dabbler-ai-router 0.21.1

Multi-provider model routing, prompt templates, session state, and metrics for the Dabbler AI-led-workflow.

Dabbler AI Orchestration

An AI-led coding-session workflow for VS Code. Structured AI sessions with cross-provider verification, automatic cost tracking, git- worktree-aware session-set state, and a Session Set Explorer in the activity bar.

---

What this repo is for

The framework treats AI coding work as a sequence of sessions — bounded slices that run to completion in one orchestrator conversation, end with a verification + commit, and stop. A session set is an ordered chain of sessions that delivers one feature, refactor, or aspect of the solution. Each set lives at docs/session-sets/<slug>/ with a small predictable shape (spec.md, session-state.json, activity-log.json, change-log.md).

Inside each session, the orchestrator (Claude Code, Codex, GitHub Copilot, or Gemini Code Assist) does mechanics — file edits, shell, git — and dispatches every reasoning task (code review, security review, analysis, architecture, documentation, test generation, end-of-session verification) through ai_router.route(). The router picks the cheapest capable model per task type, escalates on poor responses, and runs cross-provider verification by a different provider to catch provider-specific blind spots.

Every routed call is appended to ai_router/router-metrics.jsonl, so per-set, per-task, and per-model spend is fully auditable. The Session Set Explorer extension is the at-a-glance companion: it reads the same files the router writes and renders three groups in the activity bar (In Progress, Not Started, Done), with worktree auto-discovery so parallel sessions surface across sibling workspaces. Full execution mechanics live at docs/ai-led-session-workflow.md; deeper feature descriptions live at docs/repository-reference.md.

---

Highlights

• Session sets and sessions — Work is organized into bounded sessions inside ordered session sets, each with its own folder of artifacts the extension reads to render the activity-bar inventory. Deep dive.
• Cost-minded orchestration — The router routes each task to the cheapest capable tier, escalates on poor responses, and uses a per-task-type effort overrides. Real metrics from contrasting projects show 73% savings vs Opus-only on a CLI/library project (990 calls) and 32% savings on a full-stack UI app with UAT/E2E gates (370 calls) — see docs/sample-reports/ for the full reports. Deep dive.
• Cross-provider verification — Every session ends with a mandatory independent verification by a model from a different provider. The verifier returns structured JSON ({"verdict": "VERIFIED" | "ISSUES_FOUND", "issues": [...]}); the orchestrator surfaces disagreements for human adjudication rather than self-resolving. Deep dive.
• Git integration + parallel session sets — Every session ends with git add -A && git commit && git push. Multiple session sets can run in parallel via isolated git worktrees on session-set/<slug> branches, with the last session merging back into main cleanly. Deep dive.
• Robust fallbacks — Tier escalation on empty/truncated/refused responses; two-attempt verifier fallback when a provider's HTTPS layer fails; documented escalation ladder if both verifier attempts fail. The work is preserved in git for human review either way. Deep dive.
• UAT + E2E support (tri-state, opt-in). Specs declare requiresUAT and requiresE2E as true | false | "suggested". true enforces a UAT checklist + matching Playwright coverage as a close-out gate; false skips both surfaces; "suggested" asks you at session start whether you want E2E tests, UAT checklist, both, or neither, records your choice, and gates close-out accordingly. No-UI repos default to the universal core (build, test, verify, commit) with no UAT/E2E surface area. Deep dive.
• Full and Lightweight tiers. Specs declare tier: full (default) or tier: lightweight. The tier changes one thing only — whether the AI router makes metered API calls. Lightweight is router-off, not Python-off: both tiers use a .venv + dabbler-ai-router, the same session-state.json lifecycle, the same close-out gate, and the same Session Set Explorer. Full adds cost-minded routing and automatic cross-provider verification; Lightweight makes zero metered calls and verifies per-set (copyable review prompts pasted into a different assistant, a dedicated different-engine verification session, or opt out). The single source of truth is docs/concepts/tier-model.md. Deep dive.

---

Quick start

1. Install the extension from the VS Code Marketplace:
   • VS Code → Extensions view (Ctrl+Shift+X) → search Dabbler AI Orchestration → Install.
   • Or from a terminal: code --install-extension DarndestDabbler.dabbler-ai-orchestration.
   • Or directly from the Marketplace listing.
   • Offline / firewall fallback: each tagged release attaches the .vsix as a downloadable asset on the GitHub Releases page; pick the latest, then Extensions → ... → Install from VSIX....
2. Open your workspace. Any folder with — or destined for — a docs/session-sets/ directory. The activity-bar Session Set Explorer icon appears automatically once that path is present.
3. Run Dabbler: Install ai-router from the command palette (Ctrl+Shift+P). The command auto-detects (or offers to create) a workspace .venv/, runs pip install dabbler-ai-router inside it, and materializes ai_router/router-config.yaml for tuning.

Then set API keys as environment variables (one-time): ANTHROPIC_API_KEY, GEMINI_API_KEY, OPENAI_API_KEY — the Prerequisites section below has the sign-up links and notes which providers are required.

Subsequent updates: Dabbler: Update ai-router from the command palette.

CLI fallback — python -m venv .venv && .venv/Scripts/pip install dabbler-ai-router, then from ai_router import route from your orchestrator script.

---

For new projects: the Getting Started form

If you're starting a new project — greenfield, or an existing local project that hasn't yet adopted the workflow — the recommended starting point is Dabbler: Get Started from the command palette. The Session Set Explorer's Getting Started form walks you through tier choice (Full vs. Lightweight — see docs/concepts/tier-model.md), the Full-tier verification budget / NTE step (saved to ai_router/budget.yaml — schema), and a one-click project scaffold: the .venv with the router package, the AI-agent instruction files, and the docs/session-sets/ home. From there the form hands you copyable prompts for drafting docs/planning/project-plan.md and decomposing it into session sets with your AI agent. The four-tier budget mapping is documented in docs/ai-led-session-workflow.md → Cost-budgeted verification modes.

Setting up without VS Code? See the manual-setup note in docs/quick-start.md. (The former conversational "adoption bootstrap" path was retired in extension 0.32.0 once the form gained its budget step; docs/adoption-bootstrap.md remains as a redirect stub for older clients.)

---

Prerequisites: tools and accounts

You need VS Code, at least one orchestrator agent installed as a VS Code extension, and API-key accounts for all three model providers (the router calls all three so cross-provider verification has somewhere to route to).

VS Code

• Download: code.visualstudio.com
• Getting-started docs: code.visualstudio.com/docs — the Extensions view (Ctrl+Shift+X) is what you'll use to install the Session Set Explorer in the Quick start above.

Orchestrator agents (install at least one)

Pick whichever AI agent you want to drive sessions; the framework is provider-agnostic and you can switch mid-set.

• Claude Code (Anthropic) — reads CLAUDE.md. Install via claude.com/product/claude-code; docs at docs.claude.com/en/docs/claude-code/overview.
• Codex (OpenAI) — reads AGENTS.md. See openai.com/codex and the open-source CLI repo at github.com/openai/codex.
• GitHub Copilot — reads AGENTS.md. See github.com/features/copilot; Marketplace listing at GitHub.copilot.
• Gemini Code Assist (Google) — reads GEMINI.md. See codeassist.google (free tier available); docs at cloud.google.com/gemini/docs/codeassist/overview.

API keys (all three required)

The router calls all three providers and cross-provider verification needs at least two providers live to be meaningful. Expect to set up all three.

• ANTHROPIC_API_KEY — console.anthropic.com (Settings → API Keys, requires billing).
• GEMINI_API_KEY — aistudio.google.com (Get API key in the left rail; free tier is generous).
• OPENAI_API_KEY — platform.openai.com (create a project, add a payment method, mint a key).

Set each as a Windows User environment variable; macOS / Linux export them in your shell profile. Optionally, pushover.net's PUSHOVER_API_KEY and PUSHOVER_USER_KEY enable end-of-session phone notifications — if unset, the orchestrator skips the notify and prints to console as usual.

---

More

For technical reference (deep feature descriptions, the UAT/E2E flag matrix, a worked end-of-session output example, and the repository file map), see docs/repository-reference.md.

For runtime mechanics (trigger phrases, the 10-step procedure, the authoritative rule list every orchestrator obeys), see docs/ai-led-session-workflow.md.

For sample manager-report output from real projects at scale, see docs/sample-reports/.

For worked examples of cross-provider AI consultation in practice — what each provider explored, where they agreed and meaningfully differed, and what makes the pattern worth using — see docs/case-studies/.

---

License

This repo is released under the MIT License. See LICENSE for the full text. Copyright © 2026 darndestdabbler.

A duplicate LICENSE lives at tools/dabbler-ai-orchestration/LICENSE alongside the extension's package.json. The duplication is required: vsce package expects the file beside the manifest and has no flag to point elsewhere. Both files must be kept in sync.