "Total usage est:" line for actual premium request cost **4a. Scope Violation Detection (CRITICAL — run before...

**⚠️ NEVER pass prompts as heredoc or inline shell strings.** The production-safety hook scans Bash commands for forbidden strings. If your prompt text contains safety warnings like "DO NOT run migrate:fresh", the hook reads it as an actual command. **Correct pattern — Write prompts to temp files, pass via `cat`:** ```bash # Write each prompt to a temp file using the Write tool # (Write tool bypasses the Bash hook entirely) # File: /tmp/copilot-prompt-1.txt # File:...

Before writing prompts, check if target files use traits: ```bash for FILE in "${SCOPE_FILES[@]}"; do TRAITS=$(grep -oP 'use\s+\K[A-Za-z\\]+Trait' "$FILE" 2>/dev/null || true) if [ -n "$TRAITS" ]; then echo "⚠️ $FILE uses traits:" echo "$TRAITS" # Find the actual trait file paths for TRAIT in $TRAITS; do TRAIT_FILE=$(find app/ -name "${TRAIT}.php" 2>/dev/null) echo " → $TRAIT_FILE" done echo " ADD trait files to agent SCOPE (methods live there, not...

```bash REGISTRY_FILE="/tmp/copilot-orchestrator-$(date +%Y%m%d-%H%M%S).json" echo '{"agents":[],"total_cost":0}' > "$REGISTRY_FILE" ``` Track per agent: - Agent ID (incrementing integer) - Status: spawning / running / completed / failed - Assigned files (for conflict detection) - Model + cost tier - Start time + duration - Premium requests used (from output) - Output summary

**CRITICAL: File Conflict Prevention** Each Copilot agent operates in the SAME workspace (no git worktrees). Agents MUST NOT edit the same files simultaneously. ``` SAFE: Agent A edits app/Services/Foo.php, Agent B edits app/Services/Bar.php UNSAFE: Agent A and B both edit app/Services/Shared.php ``` **Decomposition rules:** 1. Each agent gets a DISTINCT set of files 2. No file appears in more than one agent's scope 3. If files overlap → serialize those tasks (sequential, not parallel) 4....

Before ANY agent is spawned, build a PROTECTED_FILES list from git state. This prevents agents from overwriting work-in-progress or today's commits. ```bash # 1. Uncommitted changes in working tree + staging area PROTECTED_UNCOMMITTED=$(git diff --name-only; git diff --cached --name-only) # 2. Files touched in the last 24 hours (committed) PROTECTED_RECENT=$(git log --since="24 hours ago" --name-only --format="" | sort -u | grep -v '^$') # 3. Merge into a single sorted, deduplicated...

| # | Task | Files | Score | Model | Tier | ~Cost | |---|------|-------|-------|-------|------|-------| | 1 | Extract FooHelper trait | 2 | 3 | claude-sonnet-4.6 | 1x | 1 req | | 2 | Rename old methods | 1 | 1 | gpt-5-mini | 0x | FREE | | 3 | Refactor BarService logic | 4 | 6 | gpt-5.3-codex | 1x | 1 req | | 4 | Update all imports | 8 | 2 | claude-haiku-4.5 | 0.33x | 0.33 req | Total estimated cost: 2.33 premium requests Parallel groups: [1,2,4] parallel → 3 sequential (shares file with...

After auto-selection, ALWAYS show the plan: ```

After initial scoring, apply these overrides: ``` IF task requires reading files > 128K total → FORCE gpt-5.2-codex or gpt-5.3-codex (272K) IF task requires output > 64K tokens → FORCE codex variant (128K output) IF task is PHP/Laravel specific → PREFER claude-sonnet-4.6 (best PHP understanding) IF task is bulk transformation (10+ files same pattern) → PREFER gpt-5-mini (free, simple pattern) IF task needs deep architectural reasoning → PREFER claude-opus-4.6 or claude-sonnet-4.6 ```

**Quality First:** ``` Score 0-1 → gpt-5-mini (0x) — trivial, save requests Score 2-3 → claude-sonnet-4.6 (1x) — best quality for standard work Score 4-6 → gpt-5.3-codex (1x) — complex, benefits from 272K context Score 7-9 → claude-sonnet-4.6 (1x) — high complexity, best reasoning Score 10+ → claude-opus-4.6 (3x) — architecture-level decisions ``` **Balanced:** ``` Score 0-1 → gpt-5-mini (0x) — free for trivial Score 2-3 → claude-haiku-4.5 (0.33x) ...

For each decomposed task, evaluate these signals: ``` SIGNAL SCORING: +0 points = baseline File count: 1 file → +0 2-3 files → +1 4-5 files → +2 6+ files → +3 Lines of change (estimated): <30 LOC → +0 30-100 LOC → +1 100-300 LOC → +2 300+ LOC → +3 Task type: rename/move/delete → +0 add field/method → +1 extract class/trait → +2 refactor logic → +3 cross-concern change → +4 Dependency complexity: Self-contained → +0 ...

``` Velg Copilot-modell: 1. claude-sonnet-4.6 (anbefalt - best kvalitet/pris) 2. gpt-5.4 (standard GPT, strong alternative) 3. gpt-5.3-codex (272K kontekst, kode-optimalisert) 4. claude-haiku-4.5 (rask, billig 0.33x) 5. gpt-5-mini (gratis, god for enkle oppgaver) 6. claude-opus-4.6 (premium 3x, dypest reasoning) ``` ---

``` Velg kostnadsstrategi: 1. Quality First (standard) - Beste modell for oppgaven, uansett pris 2. Balanced - God kvalitet, hold kostnad under 1x per agent der mulig 3. Budget - Bruk 0x/0.33x modeller mest mulig, kun 1x for komplekse oppgaver 4. Free Only - Kun 0x-modeller (gpt-4.1, gpt-5-mini) ```

``` Copilot Agent Configuration: Velg agent-tilordning: 1. Smart Auto-Select (anbefalt) - Analyserer oppgavekompleksitet, velger optimal modell+provider 2. Uniform - Samme modell for alle agenter (du velger) 3. Manual per agent - Du spesifiserer modell per oppgave ```

Before ANY agent spawning, present the user with configuration choices using AskUserQuestion:

```bash copilot --continue # Resume most recent copilot --resume # Session picker (interactive only) copilot --resume {sessionId} --yolo # Resume specific session ``` ---

- `-p "{prompt}"` - Non-interactive mode (REQUIRED for spawning) - `--yolo` - Grant all permissions (equivalent to `--allow-all-tools --allow-all-paths --allow-all-urls`) - `--no-ask-user` - Disable ask_user tool (fully autonomous, no questions) - `--model {model}` - Model selection - `--add-dir {directory}` - Add accessible directory (can use multiple times) - `--no-custom-instructions` - Skip loading AGENTS.md (faster startup for isolated tasks) - `--disable-builtin-mcps` - Disable GitHub MCP...

```bash copilot -p "{prompt}" --yolo --no-ask-user --model {model} 2>&1 ```

**Tier 0 — FREE (0x premium requests)** | Model | Context In | Output | Vision | Best For | |-------|-----------|--------|:------:|----------| | `gpt-4.1` | 111K | 16K | ✅ | Trivial ops, syntax checks | | `gpt-5-mini` | 128K | 64K | ✅ | Simple edits, large output needed | **Tier 1 — BUDGET (0.33x premium requests)** | Model | Context In | Output | Vision | Best For | |-------|-----------|--------|:------:|----------| | `claude-haiku-4.5` | 128K | 32K | ✅ | Fast parallel, reliable | |...

| Scenario | Use Copilot | Use Codex | Use Gemini | Use Claude Task | |----------|-------------|-----------|------------|-----------------| | Multi-provider model access | ✅ Best | GPT only | Gemini only | Claude only | | Cost-optimized bulk ops (0x models) | ✅ Free tier | Subscription | Free tier | Haiku 0.33x | | GitHub-native context (PRs, issues) | ✅ Native MCP | | | | | Autonomous file edits | ✅ | ✅ | ✅ | | | Tasks needing inherited context | | | | ✅ | | Coordinated multi-agent + shared...

All MCP servers are active (deferred loading — 0 cost until first use). For orchestration: **Quick Fetch (run at agent start to activate tools):** ``` ToolSearch("select:mcp__serena__find_symbol,mcp__serena__get_symbols_overview,mcp__mysql__execute_sql") ``` **NEVER use keyword search in ToolSearch** — always use `select:` with exact tool names. - **Serena** — Understand codebase structure for task delegation - **MySQL** — Check data state when coordinating data-related tasks ---

**Purpose:** Spawn, track, and coordinate multiple parallel GitHub Copilot CLI agents for autonomous code tasks. Unique advantage: access to models from ALL major providers (Anthropic, OpenAI, Google, xAI) via a single CLI. ---