feat(claude): Claude Code status line showing Fireworks savings

[tarunipaleru]

What

Adds a Claude Code status-line command that surfaces live cost savings from routing through Fireworks, closing the gap noted in the README's "Claude Code pricing estimates (important)" section: Claude Code's in-UI cost column uses Anthropic list prices and overstates spend for Fireworks users, and FireConnect can't override that column. This puts the real bill — and the savings vs. paying Anthropic directly — in the status bar instead.

packages/setup-cli/bin/cc-fireworks-savings.mjs

On each render Claude Code pipes a status-line JSON object ({ model, transcript_path, workspace, ... }) to the script on stdin. It then:

1. parses the session transcript for actual per-request token usage (input / output / cache read / cache writes),
2. computes the real Fireworks spend from the existing serverless pricing modules — the same rates fireconnect model list reports,
3. compares it to what the same token mix would cost on the Anthropic-equivalent tier at list prices (Opus for GLM, Sonnet for Kimi, Haiku for DeepSeek, Fable for FireRouter), including Anthropic cache read/write rates,
4. renders a single line:

🔥 Fireworks · $0.28 spent (25 req) vs Anthropic $1.14 saved $0.86 75% · glm-5p2 · fireworks

Why this shape

• Reuses existing modules (pricing.mjs, model-specs.mjs, usage.mjs, incumbent-detect.mjs) rather than hardcoding rates, so it tracks catalog/pricing updates automatically — no second source of truth to keep in sync.
• firerouter has no static rate, so for router rows the script prices against the concrete serverless model the transcript recorded, when available; rows it can't price are skipped (never counted as zero-cost savings).
• Pure core for testing: computation is split into an exported computeSavings(transcriptText) + renderStatusLine(ctx, text) so the logic is exercised without spawning a process. main() is a thin stdin/file wrapper.
• Never crashes the status line — malformed input or missing transcripts fall back to a safe minimal line (a crashing status line command is worse than no status line).

Wiring

Add to ~/.claude/settings.json:

{
  "statusLine": {
    "type": "command",
    "command": "node /path/to/fireconnect/cli/packages/setup-cli/bin/cc-fireworks-savings.mjs",
    "padding": 0
  }
}

Set CC_STATUSLINE_NOCOLOR=1 for non-TTY / CI captures.

Tests

New test/cli/cc-fireworks-savings.test.mjs (10 cases, all passing):

• Fireworks spend computed below Anthropic-equivalent cost, with exact expected values for a known token mix.
• Equivalence map covers each Fireworks family → Anthropic tier; unknown slugs fall back to Sonnet.
• costAtRates prices the same token mix at Anthropic cache rates.
• firerouter case: a concrete dispatched model is priced as Fireworks, not skipped.
• Dedup of repeated assistant message ids.
• Non-assistant and malformed lines ignored; empty/null/garbage input returns zeros and never throws.
• Render produces a single line with savings > 0; safe fallback with no transcript data.

node --test packages/setup-cli/test/cli/cc-fireworks-savings.test.mjs
# tests 10, pass 10, fail 0

Full-suite note: node --test test/**/*.test.mjs shows 8 unrelated failures on this machine, all environmental — deepagents-toml tests need tomllib (Python ≥3.11) and detectIncumbent tests need live Anthropic credentials. None touch the status-line code path.

Files

• packages/setup-cli/bin/cc-fireworks-savings.mjs — the status-line command (new)
• packages/setup-cli/test/cli/cc-fireworks-savings.test.mjs — unit tests (new)
• README.md — documents the status line under "Claude Code pricing estimates"

🤖 Generated with Claude Code