AI harnesses and model routing

This template repository is pre-configured to work seamlessly across three AI development environments:

1. Claude Desktop and Claude Code (CLAUDE.md and AGENTS.md)
2. ChatGPT and OpenAI Codex Desktop and command-line tool (AGENTS.md)
3. Google Antigravity Desktop (Antigravity 2.0) and Antigravity IDE (GEMINI.md, .agents/rules/routing.md, and AGENTS.md)

How each harness loads instructions

Harness	Discovery mechanism	Primary files read
Claude Desktop and Claude Code	Reads CLAUDE.md at the repository root. Line 1 of CLAUDE.md contains @read AGENTS.md (and @AGENTS.md) so Claude loads shared engineering rules before applying its subagent routing policy.	CLAUDE.md, AGENTS.md
ChatGPT and OpenAI Codex Desktop	Walks up from the working directory to the repository root and loads AGENTS.md.	AGENTS.md
Google Antigravity Desktop and IDE	Walks up to the repository root to load GEMINI.md and AGENTS.md, and scans .agents/rules/*.md for workspace-scoped rules.	GEMINI.md, AGENTS.md, .agents/rules/routing.md

Unified 4-tier model routing system

Across all three harnesses, keep the human-selected main-session model unchanged for top-level orchestration, judgment calls, and final synthesis. When spawning subagents for parallel exploration, implementation, or verification, route by failure impact, state interaction, and ambiguity rather than diff line count:

Tier	Workload profile	Google Antigravity (GEMINI.md)	Claude Desktop and Code (CLAUDE.md)	ChatGPT and Codex Desktop (AGENTS.md)
Tier 1 (bulk discovery and grep fan-out)	Per-file classification, grep-and-report symbol lookups, file inventory, and read-only formatting or configuration checks.	Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite, flash_lite) at High effort	Haiku 5.5 (haiku-5.5, tier alias haiku) at High effort	gpt-6-luna at high effort
Tier 2 (mechanical and bounded coding)	Exploration fan-out, mechanical migrations, log/PR summaries, unit test scaffolding, bounded implementation, and Vale docs updates.	Gemini 3.8 Flash (gemini-3.8-flash, flash) at Medium effort	Sonnet 5.5 (sonnet-5.5, tier alias sonnet) at Medium effort	gpt-6.1-sol at medium effort
Tier 3 (substantive and cross-module code)	Implementing features with a decided design, cross-module debugging across lib.rs and state.rs, KV lease reasoning, review finder passes, and nontrivial tests.	Gemini 3.8 Flash (gemini-3.8-flash, flash) at High effort	Sonnet 5.5 (sonnet-5.5, tier alias opus / sonnet) at High effort	gpt-6.1-sol at high effort
Tier 4 (heavy complexity and judge stages)	Architecture/decomposition decisions, KV fingerprint and version contracts, HMAC-SHA256 security triage, race debugging, and workflow verify/judge/synthesize stages.	Gemini 4 Argon (gemini-4-argon, pro / inherit) at High effort	Opus 5.5 (opus-5.5, tier alias fable / opus) at High effort	gpt-6-astra at high effort

Claude routing rules

CLAUDE.md begins with @read AGENTS.md and defines the Claude-specific subagent routing policy:

• Workflow stages: route finders and sweepers to sonnet (sonnet-5.5 at medium effort, or high / opus-5.5 when the code is subtle). Route verify, judge, and synthesize stages to fable (opus-5.5 at high effort). Never run a stage with more than 5 concurrent agents on fable (opus-5.5) unless it is a verify or judge stage.
• Prefer agents as the routing mechanism: because the session model cannot switch mid-run, spawn a subagent with an explicit model and effort whenever a unit of work is self-contained enough to hand off (exploration, migrations, review passes, per-file sweeps, or verification). Keep inline only work that depends on accumulated conversation context or edits that are too small to justify a handoff.
• Mismatch reporting: if the active session model is mismatched to the task tier or an environment variable such as CLAUDE_CODE_EFFORT_LEVEL overrides per-launch parameters, state the mismatch and suggest /model. Never proceed silently on a noticed mismatch.

Antigravity routing rules

GEMINI.md and .agents/rules/routing.md configure subagent orchestration and workspace isolation for Antigravity:

• Subagent invocation (invoke_subagent):
• Pass Model: "pro" (or "inherit" when the parent session runs Gemini 4 Argon at High effort) for Tier 4 architecture, security, and verification tasks.
• Pass Model: "flash" for Tier 2 and Tier 3 implementation, refactoring, and test tasks (Gemini 3.8 Flash).
• Pass Model: "flash_lite" for Tier 1 read-only symbol lookup and grep fan-out (Gemini 3.5 Flash-Lite).
• Workspace isolation (Workspace):
• Use Workspace: "inherit" for read-only research subagents.
• Use Workspace: "share" for concurrent self writing subagents so each writer works in an isolated working tree backed by the shared Git object store.
• Use Workspace: "branch" for sandboxed experimental branches.
• Reactive wakeup: the Antigravity runtime wakes the parent agent automatically when a subagent sends a message or a background task finishes. Never poll subagent status in a loop.