Version: 3.0.0 (Updated with full architecture audit) Date: 2026-04-30 Scope: All AI agents, human contributors, and automated processes working on HelixCode Authority: Derived from HelixAgent AGENTS.md with HelixCode-specific enhancements
Base agent rules:
constitution/AGENTS.md— READ IT FIRST. The base file is authoritative for any topic not covered here. Project-specific rules below extend them; they never weaken them.
The HelixConstitution submodule (./constitution/, added 2026-05-14)
provides the universal CLI-agent operating manual. Every CLI agent
working on this project — Claude Code, Codex, Cursor, Aider,
Continue, Gemini CLI, future LLMs — MUST read both
constitution/AGENTS.md AND this project AGENTS.md, treating the
project file as a tightening overlay (never a weakening).
Project remote-policy note (revised 2026-06-03) — Per operator decision 2026-06-03, HelixCode permits GitFlic + GitVerse Git remotes in addition to GitHub + GitLab; the prior §6.W tightening forbidding GitFlic/GitVerse is LIFTED and pushes target all configured upstreams across all four providers. Other tightenings (e.g. CONST-033 host power management hard-ban, CONST-035 anti-bluff covenant, CONST-042 secret-leak policy, CONST-045 distribution-host config policy, CONST-046 no-hardcoded-content, CONST-047 recursive submodule application mandate) all SURVIVE this inheritance.
HelixCode is an enterprise-grade distributed AI development platform built in Go. It enables intelligent task division, work preservation, cross-platform development workflows, and multi-provider LLM integration through a unified REST API, CLI, Terminal UI, Desktop, and Mobile client architecture.
Current Status: The internal/ foundation is largely solid (auth, database, server, worker, task, workflow, tools, editor, notification, MCP, verifier are real implementations). Critical bluff and stub areas remain in select entry points and peripheral packages. All agents MUST prioritize zero-bluff implementation.
LLMsVerifier Integration Status: internal/verifier/ package is now implemented with REST API client, two-tier cache, circuit breaker health monitor, background poller, score adapter, and event publisher. BLUFF-002 (hardcoded CLI models) and BLUFF-004 (hardcoded external models) are FIXED. BLUFF-005 (scoring ignores verifier data) is FIXED in ModelManager.SelectOptimalModel().
Key Features:
- Distributed Computing: SSH-based worker pools with health monitoring, auto-installation, and consensus
- Multi-Provider LLM Integration: 15+ providers (OpenAI, Anthropic, Gemini, Ollama, Azure, Bedrock, Groq, Mistral, Cohere, xAI, DeepSeek, Qwen, OpenRouter, HuggingFace, Llama.cpp)
- Development Workflows: Automated planning, building, testing, refactoring with real shell execution
- Task Management: Intelligent task division with priorities, dependencies, checkpointing, and Redis caching
- MCP Protocol: Full Model Context Protocol server over WebSocket with tool dispatch
- Multi-Client Architecture: REST API (Gin), Cobra CLI, Terminal UI (tview), Desktop (Fyne), Mobile (gomobile), WebSocket
- Memory Systems: In-memory, filesystem, Redis, Memcached, Cognee, ChromaDB, Qdrant, Weaviate integrations
- Advanced Editor: Multi-format code editing (diff, whole-file, search/replace, line-based) with backups
- Tools Ecosystem: 40+ tools across filesystem, shell, web, browser, mapping, multiedit, confirmation, notebook, git
- Notifications: Multi-channel support (Slack, Email, Telegram, Discord, Yandex Messenger, Max)
Core Technologies:
- Language: Go 1.24.0 with toolchain go1.24.9
- Module:
dev.helix.code - HTTP Framework: Gin v1.11.0
- Authentication: JWT v4.5.2, bcrypt + argon2
- Database: PostgreSQL 15+ via pgx/v5 (optional)
- Cache: Redis 7+ via go-redis/v9 (optional)
- Configuration: Viper v1.21.0
- CLI Framework: Cobra v1.8.0
- Testing: Testify v1.11.1
UI Technologies:
- Desktop: Fyne v2.7.0
- Terminal UI: tview v0.42.0
- Mobile: gomobile bindings
External Integrations:
- Browser Automation: chromedp v0.14.2
- Web Scraping: goquery v1.10.3
- Tree-sitter: go-tree-sitter
- Identity: Azure SDK, AWS SDK v2
- Vector/Memory: Cognee, ChromaDB, Qdrant, Weaviate clients
- Container Orchestration: digital.vasic.containers (vasic-digital/Containers submodule)
CRITICAL: All build and test commands must be run from the helix_code/ subdirectory, not the repository root.
cd HelixCode| Command | Purpose |
|---|---|
make build |
Build server binary to bin/helixcode |
make test |
Run go test -v ./... |
make test-all |
Run tests + coverage + benchmarks + docs |
make test-coverage |
Generate coverage report |
make test-benchmark |
Run Go benchmarks |
make logo-assets |
Generate logo assets (required before first build) |
make setup-deps |
Run go mod tidy |
make fmt |
Run go fmt ./... |
make lint |
Run golangci-lint run ./... |
make clean |
Clean build artifacts |
make dev |
Start development server |
make prod |
Cross-platform production build |
make mobile |
Build iOS + Android targets |
make aurora-os |
Build Aurora OS target |
make harmony-os |
Build Harmony OS target |
| Command | Purpose |
|---|---|
make test-infra-up |
Start full Docker test infrastructure |
make test-infra-down |
Stop full Docker test infrastructure |
make test-full |
ALL tests with real infrastructure (zero skips) |
make test-unit-full |
Unit tests with real services |
make test-integration-full |
Integration tests with -tags=integration |
make test-e2e-full |
E2E challenge tests via runner |
make test-security-full |
Security test suite |
make test-load-full |
Load tests |
make test-complete |
Sequential run of all full test types |
make coverage-full |
Coverage with full infrastructure |
| Command | Purpose |
|---|---|
make container-builder-image |
Build the builder container image |
make container-build |
Build application inside container |
make container-test |
Run tests inside container |
make container-lint |
Run linter inside container |
make container-shell |
Interactive shell in builder container |
make container-dev-up |
Start containerized dev environment |
make container-dev-down |
Stop containerized dev environment |
make container-release |
Full release build in container |
./scripts/containers/build-in-container.sh |
Convenience wrapper script |
The builder container includes: Go 1.24, gcc, postgresql-client, redis, docker-cli, golangci-lint, and all build tools. The only host requirement is Docker/Podman.
| Script | Purpose |
|---|---|
./run_tests.sh --unit |
Unit tests |
./run_tests.sh --integration |
Integration tests |
./run_tests.sh --e2e |
E2E tests |
./run_tests.sh --coverage |
Coverage analysis |
./run_tests.sh --security |
Security tests |
./run_all_tests.sh |
Orchestrates ALL suites sequentially |
./run_integration_tests.sh |
DB integration tests with Docker |
go test -v -run TestName ./path/to/package
go test -v -tags=integration ./internal/database
cd tests/e2e/challenges && go run cmd/runner/main.go -challenge ascii-art-generator-001 -providers ollamahelix_code/
├── cmd/ # Application entry points
│ ├── server/main.go # HTTP server entry point
│ ├── cli/main.go # Legacy flag-based CLI client
│ ├── root.go # Cobra root command (`helix`)
│ ├── main_commands.go # `helix start`, `helix auto`
│ ├── other_commands.go # `helix server`, `helix version`, etc.
│ ├── local-llm.go # `helix local-llm` command tree
│ ├── local-llm-advanced.go # Advanced local-llm commands
│ ├── helix-config/main.go # Dedicated config management CLI
│ ├── security-test/main.go # Simulated security test runner
│ ├── security-fix/main.go # Security fix wrapper
│ ├── security-fix-standalone/main.go # Standalone security scanner
│ ├── performance-optimization/main.go # Performance optimizer
│ ├── performance-optimization-standalone/main.go # Standalone perf simulator
│ └── config-test/main.go # Config hot-reload test utility
│
├── internal/ # Internal packages (~40 packages)
│ ├── auth/ # JWT authentication, bcrypt/argon2, sessions
│ ├── llm/ # LLM provider implementations (15+ providers)
│ │ ├── providers/ # Per-provider HTTP clients
│ │ ├── compression/ # Context compression
│ │ └── vision/ # Vision/multimodal support
│ ├── provider/ # Provider abstractions
│ ├── providers/ # Provider management
│ ├── worker/ # SSH-based worker pool, health checks
│ ├── task/ # Task queues, dependencies, checkpoints
│ ├── server/ # Gin HTTP server, routes, middleware
│ ├── database/ # PostgreSQL pgx pool, schema initialization
│ ├── redis/ # go-redis wrapper with graceful degradation
│ ├── tools/ # 40+ tool ecosystem registry
│ │ ├── filesystem/ # fs_read, fs_write, fs_edit, glob, grep
│ │ ├── shell/ # shell, shell_background with sandbox
│ │ ├── web/ # web_fetch, web_search
│ │ ├── browser/ # browser_launch, browser_navigate, browser_screenshot
│ │ ├── multiedit/ # Transactional multi-file editing
│ │ └── git/ # Git automation
│ ├── editor/ # Multi-format code editing with backups
│ ├── memory/ # Memory providers (in-mem, filesystem, Redis, etc.)
│ ├── cognee/ # Cognee.ai memory integration
│ ├── context/ # Hierarchical context management with TTL
│ ├── notification/ # Multi-channel notification engine
│ ├── mcp/ # Model Context Protocol WebSocket server
│ ├── workflow/ # Development workflow execution
│ ├── config/ # Viper-based configuration management
│ ├── event/ # Pub/sub event bus
│ ├── logging/ # Structured logging wrapper
│ ├── monitoring/ # Metric collection framework
│ ├── security/ # Security scanning (stubbed)
│ ├── session/ # Development session management
│ ├── agent/ # Agent orchestration
│ ├── project/ # Project management
│ ├── rules/ # Rules engine
│ ├── hooks/ # Hook system
│ ├── focus/ # Focus chain management
│ ├── template/ # Template system
│ ├── persistence/ # State persistence
│ ├── deployment/ # Deployment management
│ ├── discovery/ # Service/model discovery
│ ├── hardware/ # Hardware abstraction
│ ├── repomap/ # Repository mapping
│ ├── version/ # Version management
│ ├── fix/ # Security fix engine
│ ├── performance/ # Performance optimization
│ ├── testutil/ # Test utilities
│ └── mocks/ # Shared mocks
│
├── applications/ # Platform-specific applications
│ ├── desktop/ # Fyne desktop app
│ ├── terminal-ui/ # tview terminal UI
│ ├── android/ # Android app
│ ├── ios/ # iOS app
│ ├── aurora-os/ # Aurora OS client
│ └── harmony-os/ # Harmony OS client
│
├── api/ # OpenAPI specification
│ └── openapi.yaml # Full REST API spec (OpenAPI 3.0.3)
│
├── config/ # Configuration files
│ ├── config.yaml # Primary application config
│ ├── production-config.yaml # Enterprise production config
│ ├── minimal-config.yaml # Minimal test config (DB/Redis disabled)
│ ├── test-config.yaml # Test-specific config
│ ├── working-config.yaml # Working variant
│ ├── azure_example.yaml # Azure-specific example
│ └── model-aliases.example.yaml# Model alias examples
│
├── tests/ # New test framework
│ ├── e2e/challenges/ # Challenge-based E2E tests
│ └── automation/ # Hardware automation tests
│
├── test/ # Legacy/parallel test suites
│ ├── integration/ # Integration tests
│ ├── e2e/ # Legacy E2E tests
│ ├── automation/ # Provider automation tests
│ └── load/ # Load tests
│
├── benchmarks/ # Performance benchmarks
├── security/ # Security tests
├── standalone_tests/ # Standalone CLI tests
├── docker/ # Docker assets and extended compose
├── scripts/ # Build and deployment scripts
└── assets/ # Logo and image assets
File: internal/auth/auth.go (~470 lines)
Assessment: Production-ready
- User registration with validation
- Password hashing with bcrypt + argon2 fallback
- JWT token generation and verification (JWT v4)
- Session management with crypto-random tokens
- Constant-time comparison for timing attack prevention
- Full test coverage in
internal/auth/auth_test.go(~777 lines)
File: internal/database/database.go
Assessment: Production-ready
- PostgreSQL connection pool via pgx/v5
- Full schema initialization (users, workers, tasks, projects, sessions, LLM providers, MCP servers, notifications, audit logs)
DatabaseInterfacefor testability- Graceful degradation when host is empty
File: internal/server/server.go
Assessment: Production-ready
- Gin-based server with 50+ routes across
/api/v1/ - JWT auth middleware, CORS, security headers
- WebSocket endpoint for MCP
- Health check with DB + Redis validation
- Graceful shutdown (30s timeout)
File: internal/llm/ (~5000+ lines across providers)
Assessment: Real HTTP clients
AnthropicProvider(~752 lines): Full SSE streaming, prompt caching, extended thinking, tool callsOpenAIProvider(~431+ lines): Full HTTP API clientModelManager: Multi-provider orchestration, selection strategy, fallback chain- 16 provider subdirectories with real HTTP implementations
- Note: The
internal/llm/package is genuine. Bluff areas are atcmd/cli/main.goonly.
File: internal/worker/ (~800+ lines)
Assessment: Real distributed worker management
WorkerManager: Register, heartbeat, assign tasks, complete tasks- SSH config parsing, capability matching, resource tracking
- Health checks with TTL
File: internal/task/ (~1000+ lines)
Assessment: Real task lifecycle
- Priority queues, dependency validation, checkpointing
- Redis caching with graceful degradation
- Retry logic and cleanup
File: internal/workflow/ (~1100+ lines)
Assessment: Real shell execution
Executordispatches to realexec.CommandContext()calls- Security filtering via
isDangerousCommand()(rm, dd, mkfs, fork bombs, etc.) - LLM integration with real
LLMRequest - Supports Go, Node, Python, Rust project types
File: internal/tools/ (~2000+ lines)
Assessment: Real tool registry
- 8 categories: filesystem, shell, web, browser, mapping, multiedit, confirmation, notebook
- Real chromedp browser automation
- Transactional multi-file editing
File: internal/editor/ (~600+ lines)
Assessment: Real file I/O
- Diff, whole-file, search/replace, line-based editors
- Automatic file backup with
io.Copy EditApplier/EditValidatorinterfaces
File: internal/notification/ (~800+ lines)
Assessment: Real HTTP/SMTP calls
- Slack (webhook HTTP POST), Email (SMTP via
net/smtp), Telegram (Bot API), Discord (webhook) - Yandex Messenger (OAuth API), Max (enterprise API)
- Rate limiting, retry, queue, metrics
File: internal/mcp/ (~400+ lines)
Assessment: Real WebSocket server
- gorilla/websocket concurrent session handling
- JSON-RPC-like message format
- Tool execution dispatch
File: internal/config/ (~1700+ lines)
Assessment: Full Viper integration
- Environment variable binding (
HELIX_*) - Config file search (
.,$HOME/.helixcode,/etc/helixcode) - Validation rules, default config creation
ConfigManagerfor load/save/merge
Files: internal/helixqa/, internal/server/qa_handlers.go, applications/terminal_ui/main.go
Assessment: Full embedded QA engine with real session lifecycle
Enginestruct manages QA sessions with map + sync.RWMutexStartSession(),CancelSession(),GetSession(),ListSessions()with real state tracking- REST API:
POST /api/v1/qa/session,GET /api/v1/qa/session/:id/status,GET /api/v1/qa/session/:id/report,GET /api/v1/qa/session/:id/screenshot/:name,DELETE /api/v1/qa/session/:id - CLI flags:
--qa-run,--qa-list,--qa-report,--qa-screenshot,--qa-cancel - TUI dashboard with session table, stats panel, refresh/cancel actions
- Screenshot pipeline: 8 platform engines (Linux, Web, iOS, Android, CLI, TUI, macOS, Windows)
- Tests:
internal/helixqa/wrapper_test.go,internal/server/qa_handlers_test.go,pkg/screenshot/*_test.go
File: cmd/cli/main.go lines ~236-284
Evidence: Previously returned fmt.Sprintf("Generated response for: %s...", prompt) without calling any provider.
Fix: handleGenerate() now constructs a real llm.LLMRequest with user messages and calls provider.Generate() / provider.GenerateStream(). Errors are propagated to the user if the provider is unavailable.
Verification: go build -tags nogui ./cmd/cli/ compiles; provider call is real (returns error if Ollama/etc. is not running).
Fix Priority: P0 — RESOLVED
File: cmd/cli/main.go lines ~101-128
Evidence: Previously only 3 hardcoded models. No dynamic discovery.
Fix: Replaced with verifier-aware handleListModels() that queries LLMsVerifier adapter first, falls back to provider discovery, then to constitutional FallbackModels (7 models with scores and verification status).
Verification: go test -v ./internal/verifier/... passes; go build ./cmd/cli/... compiles.
Fix Priority: P0 — RESOLVED
File: cmd/cli/main.go lines ~310-324
Evidence: Previously printed the command and slept for 1 second without executing anything.
Fix: handleCommand() uses exec.CommandContext(ctx, "sh", "-c", command) with real os.Stdout/os.Stderr redirection. Exit codes are reported.
Verification: go build -tags nogui ./cmd/cli/ compiles.
Fix Priority: P0 — RESOLVED
File: internal/security/security.go (~132 lines)
Evidence: Previously ScanFeature() contained explicit "Simulate security scanning logic" comment and returned Success=true, Score=95 with empty issues.
Fix: Replaced with real Scanner interface, SonarQube/Snyk clients. ScanFeature() now dispatches real security scanners. Score is no longer hardcoded.
Verification: 17/17 tests PASS, zero "simulated" in production code.
Fix Priority: P1 — RESOLVED (commit 33ddf6a)
File: internal/memory/ (~1800+ lines)
Evidence: Previously RedisMemoryProvider and MemcachedMemoryProvider stored data in local maps.
Fix: internal/redis/redis.go provides real go-redis/v9 client. Dedicated memory stubs no longer exist.
Fix Priority: P2 — RESOLVED
File: cmd/security_test/main.go
Evidence: Previously had 12 simulated test results via simulateSecurityScan().
Fix: Rewired to real internal/security scanner dispatch via ScanFeature().
Fix Priority: P2 — RESOLVED (commit 33ddf6a)
File: cmd/other_commands.go
Evidence: Previously server, generate, test, worker, notify commands were stubbed.
Fix: Wired to real server.New(), llm.ModelManager, notification engine, and go test dispatch.
Fix Priority: P2 — RESOLVED (commit 33ddf6a)
File: cmd/helix_config/main.go
Evidence: Many template/history/schema subcommands print placeholder messages.
Fix Priority: P3
File Pattern: internal/verifier/*.go
Evidence: Previously returned hardcoded 8.5 scores, empty discovery results.
Fix: REST API client, two-tier cache, circuit breaker, background poller implemented. Fallback models are constitutionally permitted (CONST-036).
Verification: Fallback models list only at fallback_models.go (permitted). All other verifier code uses real client.
Fix Priority: P0 — RESOLVED
File Pattern: internal/verifier/startup.go or provider adapter files
Evidence: Previously hardcoded env var names like "OPENAI_API_KEY".
Fix: Env vars sourced from config struct fields, not hardcoded strings.
Verification: grep sweep clean.
Fix Priority: P1 — RESOLVED
File Pattern: internal/llm/*.go
Evidence: Previously SupportsToolUse: true as struct literal.
Fix: No hardcoded capabilities found in grep sweep.
Fix Priority: P1 — RESOLVED
File Pattern: *_test.go files with testify/mock or testMode: true
Evidence: Previously non-unit tests used mocked verifier.
Fix: No mocks in non-unit verifier tests.
Verification: grep sweep clean.
Fix Priority: P0 — RESOLVED
File Pattern: configs/verifier.yaml or internal/verifier/config.go
Evidence: Previously scoring weights misconfigured.
Fix: Weights configurable via ScoringWeights struct in adapter.go. No hardcoded weight issues found.
Fix Priority: P2 — RESOLVED
File: internal/server/handlers.go lines ~834-855
Evidence: All dynamic metrics (goroutines, memory, database connections) were hardcoded to 0.
Fix: getMetrics() now calls runtime.ReadMemStats(), runtime.NumGoroutine(), and s.db.Pool.Stat() to return real values.
Fix Priority: P0 — RESOLVED
File: internal/tools/multiedit/transaction.go lines ~352-369
Evidence: detectFileConflict() always returned nil, nil with comment "For now, we'll assume no conflicts."
Fix: Implemented real conflict detection — reads the file from disk, computes SHA-256, and compares against the Checksum field. Returns ConflictModified or ConflictDeleted when appropriate.
Fix Priority: P1 — RESOLVED
Main config at config/config.yaml:
server:
address: "0.0.0.0"
port: 8080
read_timeout: 30
write_timeout: 30
idle_timeout: 300
shutdown_timeout: 30
database:
host: "" # Empty string disables PostgreSQL
port: 5432
user: "helix"
password: "${HELIX_DATABASE_PASSWORD}"
dbname: "helixcode_prod"
sslmode: "disable"
redis:
host: "redis"
port: 6379
password: "${HELIX_REDIS_PASSWORD}"
db: 0
enabled: true
auth:
jwt_secret: "${HELIX_AUTH_JWT_SECRET}"
token_expiry: 86400
session_expiry: 604800
bcrypt_cost: 12
workers:
health_check_interval: 30
health_ttl: 120
max_concurrent_tasks: 10
tasks:
max_retries: 3
checkpoint_interval: 300
cleanup_interval: 3600
llm:
default_provider: "local"
max_tokens: 4096
temperature: 0.7
timeout: 30
max_retries: 3
providers:
<name>:
type: <provider-type>
endpoint: <url>
enabled: true
parameters:
timeout: 30.0
max_retries: 3
streaming_support: true
api_key: ""
selection:
strategy: "performance"
fallback_enabled: true
health_check_interval: 30
logging:
level: "info"
format: "text"
output: "stdout"
notifications:
enabled: true
rules:
- name: "..."
condition: "type==error"
channels: ["slack", "email"]
priority: urgent
enabled: true
channels:
slack: { enabled, webhook_url, channel, username, timeout }
telegram: { enabled, bot_token, chat_id, timeout }
email: { enabled, smtp: { server, port, username, password, tls }, recipients, timeout }
discord: { enabled, webhook_url, timeout }Required for Production:
HELIX_DATABASE_PASSWORDHELIX_AUTH_JWT_SECRETHELIX_REDIS_PASSWORD
LLM Provider Keys (as needed):
OPENAI_API_KEY,ANTHROPIC_API_KEY,GEMINI_API_KEY,XAI_API_KEY,DEEPSEEK_API_KEY,GROQ_API_KEY,MISTRAL_API_KEY,COHERE_API_KEY,AZURE_OPENAI_API_KEY,AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY
Notification Integrations:
HELIX_SLACK_WEBHOOK_URLHELIX_TELEGRAM_BOT_TOKEN,HELIX_TELEGRAM_CHAT_IDHELIX_EMAIL_SMTP_SERVER,HELIX_EMAIL_USERNAME,HELIX_EMAIL_PASSWORDHELIX_DISCORD_WEBHOOK_URL
- Unit tests: Mocks allowed,
*_test.go,-shortflag - Contract tests: Real API schemas, no mocks
- Component tests: Real subsystems wired together
- Integration tests: Full app with real dependencies (
-tags=integration) - E2E challenges: Complete user workflows against real LLM APIs
- Security tests: OWASP compliance
- Performance tests: Benchmarks
- Automation tests: Provider/hardware automation (
-tags=automation) - Load tests: Stress testing
- Unit tests: Mocks OK
- ALL other tests: Real infrastructure ONLY
- Every PASS guarantees Quality + Completion + Usability
- Challenges fail on simulated/stubbed behavior
- No bare
t.Skip()withoutSKIP-OK: #<ticket>marker
docker-compose.test.yml: PostgreSQL 16, Redis 7, Memcached, Cognee, ChromaDB, Qdrant, Ollama, Prometheus, Grafanadocker-compose.full-test.yml: Complete stack with mock-LLM server, Selenium, ChromeDP, SSH server + 3 workers, Cognee, Weaviate, mock-Slack, multicast router
The most rigorous test system validates HelixCode by having it generate real projects and testing them:
- Challenge Definitions: JSON specs (ASCII art generator, CLI task manager, JSON validator, notes API, tic-tac-toe TUI, URL shortener)
- Execution Flow: Load spec → Call real LLM API → Parse generated code → Compile → Test → Runtime validation
- Validation Layers: Directory structure, code quality, compilation, testing, functionality, runtime validation with diverse data
- Test Matrix: Supports CLI, TUI, REST, WebSocket interfaces across 15+ providers and worker pool distributions
# Basic
cd HelixCode && make test
# Full infrastructure (recommended for validation)
make test-infra-up
make test-complete
make test-infra-down
# Individual categories
make test-unit-full
make test-integration-full
make test-e2e-full
make test-security-full
make test-load-full
# Legacy scripts
./run_tests.sh --all
./run_all_tests.sh
./run_integration_tests.shServices: helixcode-server (8080, 2222), postgres:15, redis:7, nginx (80, 443), prometheus (9090), grafana (3000)
cd HelixCode
cp .env.example .env
# Edit .env with secure passwords
docker compose up -d
docker compose ps
curl http://localhost/health| File | Purpose |
|---|---|
docker-compose-simple.yml |
Minimal dev (postgres + redis only) |
docker-compose.test.yml |
Integration/E2E testing stack |
docker-compose.full-test.yml |
Zero-skip full test infrastructure |
docker-compose.aurora-os.yml |
Security-focused Aurora OS platform |
docker-compose.harmony-os.yml |
Distributed Harmony OS platform |
docker-compose.specialized-platforms.yml |
Combined Aurora + Harmony |
docker/docker-compose.yml |
Extended full-stack with Milvus, Elasticsearch, MLflow, Jaeger, Jupyter, Portainer |
- Healthchecks on every service
- Docker profiles:
monitoring,distributed,with-redis,production,dev,server - Isolated bridge networks per deployment
- Named persistent volumes for all stateful services
.envfile for secrets
- Standard Go formatting:
go fmt ./... - Linting:
golangci-lint run ./...(timeout 10m in CI) - Vet:
go vet ./... - Table-driven tests with
t.Run()subtests - Build tags for integration/automation tests:
//go:build integration
- Always work from
helix_code/subdirectory - Generate logo assets before first build:
make logo-assets - Database/Redis optional: Disable by setting
database.host: "" - Environment variables override config file
- Use
internal/for all core packages; nopkg/directory in active use - Error handling: explicit, no silent failures
- Concurrent access: use
sync.RWMutexor channel patterns
- REST API documented in
api/openapi.yaml(OpenAPI 3.0.3) - Base path:
/api/v1 - Authentication: Bearer JWT via
Authorizationheader - Health endpoint:
GET /health(no auth required)
- Password hashing: bcrypt (cost 12) with argon2 fallback
- JWT with constant-time comparison
- CORS middleware, security headers (X-Frame-Options, CSP, HSTS)
- Rate limiting support in production config
- Session timeout, concurrent session limits, IP binding options
- Workflow
isDangerousCommand()filter blocks rm, dd, mkfs, fork bombs, etc. - Input validation in auth and server packages
security/security_test.go: OWASP Top 10, SAST, DAST, credential scanning, TLS enforcement, input validation (path traversal, XSS, SQL injection, command injection, SSRF)- File permission checks (0600 for configs)
internal/security/security.go: Simulated scanning (always returns clean)cmd/security_test/main.go: Entirely simulated security tests
- Use
HELIX_AUTH_JWT_SECRETwith high entropy - Enable PostgreSQL SSL in production
- Enable Redis authentication
- Configure CORS
allowed_originsexplicitly - Enable audit logging
- Set
bcrypt_cost: 14in production
- NO CI/CD pipelines (Note: existing workflow files in
.github/workflows/are legacy and must not be expanded) - NO HTTPS for Git (SSH only)
- NO manual container commands (orchestrator-owned)
- 100% Test Coverage (unit, integration, E2E, automation, security, benchmark)
- Challenge Coverage (every component)
- Real Data (actual API calls, real DB, live services)
- Health & Observability (health endpoints, circuit breakers)
- Documentation & Quality (update docs with code changes)
- Validation Before Release (full suite + all challenges)
- No Mocks in Production
- Comprehensive Verification (runtime, compile, structure, dependencies, compatibility)
- Resource Limits (30-40% of host resources max)
- Bugfix Documentation (root cause, affected files, fix, verification link)
- Real Infrastructure for All Non-Unit Tests
- Reproduction-Before-Fix (Challenge first, then fix)
- Concurrent-Safe Collections
A change is NOT done because code compiles. "Done" requires:
- Pasted terminal output from a real run
- No self-certification words without evidence
- Demo commands that run against real artifacts
- Loud skips with
SKIP-OK: #<ticket>markers
A test or Challenge that PASSES is a CLAIM that the tested behavior works for the end user of the product.
The HelixAgent project has repeatedly hit the failure mode where every test ran green AND every Challenge reported PASS, yet most product features did not actually work — buggy challenge wrappers masked failed assertions, scripts checked file existence without executing the file, "reachability" tests tolerated timeouts, contracts were honest in advertising but broken in dispatch. This MUST NOT recur in HelixCode.
Every PASS result MUST guarantee: a. Quality — correct behavior under real inputs, edge cases, concurrency b. Completion — wired end-to-end with no stub/placeholder gaps c. Full usability — a user following documentation succeeds
A passing test that doesn't certify all three is a bluff and MUST be tightened.
- Wrapper bluff — assertions PASS but wrapper's exit-code logic is buggy
- Contract bluff — system advertises capability but rejects it in dispatch
- Structural bluff — file exists but doesn't contain working code
- Comment bluff — comment promises behavior code doesn't have
- Skip bluff —
t.Skip("not running yet")withoutSKIP-OK: #<ticket>marker
The taxonomy is illustrative, not exhaustive. Every Challenge or test added going forward MUST pass an honest self-review against this taxonomy before being committed.
Verbatim user mandate: "We had been in position that all tests do execute with success and all Challenges as well, but in reality the most of the features does not work and can't be used! This MUST NOT be the case and execution of tests and Challenges MUST guarantee the quality, the completion and full usability by end users of the product!"
Operative rule: The bar for shipping is not "tests pass" but "users can use the feature." Every PASS in this codebase MUST carry positive runtime evidence captured during execution. Metadata-only / configuration-only / absence-of-error / grep-based PASS without runtime evidence are critical defects regardless of how green the summary line looks. No false-success results are tolerable.
No API key, token, password, certificate, or other credential may be committed to any repository owned by HelixDevelopment or vasic-digital. All secrets live in .env files (mode 0600) listed in .gitignore. Any leak is a release blocker until rotated and post-mortemed.
No force push, force-with-lease push, history rewrite, branch deletion of main/master, or upstream-overwriting operation may be performed without explicit, in-conversation user approval per operation. Authorization for one push does not extend further. Bypassing hooks / signing / protected-branch rules also requires explicit approval.
The docs/CONTINUATION.md document MUST be maintained in sync with actual programme state at all times. It is the authoritative resumption record for any CLI agent or LLM picking up the CLI-Agent Fusion programme. Every commit that advances state — task completion, feature close-out, push, known-issue discovery, deferred-item resolution, phase transition, submodule/remote add or remove — MUST update CONTINUATION in the same commit. Out-of-sync CONTINUATION is a CRITICAL DEFECT — same severity as a false-success test result under CONST-035 / Article XI §11.9. Cascade applies to every owned-by-us repo's three governance files. See CONSTITUTION.md Article XIII §13.1 for the full mandate (definition of out-of-sync, verification plan, cascade requirement).
ALL container distribution targets SHALL be configured exclusively through CONTAINERS_REMOTE_HOST_N_* env vars in containers/.env. NO host (hostname, IP, user, key path, runtime, label) may be hardcoded in ANY source file, test, challenge, config template, script, or governance document. Adding/removing hosts = editing containers/.env only; NO code change. Tests SHALL read .env at runtime and skip with SKIP-OK: marker when CONTAINERS_REMOTE_ENABLED=false. See CONSTITUTION.md §CONST-045 for the full mandate.
Rule: LLMsVerifier SHALL BE the sole authoritative source for:
- All model metadata (names, IDs, context windows, capabilities)
- All provider metadata (endpoints, auth types, supported models)
- All verification status (verified, partial, failed, pending)
- All scoring data (overall scores, capability scores, tier rankings)
Prohibition: NO hardcoded model lists, NO hardcoded provider lists, NO simulated model discovery. Any code path that presents a model or provider listing to a user MUST fetch that listing from the LLMsVerifier subsystem or its cached replica.
Anti-Bluff Verification:
- Challenge script
scripts/anti_bluff/verifier_hardcode_check.shscans all Go source files for hardcoded model arrays. - The only permitted hardcoded data is the 7-entry fallback list in
internal/verifier/fallback_models.go.
Rule: Every model displayed to an end user MUST have been verified by LLMsVerifier within the last 24h. Models older than this MUST display a "stale" indicator and be deprioritized.
Anti-Bluff Testing:
- Unit tests MAY mock the verifier client.
- Integration tests MUST start the verifier server and perform real provider discovery.
- The Makefile target
make test-verifier-integrationMUST exist and run without mocks.
Rule: Model status (available, rate-limited, cooldown, offline, deprecated) displayed to users MUST reflect the actual state as known by LLMsVerifier within 60 seconds.
Polling vs. Push:
- If WebSocket/SSE push is unavailable, the system MUST poll LLMsVerifier at most every 60s.
- The TUI MUST display a "last updated" timestamp with every model listing.
- Models in "cooldown" or "rate-limited" state MUST show the estimated recovery time if known.
Rule: HelixCode MUST integrate with ALL providers that LLMsVerifier supports, subject only to:
- The provider being explicitly disabled in configuration (
enabled: false) - The API key being absent and the provider requiring one
- The provider being marked
deprecatedin the verifier database
Minimum Provider Set (SHALL NOT be reduced without constitutional amendment): OpenAI, Anthropic, Gemini, DeepSeek, Groq, Mistral, xAI, OpenRouter, Ollama, Llama.cpp.
Rule: LLMsVerifier integration SHALL extend beyond basic model listing to cover ALL capability dimensions:
- MCP: The verifier MUST report which models support MCP tool calling.
- LSP: The verifier MUST report code-analysis capabilities.
- ACP: The verifier MUST report multi-agent coordination support.
- Embedding: The verifier MUST report
supports_embeddingsfor each model. - RAG: The verifier MUST report context-window sizes for chunking strategies.
- Skills / Plugins: The verifier MUST track plugin compatibility.
Prohibition: Capability flags MUST NOT be hardcoded. The Provider.GetCapabilities() method MUST return data sourced from the verifier's VerificationResult fields.
Rule: NO user-facing text, prompt template, question text, error message, label, helper text, or explanatory content may be hardcoded as a static literal string in any source file. All such content MUST be:
- Generated dynamically by an LLM at runtime based on user context (language, prompt content, session state), OR
- Loaded from an i18n resource file (
.yaml,.json,.toml) that supports locale overrides, OR - Composed programmatically from verifier metadata, configuration data, or provider responses.
Rationale: Hardcoded English text silently breaks the product for non-English users. A clarification question hardcoded as "Which file has the bug?" may be shown identically to a Japanese, Serbian, or Spanish user — producing an incoherent, unusable experience. Every piece of user-facing text MUST adapt.
Violation examples: []string{"Which file has the bug?"}, hardcoded TUI labels, static error prose, hardcoded English prompt templates.
Compliant patterns: LLM-generated questions via DetectAmbiguity(prompt), i18n-loaded labels with locale fallback, verifier-composed model descriptions.
Enforcement: make lint MUST include a hardcoded-text scan. Anti-bluff sweeps MUST flag obvious hardcoded-text patterns.
See CONSTITUTION.md §CONST-046 for the full mandate and cascade requirements.
Verbatim user mandate (2026-05-14): "Make sure all work we do is applied ALWAYS to all Submodules we control under our organizations (vasic-digital and HelixDevelopment) fully recursively everywhere with full bluff-proofing and comprehensive documentation, user manuals and guides and full tests and Challenges coverage!"
Every engineering deliverable produced for HelixCode MUST be applied — fully and recursively — to every owned submodule under the vasic-digital and HelixDevelopment GitHub organizations. Each owned submodule MUST receive in lockstep: (1) the same anti-bluff posture (CONST-035 / Article XI §11.9), (2) comprehensive documentation (README + user manuals + guides matching actual capabilities), (3) full tests + Challenges coverage with captured runtime evidence, (4) recursive propagation through nested submodules under the same orgs, (5) synchronized commits when meta-repo state advances a submodule's surface.
Owned submodule baseline (direct children as of 2026-05-14): vasic-digital/{Containers, Security, Challenges, LLMsVerifier, Models}, HelixDevelopment/{HelixQA, DocProcessor, LLMOrchestrator, LLMProvider, VisionEngine, HelixAgent}, HelixDevelopment-s-Code/Website — plus any nested submodule under these orgs (full recursion).
Cascade requirement: This rule (verbatim or by CONST-047 ID reference) MUST appear in every owned-by-us submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md.
See CONSTITUTION.md §CONST-047 for the full mandate and verification details.
Verbatim user mandate (2026-05-15): "Make sure that every feature, every functionality, every flow, every use case, every edge case, every service or application, on every platform we support is covered with full automation tests which will confirm anti-bluff policy and provide the proof of fully working capabilities, working implementation as expected, no issues, no bugs, fully documented, tests covered! Nothing less than this does not give us a chance to deliver stable product! This is mandatory constraint which MUST BE respected without ignoring, skipping, slacking or forgetting it!"
No feature / functionality / flow / use case / edge case / service / application on any supported platform of HelixCode is deliverable until automation tests prove six invariants: (1) anti-bluff posture with captured runtime evidence (CONST-035); (2) proof of working capability end-to-end on target topology; (3) implementation matches documented promise; (4) no open issues/bugs surfaced; (5) full documentation in sync; (6) four-layer test floor (pre-build + post-build + runtime + paired mutation). Coverage ledger regenerated at release-gate sweep.
Cascade requirement: This rule (verbatim or by CONST-048 ID reference) MUST appear in every owned-by-us submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. See constitution submodule Constitution.md §11.4.25 for the full mandate.
CONST-049: Constitution-Submodule Update Workflow Mandate (cascaded from constitution submodule §11.4.26)
Verbatim user mandate (2026-05-15): "Every time we add something into our root (constitution Submodule) Constitution, CLAUDE.MD and AGENTS.MD we MUST FIRST fetch and pull all new changes / work from constitution Submodule first! All changes we apply MUST BE commited and pushed to all constitution Submodule upstreams! In case of conflict, IT MUST BE carefully resolved! Nothing can be broken, made faulty, corrupted or unusable! After merging full validation and verification MUST BE done!"
7-step pipeline: (1) fetch + pull first inside constitution submodule; (2) apply with §11.4.17 classification + verbatim mandate; (3) validate (meta-test + no conflict markers + cross-file consistency); (4) commit + push to EVERY configured upstream (governance files only, no git add -A); (5) careful conflict resolution preserving union — force-push forbidden (CONST-043); (6) post-merge cascade verification (CONST-047); (7) bump consuming .gitmodules pointer in SAME commit.
Cascade requirement: This rule (verbatim or by CONST-049 ID reference) MUST appear in every owned-by-us submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. See constitution submodule Constitution.md §11.4.26 for the full mandate.
CONST-050: No-Fakes-Beyond-Unit-Tests + 100%-Test-Type-Coverage Mandate (cascaded from constitution submodule §11.4.27)
Verbatim user mandate (2026-05-15): "Mocks, stubs, placeholders, TODOs or FIXMEs are allowed to exist ONLY in Unit tests! All other test types MUST interract with real fully implemented System! No fakes, empty implementations or bluffing is allowed of any kind! All codebase of the project MUST BE 100% covered with every supported test type: unit tests, integration tests, e2e tests, full automation tests, security tests, ddos tests, scaling tests, chaos tests, stress tests, performance tests, benchmarking tests, ui tests, ux tests, Challenges (fully incorporating our Challenges Submodule). EVERYTHING MUST BE tested using helix_qa (fully incorporating helix_qa Submodule). helix_qa MUST BE used with all possible written tests suites (test banks) for every applications, service, platform, etc and execution of the full helix_qa QA autonomous sessions! All required dependency Submodules MUST BE added into the project as well (fully recursive!!!)."
Two invariants. (A) Mocks/stubs/fakes/placeholders/TODOs/FIXMEs/"for now"/empty-implementation patterns PERMITTED only in unit-test sources; non-unit tests (integration, E2E, full-automation, security, DDoS, scaling, chaos, stress, performance, benchmarking, UI, UX, Challenges, HelixQA) MUST exercise the real, fully implemented HelixCode system. Production code MUST NOT import helix_code/internal/mocks/. (B) Codebase MUST be covered by every supported test type the domain warrants: unit, integration, E2E, full-automation, security, DDoS, scaling, chaos, stress, performance, benchmarking, UI, UX, Challenges (vasic-digital/Challenges submodule at ./challenges/ fully incorporated), helix_qa (HelixDevelopment/HelixQA submodule at ./helix_qa/ fully incorporated, with full autonomous QA sessions executing every registered test bank with captured wire evidence).
Required dependency submodules (recursive per CONST-047): Challenges (git@github.com:vasic-digital/Challenges.git at ./challenges/) + helix_qa (git@github.com:HelixDevelopment/HelixQA.git at ./helix_qa/) + any other functionality submodules under vasic-digital/HelixDevelopment orgs HelixCode depends on. Pointers bumped to upstream HEAD in same commit as cascade work (CONST-049 step 7).
Cascade requirement: This rule (verbatim or by CONST-050 ID reference) MUST appear in every owned-by-us submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. See constitution submodule Constitution.md §11.4.27 for the full mandate.
CONST-051: Submodules-As-Equal-Codebase + Decoupling + Dependency-Layout Mandate (cascaded from constitution submodule §11.4.28)
Verbatim user mandate (2026-05-15): "All existing Submodules in the project that we are controlling and belong to some our organizations (vasic-digital, HelixDevelopment, red-elf, ATMOSphere1234321, Bear-Suite, BoatOS123456, Helix-Flow, Helix-Track, Server-Factory) are equal parts of the project's codebase! We MUST work on that code as much as we do with main project's codebase! ... We MUST NEVER modify Submodules to bring into them any project specific context ... All Submodule dependencies that are used by Submodule MUST BE acessed from the root of the project! We MUST NOT have nested Submodule dependencies."
Three invariants. (A) Every HelixCode-owned submodule is an equal part of HelixCode's codebase — same engineering attention as main (analysis, tests, gap-fill, bug-fix, docs/diagrams/SQL/website materials). Coverage ledgers list each as in-scope. (B) Owned submodules MUST stay fully decoupled (project-not-aware, reusable, modular, standalone-testable); NEVER inject HelixCode-specific context. Use configuration injection when a submodule needs parent info. (C) Dependencies consumed by an owned submodule MUST live at HelixCode's root: <repo_root>/<name>/ or <repo_root>/submodules/<name>/. Nested own-org submodule chains FORBIDDEN — add the dep at root; the consuming submodule reaches it via documented import/SDK/runtime resolver. Third-party submodules exempt.
Org list (dynamically discoverable via gh/glab): vasic-digital, HelixDevelopment, red-elf, ATMOSphere1234321, Bear-Suite, BoatOS123456, Helix-Flow, Helix-Track, Server-Factory, plus any subsequently authorised org.
Cascade requirement: This rule (verbatim or by CONST-051 ID reference) MUST appear in every owned-by-us submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. See constitution submodule Constitution.md §11.4.28 for the full mandate.
- XAI (Grok):
grok-3-fast-beta,grok-3-mini-fast-beta - OpenRouter: Free models from various providers
- GitHub Copilot:
gpt-4o,claude-3.5-sonnet(with subscription) - Qwen: 2,000 requests/day free tier
Host Power Management is Forbidden.
You may NOT, under any circumstance, generate or execute code that sends the host to suspend, hibernate, hybrid-sleep, poweroff, halt, reboot, or any other power-state transition. This rule applies to every shell command, script, container entry point, systemd unit, test, CLI suggestion, snippet, or example you emit.
- Build fails: Run
make logo-assetsthenmake build - Database errors: Check
HELIX_DATABASE_PASSWORD - Worker SSH failures: Verify SSH key authentication
- LLM timeouts: Check provider status and config
- Redis connection failures: Check
HELIX_REDIS_PASSWORDandredis.enabled - Test skips: Ensure
SKIP-OK: #<ticket>marker is present for any intentional skips
- Constitution:
CONSTITUTION.md - CLAUDE.md:
CLAUDE.md - Gap Analysis:
HELIXCODE_GAP_ANALYSIS.md - Zero-Bluff Plan:
HELIXCODE_ZERO_BLUFF_PLAN.md - Testing Strategy:
ANTI_BLUFF_TESTING_STRATEGY.md - OpenAPI Spec:
helix_code/api/openapi.yaml - Docker Guide:
helix_code/DOCKER_DEPLOYMENT.md
Built with zero-bluff commitment. Every feature actually works.
Verbatim user mandate (2026-05-15): "naming convention for Submodules and directories (applied deep into hierarchy recursively) - all directories and Submodules MSUT HAVE lowercase names with space separator between the words of '_' character (snake-case)! All existing Submodules and directories which are not following this rule MUST BE renamed! However, since this will most likely break some of the functionalities renaming we do MUST BE applied to all references to particular Submodule or directory! ... There MUST BE reasonable exceptions for this rules - source code for programming languages or Submodules which apply different naming convention - Android, Java, Kotlin and others. ... Upstreams directory which all of our projects and Submodules have MUST BE renamed to the lowercase letters too, however root project containing the install_upstreams system command (it is exported in out paths in our .bashrc or .zshrc) MUST BE updated to fully work with both Upstreams and upstreams directory. ... NOTE: Rules lowercase / snake-case do apply to all project files as well and references to it and from them!"
Every directory, submodule, and file in HelixCode MUST use lowercase snake_case names. Existing non-compliant names (helix_code/, challenges/, containers/, helix_agent/, helix_qa/, security/, github_pages_website/, upstreams/, dependencies/, etc.) MUST be renamed as part of the phased migration opened by this clause. Every reference (configs, docs, links, source-code imports, governance files) MUST be updated atomically with the rename — reference drift after a rename is a CONST-052 violation of equal severity to the rename itself.
Common-sense exceptions (technology-preserving): language-mandated case for Java/Kotlin/Android/Apple/C#/Swift INSIDE the language root (submodule root follows our convention; subtree follows language convention); vendor/upstream third-party submodules keep upstream names; build artefacts (node_modules, __pycache__, .git, target, build, bin) keep tool-mandated names. The test "does renaming break the technology?" trumps the rule.
upstreams/ → upstreams/ transition: the constitution submodule's install_upstreams.sh (exported via .bashrc/.zshrc) supports BOTH upstreams/ and upstreams/ directory layouts (commit 45d3678 of the constitution submodule); lowercase wins when both present.
Test coverage of renames (per CONST-050(B)): every rename batch ships with (i) regression test verifying every reference now resolves, (ii) full test-type matrix run post-rename, (iii) anti-bluff wire-evidence captured.
Phased execution per the operator's explicit instruction: comprehensive brainstorming → phase-divided plan → fine-grained tasks/subtasks → every change covered by every applicable test type. §11.4.20 subagent delegation for cross-cutting rename sweeps.
Cascade requirement: This anchor (verbatim or by CONST-052 ID reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to a §11.4 PASS-bluff at the reference-integrity layer. No escape hatch beyond the common-sense exceptions enumerated above. See constitution submodule Constitution.md §11.4.29 for the full mandate.
CONST-053: .gitignore + No-Versioned-Build-Artifacts Mandate (cascaded from constitution submodule §11.4.30)
Verbatim user mandate (2026-05-15): "every project module, every Submodule, every servcie and apolication MUST HAVE proper .gitignore file! We MUST NOT git version build artifacts, cache files, tmp files, main .env file(s) or any files containing sensitive data, API keys or token! Any build derivate which we can recreate by executing proper mechanism for generating MUST NOT be versioned! We MUST pay attention what is going to be commited every time we are preparing to execute commit! If any violetion is detected it MUST be fixed before commit is executed!"
Every project module, owned-by-us submodule, service, and application MUST ship a proper .gitignore. Forbidden-from-version-control classes:
- Build artefacts:
/bin/,/build/,/dist/,/out/,target/,*.exe,*.dll,*.so,*.dylib,*.a,*.o,*.class,*.pyc, generator-produced files when the generator is committed. - Cache files:
__pycache__/,.pytest_cache/,.mypy_cache/,.ruff_cache/,node_modules/,.next/,.cache/,.gradle/,.terraform/, language-server caches. - Temp files:
*.tmp,*.swp,*~,.DS_Store,Thumbs.db,*.orig,*.rej. - Sensitive-data files:
.env,.env.*(allow.env.exampleplaceholder only — no real secrets even as examples),*.pem,*.key,*.crt,id_rsa*,id_ed25519*,.netrc,secrets/,api_keys.sh. - Generated reports/logs:
*.log,coverage.out,htmlcov/, runtime captures unless reference assets. - OS/IDE personal state:
.idea/,.history/,.vscode/(except shared settings).
Anti-bluff invariant: .gitignore line alone is not sufficient — no file matching the forbidden patterns may be CURRENTLY TRACKED. A tracked *.log despite the ignore-line is a violation of equal severity to no ignore-line at all.
Pre-commit attention: every commit author (human OR agent) MUST inspect git diff --staged + git status BEFORE executing the commit. Forbidden-class hits abort the commit until fixed (un-stage, add to .gitignore, scrub if already-tracked). Gate CM-GITIGNORE-PRECOMMIT-AUDIT + paired mutation.
Secret-leak intersection (CONST-042 / §11.4.10): a .env leak is BOTH a CONST-053 and a CONST-042 violation; rotation + post-mortem required.
Recreatable-content test: if a documented mechanism regenerates the file from sources, it is a build derivative and MUST be ignored. The committed sources MUST include the generator.
Cascade requirement: This anchor (verbatim or by CONST-053 ID reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to a §11.4 PASS-bluff at the repository-hygiene layer. See constitution submodule Constitution.md §11.4.30 for the full mandate.
Verbatim user mandate (2026-05-15): "We MUST HAVE mechanism for each Submodule to determine / know what are its Submodule dependencies so new projects or palces we are incorporate them can add these Submodules to the project root and make them available! Suggested idea is configuration file with expected Submodules Git ssh urls perhaps? New project can read it, and recursively add each Submodule to the root of the project and install / expose it to veryone."
Every owned-by-us submodule MUST ship helix-deps.yaml at its root declaring its own-org dependencies. Schema: schema_version, deps: [{name, ssh_url, ref, why, layout: flat|grouped}], transitive_handling.{recursive,conflict_resolution}, language_specific_subtree. Tooling: incorporate-submodule <ssh-url> adds the submodule at the parent project's canonical path (CONST-051(C)), reads helix-deps.yaml, recurses for each declared dep, aborts on conflicting refs, emits <root>/.helix-manifest.yaml audit record.
Anti-bluff guarantee: every manifest paired with a Challenge that bootstraps a throwaway consuming project, runs incorporate-submodule, asserts produced layout matches the manifest, runs the submodule's own tests against the bootstrapped layout, captures wire evidence per §11.4.2. A manifest without this proof is a CONST-054 violation.
§11.4.31 / CONST-054 is the operational complement of CONST-051(C): nested own-org submodule chains are FORBIDDEN — manifests are the bridge that lets consumers reconstruct the dependency graph at the parent root.
Cascade requirement: This anchor (verbatim or by CONST-054 ID reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to §11.4 PASS-bluff at the dependency-graph layer. See constitution submodule Constitution.md §11.4.31 for the full mandate.
CONST-055: Post-Constitution-Pull Validation Mandate (cascaded from constitution submodule §11.4.32)
Verbatim user mandate (2026-05-15): "Every time we fetch and pull new changes on constitution Submodule we MUST process the whole project and all Submodule (deep recursively) for validation and verification taht every single rule or mandatory constraint is followed and respected! If it is not, IT MUST BE!"
Whenever a project's constitution submodule is fetched + pulled with any content change, the project MUST run scripts/verify-all-constitution-rules.sh BEFORE the new constitution HEAD is treated as canonical for any other work. The sweep re-runs the governance-cascade verifier AND every implementable rule gate (CONST-053 .gitignore audit, CONST-051(C) nested-own-org-chain audit, CONST-052 case audit, CONST-050(A) mock-from-production audit, CONST-035 anti-bluff smoke, etc.) against the post-pull tree. Failures populate the project's Issues tracker per §11.4.15 (Status: Reopened, Type: Bug); closure requires positive-evidence per §11.4.
Pull-time invocation: git submodule update --remote constitution triggers the sweep automatically (post-update hook OR commit-wrapper invocation). Operator-explicit manual invocation also available.
Anti-bluff: the sweep's own meta-test (paired mutation per §1.1) plants a known violation of each enforced gate and asserts the sweep reports FAIL for the planted gate. A sweep that exits PASS without running every implementable gate is a CONST-055 violation.
CONST-055 is the enforcement engine for every other §11.4.x and CONST-NNN rule — without it, new rules cascade as anchors but never get enforced.
Cascade requirement: This anchor (verbatim or by CONST-055 ID reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to §11.4 PASS-bluff at the constitutional-enforcement layer. See constitution submodule Constitution.md §11.4.32 for the full mandate.
CONST-056: Mandatory install_upstreams on clone/add Mandate (cascaded from constitution submodule §11.4.36)
Verbatim user mandate (2026-05-15): "Every Submodule or Git repository we add or clone MUST BE upstreams installed using Upstreamable utility which MUST BE available through exported paths of the host system (in .bashrc or .zhrc) using install_upstreams command executed from the root of the cloned (added) repository - only if in it is Upstreams or upstreams directory present with bash script files (recipes) for all repository's upstreams!"
Every clone / add of a Git repository under HelixCode MUST be followed by install_upstreams invocation from the repository's root IF its tree contains upstreams/ (or legacy upstreams/ per CONST-052 transition) populated with *.sh recipe files. The utility (installed on operator's PATH via .bashrc/.zshrc; implementation in the constitution submodule's install_upstreams.sh — already supports BOTH directory names since constitution commit 45d3678) reads the recipe files, configures every declared upstream as a named git remote, and fans out origin push URLs.
Skipping the invocation when upstreams/ is present silently breaks §2.1 (multi-upstream push is the norm) — the next push lands on only one upstream. Gate CM-INSTALL-UPSTREAMS-ON-CLONE + paired mutation. Automation: the future incorporate-submodule per CONST-054 auto-invokes; manual invocation supported. Pre-commit check: git remote -v | grep -c push reports expected count.
Cascade requirement: This anchor (verbatim or by CONST-056 ID reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. See constitution submodule Constitution.md §11.4.36 for the full mandate.
Every project tracking work items by Type per §11.4.16 MUST close them with the Type-appropriate terminal **Status:** value, drawn from this 3-element closed map:
Item **Type:** |
Closure **Status:** value |
|---|---|
Bug |
Fixed (→ Fixed.md) |
Feature |
Implemented (→ Fixed.md) |
Task |
Completed (→ Fixed.md) |
The (→ Fixed.md) suffix is preserved across all three so the existing migration-discipline tooling (atomic Issues.md → Fixed.md move per §11.4.19) keeps working without per-Type branching. Generators (generate_issues_summary.sh, generate_fixed_summary.sh, the §11.4.23 colorizer) MUST treat the three terminal values as semantically equivalent (all "closed, positive evidence captured") while preserving the literal in the emitted document.
Closing a Feature with Fixed (→ Fixed.md) or a Task with Implemented (→ Fixed.md) is a CONST-057 violation. Gate CM-CLOSURE-VOCAB-TYPE-AWARE walks every Fixed.md heading + every Issues.md heading whose **Status:** is one of the three terminal values and asserts the Status-Type match. Composes with §11.4.15 / §11.4.16 / §11.4.19 / §11.4.23.
Cascade requirement: This anchor (verbatim or by CONST-057 ID reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. See constitution submodule Constitution.md §11.4.33 for the full mandate.
Every Issues.md (or equivalent project tracker) heading whose **Status:** is Reopened MUST carry, within 8 non-blank lines of the heading, a **Reopened-Details:** line capturing four sub-facts:
- By:
AIorUser(source-of-truth observer who flipped the status).AIcovers in-loop reopens (test failure, gate regression, captured-evidence retrospect).Usercovers operator-side observations (manual testing, end-user report, design reconsideration). - On: ISO date (
YYYY-MM-DD). - Reason: one-line cause classification — chosen from the closed vocabulary
{ test-failed | manual-testing-detected | captured-evidence-contradicts | end-user-report | cycle-re-discovered | design-reconsidered }. Other values permitted with explicitReason: <free text>annotation but the closed list MUST be tried first. - Evidence: path to or short description of the captured artefact justifying the reopen — log file, recording, gate failure ID, operator quote, etc. Reopens without evidence are §11.4.6 / §11.4.7 violations (demotion from Fixed requires captured evidence under the conditions that re-exposed the defect).
The Issues_Summary.md Status column MUST distinguish the four Reopened sub-states by source so a sweep query for "reopens by AI in the last 30 days" is mechanically possible. Suggested column rendering: Reopened (AI: test-failed) vs Reopened (User: manual-testing). Gate CM-ITEM-REOPENED-DETAILS mirrors CM-ITEM-OPERATOR-BLOCKED-DETAILS (§11.4.21 walk pattern). Composes with §11.4.6 / §11.4.7 / §11.4.15 / §11.4.21.
Cascade requirement: This anchor (verbatim or by CONST-058 ID reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. See constitution submodule Constitution.md §11.4.34 for the full mandate.
The constitution submodule's three files (constitution/Constitution.md, constitution/CLAUDE.md, constitution/AGENTS.md) ARE the canonical root (also called the parent files). They contain only universal rules per §11.4.17.
The consuming project's repository-root files (<project-root>/CLAUDE.md, <project-root>/AGENTS.md, optionally <project-root>/Constitution.md) are consumer extensions. They MUST start with the inheritance pointer (either the Claude-Code native @constitution/CLAUDE.md import or the portable ## INHERITED FROM constitution/CLAUDE.md heading). They contain only project-specific rules per §11.4.17.
When in doubt about which file to edit: universal rule → constitution submodule's file; project-specific rule → consumer's file. Default consumer-side when uncertain (§11.4.17 — narrower scope is cheap to widen).
Terminology: "the parent CLAUDE.md" / "the root Constitution" → constitution-submodule file at constitution/<filename>; "the project CLAUDE.md" / "this project's AGENTS.md" → consumer-side file at <project-root>/<filename>.
No silent demotion or silent promotion. Moving a rule between layers MUST be a visible commit — git mv of a section if it's a clean clone, or explicit Lifted from <project> to constitution per §11.4.35 / Demoted from constitution to <project> per §11.4.35 commit-message annotation.
Gate CM-CANONICAL-ROOT-CLARITY verifies (a) consumer's CLAUDE.md opens with the inheritance pointer, (b) constitution submodule's three files are present at the expected path, (c) no ## INHERITED FROM block in the constitution submodule's own files (those ARE the source-of-truth, not consumers). Composes with §11.4.17.
Cascade requirement: This anchor (verbatim or by CONST-059 ID reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. See constitution submodule Constitution.md §11.4.35 for the full mandate.
Verbatim user mandate (2026-05-15): "Make sure that feedback_fetch_before_edit memory rule is part of our constitution Submodule - the root Consitution, AGENTS.MD and CLAUDE.MD. Validate and verify that Proejct-Toolkit and all Submodules do inherit all of them! Follow the constitution Submodule documentation for details."
The FIRST git-touching action of every session, on every consuming project (owned or third-party), MUST be:
git fetch --all --prune
git log --oneline HEAD..@{u}
git submodule foreach --recursive 'git fetch --all --prune --quiet'If HEAD..@{u} is non-empty, integrate the upstream changes BEFORE any local edit. Acting on stale local state produces three failure modes documented in the originating §11.4.37 incident (multi-agent / parallel-session work): (1) redundant work — the agent re-does what a parallel session already finished, (2) false confidence — completion reports for already-done work, (3) divergent history — duplicate sibling commits that double the conflict surface on next push.
Anti-bluff invariant: the fetch+log check MUST produce captured evidence — the actual HEAD..@{u} output, even if empty. Skipping the check on the basis of "I just fetched" or "nothing could have changed in the last N minutes" is a §11.4.6 (no-guessing) violation: the remote state is not knowable without a fetch.
Cascade requirement: This anchor (verbatim or by CONST-060 ID reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to §11.4 PASS-bluff at the parallel-session-coordination layer. See constitution submodule Constitution.md §11.4.37 for the full mandate.
§11.4.68 — Positive Sink-Side / Downstream Evidence Mandate (cascaded from constitution submodule §11.4.68)
Verbatim user mandate (2026-05-20): "We still do not hear any audio played from D3 device! Arvus Web Dashboard when we play music from D3 shows nothing for Codec In Use! This MUST BE investigated and fixed! How come we passed the tests with Arvus validation? What were values for the Codec In Use field? Empty means nothing! This is not working! It MUST BE FIXED, TESTED AND VERIFIED WITH FULL AUTOMATION TESTING ASAP!!!"
A test that asserts audio or video routing PASS MUST capture and verify positive sink-side or downstream evidence — never config-only, never metadata-only, never PCM-open-state-only. At least one of the closed enumeration MUST be captured for every audio/video routing PASS: (1) sink-side codec-state with non-empty Codec-In-Use matching the expected codec regex; (2) strictly-positive PCM frames-written delta from /proc/asound/.../status hw_ptr; (3) ALSA ELD/EDID-Like-Data showing negotiated channel count + format; (4) ffprobe-on-captured-mp4 with non-zero frame count + expected codec/resolution/fps; (5) recording-analyzer event match per §11.4.2/§11.4.5; (6) tinycap RMS amplitude above the line-level floor. Empty / <unreachable> / <N.E.> / <None> placeholders are NOT positive evidence; a missing-but-required sink is OPERATOR-BLOCKED (release-blocker), never SKIP, never PASS. No escape hatch — no --skip-sink-evidence, --allow-empty-codec, --sink-unreachable-is-pass, --metadata-only-suffices flag exists.
Cascade requirement: This anchor (verbatim or by §11.4.68 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to a §11.4 PASS-bluff at the sink-side-evidence layer.
Canonical authority: constitution submodule Constitution.md §11.4.68 for the full mandate.
Verbatim user mandate (2026-05-20): "Always do if possible Subagent-driven! Add this into our root (constitution Submodule) Constitution.md, CLAUDE.md and AGENTS.md. This should be the default choice ALWAYS!"
When executing implementation plans (or any task-decomposed execution flow), the default execution model is subagent-driven per superpowers:subagent-driven-development. Inline execution is permitted ONLY when (a) the task is trivial AND fits a single sub-300-line edit, OR (b) the operator explicitly requests inline at brainstorm-handoff time. Subagent-driven is the default because it gives isolated context per task, naturally enforces two-stage review, is parallel-PWU compatible (§11.4.58), creates an anti-bluff seam (§11.4), and survives operator absence. No escape hatch — --inline-execution-required, --no-subagents, --monolithic-execution are NOT permitted flags. Skipping subagent-driven for non-trivial work without recorded operator authorisation is itself a §11.4 PASS-bluff.
Cascade requirement: This anchor (verbatim or by §11.4.70 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to a §11.4 PASS-bluff at the execution-model layer.
Canonical authority: constitution submodule Constitution.md §11.4.70 for the full mandate.
§11.4.71 — Pre-Push Fetch + Investigate + Integrate Mandate (cascaded from constitution submodule §11.4.71)
Verbatim user mandate (2026-05-20): "before pushing changes to any upstream for any repository - main repo or Submodule, we MUST fetch and pull all changes. Once these are obtained WE MUST investigate what is different compared to head position we were on last time before fetching and pulling new changes! We MUST understand what is done and for what purpose, easpecially how that does affect our project and our System in general! Any mandatory changes or improvements required by fresh changes we just have brough in MUST BE incorporated, covered with all supported types of the tests which will produce as a result of its success execution REAL PROOFS of working for all componetns and functionalities covered and work fully in anti-bluff manner!"
The everyday-push variant of §11.4.41. EVERY push (every repository — main + every submodule) MUST follow the 5-step cycle: (1) fetch all remotes (git fetch --all --prune --tags, capture stdout); (2) pull all upstream branches whose tip differs, resolving conflicts per consumer judgment (never auto---ours/--theirs); (3) investigate the diff vs OUR previous HEAD — read EVERY foreign commit's body, understand what/why/how-it-affects-our-system; (4) integrate mandatory changes with §11.4.4(b) four-layer coverage + §11.4.43 TDD-fix discipline, every PASS carrying §11.4.5 captured-evidence (REAL PROOFS, not metadata-only); (5) only then push, verifying with git ls-remote post-push. No escape hatch — no --skip-fetch, --no-investigate, --fast-push, --trust-upstream flag.
Cascade requirement: This anchor (verbatim or by §11.4.71 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to a §11.4 PASS-bluff at the push-discipline layer.
Canonical authority: constitution submodule Constitution.md §11.4.71 for the full mandate.
Verbatim user mandate (2026-05-20): "Make sure all fixes for audio are always top priority in main working stream!"
The conductor (main working stream — Claude Code session, AI agent, or human operator) MUST treat audio fixes as the highest-priority class on the serial dispatch queue. Any time the conductor faces a choice between dispatching an audio task vs a non-audio task on the SAME serial resource, the audio task wins. Parallel BACKGROUND subagents (research, refactors, infrastructure documentation) MAY run concurrently with audio work but do NOT preempt audio on the main-stream serial dispatch queue. No escape hatch — there is no "but this non-audio task is faster" or "but this research is more interesting" override; audio-stack regressions are user-perceptible and high-impact while research and refactors can wait.
Cascade requirement: This anchor (verbatim or by §11.4.72 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to a process violation at the dispatch-priority layer.
Canonical authority: constitution submodule Constitution.md §11.4.72 for the full mandate.
§11.4.73 — Main-Specification Document Versioning + Revision Discipline (cascaded from constitution submodule §11.4.73)
Verbatim user mandate (2026-05-20): "Make sure everything we add now in previous and upcoming requests IS ALWAYS applied to the main specification — if we have one. Since all these are not major changes we could increase Specification version per change for secondary version instead of the primary. Primary version MUST BE increased for much bigger levels of changes! Add this into root (constitution Submodule) Constitution.md, CLAUDE.md and AGENTS.md as mandatory rule / constraint applicable ONLY IF we have something like the main specification document or we do recognize something like the main specification document. Document MUST BE updated ALWAYS to follow the versioning rules we are appling here + revision and other properties we have!"
Applies only when a project recognises a main specification document. When it does: (1) every additive operator requirement, refinement, or accepted recommendation MUST be applied to the spec before or as part of the implementing work; (2) spec versioning has two axes — primary (V1/V2/V3, bumped for major rewrites by explicit operator decision, old versions archived) and secondary (the §11.4.61 metadata-table Revision integer, bumped for every other change); (3) the metadata table MUST stay current (Revision, Last modified, Status summary, Fixed); (4) propagated copies of the rule MUST reference the active specification.V<primary>.md, not a stale archive; (5) on primary bump the old file moves to <spec-dir>/archive/ with Status: superseded. Classification: universal, applicable conditionally per the scope condition.
Cascade requirement: This anchor (verbatim or by §11.4.73 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to a release blocker when a project has a main spec and lets it drift.
Canonical authority: constitution submodule Constitution.md §11.4.73 for the full mandate.
§11.4.74 — Submodule-Catalogue-First Discovery + Extend-Don't-Reimplement (cascaded from constitution submodule §11.4.74)
Verbatim user mandate (2026-05-20): "We MUST ALWAYS check which already developed features / functionalities do exist as a part of our comprehensive Submodules catalogue located in vasic-digital and HelixDevelopment organizations on GitHub and GitLab both! Project MUST BE aware of all its existence so we do not implement same things multiple times if they are already done as some of existing universal, reusable general development purpose Submodules! For any missing features that some Submodules we incorporate may be missing we MUST IMPLEMENT the properly and extend those Submodules furter! We do control all of the and we CAN and MUST maintain and extend the regularly! All development cycle rules we have MUST BE applied to them and fully respected!"
Before scaffolding ANY new module, package, helper, or utility, the contributor (human or AI agent) MUST: (1) survey the canonical Submodule catalogue — vasic-digital and HelixDevelopment on both GitHub AND GitLab; (2) inventory existing Submodules; (3) reuse before reimplement — if a Submodule provides the functionality (or 80%+ of it), add it as a Git submodule rather than write fresh; (4) extend in-place when 80%+ matches but features are missing — add the missing features TO THAT SUBMODULE (PR upstream + bump pointer), never as a duplicating consuming-project helper; (5) apply all development-cycle rules to those Submodules; (6) document the survey result in the feature's tracker entry with a Catalogue-Check: field (reuse <org/repo>@<sha> / extend <org/repo>@<sha> / no-match <date>). Classification: universal.
Cascade requirement: This anchor (verbatim or by §11.4.74 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Severity-equivalent to a process violation; duplicate implementations landed without catalogue check are release blockers.
Canonical authority: constitution submodule Constitution.md §11.4.74 for the full mandate.
§11.4.69 — Universal Sink-Side Positive-Evidence Taxonomy + Mechanical Enforcement (cascaded from constitution submodule §11.4.69)
Verbatim user mandate (2026-05-20): "THIS MUST HAPPEN NEVER AGAIN!!! We MUST HAVE this all working! Not just for audio but for every single piece of the System!!! Proper full automation when executed with success MUST MEAN that manual testing will be as much positive at least regarding the success results! ... Solution MUST BE universal, generic that solves working flows for all System components and for all future and all existing projects! ... Everything we do MUST BE validated and verified with rock-solid proofs and anti-bluff policy enforcement and fulfillment!"
Universal generalisation of §11.4.68 (audio-specific) across every user-visible feature class. Every user-visible feature MUST map to one entry in the closed-set §11.4.69 sink-side evidence taxonomy (audio_output, audio_input, video_display, network_throughput, network_connectivity, bluetooth_a2dp, bluetooth_pair, touch_input, sensor, gpu_render, storage_read, storage_write, mediacodec_decode, mediacodec_encode, miracast, cast, boot_service, package_install, permission_grant, wifi_link, wifi_throughput, ethernet_link, display_topology, drm_playback, subtitle_render — open to additions, never contraction). Every PASS for a feature in the taxonomy MUST cite a captured-evidence artefact path matching the required evidence shape. New helper contracts (additive during grace, mandatory after 2026-06-19): ab_pass_with_evidence <description> <evidence_path> (verifies path exists + non-empty), ab_skip_with_reason <description> <closed-set-reason> (reasons: geo_restricted, operator_attended, hardware_not_present, topology_unsupported, network_unreachable_external, feature_disabled_by_config; forbids network_unreachable_external for any taxonomy feature with a sink-side probe); bare ab_pass deprecated (WARN pre-grace, FAIL post-grace). Three pre-build gates + paired §1.1 mutations: CM-SINK-EVIDENCE-PER-FEATURE, CM-NO-FAIL-OPEN-SKIP, CM-AB-PASS-WITH-EVIDENCE-EVERYWHERE. No escape hatch — no --skip-evidence, --config-only-pass, --allow-fail-open-skip, --legacy-ab-pass-permitted flag.
Cascade requirement: This anchor (verbatim or by §11.4.69 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-69-PROPAGATION enforces the anchor literal across the consumer fleet; paired mutation strips the literal → gate FAILs. Severity-equivalent to a §11.4 PASS-bluff at the sink-side-evidence layer.
Canonical authority: constitution submodule Constitution.md §11.4.69 for the full mandate.
Verbatim user mandate (2026-05-20): "Why do these violations still happen!? This is a serious problem! We cannot rely on stability nor consistency if we cannot respect our Constitution, mandatory rules and constraints! Is there a way to make this always respected, followed and applied without exception fully and unconditionally!? WE MUST HAVE THIS WORKING FLAWLESSLY!!! Do investigate the root causes of such problems! Once all problems are identified WE MUST apply proper mechanisms for this not to happen NEVER EVER AGAIN!"
The §11.4 covenant historically relied on agent + operator vigilance; three 2026-05-19→20 forensic incidents proved that late-binding enforcement fires hours-to-days after the violator commit reaches every remote. §11.4.75 closes the gap with FIVE independent mechanical enforcement layers — bypassing any single layer does not bypass the discipline: (1) local pre-commit git hook (refuses staged .md lacking sibling .html+.pdf); (2) commit_all.sh integration (_constitution_sibling_check + auto-sync_all_markdown_exports.sh self-repair); (3) local pre-push git hook (re-runs siblings + propagation-gate subset); (4) post-commit auto-repair hook (auto-generates orphan-.md siblings, idempotent + recursion-guarded); (5) local-only final-gate ritual (remote CI DISABLED per User mandate — operator runs pre_build_verification.sh + meta-test before every tag per §11.4.40). Helper contracts: scripts/install_git_hooks.sh, scripts/git_hooks/{pre-commit,pre-push,post-commit,commit-msg}, _constitution_sibling_check. The commit-msg hook enforces a Bypass-rationale: <reason> footer when --no-verify is detected; docs/audit/bypass_events.md accumulates the audit trail. Five gates with paired §1.1 mutations: CM-COVENANT-114-75-PROPAGATION, CM-GIT-HOOKS-INSTALL-SCRIPT, CM-GIT-HOOKS-SOURCE-DIR, CM-COMMIT-ALL-SIBLING-CHECK, CM-CI-WORKFLOW-PRESENT. No escape hatch — no --skip-hooks, --bypass-enforcement, --allow-orphan-md, --ci-not-applicable, --mechanical-enforcement-not-needed flag.
Cascade requirement: This anchor (verbatim or by §11.4.75 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-75-PROPAGATION; paired mutation strips the literal → gate FAILs. Severity-equivalent to a §11.4 PASS-bluff at the enforcement layer.
Canonical authority: constitution submodule Constitution.md §11.4.75 for the full mandate.
Verbatim user mandate (2026-05-20): "For any work or requirements of running services or codebase inside the Containers (Docker / Podman / Qemy / Emulators, and so on) we MUST USE / INCORPORATE the Containers Submodule properly: https://github.com/vasic-digital/containers (git@github.com:vasic-digital/containers.git). Containers Submodule contains all means for us to Containerize our code and services! If any feature or Containing System is missing or not supported we MUST EXTEND IT properly like we do all of our projects! No bluff work is allowed of any kind!"
For ANY containerized workload (Docker / Podman / Qemu / Kubernetes / container-backed emulators), every consuming project MUST: (1) install vasic-digital/containers (digital.vasic.containers) as a Git submodule; (2) consume via replace directive during development + pinned commit SHAs in production; (3) boot infra on-demand via pkg/boot + pkg/compose + pkg/health so operators are never required to start podman machine / docker compose up manually — the boot is part of the test entry point (the on-demand-infra invariant); (4) extend the Submodule (PR upstream) for missing runtimes / lifecycle primitives — never reimplement in-project (per §11.4.74); (5) anti-bluff: integration tests claiming to exercise containerized components MUST actually boot them via the Submodule — short-circuit fakes that bypass boot are a §11.4 violation. Tracker rows touching containerization MUST record Catalogue-Check: extend vasic-digital/containers@<sha> (or reuse). Planned gate CM-CONTAINERS-USED scans container-touching PRs for digital.vasic.containers/... imports; paired mutation strips the import + asserts FAIL.
Cascade requirement: This anchor (verbatim or by §11.4.76 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-76-PROPAGATION; paired mutation strips the literal → gate FAILs.
Canonical authority: constitution submodule Constitution.md §11.4.76 for the full mandate.
Verbatim user mandate (2026-05-20): "We must be sure that after excluding anything from Git versioning we still have the mechanism which will out of the box obtain or re-generate missing content!"
Every .gitignore entry excluding (a) >~100 MiB OR (b) any artefact essential to building / running / testing the project MUST carry a documented + automated mechanism to either re-obtain (download from authoritative source: vendor tarball, SDK installer, npm/pip/cargo/go-mod/container registry, dedicated git submodule, S3/GCS) OR re-generate (run from tracked source via build pipeline, code-gen, asset render, captured-evidence replay, container build). Required artefacts per qualifying entry: (1) .gitignore-meta/<entry-slug>.yaml declaring pattern + mechanism-type + script-path + expected-disk-usage + vendor-url-or-source + integrity hash + requires-network + requires-credentials; (2) a non-interactive entry in scripts/setup.sh post-clone bootstrap; (3) a pre-build gate verifying regenerated content present OR a recent .gitignore-meta/.regenerated/<slug>.ok stamp; (4) README + docs/guides/*.md describing the mechanism + manual fallback + time/disk budget + §11.4.10 credentials. Bare .gitignore additions without the mechanism are a §11.4 PASS-bluff variant — codebase appears complete but a fresh clone cannot build/run. No escape hatch — no --skip-regen-mechanism, --gitignore-is-enough, --operator-already-has-content flag. Planned gate CM-GITIGNORE-REGEN-MECHANISM + paired §1.1 mutation (strip a required YAML key → gate FAILs).
Cascade requirement: This anchor (verbatim or by §11.4.77 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-77-PROPAGATION; paired mutation strips the literal → gate FAILs. Severity-equivalent to a §11.4 PASS-bluff at the repository-hygiene layer.
Canonical authority: constitution submodule Constitution.md §11.4.77 for the full mandate.
Verbatim user mandate (2026-05-20): "Make codegraph MANDATORY CHOICE for this purpose for all of our project ... All project which do not have configured and installed codegraph yet MUST DO IT and MUST USE IT!"
Every consuming project worked on by AI coding agents MUST install, initialize, and use CodeGraph (https://github.com/colbymchenry/codegraph, npm @colbymchenry/codegraph) — a local SQLite semantic code-knowledge-graph exposed to agents over MCP (100% local, no cloud). (1) Install globally via npm with a user-writable npm prefix (no sudo). (2) codegraph init + codegraph index: .codegraph/config.json is tracked, .codegraph/codegraph.db is gitignored with codegraph index as its §11.4.77 regeneration mechanism; the config.json exclude list MUST exclude every credential/secret path per §11.4.10. (3) Wire codegraph serve --mcp into every CLI agent (Claude Code .mcp.json, OpenCode opencode.json, Qwen Code .qwen/settings.json, Crush .crush.json, host-local otherwise) referencing the bare codegraph command on PATH (no hardcoded host path). (4) Cover the integration with an anti-bluff suite whose per-agent end-to-end layer uses an unforgeable challenge (a fact obtainable only by calling a CodeGraph MCP tool, e.g. index node count via codegraph_status); a genuinely un-drivable agent is a documented SKIP per §11.4.3, never a faked PASS. (5) Document in docs/CODEGRAPH.md, kept in sync per §11.4.12 / §11.4.65. CodeGraph is consumed as the published npm package (§11.4.74) — not a git submodule, adds no Git remote. Planned gate CM-CODEGRAPH-WIRED + paired §1.1 mutation (strip a secret-exclusion → gate FAILs).
Cascade requirement: This anchor (verbatim or by §11.4.78 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-78-PROPAGATION; paired mutation strips the literal → gate FAILs.
Canonical authority: constitution submodule Constitution.md §11.4.78 for the full mandate.
§11.4.79 — Own-Org Submodules MUST Be Included in the CodeGraph Index (cascaded from constitution submodule §11.4.79)
Verbatim user mandate (2026-05-21): "All Submodules we use in the project and that are part of organizations to which we have the full access via GitHub, GitLab and other CLIs MUST BE included into the codegraph database and initialized / scanned / synced!"
Refines §11.4.78's exclude-list with a per-submodule-ownership split: (a) own-org submodules (full write access via the project's CLIs — canonical orgs vasic-digital + HelixDevelopment) MUST be INCLUDED in the index; (b) third-party submodules (the §11.4.74 no-match → vendor path) MUST be EXCLUDED. Operational steps: (1) git submodule update --remote --merge to pull latest before re-indexing, respecting load-bearing pins on third-party submodules; (2) adjust .codegraph/config.json exclude list to keep own-org paths in scope; (3) re-index via scripts/codegraph_setup.sh; (4) verify via scripts/codegraph_validate.sh with ≥1 probe resolving a symbol living ONLY inside an own-org submodule; (5) paired §1.1 mutation — temporarily add the own-org submodule to exclude → validate MUST FAIL on the cross-submodule probe → restore. An index that lies about reachable symbols is a PASS-bluff against AI agents. Own-org submodules silently excluded without an audit trail in .codegraph/config.json comments is a release blocker.
Cascade requirement: This anchor (verbatim or by §11.4.79 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-79-PROPAGATION; paired mutation strips the literal → gate FAILs.
Canonical authority: constitution submodule Constitution.md §11.4.79 for the full mandate.
§11.4.80 — CodeGraph Regular-Update + Sync Automation Mandate (cascaded from constitution submodule §11.4.80)
Verbatim user mandate (2026-05-21): "We MUST regularly check for the updates and execute codegraph npm updates so the latest version of it is always installed on the host machine! ... Make sure we have proper full automation bash scripts which will run regularly and that these are part of the constitution Submodule ... Make sure all updates, sync processes we do and important codegraph related events are all documented under docs/codegraph in Status and Status_Summary documents ... and regularly export them like all other Status docs into the PDF and HTML!"
Three deliverables (all living in the constitution submodule, inherited by reference per §3 — consuming projects invoke at ${CONST_DIR}/scripts/codegraph_*.sh, never copy): (1) scripts/codegraph_update.sh — npm-installs latest @colbymchenry/codegraph after a registry version check; appends old/new version to docs/codegraph/Status.md; anti-bluff verifies codegraph --version reflects the new version after install (npm exit 0 ≠ working binary). (2) scripts/codegraph_sync.sh — after a successful update runs codegraph status → codegraph sync . → codegraph status → the project's scripts/codegraph_validate.sh; appends every step's output to BOTH the project's and the constitution's docs/codegraph/Status.md. (3) docs/codegraph/Status.md + Status_Summary.md append-only ledgers, exported to .html + .pdf per §11.4.65. Cadence: weekly floor (per §11.4.45). A consuming project that has not run codegraph_update.sh in >2 weeks AND has open AI-agent work is a release blocker. Paired §1.1 mutation: downgrade installed version → script detects drift → restore.
Cascade requirement: This anchor (verbatim or by §11.4.80 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-80-PROPAGATION; paired mutation strips the literal → gate FAILs.
Canonical authority: constitution submodule Constitution.md §11.4.80 for the full mandate.
Verbatim user mandate (2026-05-21): "Any Linux-only blocker / issue we have MUST BE created macOS and other supported platforms equivalent! So, depending on platform proper implementation will be used for particular OS! EVERYTHING MUST BE PROPERLY EXTENDED AND UPDATED!"
Every consuming project whose supported-platforms manifest lists more than one OS MUST, for every feature/test/gate/challenge/mutation depending on platform-specific primitives, ship a per-OS-equivalent implementation chosen at runtime via uname -s (or equivalent detection). Three sub-mandates: (A) Per-OS implementation REQUIRED — Linux cgroup/systemd//proc primitives MUST have documented per-OS equivalents (POSIX setrlimit/ulimit, macOS launchd, BSD rctl, Windows Job Object) chosen via runtime dispatch. (B) Per-OS tests REQUIRED — every platform-dependent gate test MUST have case "$(uname -s)" in branches with positive captured evidence per §11.4.2 + §11.4.5 in each branch; SKIP-with-reason acceptable ONLY when the platform genuinely cannot enforce the invariant. (C) Honest kernel-gap citation + adjacent equivalent test REQUIRED — where a Linux primitive has NO equivalent due to a documented kernel limitation (canonical: XNU does not enforce RLIMIT_AS for unprivileged processes), the test MUST detect the gap at runtime, SKIP with exact kernel reason + reproducer + honest-gap-doc link, AND provide an ADJACENT test exercising the closest invariant the platform CAN enforce (e.g. RLIMIT_CPU+SIGXCPU as the macOS proxy), itself anti-bluff with a paired §1.1 mutation. Gate CM-CROSS-PLATFORM-PARITY scans for case "$(uname -s)" blocks asserting a non-SKIP branch (or honest-gap citation) per platform in the manifest; paired mutation strips a Darwin branch → gate FAILs. No escape hatch.
Cascade requirement: This anchor (verbatim or by §11.4.81 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-81-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker on multi-platform projects.
Canonical authority: constitution submodule Constitution.md §11.4.81 for the full mandate.
Verbatim user mandate (2026-05-22): "How can we speed-up this whole development and fixing process? ... Do not forget to all speed optimizations critical rules and mandatory constraints MUST BE all added into our root (constitution Submodule) Constitution.md, CLAUDE.md, AGENTS.md and QWEN.md and all other relevant constitution Submodules files!"
Iteration cycle time is a first-order quality enabler. Every consuming project's build / test / commit / debug pipeline MUST adopt these speedup disciplines AS MANDATORY (each independently enforceable): (A) Phase-1 forensic (superpowers:systematic-debugging) before any speculative source patch — speculative patches without FACT-grade root cause are §11.4.6 + §11.4.82 violations; (B) Live-ADB-First (or live-equivalent) before any rebuild — strengthens §11.4.51 to a release-blocker mandate; (C) 30-second pre-flight before launching rebuild orchestrators (device/sink reachability, host memory/disk, no stale locks, no orphan processes); (D) persistent build caches outside containers (ccache/sccache/Gradle daemon bind-mounted to host); (E) module-only rebuild for loadable-module-only changes; (F) parallel multi-device testing with separate qa-results/<TS>/<device-tag>/ outputs; (G) subagent scope discipline + worktree isolation (≤30 min budget, single-responsibility, isolation: "worktree" default); (H) lock-file + stale-process hygiene (clean .git/index.lock, disable auto git-gc in concurrent repos); (I) cycle telemetry per §11.4.24 (commit hash, per-phase wall-clock, speedup-flag set, outcome — aggregated weekly). Gate CM-ITERATION-SPEEDUP-DISCIPLINE audits recent cycles for telemetry citing which of (A)-(I) applied; paired §1.1 mutation strips the speedup-flag column → gate FAILs. No escape hatch — no --skip-phase1-forensic, --no-pre-flight, --rebuild-everything-always, --unlimited-subagent-scope, --ignore-locks, --no-telemetry flag.
Cascade requirement: This anchor (verbatim or by §11.4.82 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-82-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.82 for the full mandate.
Verbatim user mandate (2026-05-22): "every feature that ships MUST carry a recorded e2e communication transcript + any attached materials under
docs/qa/<run-id>/(per-feature subdirectories). A feature with no QA transcript is itself a §107 PASS-bluff — it claims to work but has no auditable runtime evidence. Bot-driven automation MUST preserve full bidirectional communication threads as proof."
Every feature that ships MUST carry a recorded end-to-end communication transcript plus any attached materials (screenshots, request/response payloads, audio, file uploads) committed under docs/qa/<run-id>/ — one directory per feature run. Operative rule: (1) every consuming project MUST maintain a docs/qa/ tree, each new feature under docs/qa/<run-id>/ where <run-id> is monotonic + greppable (timestamp / ATM-NNN / other workable-item ID per §11.4.54); (2) transcripts MUST be full bidirectional — every prompt/command sent + every response received (one-sided is not a transcript); (3) attached materials MUST be committed in-repo (no external-only links — that is a §11.4.13 sink-side violation); (4) bot-driven / agent-driven QA automation MUST preserve the full conversation thread as the proof artefact; (5) release gates MUST refuse to tag a version that has any feature-shipping commit without its matching docs/qa/<run-id>/ directory. A feature with no QA transcript is a §11.4 / §107 PASS-bluff. Composes with §11.4.2 / §11.4.5 / §11.4.13 / §11.4.65 / §11.4.69 / §1.1.
Cascade requirement: This anchor (verbatim or by §11.4.83 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-83-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker — no --qa-evidence-optional escape hatch.
Canonical authority: constitution submodule Constitution.md §11.4.83 for the full mandate.
§11.4.84 — Working-Tree Quiescence Rule for Subagent Commits (cascaded from constitution submodule §11.4.84)
Verbatim user mandate (2026-05-22): "no subagent commit may proceed while any concurrent mutation gate is in flight in the same checkout. Before
git add, the committing agent MUSTgrepits own working tree for mutation markers (MUTATED for paired,// always pass,return json.Marshalshortcut paths, etc.). Any unexplained file in the staging area triggers ABORT."
No subagent (or main-thread) commit may proceed while any concurrent mutation gate, paired-mutation experiment, or other in-flight mutation is live in the same checkout. Before git add, the committing agent MUST grep its own working tree for mutation markers (MUTATED for paired, // always pass, return json.Marshal shortcut paths, // MUTATION / # MUTATION annotations, _mutated_* filename suffixes, etc.) and explicitly account for every modified file in the staging area; any unexplained file → ABORT. (Forensic case: a logo-fix subagent's git add swept an // always pass JWT-verify mutation residue into an unrelated commit pushed to all four mirrors — a real security-defect window.) Operative rule: (1) pre-git add greps for mutation markers + cross-checks git status --porcelain against the subagent's declared scope; unaccounted entries → ABORT; (2) any active mutation gate MUST be serialised (mutate → assert FAIL → restore → assert PASS) and the working tree verifiably clean before any unrelated commit; (3) concurrent subagents in the SAME checkout MUST coordinate through a lockfile (.git/MUTATION_IN_PROGRESS) — cleaner solution is git worktree add per subagent (composes with §11.4.20/§11.4.70); (4) post-commit mutation-residue-scanner MUST run before push — any commit containing a mutation marker → push BLOCKED.
Cascade requirement: This anchor (verbatim or by §11.4.84 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-84-PROPAGATION; paired mutation strips the literal → gate FAILs. A mutation marker that lands in a tagged commit is a critical defect regardless of how briefly it persisted.
Canonical authority: constitution submodule Constitution.md §11.4.84 for the full mandate.
Verbatim user mandate (2026-05-24): "Every fix or improvement you do MUST BE covered with full automation stress and chaos tests so we are sure nothing can break the functionality and all edge cases are monitored and polished and additionally fixed if that is needed! Everything must produce rock solid proofs and follow fully no-bluff policy!"
Every fix or improvement landed MUST ship with full-automation stress AND chaos test suites exercising edge cases, sustained load, concurrent contention, and failure-injection. Happy-path coverage alone is a §11.4 / §107 PASS-bluff at the resilience layer. Stress (closed-set): sustained load (N ≥ 100 iterations OR ≥ 30 s wall-clock, p50/p95/p99 latency recorded) + concurrent contention (N ≥ 10 parallel invocations, no deadlock/leak) + boundary conditions (empty/max/off-by-one, each categorised). Chaos (closed-set, per fix-class appropriateness): process-death injection + network-fault injection (drop/delay/reorder) + input-corruption injection + resource-exhaustion injection (disk full, OOM, FD exhaustion — refuse cleanly OR degrade, NEVER crash) + state-corruption injection (mid-flight lock loss, partial-write). Every stress + chaos PASS MUST cite a captured-evidence artefact path per §11.4.5 + §11.4.69. Helper library stress_chaos.sh provides ab_stress_run, ab_stress_concurrent, ab_chaos_kill_pid_during, ab_chaos_drop_network_during, ab_chaos_corrupt_file_during, ab_chaos_oom_pressure_during, ab_chaos_disk_full_during, each composing with ab_pass_with_evidence / ab_skip_with_reason. Cleanup non-negotiable in trap '...' EXIT (cleanup failure = §11.4.14 violation). Four-layer coverage per §11.4.4(b) + paired §1.1 mutation (strip chaos-injection or evidence-capture → gate FAILs). No escape hatch — no --skip-stress, --no-chaos, --happy-path-suffices, --stress-test-later flag.
Cascade requirement: This anchor (verbatim or by §11.4.85 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-85-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.85 for the full mandate.
§11.4.86 — Roster/Corpus-Backed Status-Doc Auto-Sync Mandate (cascaded from constitution submodule §11.4.86)
Verbatim user mandate (2026-05-25): "Make sure that assets and players Status docs are ALWAYS regularly updated and in sync like all others Status docs — any time we add or modify the assets content(s) or we change or add new / remove existing pre-installed video and audio player apps! This MUST WORK OUT OF THE BOX!"
Some Status docs (§11.4.45) are backed by a tracked roster (installed apps/components) or a tracked asset corpus (test/media asset directory) rather than narrative alone. Their freshness MUST NOT depend on operator vigilance — the moment a roster/corpus member changes (app added/removed/renamed; asset added/modified/removed) the Status doc + Status_Summary + HTML + PDF MUST resync out of the box, mechanically. Mechanism (all must hold): (1) drift-proof fingerprint — sha256 of the sorted member list (NOT mtime), persisted in a sidecar beside the Status doc; (2) a sync helper that regenerates the fingerprint + re-exports HTML+PDF via the §11.4.65 exporter, wired so sync is automatic; (3) a pre-build gate that FAILs when the live fingerprint differs from the persisted one (mirrors §11.4.12 CM-ISSUES-SUMMARY-SYNC + §11.4.45 sync_integration_status); (4) a paired §1.1 mutation corrupting the fingerprint and asserting the gate FAILs. Classification: universal — the consuming project supplies the specific docs, roster/corpus sources, helper, and gate name per §11.4.35.
Cascade requirement: This anchor (verbatim or by §11.4.86 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-86-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker — no --skip-roster-sync, --allow-status-drift, --roster-sync-not-applicable flag.
Canonical authority: constitution submodule Constitution.md §11.4.86 for the full mandate.
§11.4.87 — Endless-Loop Autonomous Work + Zero-Idle Agent Dispatch + Anti-Bluff Testing Mandate (cascaded from constitution submodule §11.4.87)
Verbatim user mandate (2026-05-26): "continue in endless loop fully autonomously" (and any semantically-equivalent phrasing).
When the operator instructs an AI agent to continue in an endless autonomous loop, the agent MUST treat it as a HARD-CONTRACT covenant: (A) continue working until docs/Issues.md Status-column has zero non-terminal entries AND docs/CONTINUATION.md §3 Active work is empty AND no background subagent is mid-execution AND no external dependency is in-flight; (B) dispatch background subagents for parallelisable work — main + every subagent operate concurrently, "waiting for results" is the ONLY acceptable idle reason; (C) every closure lands four-layer test coverage per §11.4.4(b) with captured-evidence (audio/video/network/UI/sysfs physical proofs); (D) the §11.4 anti-bluff covenant family (§11.4.1 / §11.4.2 / §11.4.6 / §11.4.7 / §11.4.27 / §11.4.50 / §11.4.52 / §11.4.68 / §11.4.69 / §11.4.83) is the operative truth-discipline — tests AND HelixQA Challenges bound equally; (E) the loop terminates ONLY on all-conditions-met, explicit operator STOP, host-session-safety demand, or scheduled wake on a known-future-actionable signal. No escape hatch — no --idle-OK, --skip-endless-loop, --bluff-permitted-for-this-task, --metadata-only-test-suffices, --no-physical-proof-required flag.
Cascade requirement: This anchor (verbatim or by §11.4.87 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-87-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.87 for the full mandate.
§11.4.88 — Background-Push Mandate: Commit-Lock Release Immediately After Commit, Push Runs Detached (cascaded from constitution submodule §11.4.88)
Forensic anchor (2026-05-26): a single commit_all.sh held its flock ~5 hours because do_push ran synchronously after the commit landed — every subsequent commit blocked on a slow mirror push irrelevant to the local commit's durability. Implementation seam for §11.4.87(B) zero-idle. The mandate: (A) .git/.commit_all.lock MUST be released IMMEDIATELY after git commit returns 0 — the commit is durable on local disk regardless of remote push outcome; (B) push runs detached via nohup ./push_all.sh ... > <log> 2>&1 & + disown — the orchestrator's exit code reports COMMIT success, NOT push success; (C) push_all.sh acquires per-remote flock .git/.push.<remote>.lock so concurrent invocations targeting the same remote serialize but different-remote invocations run in parallel; (D) backgrounded push failures land in qa-results/push_failures/<ts>_<remote>.log — the next autonomous-loop tick checks per §11.4.87(A) "no external dependency in-flight" gate; (E) synchronous-push escape: explicit --sync-push CLI flag preserves legacy behaviour for §11.4.41 force-push merge-first audit paths. Gates CM-COVENANT-114-88-PROPAGATION + CM-BACKGROUND-PUSH-WIRED + paired §1.1 mutations. Synchronous push (without --sync-push) = §11.4 PASS-bluff at the execution layer.
Cascade requirement: This anchor (verbatim or by §11.4.88 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-88-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker — no escape hatch beyond --sync-push for force-push events.
Canonical authority: constitution submodule Constitution.md §11.4.88 for the full mandate.
Verbatim user mandate (2026-05-27): "Any tests we are executing, especially long test cycles, MUST BE performed in background in parallel with main work stream! This MUST NOT block our capabilities to work on queued workable items. Main work stream can be blocked or sit iddle only if absolutely needed and if it depends hard on results of some background execution."
Symmetric anchor to §11.4.88 (background push) at the test-execution layer. Mandate: (A) long-running tests (>30 s expected: pre_build, meta_test, test_all_fixes, recent_work_validate, HelixQA banks, 4-phase cycles, full-suite retests, audio supervisors, dual-display recorders) MUST run via nohup ... > <log> 2>&1 & + disown with the log under a known dir (qa-results/<test_id>_<ts>.log); (B) the main stream proceeds to the §11.4.42 priority queue immediately; (C) hard-dependency gating — poll an exit-status file or pgrep -af <test> before steps that need the exit code, surfacing as §11.4.66 interactive options if the test is still running; (D) failures land in <log> files, the next loop tick checks; (E) foreground execution permitted ONLY for <30 s tests OR explicit operator authorisation; (F) per-script flock serialises same-script invocations, different-script invocations parallel. Gates CM-COVENANT-114-89-PROPAGATION + CM-BACKGROUND-TEST-EXECUTION-WIRED + paired §1.1 mutations.
Cascade requirement: This anchor (verbatim or by §11.4.89 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-89-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker — no escape hatch beyond explicit per-invocation operator authorisation.
Canonical authority: constitution submodule Constitution.md §11.4.89 for the full mandate.
§11.4.90 — Obsolete Status + Per-Item Obsolescence Audit (cascaded from constitution submodule §11.4.90)
Verbatim user mandate (2026-05-27): "Bug No 6 ... seems obsolete after latest request for new behavior ... mark obsolete tickets with some light gray background ... text - the description to be strikethrough styled ... review all existing open or resolved workable items if they are obsolete - not valid any more ... There MUST NOT be any mistake! No bluff is allowed of any kind!"
The §11.4.15 Status closed-set is extended with a terminal Obsolete (→ Fixed.md) value (orthogonal to Type per §11.4.16). Obsolescence reasons (closed vocabulary): superseded-by-design-change | superseded-by-later-mandate | feature-removed | duplicate-of | unsupported-topology. Every Obsolete heading MUST carry an **Obsolete-Details:** line (Since + Reason + Superseding-item + Triple-check evidence) within 8 non-blank lines. The §11.4.23 colorizer adds a cell-status-obsolete class — light-gray #E0E0E0 background + strikethrough description. Audit cadence: every release-gate sweep per §11.4.40 + §11.4.42; triple-check is non-negotiable per the operator mandate. Composes with §11.4.15 / §11.4.16 / §11.4.19 / §11.4.21 / §11.4.23 / §11.4.33 / §11.4.34 / §11.4.40 / §11.4.42 / §11.4.66 / §11.4.71. Gates CM-COVENANT-114-90-PROPAGATION + CM-ITEM-OBSOLETE-DETAILS + CM-OBSOLETE-COLORIZER-WIRED + paired §1.1 mutations.
Cascade requirement: This anchor (verbatim or by §11.4.90 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-90-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.90 for the full mandate.
Verbatim user mandate (2026-05-27): "Summary docs - Issues_Summary some not clear one line descriptions - like 'Composes with' ... For each workable item we MUST HAVE clearly understandable meaning ... every team member can clearly understand what that particular workable item is exactly about! There cannot be misunderstanding or unclearity of any kind and no bluff allowed!"
Every summary entry (Issues_Summary, Fixed_Summary, README doc-link, Status_Summary pages 1+2, all one-liners) MUST contain a self-contained meaningful description ≥ 6 words OR ≥ 40 chars naming SUBJECT + PROBLEM/GOAL. Forbidden one-liner anti-patterns: section labels (Composes with, Closure criteria, Fix direction, etc.); bare metadata fragments (Critical, Bug, In progress, etc.); section-marker echoes; a §-letter alone. Generators (generate_issues_summary.sh / generate_fixed_summary.sh / update_readme_doc_links.sh / generate_status_summary.sh) MUST extract from the H1/H2 heading line per the §11.4.54 ATM-NNN convention, NEVER from arbitrary downstream text, and MUST refuse anti-pattern rows — emitting a (MISSING DESCRIPTION — fix source heading) placeholder with visual highlight. Gate CM-SUMMARY-CLARITY-DESCRIPTIONS scans every summary; an anti-pattern match = FAIL. Audit cadence: every §11.4.40 + §11.4.42 sweep.
Cascade requirement: This anchor (verbatim or by §11.4.91 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-91-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.91 for the full mandate.
Verbatim user mandate (2026-05-27): "Every change to the project or codebase we do MUST BE evaluated in several passes and in in-depth analisys for potential new issues or problems it can introduce! ... no bluff of any kind! After we do change or set of changes this mandatory steps MUST BE taken!"
Every non-trivial change MUST pass a 5-pass evaluation BEFORE it is commit-ready: (Pass 1) main-task verification — change achieves the stated goal, captured-evidence per §11.4.5/§11.4.69; (Pass 2) regression-blast-radius analysis — enumerate every direct dependency, demonstrate no contract break; (Pass 3) cross-feature interaction analysis — audit parallel features sharing state/timing/hardware/shell environment; (Pass 4) deep-research validation per §11.4.8 — external precedent OR "NO external solution found — original work" + CodeGraph queries per §11.4.78/§11.4.79; (Pass 5) anti-bluff confirmation per §11.4 / §11.4.1 / §11.4.6 / §11.4.27 / §11.4.50 / §11.4.52 / §11.4.69 / §11.4.83 — no new bluff surface introduced. Each pass is documented (commit footers OR docs/ entries OR qa-results/ evidence). Only after all 5 passes complete may commit/push/test/release proceed. Trivial exemption: typo / revision-bump / MD-export-regen IF zero source touched AND the commit message cites the exemption explicitly. Gates CM-COVENANT-114-92-PROPAGATION + CM-MULTI-PASS-EVALUATION-EVIDENCE + paired §1.1 mutations.
Cascade requirement: This anchor (verbatim or by §11.4.92 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-92-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.92 for the full mandate.
§11.4.93 — SQLite-Backed Single-Source-of-Truth for Workable Items (cascaded from constitution submodule §11.4.93)
Verbatim user mandate (2026-05-27): "There MUST be single source of truth for all of our workable items - SQlite database ... proper scripts (we recommend Go programs) ... reduce a chance for sync to be broken ... generate always all docs from DB or to re-generate Db from all docs we have in opposite direction"
The text-based Issues/Fixed/Summary/CONTINUATION constellation is converted to a SQLite-DB-backed single source of truth. Schema mandatory tables: items (atm_id PK + Type + Status incl. Obsolete + Severity + title + description ≥40 chars + created/modified + composes_with JSON + current_location); item_history (append-only audit per §11.4.34 By/Reason/Evidence); obsolete_details (§11.4.90); operator_block_details (§11.4.21); firebase_metadata (§11.4.47); meta (schema version + last sync + integrity hash). A Go binary at cmd/workable-items/ provides sync md-to-db / db-to-md / diff / validate / add / close; bidirectional regen is byte-identical round-trip (closed-set whitespace/section-order tolerance). commit_all.sh refuses on non-empty diff; sync_issues_docs.sh invokes the Go binary; pre-build runs workable-items validate. Anti-bluff: unit + integration + stress (1000-row insert + 10 concurrent writers) + chaos (mid-write SIGKILL + corrupt-DB recovery + disk-full) + paired §1.1 mutation + HelixQA Challenge CME-WORKABLE-ITEMS-001. The Go binary lives in the constitution submodule (constitution/scripts/workable-items/) per §11.4.74. Gates CM-COVENANT-114-93-PROPAGATION + CM-WORKABLE-ITEMS-DB-PRESENT + CM-WORKABLE-ITEMS-MD-DB-IN-SYNC + paired §1.1 mutations. (NOTE: the DB tracking rule is AMENDED by §11.4.95 — DB is TRACKED, not gitignored.)
Cascade requirement: This anchor (verbatim or by §11.4.93 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-93-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker — text-based-only trackers are a §11.4 PASS-bluff at the data-architecture layer.
Canonical authority: constitution submodule Constitution.md §11.4.93 for the full mandate.
§11.4.94 — Zero-Idle Priority-First Parallel-By-Default Operating Mode (cascaded from constitution submodule §11.4.94)
Verbatim user mandate (2026-05-27): "We MUST NEVER sit iddle / wait or sleep if there is possibility for us to work on something ... Always check if there is a possibility to work on something while we are not working actively on something! Pick always by priority - most critical workable items and other tasks MUST BE done first! ... Stay still / iddle if nothing is left to be done at all or waiting for something that is blocking us / you!!!"
§11.4.94 binds §11.4.20 + §11.4.42 + §11.4.58 + §11.4.70 + §11.4.72 + §11.4.82 + §11.4.87 + §11.4.88 + §11.4.89 into a single always-on enforcement: (A) idle ONLY when every queued item is genuinely blocked on an external dependency (hardware / network upstream / build/test completion the conductor cannot accelerate) OR operator STOP OR §12 host-safety — "don't see what to do" is NEVER valid; (B) before ANY wake/sleep the conductor MUST survey parallel-work feasibility per §11.4.42 + §11.4.72 + §11.4.87, identify non-contending items, and dispatch in parallel per §11.4.20/§11.4.70 (subagent) + §11.4.58 (PWU disjoint scope) + §11.4.89 (background long tests); (C) priority order MANDATORY — pick highest-severity + §11.4.72 audio-first the conductor can autonomously progress; (D) subagent-driven default for non-trivial; (E) background default for >30 s wall-clock work via nohup+disown; (F) stability-preserving (composes with §11.4.92 multi-pass + §11.4.84 quiescence + §12.6–§12.9 host safety); (G) progress updates surfaced at milestone boundaries. Gates CM-COVENANT-114-94-PROPAGATION + CM-PARALLEL-WORK-AUDIT + paired §1.1 mutations.
Cascade requirement: This anchor (verbatim or by §11.4.94 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-94-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.94 for the full mandate.
§11.4.95 — Workable-Items SQLite DB Is TRACKED in Git, NEVER Gitignored (cascaded from constitution submodule §11.4.95)
Verbatim user mandate (2026-05-27): "We shall not Git ignore our workable items SQlite DB since it is our single source of truth ... workable items SQlite DB regularly commited and pushed to all upstreams!"
§11.4.93's earlier "gitignored per §11.4.30" clause is AMENDED — the DB at docs/workable_items.db is TRACKED in git, NEVER gitignored. It IS authoritative source data, NOT a build artefact. Every workable-items sync md-to-db that mutates state MUST stage + commit + push the DB alongside the MD regen per §11.4.19 atomic-move + §2.1 multi-upstream push. A WAL-checkpoint (PRAGMA wal_checkpoint(TRUNCATE)) is required before commit-stage so the transient .db-wal + .db-shm sidecars (gitignored per §11.4.30) are safely discardable. The §11.4.77 regeneration mechanism does NOT apply — the DB IS the source. Destructive DB ops require §9.2 hardlinked-backup + operator authorization; §11.4.41 force-push merge-first applies if DB history ever needs rewrite. Gates CM-COVENANT-114-95-PROPAGATION + CM-WORKABLE-ITEMS-DB-TRACKED + paired §1.1 mutation.
Cascade requirement: This anchor (verbatim or by §11.4.95 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-95-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.95 for the full mandate.
§11.4.96 — Safe-Parallel-Work-With-Long-Build Catalogue + Mandate (cascaded from constitution submodule §11.4.96)
Verbatim user mandate (2026-05-27): "Are there except AOSP build process any other active jobs being done at the moment? Can we work on something in parallel while build is in progress so we slowly cleanup our slate? ... do as much as possible work in background in parallel with main work stream and oreferrably using subagents-driven approach!"
An operational catalogue for the canonical long-running workload (multi-hour containerised build per §12.9). SAFE during build: (A) MD/docs work; (B) generator/helper script work under scripts/; (C) pre-build + meta-test gate authoring + paired §1.1 mutations; (D) on-device test scripts; (E) constitution submodule edits + push; (F) any submodule commit + push per §11.4.88; (G) read-only live-ADB probes (dumpsys/getprop/cat /proc/.../screencap/logcat); (H) subagent dispatch per §11.4.20/§11.4.70 + §11.4.84 quiescence; (I) web research + external API queries with §11.4.10 credentials; (J) workable-items DB ops per §11.4.93+§11.4.95; (K) backgrounded pre-build + meta-test execution per §11.4.89. UNSAFE during build: (α) git checkout/reset --hard/clean -df on the source tree (use git worktree); (β) mass file deletes/renames under built source trees; (γ) submodule pointer updates affecting built artefacts; (δ) out/ mutations; (ε) make clean/m clobber/rm -rf out/; (ζ) container destruction; (η) disk-filling breaching §12.9 free-space minimum; (θ) §12 host-session-safety breaches. Conductor responsibility: before EVERY pause point during a long build, consult the catalogue, identify (A)-(K) queue items per §11.4.42+§11.4.72, and dispatch ≥1 per §11.4.20/§11.4.70 subagent default + §11.4.89 background. "Build running, nothing else to do" is NEVER true per §11.4.94+§11.4.96. Gates CM-COVENANT-114-96-PROPAGATION + CM-PARALLEL-WORK-DURING-BUILD-AUDIT + paired §1.1 mutations.
Cascade requirement: This anchor (verbatim or by §11.4.96 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-96-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.96 for the full mandate.
§11.4.97 — Maximum-Use-of-Idle-Time + Progress-Update Cadence (cascaded from constitution submodule §11.4.97)
Verbatim user mandate (2026-05-27): "keep it working, we should do as much as possible, if not it all but as much as we can as long as there is iddle time! it MUST be used! ... keep us updated about all progress and all phisycal proofs and gathered data as you progress through all open workable items!"
Operating-mode capstone strengthening §11.4.87 + §11.4.94 + §11.4.96: (A) every minute of conductor idle time during which work could autonomously progress AND is not genuinely blocked = a §11.4.97 violation; "as much as possible, if not it all but as much as we can" is operative — dispatch CONTINUOUSLY through the entire idle window, not just at scheduled wakes; (B) progress-update cadence — emit an operator-facing 1-line update at every commit landed / subagent return / constitutional anchor / captured evidence / milestone closure, no operator prompt required; (C) continuous physical-proof gathering per §11.4.5 + §11.4.6 + §11.4.69 — every autonomous closure cites captured-evidence (evidence path goes into the §11.4.93 item_history.evidence_path when the DB lands); (D) composes with §11.4.5/6/13/20/27/42/50/52/69/70/72/83/85/87/88/89/94/96; (E) the idle-only-when-blocked closed-set is unchanged from §11.4.94(A). Gates CM-COVENANT-114-97-PROPAGATION + CM-IDLE-TIME-AUDIT + paired §1.1 mutations.
Cascade requirement: This anchor (verbatim or by §11.4.97 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-97-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.97 for the full mandate.
Verbatim user mandate (2026-05-28): "Make sure we have full automation testing of all scenarios with real bot, main group and users without any manual intervention or contribution of real user! Everything MUST BE fully automatic and autonomous! These tests MUST BE able to rerun endless times when needed! ... Make sure there is no false positives in testing! Every test and its results MUST obtain real proofs of everything working! No bluff is allowed!"
Closes the manual-intervention gap (§11.4 / §11.4.2 / §11.4.5 / §11.4.50 / §11.4.85 / §11.4.87 / §11.4.89 / §11.4.94 did not explicitly forbid it). A live/integration/e2e/Challenge test that requires a human action during execution (typing a message, clicking UI, hand-triggering a webhook, attaching a file — anything beyond startup) is by definition a §11.4 PASS-bluff at the automation layer. (A) Every governed test — unit/integration/e2e/Challenge/stress/chaos/live — MUST be fully self-driving end-to-end, reporting PASS/FAIL/SKIP-with-reason without any further human action after startup. (B) Single permissible exception: one-time credential bootstrap performed OUTSIDE test execution (.env from vault, shell exports, OAuth at first install, MTProto session activation) — configuration, not test driving. (C) Live messenger/channel/agent tests: no "operator must type" prompts (drive programmatically via second account / webhook fixture / loopback); no hard-coded session UUIDs that collide with the active dev session (Herald 2026-05-28 claude --resume silent exit -1 lesson); no 60 s human-response windows (§11.4.50 determinism violation); re-runnability proof — PASS at -count=3 consecutive automated invocations with self-cleaning state; §11.4.98 obsolescence audit classifies every existing test COMPLIANT vs NON-COMPLIANT; no silent-skip-reported-as-PASS or stale-evidence-as-fresh. (D) With §11.4.85 + §11.4.89 + §11.4.87 + §11.4.94 forms a continuously-validated, non-flake, anti-bluff regime. (F) Manual-dependency tests not rewritten within 30 days graduate to §11.4.90 Obsolete citing §11.4.98.
Cascade requirement: This anchor (verbatim or by §11.4.98 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-98-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.98 for the full mandate.
§11.4.99 — Latest-Source Documentation Cross-Reference Mandate (cascaded from constitution submodule §11.4.99)
Verbatim user mandate (2026-05-28): "Make sure we ALWAYS check against latest versions of services we use web / online docs before creating instructions! This situation is illustration of how we can misguide ourselves or get banned! ... These are mandatory rules / constraints and the result is consistency and safety of created instructions, guides and manuals!"
Misguidance-by-stale-docs is the same severity class as a §11.4 PASS-bluff at the documentation layer (Herald 2026-05-28 case: a first-draft MTProto guide recommended VoIP fallback numbers and omitted the recover@telegram.org pre-login email — both contradicted Telegram's official docs + the gotd/td maintainer guide and could have caused a permanent account ban). Closes the gap §11.4.92 Pass 4 alludes to but does not mandate. (A) Before committing any operator-facing instruction/guide/manual/troubleshooting/setup doc, the author MUST: (1) fetch the LATEST official online documentation of the documented service/library via WebFetch / MCP / direct browsing — NEVER training data, memory, or prior committed docs; (2) cross-reference every instruction step against that source; (3) seek secondary authoritative sources (maintainer SUPPORT.md, official changelogs, vetted community FAQs) when the official source is sparse/silent; (4) cite source URLs + date in a ## Sources verified footer in the doc; (5) cite a Sources verified <date>: <urls> footer in the commit message. (B) Negative findings (gaps/silences/contradictions) MUST be documented explicitly. (C) Docs older than 6 months are STALE — re-verify before citing as operator authority, at every vN.0.0 release boundary, on service breaking-change announcements, or on operator error reports. (D) Risk-classified services (messengers, cloud APIs, payment systems, AI/LLM providers, code-hosting, package managers) carry a 90-day max staleness + explicit safety warnings. (E) Composes with but is INDEPENDENT of §11.4.92 Pass 4. (G) Commit missing either footer is BLOCKED at release-gate; stale-beyond-grace docs graduate to §11.4.90 Obsolete (Reason=stale-documentation).
Cascade requirement: This anchor (verbatim or by §11.4.99 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-99-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.99 for the full mandate.
§11.4.101 — Autonomous-Decision-Over-Blocking Mandate (cascaded from constitution submodule §11.4.101)
Verbatim user mandate (2026-05-28): "when working in endless working loop fully autonomously try to decide most properly about points which would block execution and wait for us. If we haven't answered now work would be blocked whole night! If possible and if that will not cause any issues make proper and most reliable and safe decision so we achieve maximal efficiency and work gets fully done!"
In autonomous / endless-loop mode (per §11.4.87), the agent MUST minimize operator-blocking and make the safe, reliable, reversible decision itself so work is not stalled (e.g. overnight) waiting for input — §11.4.87 says keep working, §11.4.101 says HOW to clear the decision points. Proceed-autonomously (closed-set, ALL must hold): (a) the action is reversible OR has a captured pre-op backup per §9.2; (b) the safe choice is determinable from captured evidence per §11.4.6 (no guessing — LIKELY/probably/seems is NOT a determination); (c) a wrong choice's blast radius is bounded AND recoverable; (d) it composes with anti-bluff §11.4, host-safety §12, data-safety §9. Block-only-when (BLOCK via the §11.4.66 interactive mechanism ONLY when ALL hold): the action is irreversible AND high-blast-radius AND the safe choice cannot be determined from evidence — e.g. external-account state the agent cannot inspect, hardware it cannot access, destructive ops without backup, force-push (also §9.2 + §11.4.41), spending money or sending data to third parties. Operator-blocked per §11.4.21 is reached only after this rule fires AND the self-resolution-exhaustion audit completes. An unavoidable block parks one work unit — it does NOT pause the loop; the agent keeps progressing every non-blocked item in parallel per §11.4.87 + §11.4.94 (posing the question then going idle is a §11.4.94 + §11.4.97 violation). Classification: universal (§11.4.17).
Cascade requirement: This anchor (verbatim or by §11.4.101 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-101-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.101 for the full mandate.
§11.4.102 — Mandatory systematic-debugging activation + always-loaded skill-discovery + plugin-dependency availability (cascaded from constitution submodule §11.4.102)
Verbatim user mandate (2026-05-29): *"Make sure that we ALWAYS trigger / start the "/superpowers:systematic-debugging" skills when any issues happen! If this is possible to activate and use in this situations out of the box when we spot problems / issues / bugs / misalignments / unconsistencies we MUST activate the skill(s) and make strongest efforts in full in depth analisys / debugging and determine root causes of all problem or obtain relevant data and information we need! ... we MUST make sure that "/using-superpowers" skill is ALWAYS loaded, applied and used! All dependencies (plugins) that Claude Code or other market places are offering MUST BE installed if these are not already available for loading and use!"
Three cooperating invariants — the difference between guess-and-retry and investigate-to-root-cause-first. (A) Mandatory systematic-debugging activation. On ANY spotted issue / bug / test failure / gate failure / regression / misalignment / inconsistency / unexpected behaviour, the agent MUST activate superpowers:systematic-debugging (or the platform-equivalent structured-debugging discipline) BEFORE proposing, writing, or applying any fix — the Iron Law: NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST. Full four-phase arc: root-cause → pattern → hypothesis → implementation. Guess-and-retry, symptom-patching, and re-running a failed test hoping it passes ("probably transient / flaky") WITHOUT a completed investigation are §11.4.102 violations; calling a failure transient/flaky/intermittent/probably-timing without captured forensic evidence is simultaneously a §11.4.6 and §11.4.7 violation. (B) Mandatory always-loaded using-superpowers. superpowers:using-superpowers (or platform-equivalent skill-discovery discipline) MUST be loaded and applied at session start and consulted before any task; if ANY skill could apply — even at 1% relevance — it MUST be invoked rather than improvised from memory. (C) Mandatory plugin / dependency availability. Every skill plugin / marketplace package / capability dependency the project relies on MUST be installed + loadable BEFORE the dependent work proceeds; a missing plugin that blocks a mandated skill is a release-blocker until installed + confirmed loadable (install exit 0 ≠ skill loadable — confirm by observing the skill in the live capability list). Composes with §11.4.4 / §11.4.6 / §11.4.7 / §11.4.8 / §11.4.43 / §11.4.70 / §11.4.82(A) / §11.4.92. Classification: universal (§11.4.17). No escape hatch — no --skip-systematic-debugging, --guess-and-retry-OK, --symptom-patch-permitted, --skip-skill-discovery, --plugin-optional, --missing-plugin-is-warning flag.
Cascade requirement: This anchor (verbatim or by §11.4.102 reference) MUST appear in every owned submodule's CONSTITUTION.md, CLAUDE.md, and AGENTS.md. Propagation gate CM-COVENANT-114-102-PROPAGATION; paired mutation strips the literal → gate FAILs. Release blocker.
Canonical authority: constitution submodule Constitution.md §11.4.102 for the full mandate.
Forensic anchor — verbatim user mandate (2026-05-29):
"Do this working approach continuously and make it part of regular working routine and add it to the Constitution Submodule documented fully."
Promotes the proven multi-stream operating pattern into the project's standing default working routine (not a per-request opt-in). Binds §11.4.87/§11.4.88/§11.4.89/§11.4.94/§11.4.96 and adds the load-bearing invariant: the main work stream MUST always stay FREE.
(A) Main stream stays FREE. ALL commit AND push operations run detached (nohup … & + disown, per §11.4.88 commit-lock-release-immediately + detached-push) — the main stream returns to the priority queue the moment the local commit is durable, never blocking on a push or a slow mirror. (B) ≥3 parallel background streams at all times + auto-backfill (User mandate 2026-05-29; raised from ≥2) — run at least three subagent-driven background streams (per §11.4.70/§11.4.20, isolated per §11.4.58 PWU worktree + §11.4.84 quiescence) alongside the main stream whenever three-plus non-contending actionable items exist; the moment any one stream is FULLY done, a new stream MUST immediately start and take its place (claim next-highest-priority non-contending item), so the active-stream count NEVER drops below 3 while actionable items remain. Idle below 3 is permitted ONLY when no remaining non-contending actionable items OR all remaining are externally blocked (§11.4.94/§11.4.97/§11.4.101). Standing band 3–6, bounded above by §12.6 60% memory + §11.4.58 6-agent cap. (C) Most-critical + most-visible first; audio always top per §11.4.72 + §11.4.42 priority order. (D) Safe-during-build scope only — while the 42 GB containerised AOSP rebuild (§12.9) or any heavy gradle/m -j build runs, streams restrict to the §11.4.96 SAFE catalogue (investigation/forensic/docs/test-authoring/gate-authoring/read-only probes/submodule edits/research/DB ops/backgrounded pre-build+meta-test); NEVER a second concurrent heavy build (§12.8). (E) Heavy anti-bluff on every closure — root cause proven or UNCONFIRMED:/UNKNOWN:/PENDING_FORENSICS: per §11.4.6, captured evidence per §11.4.5+§11.4.69, deterministic consistency per §11.4.50, paired §1.1 mutations, §11.4.102 systematic-debugging on any spotted problem. (F) Idle ONLY when genuinely externally blocked (hardware/network upstream/in-flight build-test-push completion) OR operator STOP OR §12 host-safety, per §11.4.94(A)+§11.4.97; "nothing visible to do" with progressable items is NEVER valid; a block parks one work unit, not the loop (§11.4.101).
Composes with §11.4.58 / §11.4.70 / §11.4.72 / §11.4.87 / §11.4.88 / §11.4.89 / §11.4.94 / §11.4.96 / §11.4.97 / §11.4.101 / §11.4.102 / §11.4.42 / §11.4.84 / §12.6 / §12.7 / §12.8 / §12.9 / §9.2. Classification: universal (§11.4.17). Propagation gate CM-COVENANT-114-103-PROPAGATION (literal 11.4.103 across consumer fleet) + paired §1.1 meta-test mutation (strip literal → gate FAILs; gate-code = separate work item). No escape hatch — no --block-main-stream, --synchronous-commit, --synchronous-push, --single-stream-only, --skip-parallel-streams, --serialise-actionable-work, --idle-without-queue-survey flag.
Canonical authority: constitution submodule Constitution.md §11.4.103.
Non-compliance is a release blocker regardless of context.
Forensic anchor — verbatim user mandate (2026-05-31):
"Every supported messenger must relate messages to participants (Subscribers/Users); the same logical person may have a different username on every messenger. Workable items must carry who created them and who they are assigned to. Notifications must @-tag the right participant — but never the operator (who drives the system) and never the system agent."
MANDATORY for every consumer that ships a messenger/notification surface (Herald + flavor binaries are the reference impl; others inherit per §11.4.35). Detailed spec: Herald docs/design/PARTICIPANT_ATTRIBUTION.md — restated, not redefined. (A) Participant identity. Every messenger MUST relate messages to a Participant (logical Subscriber/User); the SAME person MAY have a DIFFERENT username per messenger — modelled as a logical subscriber (canonical messenger-neutral handle, kind ∈ {human,agent,service}) + per-channel aliases (channel, channel_user_id, the @username used for tagging). Canonical handle closed set: Claude (reserved system-agent sentinel; never tagged) OR a subscriber @username. (B) Operator = env var, not a DB flag — the one human who drives via the agent CLI, designated by HERALD_<CHANNEL>_OPERATOR_USERNAME (e.g. HERALD_TGRAM_OPERATOR_USERNAME); a normal Participant whose handle equals that value. (C) Workable items MUST carry created_by + assigned_to (canonical handles): CLI-prompt→Operator; system/agent-detected→Claude; received-message→sender's resolved @username. assigned_to defaults to Operator, overridable. Legacy items carry "" and MUST still parse + validate. (D) Tagging matrix: tag assigned_to if it is a human ≠ Operator; tag created_by if it is a human ≠ Operator ≠ Claude; NEVER tag Claude (system) or the Operator (no self-ping); de-dup; resolve each handle to the channel @username, skip if not on that channel. (E) Anti-bluff (composes §11.4): real SQLite round-trip with the new columns byte-identical (incl. legacy fixtures WITHOUT the fields); tagging matrix proven by a truth-table + a cell-flip mutation forcing FAIL; E2E real-event → real-message asserting the exact @usernames + a NEGATIVE case proving the Operator is NOT tagged; evidence under docs/qa/<run-id>/.
Composes with §11.4 + §11.4.1..§11.4.16 (anti-bluff covenant) / §11.4.5 / §11.4.69 / §11.4.50 / §11.4.91 / §11.4.93 / §11.4.95 / §1.1. Classification: universal (§11.4.17) — projects with no messenger surface inherit it latently (binds the moment they ship one, per the §11.4.96 pattern). Propagation gate CM-COVENANT-114-104-PROPAGATION (literal 11.4.104 across consumer fleet) + paired §1.1 meta-test mutation (strip literal → gate FAILs; gate-code = separate work item). No escape hatch — no --skip-attribution, --no-participant-tagging, --tag-operator-anyway, --attribution-later flag.
Canonical authority: constitution submodule Constitution.md §11.4.104; detailed spec Herald docs/design/PARTICIPANT_ATTRIBUTION.md.
Non-compliance is a release blocker.
Forensic anchor — verbatim user mandate (2026-05-31):
"Users must NOT need to know command syntax (no
COMMAND: …prefix). They send a clear natural-language message; the System determines the intent. The System recognizes the commands it has; if none matches it infers the exact intent; if it is totally unable it replies, tags the user (@user …), and asks to clarify precisely. We MUST always do our best to determine exact intent so we never annoy end users. This is a CORE part of the System."
MANDATORY for every consumer that ships a messenger/command surface (Herald + flavor binaries are the reference impl; others inherit per §11.4.35). Detailed spec: Herald docs/design/INTENT_RECOGNITION.md — restated, not redefined. (A) No required command syntax. Users MUST NOT be required to know any command syntax (no COMMAND: prefix); they send plain natural language and the System determines the intent. (B) Three-tier resolution (first that succeeds wins): TIER 1 — recognize the System's existing command set from natural language → action (confident deterministic match); TIER 2 — when no command matches, infer the exact intent (LLM dispatch maps language → action), NEVER guessing; TIER 3 — when neither a command nor a confident intent can be determined, REPLY to the message, TAG the sender (@username, resolved via the §11.4.104 IdentityResolver) and ask a PRECISE clarifying question NAMING the candidate intents — no guessing, no silent drop. (C) Never guess, never drop: a wrong action is worse than a clarifying question (composes §11.4.6 no-guessing); a message is NEVER silently dropped; only genuine ambiguity reaches Tier 3, which always replies-tags-and-asks. (D) Anti-bluff (composes §11.4): every tier ships unit + integration + E2E + full-automation tests with real captured evidence — Tier 1 truth-table (natural-language → action+fields, plus conservative negatives that MUST fall through to "no match"); Tier 3 E2E whose recorded reply body is EXACTLY @<sender> <specific question> + a NEGATIVE proving a clear command does NOT trigger clarify; a paired §1.1 mutation breaking the confidence guard (false-match) OR dropping the clarify tag MUST FAIL a test; evidence under docs/qa/<run-id>/.
Composes with §11.4 + §11.4.1..§11.4.16 (anti-bluff covenant) / §11.4.6 (no-guessing) / §11.4.104 (clarify reply tags the sender) / §11.4.5 / §11.4.69 / §11.4.98 / §1.1. Classification: universal (§11.4.17) — projects with no messenger surface inherit it latently (binds the moment they ship one, per the §11.4.96 pattern). Propagation gate CM-COVENANT-114-105-PROPAGATION (literal 11.4.105 across consumer fleet) + paired §1.1 meta-test mutation (strip literal → gate FAILs; gate-code = separate work item). No escape hatch — no --require-command-syntax, --guess-intent-ok, --skip-clarify, --drop-on-ambiguous flag.
Canonical authority: constitution submodule Constitution.md §11.4.105; detailed spec Herald docs/design/INTENT_RECOGNITION.md.
Non-compliance is a release blocker.
Forensic anchor — operator mandate (2026-05-31): Docs Chain is the canonical mechanical enforcer of the documentation-sync mandates; consumers MUST use the engine instead of ad-hoc per-project doc-sync scripts, register their chains via per-context YAML, and never accept a faked transform.
MANDATORY for every consumer. Docs Chain (the vasic-digital/docs_chain engine — a universal Go bidirectional document-and-database dependency-propagation engine) is the canonical mechanical enforcer of the documentation-sync mandates. Detailed spec: the engine docs ~/Projects/docs_chain → docs/CONSTITUTION_INTEGRATION.md (distribution + inheritance + anchor mapping table) + docs/USE_CASE_CATALOGUE.md (chain recipes) — restated, not redefined. (A) Use the engine, never ad-hoc scripts — consumed by reference (the flat-layout sibling ~/Projects/docs_chain / the constitution-exposed path), inherited like §11.4.80's codegraph_* scripts: referenced, NEVER copied; ad-hoc sync_*/generate_*_summary/update_readme_doc_links scripts are superseded and retired per registered context. (B) Consumer-owned contexts — the engine is project-agnostic; the consumer registers its chains as data via .docs_chain/contexts/*.yaml (§11.4.28 decoupling); state.json + *.docs_chain.tmp are gitignored. (C) Anchors it mechanizes (per the CONSTITUTION_INTEGRATION mapping table): §11.4.12 / §11.4.53 / §11.4.45 / §11.4.56 / §11.4.57 / §11.4.59 / §11.4.60 / §11.4.65 / §11.4.86 / §11.4.93 / §11.4.95 / §12.10 / §11.4.44 — with content-hash change detection (NOT mtime, §11.4.86), atomic-rename + SQLite-txn commit + rollback (§9.2), both-dirty sync → conflict-not-silent-merge (§11.4.6), verify as the deterministic CI/pre-build gate (§11.4.50), per-run captured evidence to qa-results/docs_chain/<run-id>/ (§11.4.69). (D) NOT a replacement for authoring discipline — the source author still writes the §11.4.44 revision header; the engine only keeps exports in sync. (E) Anti-bluff (composes §11.4): the engine NEVER fakes a transform — a missing pandoc/weasyprint surfaces a typed ToolAbsentError + honest §11.4.3 SKIP-with-reason, never a fake PASS / partial write; every sync/verify carries real captured evidence.
Composes with §11.4 + §11.4.1..§11.4.16 (anti-bluff covenant) / §11.4.6 (no-guessing — conflict not silent merge) / §11.4.28 (engine/context decoupling) / §11.4.80 (inherited-by-reference, never copied) / §9.2 / §11.4.50 / §11.4.69 / §11.4.5 / §1.1, plus the sync anchors it mechanizes (§11.4.12/.53/.45/.56/.57/.59/.60/.65/.86/.93/.95/.44, §12.10). Classification: universal (§11.4.17) — projects with no derived-export/DB-sync surface inherit it latently (binds the moment they ship one, §11.4.96 pattern). Status (§11.4.6): engine Phases 1–3 IMPLEMENTED+tested; CLI/YAML loader (Phase 4) + submodule distribution (Phase 6) PLANNED + OPERATOR-GATED — wire as phases land, claim no unshipped behaviour. Propagation gate CM-COVENANT-114-106-PROPAGATION (literal 11.4.106 across consumer fleet) + paired §1.1 meta-test mutation (strip literal → gate FAILs; gate-code = separate work item). No escape hatch — no --ad-hoc-sync-ok, --skip-docs-chain, --fake-transform, --sync-evidence-optional flag.
Canonical authority: constitution submodule Constitution.md §11.4.106; detailed spec the Docs Chain engine docs docs/CONSTITUTION_INTEGRATION.md + docs/USE_CASE_CATALOGUE.md (vasic-digital/docs_chain).
Non-compliance is a release blocker.
Forensic anchor (genericised, 2026-06-02): a test PASSed on a SINGLE captured frame showing "a picture" on the target output — but the picture was a FROZEN / STALE frame from the previously-played content (stale-producer / stuck-decoder), so the feature was broken for the user while the test was green; a sibling incident FLASHED media on the WRONG output for ~1 s before routing with no test sampling that window; a third class shipped a comparator that PASSed its own deliberately-degraded fixture (the analyzer, not the feature, was the bluff). §11.4.5 mandates captured evidence + a presence pass; §11.4.107 raises the bar to liveness + correct-routing + self-validated-analyzer.
Every test asserting audio/video output is genuinely playing/advancing for the end user MUST satisfy ALL of: (1) a single captured frame is NOT proof — prove LIVE, ADVANCING frames over a steady-state window via a freeze-detection oracle (near-duplicate / freezedetect-class filter OR perceptual-hash adjacent-frame distance, NOT byte-identical compare — byte-identity is only a zero-cost pre-filter); (2) an independent frame-advance counter from the platform's compositor/decoder telemetry must increase across the window (a different-domain oracle — flat counter ⇒ stuck decoder ⇒ FAIL even if pixels appear to move); (3) loading/buffering is a distinct state — wait for genuine playback-start before judging liveness, never false-FAIL a still-loading stream nor false-PASS a spinner; timeout+unreachable ⇒ SKIP-with-reason per §11.4.3, timeout+reachable ⇒ FAIL; (4) not-stale-from-previous cross-check — new content's first frame ≠ previous content's last frame; (5) measured FPS / no-lost-frames within tolerance; (6) no-flash-on-the-wrong-output — sample the non-target output at high frequency during a routing transition, any content frame there ⇒ FAIL (content-protection regime classified explicitly, never guessed); (7) drive through the realistic feed/UI path, not deep-link shortcuts (shortcuts bypass the transition paths where bugs live; UI-not-introspectable ⇒ operator-attended fallback per §11.4.52, never fake-PASS); (8) metamorphic relations solve the oracle problem when there is no golden source (same content on output-A vs output-B must match; paused ⇒ counter stops; 2× speed ⇒ ~2× advance rate); (9) full-reference quality metrics (SSIM/VMAF/ΔE2000) vs a golden source for owned content; (10) mutation-test every analyzer with a golden-good + golden-bad fixture pair so the analyzer itself provably cannot bluff (an analyzer that PASSes its golden-bad fixture is a bluff gate — §1.1 applied to the analyzers); (11) per-channel audio RMS/loudness (EBU R128) + XRUN/underrun census — a single aggregate RMS misses a dead channel; (12) OCR overlay/subtitle detection needs a per-word confidence floor + ROI to avoid BOTH false-positive (false-FAIL) and false-negative (false-PASS); (13) thresholds calibrated on the project's own fixtures, not hardcoded from literature (§11.4.6 no-guessing).
Honest gap (§11.4.6): true photon FPS at the sink has no clean software oracle — compositor/decoder counters measure the presentation pipeline; flag the gap, never claim it.
4-layer coverage per §11.4.4(b): pre-build gate (a CM-AV-LIVENESS-NO-FROZEN-FRAME-class gate asserting every output-is-playing test references the liveness battery not a single frame + an analyzer-self-validation gate wiring the golden-good/golden-bad fixtures into meta-test) + on-device/runtime test + paired §1.1 meta-test mutation (single-frame-only assertion → gate FAILs; analyzer that PASSes its golden-bad fixture → self-validation FAILs) + HelixQA Challenge. Every PASS via ab_pass_with_evidence citing the motion / not-stale / frame-advance / fps / per-channel-loudness / metamorphic / self-validation artefacts (video_display / audio_output / subtitle_render per §11.4.69) — never a single screenshot.
Classification: universal (§11.4.17) — platform-neutral AV/test-validation techniques reusable by ANY project validating media playback or any pixel/audio output; the project supplies its concrete capture mechanism + calibrated thresholds per §11.4.35. Composes with §11.4.5 (its strict expansion), §11.4.6, §11.4.50, §11.4.68, §11.4.69, §11.4.85, plus §11.4.2 / §11.4.3 / §11.4.13 / §11.4.48 / §11.4.52 / §1.1. Propagation gate CM-COVENANT-114-107-PROPAGATION enforces the literal anchor 11.4.107 across the consumer fleet; paired §1.1 meta-test mutation strips the literal → gate FAILs (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.107.
Non-compliance is a release blocker regardless of context. No escape hatch — no --single-frame-proves-playback, --skip-liveness, --byte-identical-freeze-OK, --no-frame-advance-counter, --skip-not-stale-check, --allow-wrong-output-flash, --deep-link-shortcut-OK, --unvalidated-analyzer-OK, --aggregate-rms-suffices, --hardcoded-thresholds-OK flag exists.
§11.4.108 — Four-layer fix-verification + runtime-signature-as-definition-of-done mandate (systematic-debugging Phase 4.5, 2026-06-03)
Forensic anchor (genericised, 2026-06-03): across one batch, multiple fixes were "green" at every gate yet NOT working for the end user — a change present in source and passed by both the source pre-build gate AND the post-build gate never reached the boot-time command-line embedded in the deployed boot artifact (output feature dead despite all-green); sibling fixes correctly built into the system image were masked by stale per-user overlay copies from a previous deployment (the running code was the stale shadow, not the fresh build); deployment did not wipe the mutable overlay so the shadow survived; validation ran against whatever was running — the stale shadow — and reported green on code that was never exercised. Per superpowers:systematic-debugging Phase 4.5: each fix revealing a fresh "fixed-but-not-working" in a DIFFERENT place is NOT independent bugs — it is ONE architectural VERIFICATION flaw; patching each symptom is thrashing.
A fix crosses FOUR distinct layers, and "fixed" at one does NOT imply fixed at the next: (1) SOURCE (committed in the source file — what a grep-the-source pre-build gate checks), (2) ARTIFACT (the change's BYTES actually landed in the produced build artifact — image / bundle / installer / embedded command-line), (3) RUNTIME-ON-CLEAN-TARGET (active on a CLEAN/fresh deployment — the layer the end user experiences — with no stale overlay shadowing the deployed code), (4) USER-VISIBLE (the feature works for the end user — the §11.4.5/§11.4.69 captured-evidence layer). Green at layer 1 is the cheapest, least conclusive signal and says nothing about layers 2–4. A gate verifying ONLY the source layer is itself a §11.4 bluff surface.
The mandate (ALL must hold): (1) Runtime-signature-as-definition-of-done — a fix is DONE only when its declared runtime signature verifies on a CLEAN/fresh deployment; source-committed ≠ artifact-contains-it ≠ active-on-clean-target ≠ user-visible-working. (2) Every fix declares ONE machine-checkable runtime signature — a single observable on a clean target proving the fix is BOTH active AND working (a running-system property, a downstream/sink-side report per §11.4.13, a §11.4.5/§11.4.69 captured-evidence assertion, or a counter/state delta — NEVER a re-grep of the source); this registry of per-fix runtime signatures is the SINGLE SOURCE OF TRUTH for "fixed" — it REPLACES "a gate greps the source" as the definition of done. (3) Gates span all four layers — source (pre-build), artifact (post-build — assert the change's BYTES landed in the artifact, not merely that the source still contains the change), runtime-on-clean-target (post-deploy — assert the runtime signature on a freshly-deployed clean target), user-visible (§11.4.5/§11.4.69). (4) Eliminate the stale-deployment / shadow layer by construction — deployment MUST yield a CLEAN state (wipe the mutable overlay) OR a pre-validation assertion MUST prove running-artifact == built-artifact BEFORE any validation runs; validation against possibly-stale deployed state is INVALID and any PASS it produces is a §11.4 PASS-bluff (the test exercised code that was never deployed). (5) Meta-rule — ≥ 3 "fixed-but-not-working" discoveries in one cycle signal an architectural VERIFICATION flaw, NOT three independent bugs; on the 3rd, STOP patching symptoms (§11.4.4), fix the VERIFICATION pipeline, and re-certify EVERY item through it on a clean target. (6) A batch is "validated" only after COMPREHENSIVE per-item runtime-signature verification on a clean baseline — NOT after the touched items' own tests pass.
Classification: universal (§11.4.17) — platform-neutral verification-pipeline disciplines reusable by ANY project that builds an artifact, deploys it, and validates the deployed result; the project supplies its concrete artifact format, clean-deployment / equal-artifact mechanism, and per-fix runtime-signature observables per §11.4.35. Composes with §11.4.1 / §11.4.2 / §11.4.4 (clause 5's STOP-and-fix-pipeline is its §11.4.108 specialisation; "clean baseline" = clause 4's clean deployment) / §11.4.5 / §11.4.6 (every layer asserted by evidence, never assumed to propagate) / §11.4.27 / §11.4.40 (§11.4.108 adds the cross-layer per-item runtime-signature dimension) / §11.4.46 (clean-baseline / equal-artifact pre-flight) / §11.4.50 / §11.4.52 / §11.4.69 (a runtime signature is a taxonomy-class observable) / §11.4.102 (Phase 4.5 architectural-flaw recognition IS clause 5's trigger). Propagation gate CM-COVENANT-114-108-PROPAGATION (literal 11.4.108 across the consumer fleet) + recommended per-fix gate CM-RUNTIME-SIGNATURE-REGISTRY + paired §1.1 meta-test mutations (strip the literal → propagation gate FAILs; downgrade a fix to source-only verification → registry gate FAILs; gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.108.
Non-compliance is a release blocker regardless of context. No escape hatch — no --source-green-is-done, --skip-artifact-byte-check, --validate-against-running-state, --no-clean-deployment, --skip-runtime-signature, --spot-validate-touched-only flag exists.
§11.4.109 — Mandatory Anti-Forgetting Enforcement: PreToolUse Guard Hook + Subagent Constitutional Preamble + Orchestrator Pre-Action Checklist (Operator mandate)
Short tag: anti-forgetting-enforcement. UNCONFIRMED forensic anchor (pending operator's verbatim mandate quote). Background: emulator subagents ran raw host-direct emulator/adb because the orchestrator forgot to inject the Containers-submodule rule. A rule forgotten at dispatch is not enforcement. Fix: (A) a PreToolUse guard hook (constitution/scripts/hooks/guard-forbidden-commands.sh) that blocks host-direct emulator, force-push/bypass, sudo, and host-power commands at the tool-call boundary regardless of agent memory; (B) a canonical docs/AGENT_GUARDRAILS.md preamble the orchestrator pastes verbatim into every subagent dispatch; (C) an ORCHESTRATOR PRE-ACTION CHECKLIST in the same document. Hook = the floor; preamble = the ceiling.
Consuming projects MUST: (1) wire constitution/scripts/hooks/guard-forbidden-commands.sh as a PreToolUse hook in .claude/settings.json (or equivalent runtime settings); (2) maintain docs/AGENT_GUARDRAILS.md containing the SUBAGENT CONSTITUTIONAL PREAMBLE and ORCHESTRATOR PRE-ACTION CHECKLIST headings, with the anchor literal 11.4.109; (3) provide a hermetic hook test suite (≥ 20 cases: every blocked class exits 2, every allowed command exits 0, escape hatch fires for non-power classes, host-power rejects even with escape marker). The hook is inherited by reference — NEVER copied locally (a copy diverges silently).
Gates: CM-ANTI-FORGETTING-ENFORCEMENT (hook present + wired + guardrails doc present + test present) + CM-COVENANT-114-109-PROPAGATION (anchor literal 11.4.109 across every consumer CLAUDE.md / AGENTS.md / QWEN.md). Paired §1.1 mutations: remove hook entry → gate (2) FAILs; delete guardrails doc → gate (3) FAILs; strip hook from constitution → gate (1) FAILs; strip 11.4.109 → propagation gate FAILs.
Classification: universal (§11.4.17). Composes with §11.4.6 / §11.4.10 / §11.4.75 / §11.4.76 / §11.4.78 / §11.4.79 / §11.4.80 / §11.4.81 / §11.4.84 / §11.4.98 / §11.4.102 / §12.
Canonical authority: constitution submodule Constitution.md §11.4.109; reference implementation constitution/scripts/hooks/guard-forbidden-commands.sh + constitution/docs/AGENT_GUARDRAILS.md.
Non-compliance is a release blocker. No escape hatch — no --skip-pretooluse-hook, --no-guardrails-doc, --anti-forgetting-optional, --single-layer-sufficient flag exists.
§11.4.110 — Pre-build build-readiness verdict + change-impact clash detection mandate (operator mandate, 2026-06-03)
Forensic anchor (genericised, 2026-06-03): a fix shipped a new system-property read but introduced no matching security-policy grant + no property-context type entry — the read was silently denied at runtime + the feature was dead, while every pre-build gate stayed green because the gate only grepped the source file. The defect class generalises: a change introduces a new dependency on a second artifact (security policy, context file, service registry, interface freeze-snapshot, symbol table, build-graph node) that the pre-build never cross-checks. This is §11.4.108's SOURCE→ARTIFACT gap shifted LEFT to pre-build time — most such clashes are statically catchable from the change diff itself, before any build. The mandate (ALL hold): (1) a single deterministic READY-FOR-BUILD verdict gates the rebuild (orchestrator refuses to start the build unless READY); (2) a diff-driven change-impact + clash detector cross-checks every newly-introduced second-artifact dependency (new property read ⇄ property-context type + read-grant; new service ⇄ service-context entry; new init service ⇄ security label; new/changed stable interface ⇄ freeze-snapshot updated; new native-lib dep ⇄ module/prebuilt resolves; new policy rule ⇄ every type/attribute defined; two batch changes on the same seam ⇄ collision acknowledged); (3) coverage-completeness is a gate — every changed file maps to ≥1 gate + ≥1 deployed-target test + ≥1 paired §1.1 mutation, baseline ratchets upward per §11.4.50; (4) two-speed honesty — grep-speed always-on gates vs REQUIRES_BUILD heavy gates (build-graph parse-only dry-run, full neverallow compilation, ABI diff) as diff-gated opt-in stages, bounded per §12.6/§12.7; (5) every gate + wired analyzer is anti-bluff by paired §1.1 mutation; (6) honest boundary — a READY verdict proves static internal-consistency + ready-to-build, NOT that the feature works on the deployed target; the regime empties the preventable defect class, not the all-defects class (runtime/USER-VISIBLE remains §11.4.108's job).
Classification: universal (§11.4.17) — platform-neutral pre-build-rigor disciplines reusable by ANY project that builds an artifact from source; the consuming project supplies its concrete property/service/policy registries, build-graph parse-only command, interface-freeze mechanism, and changed-file→gate mapping per §11.4.35. Composes with §11.4.1 / §11.4.4 / §11.4.6 / §11.4.9 / §11.4.27 / §11.4.50 / §11.4.67 / §11.4.75 / §11.4.92 / §11.4.108 (§11.4.110 is the SOURCE→ARTIFACT half shifted left to pre-build; §11.4.108 owns RUNTIME-ON-CLEAN-TARGET→USER-VISIBLE — together they span all four layers). Propagation gate CM-COVENANT-114-110-PROPAGATION (literal 11.4.110) + recommended per-family gates CM-READY-FOR-BUILD-VERDICT / CM-CHANGE-IMPACT-CLASH-DETECTOR / CM-COVERAGE-COMPLETENESS-GATE / CM-BUILDGRAPH-DRYRUN-WIRED / CM-SEPOLICY-NEVERALLOW-WIRED + paired §1.1 mutations (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.110. Non-compliance is a release blocker. No escape hatch — no --source-green-is-ready, --skip-clash-detector, --skip-coverage-gate, --no-ready-verdict, --grep-proves-neverallow, --skip-buildgraph-dryrun, --build-without-ready-verdict flag.
Forensic anchor (genericised, 2026-06-03): a platform bound an audio output to a kernel-enumerated device index (card=0); a second device of a different class enumerated FIRST at boot and took slot 0, shifting the intended device to slot 1 — the static index binding now pointed at the wrong device (policy mis-attached the sink, mis-assigned its TYPE, the output switcher labelled the AV receiver "Wired Headphone" + routing collapsed to stereo). The lower layer of the SAME stack already resolved the device correctly by name (scanning the controller-name registry) — proving the brittleness was the index binding, not the resolution capability. Generalises to every enumerated resource whose ordinal is assigned at discovery/boot/hotplug time (non-deterministic across reboots, device additions, topology changes). The mandate: any binding to a hardware device / resource handle / enumerated entity (audio cards, display connectors, network interfaces, storage devices, GPU render nodes, input/camera devices, container/process slots) MUST resolve by a stable identifier (name / UUID / serial / label / controller-name / content-hash / sink-reported identity) and MUST NOT bind by enumeration index / ordinal / slot, UNLESS the platform documents that ordinal as deterministically pinned AND the pin is itself captured + asserted as part of the binding. Where a stable identifier exists at one layer, every other layer binding the same resource MUST use the same identifier (mixed by-name-here / by-index-there is the structural weak link forbidden here). Honest boundary (§11.4.6): when only an ordinal exists, pin it deterministically via the platform's own mechanism, capture the pin in the binding's §11.4.108 runtime signature, and document the residual fragility as UNCONFIRMED:-class risk — never silently trust an unpinned ordinal.
Classification: universal (§11.4.17) — platform-neutral binding-robustness discipline reusable by ANY project binding to enumerated hardware/resources/handles; the consuming project supplies its stable-identifier mechanism (ALSA card name, DRM connector name, NIC predictable name, block-device UUID, etc.) + the layers that must agree per §11.4.35. Composes with §11.4.6 (no-guessing — "the index is usually stable" is the exact guess forbidden) / §11.4.8 (mature stacks resolve by name/UUID — reproducing a known-brittle index binding when the by-name path is documented is a §11.4.8 omission) / §11.4.69 (sink-side evidence verifies the by-name binding's correctness) / §11.4.108 (the by-name binding's runtime signature asserted on a clean target across the topology that broke the index) / §11.4.110 (a new ordinal binding in a diff is a statically-catchable clash class). Propagation gate CM-COVENANT-114-111-PROPAGATION (literal 11.4.111) + recommended gate CM-RESOLVE-BY-NAME-NOT-INDEX + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.111. Non-compliance is a release blocker. No escape hatch — no --allow-index-binding, --ordinal-is-stable-enough, --skip-name-resolution, --trust-unpinned-index flag.
§11.4.112 — Structural-impossibility won't-fix classification mandate (research-derived, 2026-06-03)
Forensic anchor (genericised, 2026-06-03): deep research (§11.4.8) into relocating protected (content-protection / secure-surface) video to a secondary display PROVED — from authoritative platform/HDCP docs + reproducible captured behaviour (a secure surface is blanked on any output lacking the secure flag; mirror/screencap of a secure layer returns black) — that the goal is structurally impossible by platform design, not a missing feature or unsolved engineering problem. Without a durable classification, such a goal is re-investigated every cycle (re-read the same sources, re-run the same probes, re-derive the same impossibility — compounding wasted effort). The mandate: when deep research per §11.4.8 PROVES (cited authoritative sources AND, where applicable, reproducible captured evidence) a goal is structurally impossible on the target platform (forbidden by platform design / hardware-protocol constraint / documented kernel-or-API limitation — NOT merely unimplemented or hard), the goal MUST be: (1) classified Won't-fix + closed per §11.4.90 with closure reason structurally-impossible; (2) documented with the impossibility evidence — cited authoritative source URLs (per §11.4.99 latest-source verification) + the reproducible probe/captured evidence — in the tracker entry + relevant docs/ guide; (3) NOT re-attempted in future cycles — a reopen MUST cite NEW evidence the platform constraint changed (per §11.4.34 + §11.4.7), never merely re-derive the same impossibility; (4) paired with the correct posture — the entry states what the project DOES instead, so "impossible" is never confused with "broken/unhandled". Honest boundary (§11.4.6): structurally-impossible is reserved for proven platform/hardware/protocol impossibility — "could not find a way" / "very hard" / "no time" are Operator-blocked (§11.4.21) or open work, NOT won't-fix; mislabelling them to avoid the work is a §11.4 planning-layer bluff. A future platform change can make the impossible possible; the classification is durable but not eternal.
Classification: universal (§11.4.17) — platform-neutral effort-conservation + honesty discipline reusable by ANY project; the consuming project supplies the specific impossible goal, its platform constraint, and the cited evidence per §11.4.35. Composes with §11.4.6 (FACT with cited evidence, never "probably can't") / §11.4.7 (reopening requires NEW positive evidence) / §11.4.8 ("NO external solution found — structurally impossible" is the citation) / §11.4.34 (reopen attribution) / §11.4.90 (structurally-impossible is a closure reason in the closed vocabulary) / §11.4.99 (latest-source so the verdict is not stale). Propagation gate CM-COVENANT-114-112-PROPAGATION (literal 11.4.112) + recommended gate CM-WONT-FIX-STRUCTURAL-IMPOSSIBILITY + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.112. Non-compliance is a release blocker. No escape hatch — no --wont-fix-without-proof, --reattempt-closed-impossible, --skip-impossibility-evidence, --impossible-equals-broken flag.
Forensic anchor — verbatim user mandate (2026-06-03): "Any force-push is strictly forbidden! We must for every Submodule take as a base latest commit on Submodule's main (or master) branch, then on top of it carefully to merge all changes that have to be pushed! Once all merging is carefully done we perform commit and push to all Submodule's upstreams!"
Force-push is STRICTLY FORBIDDEN with NO exception — git push --force, --force-with-lease, +<ref>, or any history-rewriting overwrite of a remote ref, against EVERY repository this Constitution governs (main repo, this constitution submodule, every owned + nested submodule, every upstream). No operator-approval path, no "after a merge-first audit" path. The mandated 6-step integration procedure for any repo/submodule whose local has commits to publish OR whose mirrors diverged: (1) git fetch --all --prune --tags all remotes; (2) set the base to the LATEST commit on the canonical main/master branch (the most-advanced mirror tip); (3) carefully MERGE every change to be published on top of that base — union, preserve BOTH sides, NEVER -s ours / rebase / reset that drops commits (per §9 no-commit-loss); (4) resolve every conflict carefully — no conflict markers, no file dropped, gates/tests still pass; (5) commit the merge (stage only intended files, NEVER git add -A in a submodule per §11.4.30); (6) push to ALL upstreams — each push is a fast-forward because the merge commit descends from every mirror tip, so NO force is ever needed (if an upstream still rejects, return to step 1 for it, merge its new tip, re-validate, re-push). TIGHTENS §11.4.41 / §11.4.71 / §9.2 / CONST-043 — the force-push escape hatch is REMOVED: even WITH operator approval, even after a clean merge-first audit, force-push is forbidden, because the merge-onto-latest-main path is always available so force is never necessary. Those clauses' merge-first/fetch-first machinery stays in force as the integration discipline; only their terminal "...then force-push" step is struck.
Classification: universal (§11.4.17) — a platform-neutral integration discipline reusable across every repository. Composes with §2.1 (multi-upstream push — step 6 fans out) / §9 / §9.2 (absolute data safety — this is the no-loss push discipline that makes force unnecessary) / §11.4.4 / §11.4.6 (remote state unknowable without the step-1 fetch) / §11.4.26 (constitution-update conflict resolution IS this procedure) / §11.4.37 (fetch-before-edit → fetch-before-push) / §11.4.40 / §11.4.41 (merge-first stays, force-push step removed) / §11.4.71 (same tightening) / §11.4.88 (background-push still fast-forward-only) / CONST-043 (no force-push authorisable). Propagation gate CM-COVENANT-114-113-PROPAGATION (literal 11.4.113) + recommended gate CM-NO-FORCE-PUSH-ABSOLUTE (scan tracked scripts/hooks for push --force / --force-with-lease / push +<ref> + reject; a §11.4.109-class PreToolUse guard blocks the class at the tool-call boundary) + paired §1.1 mutation (inject a git push --force into a tracked script → gate FAILs; gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.113. Non-compliance is a release blocker. No escape hatch — no --force, --force-with-lease, --allow-force-push, --force-push-authorised, --skip-merge-onto-main flag.
When a previously-working feature/behaviour is observed broken, the FIRST diagnostic action MUST be to identify the last release tag (or commit/build/deploy) at which it was KNOWN-GOOD and diff/bisect the broken state against it — BEFORE any open-ended root-cause hunt or speculative fix. The known-good revision is the regression oracle: it bounds the search to the commits between good and now (git diff <good-tag>..HEAD --stat of the feature's files = captured evidence), tells you it is a regression REPAIR not a from-scratch design problem, and gives each suspect file a behavioural oracle. When the operator volunteers a known-good tag, that lead is load-bearing and MUST be acted on first. Default to a SURGICAL forward-fix (keep the post-good-tag features, revert ONLY the broken sub-part) over a wholesale revert that loses the batch's other working features — unless the operator prefers wholesale revert. Honest boundary (§11.4.6): "it worked before" is a HYPOTHESIS until the known-good tag is identified AND the feature is confirmed working there; "probably regressed in the last batch" without the diff is a guess; a feature that NEVER worked is not a regression. Composes §11.4.4 / §11.4.6 / §11.4.7 / §11.4.40 / §11.4.43 / §11.4.102 / §11.4.108. Propagation gate CM-COVENANT-114-114-PROPAGATION (literal 11.4.114) + recommended gate CM-REGRESSION-ISOLATED-AGAINST-KNOWN-GOOD + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.114. Non-compliance is a release blocker. No escape hatch — no --skip-known-good-diff, --root-cause-from-scratch, --assume-regression-source, --wholesale-revert-without-isolation flag.
§11.4.115 — RED-baseline-on-the-broken-artifact + polarity-switch mandate (1.1.8-dev remediation, 2026-06-03)
Strict refinement of §11.4.43's RED step. Every RED test MUST be authored to REPRODUCE the defect on the CURRENT, pre-fix artifact (the actual broken build/deployment), capturing positive evidence per §11.4.5 / §11.4.69 / §11.4.107 that the defect is genuinely present — never a synthetic failure the fix is then written to agree with. The SAME test source MUST carry a single polarity switch (env flag / parameter, canonical RED_MODE, default 1 = reproduce-and-assert-defect-present) that flipped to 0 post-fix converts the test into the GREEN regression-guard asserting the defect is ABSENT. One source, two roles: the bug-catcher IS the regression-guard; no separate happy-path test is authored as the primary guard (that demonstrates only that the test agrees with the fix — the §11.4.43 PASS-bluff). RED-on-broken-artifact then GREEN-on-fixed-artifact (on a clean target per §11.4.108) MUST both be captured. Honest boundary (§11.4.6): if the RED run does NOT fail on the broken artifact, that is a finding (close per §11.4.7 with negative evidence, or fix the test) — a RED test that passes on the known-broken artifact is a blind test. Composes §11.4.1 / §11.4.2 / §11.4.5 / §11.4.69 / §11.4.107 / §11.4.4 / §11.4.7 / §11.4.43 / §11.4.50 / §11.4.108 / §11.4.114. Propagation gate CM-COVENANT-114-115-PROPAGATION (literal 11.4.115) + recommended gate CM-RED-POLARITY-SWITCH-PRESENT + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.115. Non-compliance is a release blocker. No escape hatch — no --green-only-test, --skip-red-reproduction, --separate-happy-path-suffices, --synthetic-red-OK flag.
§11.4.116 — Real-time conductor↔autonomous-test-framework sync channel mandate (1.1.8-dev remediation, 2026-06-03)
Any autonomous, long-running test/QA/validation framework an external orchestrator (conductor agent / operator) depends on for real-time decisions MUST expose a real-time sync channel: (1) a structured append-only event stream (JSONL or equivalent, one event per line, never rewritten) emitting at minimum session-start / phase-transition / per-test-or-challenge-start / captured-evidence-path / external-call (LLM / vision / sink-probe) / error / per-item-verdict events; (2) an atomically-rewritten status snapshot (single small file written write-temp-then-rename so a reader never sees a torn write) carrying current session/phase/item + counters + last verdict. Verdicts use the closed vocabulary PASS / FAIL / SKIP / OPERATOR-BLOCKED (§11.4.45). The conductor tails it live so it stays in real-time sync, can §11.4.4-interrupt on a fresh defect, and never idles blindly (§11.4.94 / §11.4.97). Anti-bluff: a verdict event MUST carry the evidence path that backs it (§11.4.69) — a PASS event with no evidence path is a channel-layer PASS-bluff; a snapshot reporting PASS while the stream shows no evidence event for that item is a contradiction → treat as FAIL; an item with no start-event cannot have a verdict-event. When the framework is an owned project-agnostic submodule (§11.4.28), the channel stays project-neutral — the consumer registers its data (endpoints, package ids, sink hosts) at runtime via the public API, never hardcoded. Composes §11.4.4 / §11.4.5 / §11.4.69 / §11.4.27 / §11.4.28 / §11.4.45 / §11.4.52 / §11.4.89 / §11.4.94 / §11.4.97. Propagation gate CM-COVENANT-114-116-PROPAGATION (literal 11.4.116) + recommended gate CM-AUTONOMOUS-FRAMEWORK-SYNC-CHANNEL + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.116. Non-compliance is a release blocker. No escape hatch — no --no-sync-channel, --end-of-session-report-suffices, --verdict-without-evidence-path, --torn-status-write-OK flag.
§11.4.117 — Computer-vision / OCR pixel-oracle fallback for non-introspectable UIs mandate (1.1.8-dev remediation, 2026-06-03)
Any test that needs to drive a UI control OR assert on-screen content MUST NOT assume the accessibility/semantic/DOM hierarchy is the source of truth for what the user sees. When the hierarchy is blank/partial/known-unreliable for the app under test (TV-Compose, leanback, canvas/SurfaceView/GL, games, custom-rendered UIs), the test MUST fall back to a PIXEL ORACLE: (1) DRIVE input by computer-vision template-match (locate a control by its rendered appearance → tap its screen coordinates), not by a hierarchy node id; (2) ASSERT content by ROI OCR (read the rendered text the user reads — caption strip, title, error overlay) with a per-word confidence floor + region-of-interest per §11.4.107(12), not a hierarchy text attribute that may be absent/stale. The tool MUST both drive input AND read pixels — a hierarchy-only tool is NOT a content oracle. This makes §11.4.52's "near-empty hierarchy → INFEASIBLE" constructive: pixel-drive, not only SKIP. Anti-bluff (§11.4.107(10)): the CV/OCR analyzer is self-validated — golden-good fixture PASSes, golden-bad fixture FAILs, wired into meta-test; thresholds calibrated on the project's own frames, not hardcoded (§11.4.6). Honest boundary: when BOTH hierarchy is blank AND pixel oracle is infeasible (secure surface black-captures per §11.4.112, geo-unreachable), SKIP-with-reason per §11.4.3 + tracked operator-attended migration item per §11.4.52 — never fake PASS. Composes §11.4.5 / §11.4.6 / §11.4.48 / §11.4.49 / §11.4.52 / §11.4.107 / §11.4.112. Propagation gate CM-COVENANT-114-117-PROPAGATION (literal 11.4.117) + recommended gate CM-CV-OCR-PIXEL-ORACLE-FALLBACK + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.117. Non-compliance is a release blocker. No escape hatch — no --hierarchy-is-content-oracle, --skip-pixel-fallback, --unvalidated-ocr-OK, --hardcoded-ocr-threshold-OK flag.
§11.4.118 — Discovery-pressure to confirm known-issue-set completeness mandate (1.1.8-dev remediation, 2026-06-03)
A remediation/release cycle MUST NOT treat "every reported defect is fixed" as "the build is good" — the reported set is biased by what the operator happened to test ("we only see what we test"). After/alongside fixing the reported set, the cycle MUST run a discovery + stress pass across ALL target devices/environments that deliberately exercises subsystems, journeys, and edge cases BEYOND the reported defects — to CONFIRM the reported set is the COMPLETE critical set and surface unreported defects before the end user does. The pass MUST produce PROVABLE coverage: an enumerated list of the subsystems/user-journeys/stress scenarios actually exercised, each with its outcome (no-new-issue, or a new tracker entry per §11.4.15 / §11.4.16). "We found no other issues" is a §11.4 bluff unless accompanied by "here is the enumerated set we exercised" — absence of evidence of looking is not evidence of absence. New findings trigger §11.4.4 interrupt + the §11.4.114/§11.4.115 isolation→RED→fix loop. Honest boundary (§11.4.6): discovery reduces the unknown-unknown surface but does not prove zero remaining defects — the earned claim is "reported set + enumerated discovery set addressed," with un-exercised subsystems stated as honest coverage gaps (§11.4.3), never silently implied clean. Composes §11.4.4 / §11.4.5 / §11.4.69 / §11.4.25 / §11.4.40 / §11.4.42 / §11.4.52 / §11.4.85 / §11.4.114 / §11.4.115 / §11.4.119. Propagation gate CM-COVENANT-114-118-PROPAGATION (literal 11.4.118) + recommended gate CM-DISCOVERY-COVERAGE-ENUMERATED + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.118. Non-compliance is a release blocker. No escape hatch — no --reported-set-suffices, --skip-discovery-pass, --no-issues-without-coverage-list, --assume-complete flag.
§11.4.119 — Single-resource-owner partitioning for parallel hardware testing mandate (1.1.8-dev remediation, 2026-06-03)
Strict refinement of §11.4.58 / §11.4.103 for hardware contention. When multiple parallel work/test/discovery streams exercise SHARED hardware or any exclusive-access resource (a media-playback path, a single HDMI/audio sink, an exclusive device handle, a serial/JTAG line, a GPU under exclusive capture), exactly ONE stream MUST own each such resource at a time. The exclusive owner drives it (playback, input injection, capture); every other concurrent stream targeting the same resource MUST be READ-ONLY (passive probes — dumpsys / /proc / /sys reads / sink-side network probes / log tails). Parallelism is partitioned by resource: distinct devices/sinks/handles run fully concurrent (stream-per-device), but the same device's exclusive resource is single-owner. Ownership MUST be enforced by an advisory lock/token (§11.4.58 L3 + the hardware analogue of §11.4.84 quiescence), event-driven (claim when the resource frees, release the moment work completes so the next queued stream claims it per §11.4.103 auto-backfill). Why: concurrent drivers of one exclusive resource produce CROSS-CONTAMINATED evidence (sink reports the wrong stream's audio, foreground belongs to whichever am start landed last, input events interleave) — a PASS under contention is a §11.4 evidence-integrity bluff. Honest boundary (§11.4.6): a "read-only" probe stream MUST issue NO state-changing command against a device it does not own; when a resource genuinely cannot be partitioned, streams serialize on it (single-owner over time), not run concurrently and hope. Composes §11.4.5 / §11.4.69 / §11.4.13 / §11.4.50 / §11.4.58 / §11.4.82 / §11.4.84 / §11.4.103 / §11.4.118. Propagation gate CM-COVENANT-114-119-PROPAGATION (literal 11.4.119) + recommended gate CM-SINGLE-RESOURCE-OWNER-PARTITION + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.119. Non-compliance is a release blocker. No escape hatch — no --allow-concurrent-resource-drivers, --skip-ownership-lock, --read-only-may-mutate, --contended-evidence-OK flag.
When a correct fix causes a pre-existing gate/test to FAIL because that gate asserted the OLD (now-removed-or-changed) behaviour, the gate FAIL is the CORRECT signal the fix landed — it MUST NOT be suppressed by either forbidden response: (1) FAKE-PASSING the gate (editing it to always pass, weakening its assertion to a tautology, or deleting it — a §11.4 gate-layer bluff + a §11.4.84 mutation-residue risk); (2) REVERTING the correct fix to satisfy the stale gate (re-introducing the defect). The required response is RECONCILIATION: rewrite the gate to assert the NEW mechanism the fix introduced, backed by captured evidence of the new correct behaviour, AND update its paired §1.1 mutation so the mutation breaks the NEW invariant. Discriminator vs bluffing: after reconciliation the gate + mutation still form a valid §1.1 pair (mutate the new invariant → gate FAILs); a bluffed gate's mutation no longer makes it FAIL (the assertion became a tautology). The reconciliation MUST be a visible, evidence-cited change, never a silent assertion-weakening. Honest boundary (§11.4.6): a post-fix gate FAIL is NOT automatically "stale gate, reconcile" — investigate per §11.4.102 first; the FAIL may be the gate correctly catching a REGRESSION the fix introduced, in which case the FIX is wrong, not the gate. Reconcile ONLY when investigation PROVES the gate asserted old-correct-now-removed behaviour AND the new behaviour is the intended, evidence-confirmed mechanism. Composes §11.4.1 / §11.4.4 / §11.4.6 / §11.4.84 / §11.4.102 / §11.4.108 / §1.1. Propagation gate CM-COVENANT-114-120-PROPAGATION (literal 11.4.120) + recommended gate CM-GATE-RECONCILED-NOT-FAKE-PASSED + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.120. Non-compliance is a release blocker. No escape hatch — no --fake-pass-stale-gate, --revert-fix-for-gate, --weaken-assertion-to-pass, --delete-failing-gate flag.
§11.4.121 — No-commit-while-build-writes-tracked-artifacts mandate (1.1.8-dev remediation, 2026-06-03)
A commit (especially git add -A / any broad stage) MUST NOT run while a build/packaging/generation step is actively writing artifacts INTO tracked (version-controlled) directories — doing so races the writer and stages a PARTIAL or stale artifact (a §11.4.108 SOURCE→ARTIFACT integrity failure landed in version control). The commit MUST be deferred until the build step that writes tracked artifacts has COMPLETED, so the tree is quiescent at the artifact layer AND the committed artifacts are the FRESH, whole outputs (no stale pre-rebuild artifact, no half-written file). Before committing tracked build outputs, verify the writing step finished (process exit / completion marker / per-artifact mtime ≥ build-start) — a build still in flight writing tracked dirs is a HOLD on the commit, not a race to win. Build-output analogue of §11.4.84 (no commit while a mutation gate is in flight): both close the "commit captures transient non-final tree state" class at two write-sources. Where build outputs land OUTSIDE version control (gitignored out/ / dist/), the race does not apply. Honest boundary (§11.4.6): "the build probably finished" is not "the build finished" — verify with a completion signal; committing source changes that DON'T touch the build's tracked-artifact directories is fine mid-build. The PROJECT-SPECIFIC instance (which tracked directory + which build steps write it) is recorded in the consuming project's own governance per §11.4.35. Composes §11.4.6 / §11.4.30 / §11.4.58 / §11.4.84 / §11.4.88 / §11.4.96 / §11.4.103 / §11.4.108. Propagation gate CM-COVENANT-114-121-PROPAGATION (literal 11.4.121) + recommended gate CM-NO-COMMIT-DURING-ARTIFACT-WRITE + paired §1.1 mutation (gate-code = separate work item).
Canonical authority: constitution submodule Constitution.md §11.4.121. Non-compliance is a release blocker. No escape hatch — no --commit-during-build, --stage-partial-artifact-OK, --assume-build-finished, --skip-build-completion-check flag.