V5.9.2 retains the V5.9.1 feasibility/planning-compiler/reverification hardening and adds hierarchical capability-class model routing: a FRONTIER main orchestrator remains spawn-only, bounded coding defaults to EFFICIENT workers, complex work/review promotes to BALANCED, and critical review/debug escalation uses a fresh FRONTIER agent. Strategic research drift still triggers scoped replanning instead of silent plan deviation.
The canonical library remains exactly 154 skills. V5.9.2 adds no new skills; it makes existing job contracts cheaper to execute through FRONTIER | BALANCED | EFFICIENT routing. For OpenAI GPT-5.6 the reference mapping is Sol / Terra / Luna at high, and Sol is never requested above high.
An evidence-driven software planning and autonomous-delivery system for modern coding agents.
CtrlAltDelegate now distinguishes implementation progress from product completion. Mandatory user journeys and live integrations can require real runtime/provider/browser/native evidence; mocks and scaffolds remain useful feedback but cannot close those claims. Verification-only blockers are deferred while dependency-ready implementation continues. An upstream job can remain IMPLEMENTED_UNVERIFIED and still unblock normal implementation dependencies; final DONE/COMPLETED still waits for real proof. This lets the coding agent build as far as safely possible before asking for human/device/credential validation.
CtrlAltDelegate turns a software idea, existing codebase, migration, audit, or infrastructure objective into an implementation-ready engineering plan, then hands that plan to a capable coding agent for autonomous implementation and verification.
The recommended workflow deliberately separates expensive coding-agent execution from deep planning:
IDEA / EXISTING PROJECT
↓
CTRLALTDELEGATE CUSTOM GPT
↓
DISCOVERY + RECURRING RESEARCH
↓
REQUIREMENTS + CRITICAL FEASIBILITY GATE
↓
STACK + ARCHITECTURE + PROGRAM DESIGN
↓
PLANNING COMPILER + PLAN CHECKER
↓
ZERO-CONTEXT JOB CONTRACTS + VERIFICATION GATES
↓
ctrlaltdelegate-delivery.zip
↓
CODING AGENT
↓
IMPLEMENT → VERIFY → REPAIR → CONVERGE
↓
COMPLETED
Planning in the Custom GPT does not consume tokens, credits, or API budget from the later coding-agent account. This makes it practical to use a strong ChatGPT model and high available reasoning/thinking settings for discovery, research, architecture, content, and execution design before Codex, Command Code, Claude Code, DeepSeek Harness, OpenCode, or another coding environment starts spending its own allowance.
CtrlAltDelegate is not a fixed prompt or a predetermined technology stack. It is a persistent engineering methodology with specialist skills, planning state, capability-aware stack selection, safe handoff, execution control, and evidence-based completion.
- Open the CtrlAltDelegate Custom GPT and describe what you want to build, change, migrate, audit, or repair.
- Work through focused discovery rounds. The planner does not force one giant questionnaire and does not repeat facts already resolved.
- CtrlAltDelegate researches drift-sensitive technical choices, maps product requirements to capabilities, selects an appropriate stack, designs architecture and program boundaries, and creates a right-sized execution/verification plan.
- The planner exports one
ctrlaltdelegate-delivery.zipcontaining a private.ctrlaltdelegate/control package and an exact coding-agent start prompt. - Copy the ZIP into the target repository root. Do not manually extract it.
- Open the project with a supported coding agent and paste the generated start prompt.
- The agent validates and safely imports the package, preserves existing project work, then implements toward
COMPLETEDwith verification and recovery loops.
By default the imported control plane is LOCAL_PRIVATE and remains outside the application's Git history:
PROJECT_ROOT/
├── .ctrlaltdelegate/ # planning, skills, contracts, runtime state
├── src/ # product code stays outside the control root
├── tests/
└── ...
A GitHub-native edition is also provided for users who prefer to run discovery, planning, and execution directly inside a coding-agent repository without the Custom GPT front end.
CtrlAltDelegate is project-type neutral. Representative workloads include:
- Websites & web applications — marketing/content sites, SaaS, B2B SaaS, dashboards, portals, marketplaces, authentication-heavy and real-time applications.
- APIs & backend systems — REST/OpenAPI, GraphQL, gRPC/Protobuf, AsyncAPI/event contracts, webhooks, background jobs, distributed services and integrations.
- Mobile & desktop — native iOS/Android, React Native/Expo, Flutter, Tauri/Electron/Qt/Wails/Avalonia and local-first applications.
- Data & analytics — transactional databases, embedded/local data, search, vector retrieval, OLAP, time series, ETL/ELT and migrations.
- AI systems — LLM applications, RAG, tool-using agents, MCP, local/model serving, evaluation, memory and retrieval architectures.
- Realtime & collaboration — WebSockets/SSE, presence, synchronization, CRDT/collaborative editing and event-driven workflows.
- CMS & commerce — content platforms, headless CMS, self-hosted commerce, payment/integration workflows and SEO/SXO planning.
- IoT & telemetry — MQTT, device/telemetry platforms, time-series systems, LoRaWAN/industrial protocol integration.
- Infrastructure & networking — CI/CD, containers, Kubernetes, Terraform, Cloudflare, reverse proxies, VPN/network automation, UniFi, OPNsense and OpenWrt.
- Brownfield work — features, debugging, refactoring, security hardening, upgrades, migrations, audits, modernization and release engineering.
Project profiles route expertise; they do not restrict what the system can plan.
V5.9 retains the V5.8 capability-driven approach and selects technology from required capabilities rather than from a fashionable default stack:
REQUIREMENTS
↓
CAPABILITY MODEL
↓
EXISTING INFRASTRUCTURE / BROWNFIELD CONSTRAINTS
↓
SMALL CREDIBLE CANDIDATE SET
↓
CURRENT AUTHORITATIVE RESEARCH
↓
COMPATIBILITY + SECURITY + OPERATIONS + COST + LOCK-IN
↓
SOLUTION MINIMIZATION
↓
STACK DECISION + REJECTED ALTERNATIVES
Self-hostable, open and portable solutions receive preference when they are technically and operationally suitable, but SaaS is not rejected merely for being SaaS. The planner considers exit paths, existing infrastructure, maintenance burden, ecosystem maturity and recurring cost.
The catalog covers capability families rather than forcing individual products: frontend/rendering, API contracts and gateways, backend frameworks and backend platforms, authentication/IAM, relational/document/analytical/time-series databases, cache, object storage, search/vector, messaging/streams/durable workflows, realtime/collaboration, mobile/desktop, CMS/commerce, IoT, observability, AI/model serving and deployment.
This lets a capability bundle such as AUTH + DATABASE + STORAGE + REALTIME legitimately select a compact platform such as Supabase when it beats a fragmented multi-service design, while a project with enterprise SSO, existing object storage, or unusual scaling constraints can select different components.
CtrlAltDelegate contains a broad canonical specialist library, but does not preload the library.
PROJECT SIGNALS
→ CANDIDATE PLANNING SKILLS
→ LOAD SMALLEST RELEVANT DECISION SURFACES
→ APPLY THEM DURING PLANNING
→ PERSIST DECISIONS
→ REFRESH ROUTING
→ HAND JOB-RELEVANT SKILLS TO EXECUTION
Relevant skills influence discovery questions, research, requirements, architecture, UI/UX, accessibility, SEO/SXO, content, security, program design and verification before coding begins. Execution workers then receive only the exact subset needed for their jobs.
For public websites the planning system can also produce final approved page copy, design direction, design-system decisions, search-intent mapping, structured-data planning and verification criteria. Approved planning content is authoritative to the coding agent unless a scoped evidence-based change is required.
The coding-agent runtime distinguishes "a tool is installed" from "the required capability is verified".
For a required capability it follows approximately:
NATIVE / EXISTING VERIFIED CAPABILITY
→ EXISTING PROJECT TOOL
→ CURRENT PROVIDER RESEARCH
→ SAFE PROJECT-LOCAL BOOTSTRAP IF NEEDED
→ RELOAD/REGISTER
→ SMOKE TEST
→ CAPABILITY STATE + TOOL LOCK
V5.9 includes provider-neutral tooling policies and first-class candidate mappings for:
- CRW / fastCRW — preferred candidate for web search, scrape, map, crawl and structured extraction.
- Obscura — lightweight interactive agent browser for JavaScript, DOM interaction, sessions, forms, screenshots/PDF and MCP.
- Playwright / Playwright MCP — reference path for real-browser interaction and production browser/UI acceptance.
Equivalent capabilities already supplied by the active harness or project are reused instead of installing duplicates. Safe automatic installs stay project-local, avoid sudo and global configuration, record source/version/license/hash evidence, and fail closed when credentials, paid services, system-wide changes or security-policy exceptions are required.
The execution system keeps process depth proportional to the work while preserving engineering quality. Core mechanisms include:
MICRO | SMALL | STANDARD | HIGH_RISKexecution rightsizing;- independent
NORMAL | ELEVATED | HIGH | CRITICALassurance profiles; - brownfield-first architecture preservation and a Solution Minimization Gate;
- program-design and vertical-slice-first implementation for substantive changes;
- machine-readable dependency/job state and persistent continuation state;
- behavior-first regression oracles and root-cause-depth checks for bugs;
- blind independent verification for material claims;
- hash-bound worker briefs and capability attestations;
- protected/editable/append-only/human-controlled surface policy;
- anti-thrashing progress signatures and scoped repair/invalidation;
- checkpoint/resume and recovery after context loss, provider limits or crashes;
- requirement → job → code → evidence → documentation convergence;
- runtime, migration, browser, network and domain-appropriate acceptance evidence.
A worker saying "done" is a claim. Completion requires evidence appropriate to the claim.
V5.9 adds a deterministic runtime skill-escalation ladder. A worker that discovers missing expertise can request one progressive reference or one canonical skill without forcing a whole-project replan when the job's objective, requirements, acceptance, scope and authority are unchanged. Larger semantic changes escalate to job rebrief or normal scoped change control. The original worker brief remains immutable and every grant is SHA-bound.
Skill maintenance is also usage-aware: all 154 skills keep structural QA and remain selectable, while expensive research/behavior campaigns can focus on frequently used or safety-critical skills using explicit local/export history. CtrlAltDelegate adds no network telemetry for this.
The preferred execution topology is frontier lead + cheaper fresh workers + independent reviewers:
FRONTIER ORCHESTRATOR
→ EFFICIENT default coding workers
→ BALANCED complex workers / semantic reviewers
→ FRONTIER critical fresh reviewer / final debugger escalation
→ parent re-verification → integrate → next wave
The orchestrator normally does not write product code when suitable delegation exists. A worker cannot promote itself; routing changes create a fresh attempt and preserve failure evidence. The main orchestrator can reject/adjudicate work but cannot satisfy independent-review gates. See docs/system/MODEL-ROUTING-AND-HIERARCHICAL-ORCHESTRATION.md.
CtrlAltDelegate keeps one canonical methodology and .agents/skills library while adapting to the capabilities of the active harness.
| Harness | V5.9 support |
|---|---|
| OpenAI Codex CLI | FIRST_CLASS |
| Command Code | FIRST_CLASS_PREVIEW — .agents/skills, AGENTS.md, subagents, headless/resume, tasks, MCP, hooks, permissions and worktrees are mapped; promotion requires the runtime conformance lane |
| DeepSeek Harness | FIRST_CLASS_PREVIEW while its upstream interface remains pre-stable |
| Pi | REFERENCE harness |
| Claude Code | COMPATIBLE with CLAUDE.md and thin skill adapters |
| OpenCode | COMPATIBLE with capability detection |
V5.9.2 routes abstract model classes rather than blindly using one model everywhere. FRONTIER owns orchestration/critical judgment, BALANCED owns complex implementation/semantic review, and EFFICIENT is the default bounded implementation/research/validation worker. Harnesses resolve the classes when per-subagent model selection exists; otherwise they inherit the active model while preserving role separation. The OpenAI GPT-5.6 reference mapping is Sol / Terra / Luna, all at high, with a hard methodology ceiling that Sol is never requested above high.
The GitHub-native package can run the complete lifecycle directly in a coding-agent repository:
DISCOVERY → RESEARCH/FEASIBILITY → STACK → ARCHITECTURE → PROGRAM DESIGN
→ PLANNING COMPILER → PLAN CHECKER → JOB/INTEGRATION GRAPH
→ IMPLEMENTATION → PARENT REVERIFY → VERIFICATION → GIT/GITHUB → COMPLETED
It includes AGENTS.md, CLAUDE.md, canonical .agents/skills, planning templates/state, harness contracts, validation scripts and release infrastructure.
Every release also contains release-handoff/UPDATE-PUBLIC-GITHUB-REPO.md. Give the GitHub-native release ZIP and that prompt to a capable coding agent in the public CtrlAltDelegate repository; it will extract the release safely, merge only release-owned changes, preserve repo-specific assets, run QA, commit and push the update.
Core release/runtime checks include:
python3 scripts/validate_system.py
python3 scripts/validate_skill_evals.py
python3 scripts/validate_control_plane.py
python3 scripts/validate_assurance_control.py
python3 scripts/validate_v58_architecture.py
python3 scripts/harness_preflight.py --jsonREADME.md, the harness matrix, manifests and release handoff are release-controlled artifacts and are validated for consistency.
.
├── AGENTS.md
├── CLAUDE.md
├── README.md
├── START-HERE.md
├── .agents/skills/ # canonical specialist library
├── .claude/skills/ # thin adapters
├── adapters/command-code/ # V5.9 harness mapping
├── config/ # routing, technology, tooling, assurance
├── docs/system/ # detailed engineering contracts
├── docs/templates/
├── evals/
├── planning/ # GitHub-native persistent planning state
├── scripts/
├── delivery-template/
└── release-handoff/
The Custom-GPT delivery uses the same core contracts under a private .ctrlaltdelegate/ control root instead of requiring the project's application code to adopt this repository layout.
CtrlAltDelegate is independently authored. External repositories, standards and vendor documentation are research inputs used to identify durable mechanisms and current facts; their prompt bodies and runtimes are not vendored into the methodology. Drift-sensitive facts are rechecked against current first-party authority when they materially affect a project decision.
See CREDITS-AND-PROVENANCE.md and .agents/skills/SOURCE-RESEARCH-MATRIX.yaml for the maintained research record.