ycmjason/aibridge

aibridge は何をしますか？

aibridge lets you drive models from other providers: it spawns the AI CLIs installed on this machine and verifies their output. You supply the judgment and prompt-craft; the CLI owns the brittle execution. Delegates bring what you may not have — different training and blind spots (real cross-checking), image rendering, heavyweight reasoning seats — and run concurrently on the backing CLIs' existing logins, no API keys. Each delegation spends that backing CLI's own quota (see Model seats for the one case where that quota is yours).

Running it

aibridge <command> [options]              # if `aibridge` is on PATH
npx -y @aibridge/cli <command> [options]  # zero-install — works everywhere

No installation step is required: the second form fetches the CLI on demand from npm. The reference docs below write every command as aibridge <command> — if aibridge is not on PATH, substitute npx -y @aibridge/cli for aibridge; everything else is identical. Requires Node ≥24.11.

Setup (a one-time probe, not an install)

• Minimum CLI version: 0.1.0.
• Probe: run aibridge --version; if the command is missing, run npx -y @aibridge/cli --version and use the npx form from then on. Compare the printed version against 0.1.0 (semver).
• Below minimum via npx (stale cache): re-run as npx -y @aibridge/cli@latest. Below minimum via a global install: suggest the user update with npm i -g @aibridge/cli.
• If node --version is below 24.11: the CLI cannot run — ask the user to upgrade Node (e.g. nvm install 24 / mise use node@24).
• Optional, for users who want the aibridge command directly on PATH: npm i -g @aibridge/cli. Ask the user before running global installs.

Requires the backing CLIs on PATH and authed: grok (recommended planner/reviewer — pass it explicitly), agy (recommended implementer — pass it explicitly), codex (recommended image-gen seat — pass it explicitly, plus openai-codex/* delegation), and optionally claude for the anthropic-claude/* delegation tier.

Subcommands

Command	Description	Reference
plan	Expand a task into a detailed implementation plan file	reference/plan.md
implement	Implement a plan file in place (edits the working tree, runs real gates)	reference/implement.md
review	Review the working-tree diff (or a plan) against a plan contract	reference/review.md
subagent	Delegate a self-contained task to another model	reference/subagent.md
image-gen	Generate a raster image via a model seat (codex, agy, or grok backend)	reference/image-gen.md
runs	Monitor and inspect execution runs	—
quota	Show backend quota and reset times (grok, agy, codex, claude)	—
models	List every model seat in the registry (slug, efforts, image format)	—

Routing

1. First word is a subcommand name → load that reference file and follow it. This is non-negotiable: the reference holds the prompt-craft that makes the output good — skipping it produces generic results.

2. No subcommand given → infer from the request:

   • an image / icon / graphic to make → image-gen;
   • a clearly-specified, self-contained task to hand to another model, or a cross-model second opinion / red-team → subagent;
   • a sizeable or risky piece of implementation work → the three-verb workflow: plan → you read, edit, and approve the plan file → implement → review --plan <file>. For high-risk designs, add a pre-implementation gate: review --plan on a clean tree reviews the plan itself.

   If genuinely ambiguous, show the table above and ask.

3. Unsure of the current flags? Run aibridge <command> --help — the surface grows over time (new models, new flags).

Model seats (recommended choices — pass them explicitly)

--model is required on every command; nothing is chosen for you. Rough starting points — not benchmarks, just what tends to work:

slug	plan	implement	review	image-gen
xai-grok/grok-4.5	✅ small–mid, well-scoped	✅ any fidelity	✅	○ JPEG
openai-codex/gpt-5.6-sol	✅ mid–big, ambiguous	○	✅	✅ PNG
anthropic-claude/opus-5	✅ mid–big, ambiguous	○	✅	✗
google-antigravity/gemini-3.6-flash	○	✅ needs high–xhigh detail	○	○ JPEG
anthropic-claude/sonnet-5	○	✅ needs high detail	○	✗

✅ recommended · ○ works · ✗ not supported. For plan the qualifier is how much ambiguity the seat absorbs; for implement it is how specified the plan must be.

These are rough recommendations, not measurements. If the task doesn't clearly fit a row — or the user has said how they want work routed — ask the user rather than guessing.

Other registered seats, same rules: openai-codex/gpt-5.6-terra / -luna (cheaper coding tiers), anthropic-claude/fable-5 (hardest, longest-running work), anthropic-claude/haiku-4-5 (quick answers), google-antigravity/gemini-3.1-pro (-high/-low only), and agy's claude-sonnet-4-6 / claude-opus-4-6-thinking / gpt-oss-120b-medium — Claude- and GPT-grade work on Antigravity's quota, which matters when the same-provider seats are the ones you are avoiding. Run aibridge models [--json] for exact per-seat facts (accepted efforts, image format, pinned model ID). aibridge <command> --help lists every seat.

• grok is aggressively capped (~30 req/min, ~1k msgs/day; aibridge quota tracks weekly credit usage): run ONE grok stage at a time. It is the only seat that tolerates a vague brief, so spend it where the input is unclear rather than where a detailed plan already exists.
• Keep the reviewer cross-model from the implementer — a model reviewing its own diff shares its own blind spots. The recommended seats already differ; if you override one seat, check the other. The same logic applies to you: for work you authored yourself, prefer a reviewer outside your own model family.
• Quota is orchestrator-relative. Every backend spends its own CLI's login/quota. A backend that shares the quota of the agent running this skill (anthropic-claude/* for Claude-based agents — it bills the claude CLI's subscription — google-antigravity/* for Antigravity-based agents, …) is a LAST RESORT: say so before using it.
• Swap seats on quota: openai-codex/gpt-5.6-sol[-<effort>] (own ChatGPT login) is the usual alternate for any seat.
• Quota preflight runs automatically before every delegation; aibridge quota is the manual two-second check when planning a pipeline.

Security & trust model

• aibridge performs real delegation by design: in tools mode the backing CLIs read/write files and run shell — at the SAME trust level as the agent running this skill. Nothing here grants more access than the user already granted their agent and those CLIs when they installed and authed them.
• Delegation sends the task prompt (and whatever the delegate chooses to read) to the backing CLI's provider. Do not delegate content the user would not send to that provider; use --no-tools for untrusted input — the delegate gets no file or shell access, reasoning only.
• The CLI is @aibridge/cli on npm, published with SLSA provenance via OIDC from the public repo (https://github.com/ycmjason/aibridge). Its packages have no install-time scripts and a single runtime dependency tree you can audit.
• Nothing runs until the user has installed this skill and the CLI executes via npx/npm under their own account.