Describe the solution you'd like

Make this repo route tasks to the most appropriate model/harness for contributors using CLI agents (Claude Code, Codex, Gemini CLI, others), by adding three thin layers on top of the existing .agents/skills/ and .claude/ infrastructure:

1. A small set of model-specific Claude Code subagents under .claude/agents/ with sharp description fields that route work to the right model:

   • claude-haiku-mechanical — rote rename / format-only / single-test fixup
   • claude-opus-architect — multi-crate refactors, lock-stack reasoning, WarpUI Entity-Handle design work
   • gemini-long-context — whole-spec audits, large diff review, Cargo.lock analysis
   • (Codex is already wired through the existing codex:codex-rescue agent — reuse, don't replace.)

   Equivalent dispatch where other harnesses expose comparable surfaces; document the harness-specific format in each case.

2. A harness-agnostic /route skill under .agents/skills/route/ that, given a task description, recommends a model with reasoning. Each harness reads the same recommendation logic; only the final dispatch step is harness-specific.

3. A pair of Claude Code PreToolUse hooks in .claude/settings.json that warn (don't block) on cost/scope mismatches — Opus on a 1-line edit, Haiku on a 30-file refactor. Disabled by default; opt-in via env var. Mirrored on other harnesses if/when they expose equivalent hook surfaces.

Is your feature request related to a problem?

Contributors using any CLI agent to develop Warp can't easily match tasks to the right model. Mechanical edits get sent to expensive models; deep architectural work like the TerminalModel lock-stack rules in WARP.md gets sent to fast/cheap ones. The harness has no signal about which model fits a task, so contributors either pick wrong or babysit. This wastes money on trivial work and produces worse results on hard work.

The repo already invests in agent-readable context (.agents/skills/, specs/, WARP.md) explicitly so any harness can pick it up. This proposal extends that investment with the missing piece: routing.

Additional context

Open question for the maintainers before spec work begins: should these artifacts live under .claude/ (Claude-Code-specific) or be promoted to harness-agnostic patterns under .agents/? The skill (#2) is naturally agnostic; the subagents (#1) and hooks (#3) are Claude Code primitives. Happy to scope either way — please flag preference on this issue.

Out of scope (separate issues to follow):

• An MCP server exposing Warp's terminal blocks / AI command outputs / Drive objects to any external agent (Tier 2)
• Ensemble / parallel-model dispatch for hard correctness work (Tier 3)
• Reducing Warp's runtime memory footprint (separate, performance track)

Operating system (OS): macOS (development host; the artifacts themselves are platform-agnostic)

How important is this feature to you? 3