From Extension to Plugin: Why It Matters
Spec-Driven Development (SDD) has a simple premise: your architecture should live in version-controlled markdown, not in ephemeral chat logs. Conductor proved this thesis over the past year by helping developers write spec.md and plan.md files before touching production code.
The catch? Conductor was locked to a single host — the Gemini CLI extension. That meant your AI-assisted workflow died the moment you switched tools.
The new Conductor Plugin breaks that lock. Plugins bundle skills, rules, MCP servers, and hooks into one portable package. The result: your SDD workflow becomes ecosystem-wide instead of tool-specific.
The evidence for this shift comes from Google's own engineering blog — see the original announcement for the official roadmap.
What Actually Changed Under the Hood
The biggest shift isn't packaging — it's interaction model. Extensions typically force you into rigid command sequences (/spec, /plan, /next). Plugins can be conversational.
Here's the practical difference:
# Old flow: rigid commands
conductor /spec "add OAuth login"
conductor /plan
conductor /task complete 3
# New flow: just talk to your agent
# "I want to add OAuth login with Google and GitHub providers"
# → Conductor generates spec.md, updates plan.md, tracks task state
Under the hood, the plugin uses hooks to detect when a spec needs regeneration and MCP servers to expose project state as a queryable resource. Your markdown artifacts (spec.md, plan.md) remain the source of truth — they're just now maintained by an agent that reads the conversation instead of waiting for slash commands.
Installing on Antigravity CLI
# One-line install from the official repo
agy plugins install https://github.com/gemini-cli-extensions/conductor
Once installed, your spec.md and plan.md files persist across sessions, tools, and even collaborators. Start a feature in Antigravity CLI, finish it in Gemini CLI — the context carries over.
The Trade-offs You Should Know Before Adopting
Portability sounds like a pure win, but there are real costs.
| Aspect | Extension (old) | Plugin (new) |
|---|---|---|
| Tool lock-in | High (Gemini CLI only) | Low (multi-CLI) |
| Command discipline | Enforced by design | Relies on agent behavior |
| Onboarding speed | Steeper (learn commands) | Faster (just chat) |
| Reproducibility | Deterministic | Depends on LLM consistency |
| Debugging | Predictable command traces | Harder to trace agent decisions |
The honest caveat: conversational SDD is more ergonomic, but it's also less deterministic. When your spec is generated from a chat, the same prompt can produce subtly different spec.md files across runs. For solo projects this is fine. For regulated environments or team-wide architectural contracts, you'll want to review every generated artifact before committing.
Also worth flagging: the plugin ecosystem is young. Expect breaking changes in hook APIs and MCP server contracts over the next 6–12 months.
What to Do Next
If you're already running Gemini CLI, migration is a one-liner — install the plugin and your existing spec.md files keep working. If you're on Antigravity CLI or another host, this is your first real opportunity to adopt SDD without committing to Google's terminal.
Concrete next steps:
- Run the install command in a throwaway repo and inspect the generated
spec.md. - Compare the plugin's conversational output against a hand-written spec for the same feature.
- If you're evaluating AI-assisted architecture at scale, pair this with broader patterns like vertical microfrontends on Cloudflare where team autonomy and boundary discipline matter just as much.
- For a data-driven reality check on when simple approaches beat complex ones, read XGBoost vs Logistic Regression on small datasets — the same bias-variance logic applies to choosing between rigid commands and conversational agents.
The mission is sound: make AI development safe, predictable, and architecturally grounded. The plugin is the vehicle. Your job is to keep the specs honest.