Add Claude Code statusline estimating smart-routing savings (client side)

[JJ27]

Summary

Adds a Claude Code statusline row for smart routing: it shows whether routing is on/off and an estimate of what routing saved this session, priced from the AI Gateway pay-per-token (PPT) endpoint. Default-on (ENABLE_SMART_ROUTING_SAVINGS=0 opts out).

The row, by state:

💰 Est. saved with smart routing: $0.18 (48%) · smart router plugin v0.4.6 · /smart-router savings for details
Smart routing on · smart router plugin v0.4.6
Smart routing off · smart router plugin v0.4.6

• The 💰 estimate appears once a subagent has run on a cheaper model; on until there's anything to show; off when routing is disabled. A cost increase reads Smart routing cost $X more (Y%).
• The version is the installed model-orchestrator plugin (shown when present, omitted otherwise). /smart-router savings — @lilly's upcoming skill — will give the granular breakdown.

Screenshots

Smart routing on, before any prompt has run:

Estimated savings after a routed subagent:

How it works

• Estimate: every token the session used (main agent + subagents) is priced at the baseline main model vs the model that actually served it — saved = Σ tokens × price(baseline) − Σ tokens × price(served) — summed per response and per token class (input, output, cache read, 5-minute and 1-hour cache writes). Computed from the Claude Code transcripts (main + <session>/subagents/agent-*.jsonl), since cost.total_cost_usd excludes subagents.
• Prices are fetched at launch and cached per workspace; the row reads the cache (a statusline refresh can't wait on the network). When prices are missing or any response can't be priced, the row shows on rather than undercount.
• Wraps any existing statusline (runs it first, prints the row below), is stdlib-only for fast startup, reads transcripts incrementally, and refreshes on main-agent turns.

Price source

GET /api/ai-gateway/v2/endpoint-rates:batchGet with repeated databricks_hosted_model_services query params. The client reads the response's USD cost group — per-million-token token_costs for input, output, and cache (5m/1h writes and reads) — and caches it per workspace (dollar rates depend on the org's DBU→$ conversion; models without it stay unpriced rather than shown in DBUs). Served ids are matched back to system.ai names (incl. Bedrock-style ids like anthropic.claude-haiku-4-5-…). Validated end-to-end against the staging endpoint.

Scope & notes

• Claude Code only for now; the pricing/rate-fetch layer is agent-agnostic, so a Codex surface is a straightforward follow-up.
• Cache rates are priced per class (they're ~87% of a session's cost, and fixed multipliers don't hold across models — e.g. Opus 5.5 cache reads bill at 0.05× input, Opus 4.8's at 0.1×).
• The client sends no LiteSwap header; once the endpoint is live on a workspace the row populates automatically.
• Refresh fires on main-agent turns (not during a long subagent); the router offering GLM/Kimi is rejected by the test workspace's Anthropic endpoint (pre-existing); a router fallback to a pricier tier is shown honestly as a cost increase.

Testing

uv run pytest tests/test_smart_routing_pricing.py tests/test_claude_statusline.py tests/test_claude_smart_routing_v2.py tests/test_databricks.py tests/test_agent_claude.py covers the cost math per token class, the off/on/estimate states, transcript accounting (split records, incremental reads, rewrites), the wrapped command through real sh/bash/zsh, the stdlib-only import guard, the endpoint request/response mapping, and launch wiring (routed + vanilla-off). Full unit suite is green; ruff/ty clean. End-to-end verified live against the staging PPT endpoint.

Also in this PR: tracing CI fix

The Claude and Codex tracing journeys (test_ug_*_exports_trace_to_configured_table) were failing with SQL statement did not finish within 180 seconds because tests/integration/utils/sql.py picked the first warehouse when none was running and counted warehouse startup against the query's 180s budget. Fix: prefer a serverless warehouse, allow up to 10 minutes in PENDING (warehouse startup) before the 180s execution limit, and report which phase timed out. Can move to its own PR if reviewers prefer.

This pull request and its description were written by Isaac.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.