docs: propose advanced AI workflow extensions roadmap

[Jarvis2021]

Summary

• add a focused RFC-style proposal for advanced, opt-in AI workflow extensions in Spec Kit
• document major gaps where optional community-driven capabilities could complement the core spec -> plan -> tasks -> implement flow
• link to a working open-source reference implementation for maintainers and community reviewers

Why this PR

Spec Kit already has strong foundations in its core workflow and extension-system direction, but there is not yet a single maintainers-friendly proposal that frames advanced AI-native workflows as optional, extension-oriented additions rather than core complexity.

This PR proposes a phased roadmap for:

• AI-assisted debugging with typed evidence capture
• guided self-healing loops after failed implementation or test runs
• multi-agent orchestration patterns for bounded parallel work
• workflow packs and plugin bundles
• resumable collaboration metadata and lightweight workflow state
• lightweight semantic code operations as optional higher-confidence helpers

Scope

This is intentionally a docs-first proposal.

It does not:

• change the Spec Kit CLI behavior
• add new required runtime services
• introduce vendor-specific orchestration into core
• force any new workflows onto existing users

Files changed

• extensions/RFC-ADVANCED-AI-WORKFLOWS.md
• README.md
• docs/index.md

Reference implementation

This proposal links to Jarvis2021/agentic-sdlc-development as a working open-source reference for maintainers who want concrete examples of:

• runtime-backed planning and resume flows
• typed debugging evidence capture
• plugin-capability packs
• resumable workflow state

Validation

• docs were aligned against the current README.md, CONTRIBUTING.md, extension RFC, and command template structure in spec-kit
• no CLI or runtime behavior changed in this PR
• no automated tests were run because this is a documentation-only proposal

AI assistance disclosure

This PR was created with AI assistance for repo analysis, drafting, and editing. The proposal scope, rationale, and final content were reviewed and directed by a human before submission.

[mbachorik]

Interesting. I'll try to take a look over the weekend.

[mnriem]

Thanks for the thoughtful write-up — it's clear you've spent real time studying how Spec Kit works. The good news is that the capabilities you're describing can already be built today using the existing extension, preset, and hook infrastructure — and in several cases the community has already built extensions that explore exactly these areas.

Here's how each area maps:

Debugging / evidence capture & self-healing — The extension system's after_implement hook (see RFC-EXTENSION-SYSTEM.md and implement.md) lets you register post-implementation verification. Community extensions already exploring this space include:

• cleanup — post-implementation quality gate that reviews changes and fixes small issues
• verify — validates implemented code against specification artifacts
• verify-tasks — detects phantom completions (tasks marked done with no real implementation)
• reconcile — reconciles implementation drift by updating spec, plan, and tasks
• cognitive-squad — multi-agent system with quality gates, backpropagation verification, and self-healing

Resumable workflow state — Spec Kit is filesystem-first: specs, plans, tasks, and constitutions are all plain files in .specify/. Community extensions that build on this:

• speckit-utils — resume interrupted workflows, validate project health, and verify spec-to-task traceability
• status — shows current workflow progress, active feature, artifact status, and task completion

Multi-agent orchestration — The README documents "Creative Exploration" as a development phase for parallel implementations. Extensions already tackling this:

• fleet — orchestrates a full feature lifecycle with human-in-the-loop gates across all phases
• conduct — executes phases via sub-agent delegation to reduce context pollution
• ralph — autonomous implementation loop using AI agent CLI

Workflow packs — This is the presets system: composable, stackable bundles of templates, commands, and customizations with priority ordering.

Review & drift analysis — Also well-covered by existing extensions:

• review — post-implementation code review with specialized agents (quality, tests, error handling, type design, simplification)
• retrospective — spec adherence scoring, drift analysis, and human-gated spec updates
• sync — detects and resolves drift between specs and implementation with AI-assisted resolution
• docguard — validates, scores, and traces documentation with automated checks and spec-kit hooks

Semantic code operations — Symbol search, usage lookup, and rename preview are IDE/agent-level capabilities that each supported agent already has access to through its host environment.

Regarding the phased roadmap — the community has already been organically building these capabilities as independent extensions without needing a formal plan in core docs. That's the extension system working as intended.

If you're interested in contributing, the most impactful path would be to build one of these ideas as a concrete extension — the Extension Development Guide has everything you need to get started

[Jarvis2021]

Thanks @mnriem for getting back. Sure Im looking forward to contribute.