Automated Adapter Builder

The repo ships a contributor tool that automates the entire adapter-creation workflow described in the Adapter Creation Guide: give it a HuggingFace architecture class name, and a pair of Claude Code agents analyzes the HF source, plans the adapter, implements it phase-by-phase with independent code review at every checkpoint, and verifies it against real models.

It lives at devtools/adapter_builder/ — it is developer tooling, not part of the shipped transformer_lens package. This page is an orientation; the authoritative reference is the tool’s own README and the docs beside it.

When to use it

• Manual path: for a single adapter you want to write yourself, follow the Adapter Creation Guide (Claude Code users: the /add-model-support slash command walks the same checklist).

• Automated path: for batch adapter work, or when you want an autonomously built and reviewed first draft, use the builder. The output is left uncommitted in an isolated git worktree for human review — the agents cannot commit, push, or open PRs.

How it works, briefly

The builder launches a Programmer and a Reviewer agent that work in a dedicated git worktree of the repo (your main working tree is never touched). Every lifecycle step is gated: the architecture brief is fact-checked against the HF source before planning, each implementation phase is code-reviewed before the next begins, and verification requires full HuggingFace output parity — model targets are chosen for config diversity (always including the architecture’s smallest model, verified first) so numerical edge cases fail in seconds.

Two coordination modes exist: agent-teams (an orchestrator session; requires Claude Code Max) and solo (two independent sessions coordinated by a signal-routing daemon; works on any tier).

Critical rules — no git writes, no oversized or structural-only verification runs, mandatory mypy/format gates — are enforced by Claude Code hooks at the framework level, not just by prompt instructions.

Launching

From the repo root:

# Agent-teams mode (Max tier)
devtools/adapter_builder/agents/launch-agent-pair.sh --architecture CohereForCausalLM

# Solo mode (any tier)
devtools/adapter_builder/agents/launch-agent-pair.sh --mode solo --architecture CohereForCausalLM

# Check everything without creating anything
devtools/adapter_builder/agents/launch-agent-pair.sh --architecture CohereForCausalLM --dry-run

When a run completes, the worktree (default: ../worktrees/<branch>/) contains the adapter, its unit tests, registry entries, and a full audit trail (.adapter-workspace/: brief, plan, every review round, verification results, and a structured timeline of every tool call). Review it, then commit and open the PR yourself — registration and testing expectations are the same as the manual path.

Requirements

Claude Code on PATH (Max tier for agent-teams mode; any tier for solo), plus git, jq, tmux, and uv. Verification loads real model weights: budget memory accordingly (--max-memory, default 96GB).