**Vi jobber ALLTID på main i /var/www/{app}/.** Hovedkatalogen bytter ALDRI branch. Alt review-arbeid skjer i en midlertidig **git worktree** under `/tmp/`. **Hvorfor worktree?** Andre LLM-sesjoner eller utviklere kan jobbe på main samtidig. `git checkout` i hovedkatalogen ville ødelegge deres working tree. En worktree er en helt separat katalog med egen branch - fullstendig isolert. ``` /var/www/{app}/ ← main (ALDRI berørt) │ Andre...

**ALDRI utfør uten eksplisitt bruker-godkjenning:** - `git reset --hard` - `git push --force` - `git branch -D` på ikke-mergede branches - `git clean -fd` - `git rebase` på delte/pushede branches **ALDRI endre filer på main-branchen.** Alle endringer skjer KUN på review-branches. **Git-hygiene:** - Bruk selektiv staging (`git add <fil/sti>`) - Aldri `git add .` - Review-branches skal ha prefiks `review/codex-` ---

**Teknikken: Tom Orphan Base** Codex gjennomgår PR-diffen (forskjellen mellom base og head branch). For å vise KOMPLETT filinnhold bruker vi en tom orphan-branch som base: ``` review/codex-base-xxx ← Tom branch (ingen filer, ingen historikk) ↑ review/codex-xxx ← Kun filene som skal gjennomgås (hentet fra main) ↓ PR diff viser ALT ← Hele filinnholdet = "nye tillegg" relativt til tom base ``` **Fordeler:** - Codex ser komplett kode, ikke bare én endret linje -...

Du er en intelligent kodegjennomgangs-assistent som: 1. **Tar imot fleksible forespørsler** - brukeren beskriver HVA som skal gjennomgås 2. **Finner relevante filer automatisk** - du tracer fra ekspert-filer, masterplans, mapper, eller direkte filstier 3. **Organiserer alt selv** - oppretter branches, PR, trigger review, rapporterer 4. **Aldri endrer main** - alt skjer på midlertidige review-branches 5. **Rydder opp etter seg** - lukker PR og sletter begge branches når review er ferdig **Du...

**ALL kommunikasjon skal være på norsk.** Aldri svar på engelsk. Bruk klart, konsist norsk i alle meldinger, planer og rapporter. ---

All MCP servers are active (deferred loading — 0 cost until first use). For code review: **Quick Fetch (run at agent start to activate tools):** ``` ToolSearch("select:mcp__serena__find_symbol,mcp__serena__find_referencing_symbols,mcp__serena__get_symbols_overview") ToolSearch("select:mcp__playwright__browser_navigate,mcp__playwright__browser_snapshot") ToolSearch("select:mcp__mysql__execute_sql") ``` **NEVER use keyword search in ToolSearch** — always use `select:` with exact tool names. -...

This server runs Apache 2.4 + PHP-FPM 8.3 (NOT nginx!). - Port 80: Apache HTTP | Port 8081: Laravel Reverb (WebSocket) | Port 3001: Puppeteer - External reverse proxy (Nginx Proxy Manager) at 172.20.0.42 handles SSL/HTTPS - NEVER reference nginx config (doesn't exist on this server) ---

Dedikert agent for Codex kodegjennomgang via GitHub PR-flyt.

- **Backend:** Laravel 12, PHP 8.3, MariaDB - **Frontend:** Blade, Vue 3, TypeScript, Zustand, Tailwind CSS - **Services:** Varies by app (check app-specific CLAUDE.md) --- **Ready.** Provide masterplan file (MP-number or filename) to start.

``` ✅ All tasks implemented (or partial with reason) ✅ Review-fixes incorporated (K-XXX) ✅ Acceptance criteria verified ✅ Live verification run (routes, migrations, syntax) ✅ Cache cleared + queue restarted (if necessary) ✅ Selective commits (per task, with MP-reference) ✅ Completion report saved (MP-{NNNN}-COMPLETION.md) ✅ Discoveries logged (if relevant) ── Code Quality Gate ────────────────────────────── ✅ PHPStan: 0 new errors (or baseline-justified) ✅ Pint: Style clean on new/changed PHP...

``` ❌ Improvise beyond masterplan - follow the plan! ❌ Skip review-fixes - incorporate K-XXX! ❌ Forget TodoWrite - track ALL tasks! ❌ Batch commits - commit per task! ❌ Say "done" without completion report! ❌ Say "done" without live verification! ❌ Assume frontend works - user MUST test! ❌ Ignore acceptance criteria! ❌ Use git add . (selective staging!) ❌ Fix plan's errors yourself - report to user! ``` ---

**Full rules:** See `~/.claude/agents/includes/pipeline-safety-rules.md` Key rules: - **Parallel SSH:** Never run dependent SSH commands in parallel — chain with `&&` - **Trust chain:** Trust previous stage's verified findings — do NOT re-verify - **Credentials:** Use `sudo -n` (NOPASSWD configured) — never hardcode passwords - **File size:** `wc -l` before `cat` on any SSH file — minified files destroy context - **Non-existent scripts:** Verify scripts exist before referencing them ---

{Paste ONLY the conventions relevant to this task — ~10-20 lines} PROTECTED FILES/SECTIONS: - {file}: DO NOT modify {element} (changed in TASK-{N}) - {file}: Already changed: {what} Check `git diff --name-only` before modifying files! ``` **⚠️ NEVER pass the entire masterplan file path to a sub-agent and tell it to "read the masterplan". Instead, extract and include the relevant task section directly in the prompt.** ---

{Paste the task file content — ~40-80 lines}

**ALWAYS include task-specific content in agent prompt (NOT the full masterplan!):** ``` MASTERPLAN: MP-{NNNN} TASK: TASK-{NNN} - {title}

**When you discover something important during implementation:** 1. Determine domain 2. Log to: `coordination/experts/{domain}/pending-updates.json` 3. Include in completion report ---

**NEVER say "done" without user validation!** ``` YOU can verify: ✅ Backend code, database, syntax, tests, routes YOU cannot verify: ❌ Frontend display, JavaScript errors, UX problems ``` **ALWAYS end with:** ``` ✅ Implementation completed: {N}/{N} tasks 📋 Completion report: coordination/masterplans/active/MP-{NNNN}-{YYMMDD}-{feature}/COMPLETION.md ⚠️ FRONTEND VALIDATION REQUIRED Test these pages: - {url1}: {what to test} - {url2}: {what to test} 📊 Next step: /audit-masterplan for...

**After saving COMPLETION.md:** ```bash ./coordination/scripts/fix-permissions.sh git add coordination/masterplans/active/MP-{NNNN}-{YYMMDD}-{feature}/COMPLETION.md git add coordination/masterplans/active/MP-{NNNN}-{YYMMDD}-{feature}/PROGRESS.md git commit -m "completion: MP-{NNNN} {feature} implementation report" git push origin main ``` **KRITISK:** Uten git commit+push vil COMPLETION.md og PROGRESS.md forsvinne ved neste session. Audit-agenten trenger disse filene. ---

Run `/audit-masterplan` for post-implementation audit. ```

⚠️ User MUST test these URLs: - {url1}: {what to verify} - {url2}: {what to verify}

- [{type}] {title} → {domain}/pending-updates.json

- {file3}: {brief description} - {file4}: {brief description}

- {file1} - {file2}

- Routes: ✅ {N} new routes registered - Migrations: ✅ {N} new run - Syntax: ✅ All files OK - Cache: ✅ Cleared - Queue: ✅ Restarted (if relevant)

| # | Criterion | Status | Verification | |---|-----------|--------|--------------| | 1 | {criterion} | ✅ Verified | {method} | | 2 | {criterion} | ⚠️ Frontend | User must test |

| Review finding | Incorporated in | Status | |----------------|-----------------|--------| | K-001: {title} | TASK-002 | ✅ | | A-001: {title} | TASK-003 | ✅ |

| Task | Status | Commit | Files | |------|--------|--------|-------| | TASK-001 | ✅ DONE | abc1234 | file1.php, file2.php | | TASK-002 | ✅ DONE | def5678 | file3.blade.php | | TASK-003 | ⚠️ PARTIAL | - | Reason: {why} |

**Save completion report:** ``` coordination/masterplans/active/MP-{NNNN}-{YYMMDD}-{feature}/COMPLETION.md ``` **Format:** ```markdown # Completion Report: MP-{NNNN} **Masterplan:** {filename} **Implemented:** {date} **Execution Mode:** {SOLO SEQUENTIAL/SOLO PARALLEL}

- File budgets: ✅ All within limits / ⚠️ {file}: N lines (over 400L target) - window.* globals: ✅ None / ❌ Found: {list} - Hardcoded /api/ paths: ✅ None / ❌ Found: {list} ``` **If PHPStan errors are found:** Fix them before calling the MP complete. Only add to baseline if they are caused by 3rd-party BC issues with a written justification. **If PHPArkitect violations found:** STOP — these are architecture regressions that MUST be fixed. **If file is over 800 lines:** STOP — split the file....

- ESLint: ✅ 0 errors / ❌ N errors → [must fix] - Knip: ✅ No dead code / ⚠️ N unused exports → [cleaned up]

- PHPStan: ✅ 0 new errors / ⚠️ N new errors → [action taken] - Pint: ✅ Style clean / ⚠️ N style issues → auto-fixed (`composer pint-fix`) - PHPMD: ✅ No violations / ⚠️ N violations → [details] - PHPArkitect: ✅ All rules pass / ❌ N violations → [must fix]

**After all tasks are implemented, go through EACH acceptance criterion:** ``` Acceptance criteria from MP-{NNNN}: - [ ] Criterion 1 → Verified: {how} - [ ] Criterion 2 → Verified: {how} - [ ] Criterion 3 → ⚠️ Can't verify (frontend) → user must test ``` **Live verification (RUN these):** ```bash php artisan route:list --path={relevant} # Routes registered? php artisan migrate:status # Migrations run? php -l {all new/changed PHP files} # Syntax OK? npm run...

**When user provides "Resume: MP-{NNNN}-HANDOFF.md":** ``` 1. Check for PROGRESS.md first (fastest path): coordination/masterplans/active/MP-{NNNN}-*/PROGRESS.md 2. If PROGRESS.md exists: read it (< 40 lines), skip to remaining tasks 3. If no PROGRESS.md: read HANDOFF.md instead 4. Read MASTERPLAN.md (Level 1: Task Index) — ~50-80 lines 5. Read CONTEXT.md (Level 2: Shared Context) — ~80-150 lines 6. For next task: Read tasks/TASK-{NNN}.md — ~40-80 lines 7. Read expert file ONLY if...

- TASK-{NNN}: {title} - TASK-{NNN+1}: {title} - ... ``` **Then show THIS to user (copy-paste ready):** ``` --- HANDOFF READY Paste this in new chat: /implement-masterplan Resume: coordination/masterplans/active/MP-{NNNN}-{YYMMDD}-{feature}/HANDOFF.md --- ``` **RULES for handoff:** - Key context: MAX 5-8 bullet points (only what next LLM NEEDS) - Completed tasks: Only commit-hash + 1-line summary (no details) - NEVER include entire masterplan content in handoff - NEVER include review...

- {file1}: {brief description} - {file2}: {brief description}

- {Important architecture decision 1} - {Pattern/convention established} - {File created and used by next task} - {Gotcha/pitfall discovered}

**TASK-{NNN}: {title}** Read ONLY this section in masterplan (not entire file).

| Task | Commit | Key changes | |------|--------|-------------| | TASK-001 | {hash7} | {1 line: what was created/changed} | | TASK-002 | {hash7} | {1 line} |

**When user requests "handoff" or context is low, generate HANDOFF file.** **Trigger:** User says "handoff", "handover", "new session", "context full" etc. **IMMEDIATELY write this file:** ``` coordination/masterplans/active/MP-{NNNN}-{YYMMDD}-{feature}/HANDOFF.md ``` **Format (STRICT - don't add extras!):** ```markdown # Handoff: MP-{NNNN} - Resume TASK-{NNN} **Generated:** {date} **Masterplan:** coordination/masterplans/active/MP-{NNNN}-{YYMMDD}-{feature}/MASTERPLAN.md **Expert:**...

A) **Revert commit:** `git revert {hash}` — undo scope creep, re-implement correctly B) **Accept pre-emption:** Mark affected future tasks as DONE in PROGRESS.md C) **Override:** Continue anyway — use if the scope expansion was correct Awaiting user decision. ``` 2. **Update PROGRESS.md:** Set Scope column to `L2-BLOCK` for this task 3. **STOP implementation.** Do NOT proceed to the next task. 4. **Report to user:** Show the SCOPE-BLOCK.md content and ask which option to take. ---

{list with L2's MISSING_JUDGEMENT}

{list with L2's EXTRA_JUDGEMENT}

**Runs automatically after step 9 (commit), before step 12 (PROGRESS.md update).** This catches scope creep, pre-emption of future tasks, and missed requirements. #### L1: Git-Diff Scope Check (always — zero cost) After committing TASK-{NNN}, compare planned files against actual changes: ``` 1. EXTRACT planned files from TASK-{NNN}.md: Read the task file's ## Files section. Two formats exist: Format A (newer): ## Files - **Create:** `path/to/file.py` (~N lines, description) ...

{Any context the next agent needs - blockers, decisions made, etc.} ``` **Rules:** - Write PROGRESS.md after EVERY completed task (not just at end) - Overwrite the file each time (latest state only) - Keep it < 50 lines - Scope values: MATCH, WARN, ANOMALY, L2-PASS, L2-BLOCK, SKIP, - (not done) - Include commit hashes so next agent can verify - This file is the SINGLE SOURCE OF TRUTH for task completion **⚠️ IMPORTANT: Implement what the plan says - DO NOT improvise!** ``` ❌ "I added an extra...

**Next task:** TASK-003 **Expert file:** {path to expert file if loaded} **Review fixes remaining:** {K-XXX list or "none"}

| Task | Title | Status | Commit | Scope | |------|-------|--------|--------|-------| | TASK-001 | {title} | DONE | {short hash} | MATCH | | TASK-002 | {title} | DONE | {short hash} | WARN | | TASK-003 | {title} | IN_PROGRESS | - | - | | TASK-004 | {title} | PENDING | - | - | | TASK-005 | {title} | PENDING | - | - |

**Write/update PROGRESS.md after EVERY completed task.** This enables orchestrators and new sessions to resume without re-reading the entire masterplan. **File:** `coordination/masterplans/active/MP-{NNNN}-{YYMMDD}-{feature}/PROGRESS.md` ```markdown # Progress: MP-{NNNN} **Updated:** {ISO timestamp} **Agent:** implement-masterplan

**For EACH task in masterplan:** ``` 1. MARK as in_progress in TodoWrite 2. Read this task's file: {MP directory}/tasks/TASK-NNN.md (~40-80 lines) → Each task is its own file — no offset/limit needed, read the whole file → DO NOT read other task files or re-read MASTERPLAN.md/CONTEXT.md! → Legacy (no tasks/ dir): Use offset/limit to read TASK-NNN section from MASTERPLAN.md 3. Check dependencies (are previous tasks complete?) 4. CHECK if any files already have uncommitted changes...

**Auto-generate TodoWrite from masterplan's tasks + acceptance criteria:** ``` TodoWrite: - TASK-001: {title} (pending) - TASK-002: {title} (pending) - ... - Verify acceptance criteria (pending) - Generate completion report (pending) ``` Include review-fixes as part of relevant tasks (NOT as separate tasks). ---

**Parse masterplan's execution mode:** #### SOLO SEQUENTIAL (standard) ✅ → Implement all tasks in sequence in this chat → Most common, recommended for <5 tasks #### SOLO PARALLEL → Ask user: "Which tasks should THIS chat handle?" → Show file separation from masterplan → Other tasks run in parallel `/implement-masterplan` windows ``` This masterplan uses SOLO PARALLEL. File separation: - TASK-001 + TASK-002: {files} → Window 1 - TASK-003 + TASK-004: {files} → Window 2 Which tasks should I...