Add Ramp Router price mode

[anassg-lago]

Price Router calls from Router's own /v1/models catalog, on both the Responses surface (OpenAI client) and the Messages surface (Anthropic client). Learn the Router key from the wrapped client; price only the default tier; bill Anthropic cache writes at their 5m/1h rate on /v1/messages; refuse Fireworks entries served through Baseten rather than misprice them. Reconciled against the Router dashboard across three sweeps; documented mismatches (eight OpenAI models, two legacy snapshots) are recorded in docs/ramp-router.md with a markup recommendation instead of code.

[claude]

Verdict: APPROVE (no blockers found). Posting as a COMMENT review because this session is not permitted to submit APPROVE/REQUEST_CHANGES reviews on this repo.

No blockers.

What I checked:

• Ran the full unit suite: uv run pytest tests/unit -q -> 930 passed. ruff check and mypy clean on src/lago_agent_sdk.
• Manually re-derived the core money math instead of trusting the checklist: the $0.02518225 claim (16 input + 20113 cache_write_5m + 5 output on claude-haiku-4-5 pricing) reconciles exactly: 16e-6 + 20113 x 1.25e-6 + 5 x 5e-6 = 0.02518225.
• Traced the token_semantics ordering hazard the code itself flags in a comment: api gets reassigned to "ramp_router" before the total_tokens consistency guard runs in extract_openai_native, so the guard and compute_cost/_split_cache_write agree on the de-overlap convention. Confirmed the reassignment sits above the guard, not below. This is exactly the kind of divergence token_semantics.py exists to prevent, and it is handled correctly.
• Confirmed TOKEN_BILLED_PROVIDERS no longer lists ramp_router (superseded by the catalog price source) and that PricingUnavailableError's new detail kwarg is optional, so no other call site breaks.
• Grepped fixtures, the capture script, and full branch history (git log -p main..head) for leaked identifiers: nothing. The "Bearer sk-router-learned" / "learned-key-123" strings in tests are fixture placeholders asserting header construction, not real keys, and capture_ramp_router.py already redacts real Bearer tokens on capture.
• Checked this repo's other open PRs: only a dependabot PR, no file overlap.

Non-blocking notes:

• The live-account dashboard reconciliation percentages (measured 2026-09-04/2026-09-07) aren't independently verifiable from here; they're dated, specific, and falsifiable if wrong, so I'm taking them at face value rather than treating them as unverified.
• _is_foreign_backend_alias knowingly turns a few Baseten-served calls that happen to match the Fireworks rate into misses instead of prices. Correct, documented tradeoff (a mispriced hit is worse than an honest miss), not a bug.

Merge notes:

• No overlapping open work in this repo.
• Companion PR getlago/lago-agent-sdk-js#55 mirrors this change. Diffed the two: computeCost/splitCacheWrite/parseRampRouter/isForeignBackendAlias/rampRouterUnpricedTier are behaviorally identical, same golden money test values, same fixtures. Either can merge first.

Genuinely good work on the depth of measurement: the documented table of Router-vs-catalog rate mismatches with dates, and the explicit refusal to bake a time-limited or undocumented factor into the SDK itself, is the right call for a billing library.

---

Generated by Claude Code