@the-bearded-bear/claude-craft 8.26.0 → 8.27.0

package/Dev/i18n/de/Workflow/commands/auto-sprint.md

@@ -1,6 +1,6 @@
 ---
 description: End-to-End-Sprint-Orchestrator (Start -> Zerlegung -> Validierung -> Implementierung -> PR -> CI -> Review -> Retro -> Merge)
-argument-hint: "<N> [--from-prd] [--auto-merge] [--max-fix-attempts=2] [--max-workers=3] [--base=main] [--dry-run] [--overnight]"
+argument-hint: "<N> [--from-prd] [--auto-merge] [--max-fix-attempts=2] [--max-workers=3] [--base=main] [--dry-run] [--overnight] [--engine=subagents|teams]"
 ---
 
 # Auto Sprint — End-to-End-Sprint-Orchestrator
@@ -31,14 +31,15 @@ $ARGUMENTS
   Regel 09, und das Karpathy-Prinzip „kein Auto-Merge ohne menschliche Review").
 - `--max-fix-attempts=2` : Maximale automatische Korrekturversuche pro fehlgeschlagenem Gate vor dem Abbruch (Standard: 2).
 - `--max-workers=3` : Maximale parallele Entwickler-Worker in der Implementierungsphase (Standard: 2, Maximum: 3).
+- `--engine=subagents`: Engine der Implementierungsphase (Standard: `subagents`, Delegation über das `Agent`-Tool). `teams`: experimentelle Agent Teams, mit Fallback auf `subagents`, wenn nicht verfügbar. Siehe `/team:sprint`, Abschnitt „Ausführungs-Engine“
 - `--base=main` : Basis-Branch für den PR (Standard: `main`).
 - `--dry-run` : Gibt die geplanten 9 Phasen und den aufgelösten Sprint-Kontext aus und hält dann an. **Keine Schreibvorgänge.**
 - `--overnight` : Wird an die Implementierungsphase weitergegeben (begrenzt, stoppt um 6 Uhr morgens).
 
 ## Voraussetzungen
 
-- Claude Code v2.1.32+ mit Agent Teams Unterstützung
-- `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` gesetzt
+- Claude Code v2.1.32+ mit Agent Teams Unterstützung — nur mit `--engine=teams`
+- `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` gesetzt — nur mit `--engine=teams`
 - `gh` CLI authentifiziert (PR erstellen / Prüfungen / Merge)
 - Docker verfügbar (alle Tests laufen über Docker — siehe Projekt-CLAUDE.md)
 - BMAD v6-Projekt: `.bmad/sprint-status.yaml` wird bei Bedarf von `/workflow:start` (Phase 1) angelegt
@@ -106,6 +107,8 @@ Einen isolierten Unter-Agenten starten:
 
 Direkt die **`/team:sprint`-Dirigenten-Rolle übernehmen** (kein verschachteltes Agent Team starten):
 
+**Engine**: Mit `--engine=subagents` (Standard) in Schritt 4 pro Story einen Sub-Agenten über das `Agent`-Tool starten (`isolation: "worktree"`, alle in einer Nachricht) statt `TaskCreate`, und in Schritt 5 auf ihre Rückkehr warten statt `TaskList` abzufragen (vollständige Zuordnung in `/team:sprint`, Abschnitt „Ausführungs-Engine“). Mit `--engine=teams` die Schritte wie beschrieben befolgen.
+
 1. `.bmad/sprint-status.yaml` lesen; Stories mit Status `ready-for-dev` filtern.
 2. Datei-Domain-Unabhängigkeit analysieren (`**/Shared/**`, `**/Common/**`, `**/Utils/**`,
    `**/Helpers/**` Überschneidungen markieren → beim gleichen Worker sequenzieren).
@@ -195,7 +198,7 @@ Phase                  | Status  | Anmerkungen
 | Story-DoD-Verfehlung nach Wiederholungen | Als `blocked` markieren, fortfahren, am Ende berichten |
 | CI rot nach Wiederholungen | Abbruch mit Bericht über fehlgeschlagene Prüfungen |
 | Merge blockiert (Schutz / Scope) | Fehler melden + vorgeschlagenes Flag, nicht erzwingen |
-| Agent Teams nicht verfügbar | Abbruch von Phase 4 mit Setup-Hinweis (`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`) |
+| Agent Teams mit `--engine=teams` nicht verfügbar | Fallback auf `--engine=subagents` (kein Abbruch); `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` setzen, um `--engine=teams` zu nutzen |
 
 ## Hinweise
 

package/Dev/i18n/de/Workflow/commands/plan.md

@@ -69,7 +69,7 @@ Früheres `/workflow:plan --research`. Recherche, Exploration und Constraints vo
 | Markt | Bestehende Lösungen? Wettbewerber? Best Practices der Branche? Neue Trends? |
 | Technik | Mögliche Technologien? Integrationen? Skalierbarkeit? Sicherheit / Compliance? |
 
-Ist der MCP Context7 konfiguriert, ihn für aktuelle technische Dokumentation nutzen (`/common:research-context7` für tiefgehende Recherche).
+Ist der MCP Context7 konfiguriert, ihn für aktuelle technische Dokumentation nutzen (`/deep-research` für tiefgehende Webrecherche; außerhalb von claude.ai der Agent `@research-assistant`).
 
 **Constraints**: technische (bestehendes Backend, Datenbank), geschäftliche (Budget, Zeitplan, Abwärtskompatibilität), regulatorische (DSGVO, PCI DSS) und Ressourcen-Constraints (Teamgröße, kein DevOps) dokumentieren.
 
@@ -213,4 +213,3 @@ Bei Unterbrechung erkennt `/workflow:plan --continue`, was vorhanden ist (PRD fe
 - `/workflow:status` - Fortschritt prüfen
 - `/project:backlog` - Stories danach hinzufügen, auflisten, bearbeiten und verschieben
 - `/gate:check` - Das Backlog validieren
-- `/common:research-context7` - Tiefgehende Recherche mit Context7 MCP

package/Dev/i18n/en/Common/agents/devops-engineer.md

@@ -17,20 +17,6 @@ permissionMode: default
 skills:
   - git-workflow
   - security
-disallowedTools:
-  - "Bash(rm -rf:*)"
-  - "Bash(dd:*)"
-  - "Bash(mkfs:*)"
-  - "Bash(kubectl delete:*)"
-  - "Bash(helm uninstall:*)"
-  - "Bash(terraform destroy:*)"
-  - "Bash(tofu destroy:*)"
-  - "Bash(docker system prune -a*)"
-  - "Bash(docker volume rm:*)"
-  - "Bash(curl * | sh*)"
-  - "Bash(wget * | sh*)"
-  - "Bash(git push --force*)"
-  - "Bash(git push -f*)"
 ---
 
 # DevOps Engineer Agent
@@ -242,3 +228,22 @@ strategy:
 3. Configure health checks
 4. Prepare rollback
 5. Monitor deployment
+
+## Destructive commands guard
+
+An agent cannot block a single command: `disallowedTools: Bash(...)` removes all of Bash. These commands therefore go through the permission prompt — always ask for explicit confirmation before running them. To block them in a project, add to its `.claude/settings.json` (the rules also apply to subagents):
+
+```json
+{
+  "permissions": {
+    "deny": [
+      "Bash(kubectl delete:*)",
+      "Bash(helm uninstall:*)",
+      "Bash(terraform destroy:*)",
+      "Bash(tofu destroy:*)",
+      "Bash(docker volume rm:*)",
+      "Bash(docker system prune -a*)"
+    ]
+  }
+}
+```

package/Dev/i18n/en/Common/agents/migration-specialist.md

@@ -6,22 +6,13 @@ maxTurns: 6
 effort: xhigh
 memory: user
 tools: [Read, Glob, Grep, Bash, WebFetch, WebSearch]
-# Audit 2026-05-18 QW-15 — migrations touch shared/prod databases. Block
-# destructive shell verbs and database drop/truncate. Investigate-then-output
-# is fine; actual destructive execution must require an explicit user opt-in.
-disallowedTools:
-  - "Bash(rm -rf:*)"
-  - "Bash(dd:*)"
-  - "Bash(mkfs:*)"
-  - "Bash(:(){:|:&};:*)"
-  - "Bash(DROP DATABASE:*)"
-  - "Bash(DROP TABLE:*)"
-  - "Bash(TRUNCATE:*)"
-  - "Bash(pg_dump:*)"
-  - "Bash(mysqldump:*)"
-  - "Bash(curl * | sh*)"
-  - "Bash(wget * | sh*)"
 permissionMode: default
+hooks:
+  PreToolUse:
+    - matcher: "Bash"
+      hooks:
+        - type: command
+          command: "grep -qiE '(^|[^A-Za-z_])(DROP[[:space:]]+(DATABASE|SCHEMA|TABLE)|TRUNCATE)([^A-Za-z_]|$)' && { echo \"BLOCKED: destructive SQL (DROP DATABASE/SCHEMA/TABLE, TRUNCATE). Hand the command to the user to run it themselves.\" >&2; exit 2; } || exit 0"
 ---
 
 # Migration Specialist Agent
@@ -140,3 +131,7 @@ You are a **Senior Migration Specialist** with 12+ years of experience in critic
 - [Stripe - Online migrations at scale](https://stripe.com/blog/online-migrations)
 - [Shopify - Sharding playbook](https://shopify.engineering/learnings-from-shopifys-largest-database-sharding-project)
 - [Strangler Fig - Martin Fowler](https://martinfowler.com/bliki/StranglerFigApplication.html)
+
+## Destructive SQL guard
+
+A `PreToolUse` hook declared in the frontmatter blocks, while this agent runs, any Bash command containing `DROP DATABASE`, `DROP SCHEMA`, `DROP TABLE` or `TRUNCATE` (for instance `psql -c "DROP TABLE …"`): hand the command to the user instead. Migration files written with Edit/Write are not affected. Limits: the hook only runs once the folder is trusted (workspace trust), and it is ignored when the agent is loaded from a plugin.

package/Dev/i18n/en/Common/agents/mlops-engineer.md

@@ -6,21 +6,6 @@ maxTurns: 6
 effort: medium
 memory: user
 tools: [Read, Glob, Grep, Edit, Write, Bash, WebFetch, WebSearch]
-# Audit 2026-05-18 QW-15 — MLOps touches feature stores, model registries
-# and inference endpoints (often production). Block destructive verbs on
-# data/model artefacts and on the underlying infra.
-disallowedTools:
-  - "Bash(rm -rf:*)"
-  - "Bash(dd:*)"
-  - "Bash(mkfs:*)"
-  - "Bash(kubectl delete:*)"
-  - "Bash(helm uninstall:*)"
-  - "Bash(mlflow models delete:*)"
-  - "Bash(mlflow registered-models delete:*)"
-  - "Bash(aws s3 rm:*)"
-  - "Bash(gsutil rm:*)"
-  - "Bash(curl * | sh*)"
-  - "Bash(wget * | sh*)"
 permissionMode: default
 ---
 
@@ -262,3 +247,22 @@ if report.as_dict()["metrics"][0]["result"]["dataset_drift"]:
 - [Google MLOps Guide](https://cloud.google.com/architecture/mlops-continuous-delivery-and-automation-pipelines-in-machine-learning)
 - [Book: Introducing MLOps](https://www.oreilly.com/library/view/introducing-mlops/9781492083283/)
 - [ML Model Governance](https://learn.microsoft.com/en-us/azure/machine-learning/concept-model-management-and-deployment)
+
+## Destructive commands guard
+
+An agent cannot block a single command: `disallowedTools: Bash(...)` removes all of Bash. These commands therefore go through the permission prompt — always ask for explicit confirmation before running them. To block them in a project, add to its `.claude/settings.json` (the rules also apply to subagents):
+
+```json
+{
+  "permissions": {
+    "deny": [
+      "Bash(kubectl delete:*)",
+      "Bash(helm uninstall:*)",
+      "Bash(mlflow models delete:*)",
+      "Bash(mlflow registered-models delete:*)",
+      "Bash(aws s3 rm:*)",
+      "Bash(gsutil rm:*)"
+    ]
+  }
+}
+```

package/Dev/i18n/en/Common/agents/ralph-conductor.md

@@ -5,17 +5,6 @@ model: opus
 effort: xhigh
 maxTurns: 10
 memory: user
-disallowedTools:
-  - "Bash(rm -rf:*)"
-  - "Bash(dd:*)"
-  - "Bash(mkfs:*)"
-  - "Bash(:(){:|:&};:*)"
-  - "Bash(sudo:*)"
-  - "Bash(chmod 777:*)"
-  - "Bash(curl * | sh*)"
-  - "Bash(wget * | sh*)"
-  - "Bash(git push --force*)"
-  - "Bash(git push -f*)"
 ---
 
 # Ralph Conductor Agent v2.0
@@ -186,11 +175,11 @@ Summary:
 
 ## Agent Teams Coordination Mode
 
-When operating in Agent Teams mode (activated via `--ralph-mode` on `/team:sprint`), the conductor takes on the role of **team lead** and coordinates a dev teammate through the Claude Code Agent Teams API instead of bash process management.
+When operating in Agent Teams mode (activated via `--ralph-mode` on `/team:sprint --engine=teams`), the conductor takes on the role of **team lead** and coordinates a dev teammate through the Claude Code Agent Teams API instead of bash process management.
 
 ### Prerequisites
 
-- `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` environment variable
+- `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` environment variable (only required by `/team:sprint --engine=teams`)
 - Claude Code v2.1.32+
 - Adapter library: `Tools/AgentTeams/lib/ralph-teams-adapter.sh`
 
@@ -227,8 +216,8 @@ The conductor runs periodic health checks through the adapter's `teams_watchdog(
 ### Keeping Bash Mode Intact
 
 All existing bash-mode orchestration remains unchanged. The Agent Teams mode is activated only when:
-1. The `--ralph-mode` flag is passed to `/team:sprint`
-2. The `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` env var is set
+1. The `--ralph-mode` flag is passed to `/team:sprint --engine=teams`
+2. The `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` env var is set (only required by `/team:sprint --engine=teams`)
 3. The adapter library is available
 
 Without these conditions, the conductor operates exactly as before.

package/Dev/i18n/en/Common/agents/refactoring-specialist.md

@@ -17,15 +17,6 @@ skills:
   - solid-principles
   - kiss-dry-yagni
   - testing
-disallowedTools:
-  - "Bash(rm -rf:*)"
-  - "Bash(dd:*)"
-  - "Bash(mkfs:*)"
-  - "Bash(:(){:|:&};:*)"
-  - "Bash(sudo:*)"
-  - "Bash(chmod 777:*)"
-  - "Bash(curl * | sh*)"
-  - "Bash(wget * | sh*)"
 ---
 
 # Refactoring Specialist Agent

package/Dev/i18n/en/Common/agents/tdd-coach.md

@@ -17,15 +17,6 @@ permissionMode: default
 skills:
   - testing
   - solid-principles
-disallowedTools:
-  - "Bash(rm -rf:*)"
-  - "Bash(dd:*)"
-  - "Bash(mkfs:*)"
-  - "Bash(:(){:|:&};:*)"
-  - "Bash(sudo:*)"
-  - "Bash(chmod 777:*)"
-  - "Bash(curl * | sh*)"
-  - "Bash(wget * | sh*)"
 ---
 
 # TDD/BDD Coach Agent

package/Dev/i18n/en/Common/agents/uiux-orchestrator.md

@@ -14,15 +14,6 @@ tools:
   - WebFetch
   - WebSearch
 permissionMode: default
-disallowedTools:
-  - "Bash(rm -rf:*)"
-  - "Bash(dd:*)"
-  - "Bash(mkfs:*)"
-  - "Bash(:(){:|:&};:*)"
-  - "Bash(sudo:*)"
-  - "Bash(chmod 777:*)"
-  - "Bash(curl * | sh*)"
-  - "Bash(wget * | sh*)"
 ---
 
 # UIUX Orchestrator Agent

package/Dev/i18n/en/Common/commands/ralph-run.md

@@ -51,6 +51,12 @@ Execute Claude in a continuous loop until the task is complete or the Definition
    - Set max iterations, timeouts, DoD criteria
    - Initialize hooks if enabled
 
+#### Session and parallel worker isolation
+
+- **Explicit Claude session**: each Ralph session pins its own Claude conversation ID — `claude --session-id <uuid>` on the first iteration, then `claude --resume <uuid>`. The UUID is stored in `.ralph/sessions/<session>/claude_session_id`, so `--continue=<session>` resumes the right conversation. Ralph no longer uses `claude --continue`, which resumed "the most recent conversation in the directory" (shared between workers, or even with the user's interactive session).
+- **One git worktree per worker** (parallel mode, `lib/parallel-manager.sh`): each story runs in a dedicated worktree `.ralph/parallel/worktrees/<session>` on branch `ralph/<session>`, with its Ralph state under `.ralph/parallel/sessions/<session>/ralph/`. At the end of the run, leftover changes are committed as `wip(ralph): <story>` on that branch, then the worktree is removed (`git worktree remove`). If that commit fails, the worktree is kept (no data loss) and reported in `output.log`. Collect the work by merging the `ralph/*` branches.
+- **Outside a git repository**: no worktree; the worker runs in the current directory (a warning is printed), Claude session isolation still applies.
+
 ### 2. Main Loop with Dashboard
 
 ```
@@ -76,6 +82,35 @@ The DoD system validates completion through multiple criteria:
 | `hook` | Run existing Claude hook |
 | `human` | Interactive human validation |
 
+#### `command` validator limit → use `type: hook`
+
+`type: command` only runs commands that pass the allowlist in `Tools/Ralph/lib/dod-validator.sh`: letters, digits, space and `_ . / " ' -`. Any command containing `=`, `|`, `&` or `:` (as well as `;`, `>`, `$`, etc.) is rejected with `ERROR: Invalid command`: the criterion is then **never** met and the loop runs until the circuit breaker trips.
+
+| Rejected command | Offending character |
+|------------------|---------------------|
+| `dotnet build /p:TreatWarningsAsErrors=true` | `:` and `=` |
+| `npm test && npm run lint` | `&` |
+| `pytest \| tee out.log` | `\|` |
+
+For these cases, put the command in a script and declare it as `type: hook`:
+
+```yaml
+definition_of_done:
+  checklist:
+    - id: build
+      type: hook
+      script: ".ralph/dod/build.sh"
+      required: true
+```
+
+```bash
+#!/usr/bin/env bash
+# .ralph/dod/build.sh — receives {"cwd": ..., "output": ...} as JSON on stdin
+dotnet build /p:TreatWarningsAsErrors=true || exit 2
+```
+
+The script returns `0` (criterion met) or `2` (criterion failed). Keys read per type: `command` → `command:`, `output_contains` and `file_changed` → `pattern:`, `hook` → `script:`, `human` → `prompt:`.
+
 ### 4. Adaptive Circuit Breaker (v2.0)
 
 Automatically selects profile based on task keywords:
@@ -100,6 +135,23 @@ Claude works...
 Stop → stop-dod-gate.sh → Block if DoD not satisfied (exit 2)
 ```
 
+## /goal (interactive) or Ralph (autonomous/CI)?
+
+`/goal` is a native Claude Code command typed inside a session: it is not a `ralph.sh` option — `ralph.sh` has none for it and never generates a `/goal`. The two tools are complementary:
+
+| Criterion | `/goal` (native) | Ralph (`ralph.sh`) |
+|-----------|------------------|--------------------|
+| Completion check | LLM judgment: a small fast model evaluates, after each turn, what is visible in the conversation — it runs no command and reads no file | Deterministic: commands, patterns and hooks run by `dod-validator.sh` |
+| Iteration cap | No fixed cap: ends when the condition is met or judged impossible, on `/goal clear`, or if Claude answers several turns in a row without tools; a bound can be written into the condition | `max_iterations` + adaptive circuit breaker |
+| Headless / CI | `claude -p "/goal ..."` loops to completion in a single invocation; CI usage not documented | Built for it: autonomous runs (`--autonomous`, `--overnight`), `headless` dashboard |
+| `disableAllHooks: true` | Unavailable: `/goal` wraps a session-scoped Stop hook (same with `allowManagedHooksOnly`) | The loop and the DoD run inside `ralph.sh`; only the hooks integration (`stop-dod-gate.sh`) is disabled |
+| Configuration | None; one goal per session, cleared by `/clear` | `ralph.yml` |
+
+- **Choose `/goal`** for an interactive task whose completion shows up in the conversation (print the test output: the evaluator does not run it).
+- **Choose Ralph** when completion must be proven by commands, with an iteration cap, unattended or in CI.
+
+Reference: [`/goal` documentation](https://code.claude.com/docs/en/goal).
+
 ## Quick Start Examples
 
 ```bash

package/Dev/i18n/en/Common/commands/research-context7.md

@@ -1,257 +1,9 @@
 ---
-description: Research with Context7 and Web
-argument-hint: [arguments]
+description: "Deprecated (removed in 9.0): use the context7 MCP or /deep-research"
+deprecated: true
+replaced-by: mcp:context7
 ---
 
-# Research with Context7 and Web
+> ⚠️ `/common:research-context7` is deprecated and will be removed in 9.0. Use the context7 MCP or `/deep-research`.
 
-You are an expert research assistant. You must use MCP Context7 to access library documentation and web search to find up-to-date information on a technical topic.
-
-## Arguments
-$ARGUMENTS
-
-Arguments:
-- Research topic or technical question
-- (Optional) Specific libraries to consult
-
-Example: `/common:research-context7 "How to implement OAuth2 authentication with NextAuth.js"` or `/common:research-context7 "React 19 best practices" react,nextjs`
-
-## Plan Mode
-
-> Plan mode is activated automatically when the scope spans multiple modules or requires cross-cutting investigation.
-
-## MISSION
-
-### Step 1: Analyze the Request
-
-Identify:
-- Main research topic
-- Concerned technologies/libraries
-- Required level of detail
-- Specific questions to answer
-
-### Step 2: Use Context7 (MCP)
-
-**Context7 provides access to up-to-date library documentation.**
-
-#### Search documentation
-
-```
-Use MCP context7 tool to:
-1. Search official library documentation
-2. Get up-to-date code examples
-3. Consult official guides and tutorials
-4. Check available APIs
-```
-
-#### Libraries supported by Context7
-
-Context7 indexes documentation from many popular libraries:
-- React, Next.js, Vue, Nuxt, Svelte
-- Node.js, Express, Fastify, NestJS
-- Python (Django, FastAPI, Flask)
-- TypeScript, Tailwind CSS
-- And many others...
-
-#### Context7 query format
-
-To use Context7, I must:
-1. Identify the exact library
-2. Formulate a precise query
-3. Request code examples if relevant
-
-### Step 3: Complementary Web Search
-
-**Use web search for:**
-
-1. **Recent information** (after Context7's cutoff date)
-   - New versions
-   - Breaking changes
-   - Official announcements
-
-2. **Community discussions**
-   - GitHub issues
-   - Stack Overflow discussions
-   - Expert blog articles
-
-3. **Comparisons and alternatives**
-   - Benchmarks
-   - Solution comparisons
-   - Feedback experiences
-
-4. **Specific use cases**
-   - Production examples
-   - Advanced patterns
-   - Solutions to common problems
-
-### Step 4: Synthesize Results
-
-#### Response Format
-
-```
-══════════════════════════════════════════════════════════════
-🔍 RESEARCH: [Topic]
-══════════════════════════════════════════════════════════════
-
-──────────────────────────────────────────────────────────────
-📚 OFFICIAL DOCUMENTATION (Context7)
-──────────────────────────────────────────────────────────────
-
-### [Library 1]
-
-**Current version**: X.Y.Z
-
-**Summary**:
-[Summary of found information]
-
-**Code example**:
-```[language]
-// Example code from documentation
-```
-
-**Useful links**:
-- [Link 1]
-- [Link 2]
-
-### [Library 2]
-...
-
-──────────────────────────────────────────────────────────────
-🌐 WEB SEARCH
-──────────────────────────────────────────────────────────────
-
-### Recent Information
-
-- [Date]: [Found information]
-- [Date]: [Found information]
-
-### Relevant Articles
-
-1. **[Article title]**
-   - Source: [URL]
-   - Summary: [Key points]
-
-2. **[Article title]**
-   ...
-
-### Community Discussions
-
-- **GitHub Issue**: [Link] - [Summary]
-- **Stack Overflow**: [Link] - [Summary]
-
-──────────────────────────────────────────────────────────────
-💡 SYNTHESIS AND RECOMMENDATIONS
-──────────────────────────────────────────────────────────────
-
-### Answer to Question
-
-[Synthetic answer based on research]
-
-### Recommended Approach
-
-1. [Step 1]
-2. [Step 2]
-3. [Step 3]
-
-### Points of Attention
-
-- ⚠️ [Attention point 1]
-- ⚠️ [Attention point 2]
-
-### Complete Code Example
-
-```[language]
-// Example code compiling found best practices
-```
-
-──────────────────────────────────────────────────────────────
-📋 SOURCES
-──────────────────────────────────────────────────────────────
-
-Documentation:
-- [Source 1]
-- [Source 2]
-
-Web:
-- [Source 1]
-- [Source 2]
-```
-
-### Step 5: Validation
-
-#### Verify Source Quality
-
-- [ ] Official sources prioritized
-- [ ] Up-to-date information (< 1 year ideally)
-- [ ] Consistency between sources
-- [ ] Testable code examples
-
-#### Verify Relevance
-
-- [ ] Answers initial question
-- [ ] Appropriate level of detail
-- [ ] Practical examples provided
-- [ ] Alternatives mentioned if relevant
-
-### Typical Use Cases
-
-#### 1. New Library
-
-```
-Question: "How to use [new library]?"
-
-→ Context7: Documentation, API, basic examples
-→ Web: Tutorials, feedback, gotchas
-```
-
-#### 2. Technical Problem
-
-```
-Question: "Why [error] with [library]?"
-
-→ Context7: Error documentation, troubleshooting
-→ Web: GitHub issues, Stack Overflow, forums
-```
-
-#### 3. Comparison
-
-```
-Question: "[Lib A] vs [Lib B] for [use case]?"
-
-→ Context7: Features of each lib
-→ Web: Benchmarks, comparisons, expert opinions
-```
-
-#### 4. Best Practices
-
-```
-Question: "Best practices for [topic]?"
-
-→ Context7: Official guidelines
-→ Web: Expert articles, popular patterns
-```
-
-#### 5. Migration
-
-```
-Question: "Migrate from [v1] to [v2]?"
-
-→ Context7: Official migration guide
-→ Web: Feedback experiences, real breaking changes
-```
-
-### Important Guidelines
-
-1. **Always cite sources** - Never invent information
-2. **Prioritize official documentation** - Context7 first
-3. **Check information date** - Web may have obsolete content
-4. **Provide testable code** - Examples must work
-5. **Be honest about limitations** - If information not found, say so
-
-### In Case of Doubt
-
-If I don't find the information:
-- Clearly indicate what wasn't found
-- Propose alternative paths
-- Suggest where to search manually
-- NEVER invent or hallucinate information
+For "$ARGUMENTS": library docs via the context7 MCP (`resolve-library-id` then `query-docs`), web research via `/deep-research`; outside claude.ai, WebSearch/WebFetch or the `@research-assistant` agent.

package/Dev/i18n/en/Common/rules/12-context-management.md

@@ -352,13 +352,15 @@ The `/context` command provides actionable suggestions to optimize context usage
 
 Adjust the model's effort level based on task complexity:
 
-| Command | Model | Usage |
-|---------|-------|-------|
-| `/effort low` | Haiku 5.5 | Simple tasks, lookups, classification |
-| `/effort medium` | Sonnet 5.5 | Standard implementation |
-| `/effort high` | Opus 5.5 | Complex reasoning, architecture |
-| `/effort xhigh` | Opus 5.5 (extended thinking) | Critical decisions, complex migrations, ADR |
-| `/effort ultracode` | Opus 5.5 (separate toggle since v2.1.284, Dynamic Workflows) | Maximum code throughput — automated pipelines, massive generation |
+| Command | Effect | Usage |
+|---------|--------|-------|
+| `/effort low` | Minimal reasoning, fast answers | Simple tasks, lookups, classification |
+| `/effort medium` | Standard depth | Standard implementation |
+| `/effort high` | Deeper reasoning | Complex reasoning, architecture |
+| `/effort xhigh` | Maximum adaptive thinking | Critical decisions, complex migrations, ADR |
+| `/effort ultracode` | Separate toggle (v2.1.284+), Dynamic Workflows | Maximum code throughput — automated pipelines, massive generation |
+
+> **Effort does not switch models:** it sets the reasoning depth of the current model; pick the model with `/model`. Recommended pairs: `low` on Haiku 5.5, `medium` on Sonnet 5.5, `high`/`xhigh` on Opus 5.5.
 
 > **Default effort per model:** medium on Opus 5.5 and Haiku 5.5; high on Sonnet 5.5, Fable 5.1 and legacy models.
 > **Fable 5.1** (`claude-fable-5-1`, $10/$50 per Mtok) is a current model, suitable for creative/narrative agents. **Mythos** is limited availability (Cyber Verification Program only).
@@ -493,7 +495,7 @@ During compaction, always preserve:
 | Variable | Description |
 |----------|-------------|
 | `CLAUDE_CODE_SUBAGENT_MODEL` | **Default** model for untyped sub-agents (e.g., `sonnet`). **Since Claude Code 2.1.251**, agent frontmatter `model:` takes precedence: the 11 reviewers with `model: haiku` stay on Haiku. **Before 2.1.251 it was the other way round** — this variable overrode the frontmatter. |
-| `CLAUDE_CODE_FORK_SUBAGENT` | `1` to isolate sub-agent contexts (forked subagents, v2.1.117+) |
+| `CLAUDE_CODE_FORK_SUBAGENT` | `1` to isolate sub-agent contexts (forked subagents, v2.1.117+); on by default in `-p`, the SDK, desktop and VS Code since 2.1.302, `0` to turn it off |
 | `ENABLE_PROMPT_CACHING_1H` / `FORCE_PROMPT_CACHING_5M` | Extend/force prompt caching (−40% on repetitive sessions) |
 | `CLAUDE_CODE_DISABLE_AUTO_MEMORY` | Set to `1` to disable automatic memory |
 
@@ -766,7 +768,7 @@ For the 5 `opus` agents (security-auditor, database-architect, migration-special
 
 **Auto-compaction and skills:** After compaction, skills auto-reload (5K tokens/skill, 25K total max).
 
-**Skills with `context: fork` (24 total):** Heavy (>100 lines, 17): `architect`, `debug-methodical`, `atomic-tasks`, `socratic-brainstorm`, `architecture-clean-ddd`, `parallel-worktrees`, `event-driven`, `cqrs`, `async`, `multitenant`, `testing`, `testing-symfony`, `testing-python`, `testing-react`, `testing-flutter`, `testing-reactnative`, `design-md-convention`. Medium (60-100 lines, 7): `edge-computing`, `wasm`, `graphql`, `monorepo`, `observability`, `docker-hadolint`, `api-gateway`.
+**Skills with `context: fork`:** every Claude Craft skill over 60 lines uses it (checked by the framework's internal conformance audit).
 
 ---
 
@@ -797,7 +799,7 @@ Templates available in `.claude/templates/hooks/`: `output-filter.json`, `pre-co
 | fallbackModel | Opus reliability — no session interruptions (active by default in settings.json) |
 | PostToolUse hook (Bash+Read) | Reduces context pollution (5K-50K tokens/session on large files) |
 | PreCompact hook | Prevents context loss |
-| context:fork skills (24 total) | 2K-15K tokens per long session chaining multiple skills |
+| `context: fork` skills (>60 lines) | 2K-15K tokens per long session chaining multiple skills |
 | **Combined total** | **60-75% overall reduction** |
 
 > Detailed examples and templates: see `@.claude/references/base/context-management.md`