The week we shipped the ecosystem

Last week we shipped a VS Code extension. This week we shipped an ecosystem. Same product, richer surface. Here is what actually landed, in order.

The extension launched with four slash commands. That was enough to prove the participant pattern; not enough to be a daily driver. v0.2.0 adds ten more, plus two for the shareable file:

Coding. /debug walks a stack trace to the root cause on T1 Max. /review hands you a staff-engineer code review with findings ranked by blast radius. /optimize flags real performance smells, not micro-optimisations. /simplify rewrites the selection with fewer lines. /test generates high-signal unit tests in the framework the file already imports. /type infers strict TypeScript for untyped JavaScript. /name suggests three better names for the selected symbol. /comment writes doc comments that explain the WHY.

Communication. /commit drafts a Conventional Commits message from the current diff. /pr writes a full pull-request description with a test plan.

Learning. /eli5 explains the selection like the reader is five. /rubberduck flips the participant into question-asking mode; it will not give the answer directly.

Sharing. /roast roasts the selected code with actual wit, then closes with one line of respect. /haiku writes a five-seven-five about what the function does in its soul.

Each command is a tight, opinionated system prompt bound to a specific T1 tier. The prompts were tuned to feel like distinct personalities, not one generic AI voice.

The slash commands cover intent. But some users want a specific persona every message, without having to remember the slash. So v0.3.0 adds three sub-participants alongside @taskclan:

• @taskclan-reviewer — locked to T1 Max, always in staff-engineer code-review mode. Ranks findings by blast radius. Says "no findings" and stops when the code is fine.
• @taskclan-guru — locked to T1 Flow, answers in the manner of a Zen teacher. Brief, oblique, teaches by asking.
• @taskclan-sensei — locked to T1 Flow, Socratic mode. Never gives the answer directly. Gives a straight answer only after "just tell me" three times.

Each is its own registered chat participant, so all four show up in VS Code's participant picker. Users pick the vibe. Slash commands are ignored on the sub-participants; the whole point is that the voice is fixed.

Shipping fourteen commands and three participants means very little if nobody can find them. So we built taskclan.com/agents. Every first-party listing we ship, in one browsable page.

The gallery pulls from the engine's marketplace catalog at build time. Skills, agents, templates, workflows, and prompts each get their own section. Every listing has a detail page with the exact install or invocation command copy-ready.

The catalog is read from live registries. When we register a new Hive agent, add a Studio template, or ship a new slash command, the gallery refreshes within an hour of the deploy. Nothing to manually update, nothing to fall behind.

The whole thing sits on top of infrastructure that was already in the engine. That was the interesting part of building it. We did not need a new agent runtime, a new marketplace, or a new manifest format. It was all there:

• The Hive intent registry and the tool-loop primitives host the agents.
• The marketplace catalog auto-lists every registered agent, workflow, prompt, and template as an installable listing. We just added skill as a new kind and populated it with the 14 slash commands.
• The Studio App Spec IR handles the templates.
• The VS Code extension became the sixth first-party surface (Cursor, Continue, Zed, Claude Code, MCP, VS Code) that speaks T1.

The work was mostly content. Prompts tuned for personality. Tier bindings that reflect the actual task. Descriptions written for browsable reading, not marketing copy.

Three tracks queued.

Studio Template Pack. Six AI-app blueprints in the Studio gallery: support triage, content moderation, meeting notes to action items, lead scoring, personal finance categorisation, recipe generation. Each a full App Spec with data models, screens, and agent binding. Clone one and ship.

Flagship hosted agents. Three public demos at demo.taskclan.com: a codebase archaeologist that answers questions about any public repo, a hosted support ticket triager, and a landing-page critic. Anyone can try each without signing up; ten free invocations per day per IP.

Continue.dev upstream PR. Add taskclan as a first-class provider in Continue's source so every Continue user sees T1 as a peer of openai, anthropic, gemini. Community-native, benefits everyone.

Install the extension: code --install-extension taskclan.taskclan-intelligence. Browse the ecosystem: taskclan.com/agents. Start building: the quickstart.