feat(analytics,proxy): routing headers visualization and routing mode query param

[yanaSelin]

Summary

Adds Switchyard/LiteLLM routing awareness across the Claude plugin: the statusline shows what a
router alias actually dispatched to, codemie models list distinguishes router aliases from
concrete deployments, and the analytics report prices and visualizes routing/classifier cost per
turn and per session. Also converts the Claude statusline to a bundled TypeScript artifact and
tightens explicit-model-choice handling across the CLI.

Changes

Statusline

• Convert the Claude statusline from a loose .mjs script to TypeScript with normal project
  imports, then bundle it into a single esbuild artifact; deploy only the bundled artifact
  (drop shim-file deployment, exclude .ts sources from the Claude plugin asset copy).
• Add a "routed to" widget: when the configured model alias is a Switchyard virtual router or a
  declared LiteLLM auto-router (isRouterCatalogEntry), show the actual backend model the last
  turn dispatched to instead of just the alias name.
• Surface classifier (routing LLM) cost and tokens, and a routing-tier indicator, sourced from
  the same transcript-based cost accounting the analytics report uses.

Model list & resolution

• codemie models list: add a TYPE column (router vs. model) and widen the ID column so
  long free-form router alias names no longer truncate.
• codemie-codex: accept an explicitly requested --model/env model even when it is not listed
  in the live catalog (router aliases resolve per request and are not always published as
  deployments), gated to an explicit choice via a new CODEMIE_MODEL_SOURCE (cli/env/default)
  marker set by AgentCLI; also recognize gpt-smart-router/gpt-fast-router-style aliases as
  Codex-compatible.
• Fix subagent-default-tier pinning to skip itself when the session model differs from the only
  provisioned tier, so a router alias assigned to CODEMIE_HAIKU_MODEL/CODEMIE_OPUS_MODEL
  no longer gets silently overridden for subagents declaring model: inherit.
• Fetch the full model descriptor (fetchCodeMieLlmModels) instead of ID-only, so router entries
  carry their is_router/litellm_router.is_router flag through to both the CLI model list and
  the statusline's routing signal.

Analytics (cost & routing report)

• Report: per-turn model/routing timeline, a "Routing" KPI section (sessions routed, classifier
  cost), and a routing-tier chart on the session page; classifier cost/tokens roll into session
  and summary totals.
• New bedrock-pricing.mjs: detects Amazon Bedrock's regional/cross-region-inference pricing
  premium from the raw backend model id and applies it to the rate card.
• Fix model-name normalization to also strip own Bedrock alias shape
  (bedrock/us.anthropic.claude-sonnet-5, no version suffix), which previously slipped through
  unnormalized and showed up as a raw deployment path next to a clean model name in the same
  session's model breakdown.
• New background-request-normalizer proxy plugin: force a concrete haiku deployment for Claude
  Code's own background requests (e.g. auto-title generation) instead of deferring to the
  upstream classifier, which sometimes misrouted these tiny calls to the expensive tier — one
  observed case billed a title-gen call on par with the session's most expensive real turn.

Other

• Fix local-dev SSO API base resolution to not require the code-assistant-api prefix.

Testing

• npm run typecheck
• npm run lint
• Manual testing: generated HTML analytics report against real routed/unrouted sessions;
  verified statusline "routed to" widget and codemie models list TYPE column locally.

Checklist

• Code follows project standards
• CI is green (npm run ci)
• No merge conflicts with main

EPMCDME-14886

EPMCDME-14887

[codemie-ai]

[AUTO_CLOSE_WARNING] ⏰ This pull request has had no activity for 7 days (last activity: 2026-09-11).

It will be automatically closed no earlier than 2026-09-25 (UTC) (~6 more days).

Any activity from anyone — a comment, a push, a review, reopening — restarts the countdown from that moment. This warning will be deleted automatically once that happens.