Drop kimi-k2-thinking from the catalog; remove temporary e2e trigger

- Moonshot serves newer platform accounts exactly one model, kimi-k3:
  K2-era ids and the kimi-latest alias return resource_not_found_error
  on both API surfaces (verified live 2026-07-21). Drop kimi-k2-thinking
  from the kimi_api catalog entry and the pricing defaults so the
  dashboard doesn't advertise a model that 404s; grandfathered accounts
  can still type K2 ids on the provider's model rows.
- Remove the temporary push trigger from agent-network-e2e.yml now that
  the branch validated green (run #76); nightly + workflow_dispatch
  remain.
- Remove CLAUDE.md from the repository.

.github/workflows/agent-network-e2e.yml

@@ -5,10 +5,6 @@ on:
   schedule:
     - cron: "0 3 * * *"
   workflow_dispatch:
-  # TEMPORARY: run on pushes to the Kimi feature branch to exercise the new
-  # provider scenarios with E2E_KIMI_TOKEN. Remove before merging.
-  push:
-    branches: [feat/kimi-3-agent-networks]
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}

CLAUDE.md

@@ -1,6 +0,0 @@
-# Repository guidance for Claude
-
-## Git conventions
-
-- Do not add `Co-Authored-By` or other AI attribution trailers to commit messages.
-- Do not use `claude/` (or otherwise Claude-named) branch names; use conventional prefixes like `feat/`, `fix/`, `chore/`.

management/internals/modules/agentnetwork/catalog/catalog.go

@@ -449,14 +449,16 @@ var providers = []Provider{
 		// meterable platform key, so it's deliberately not the default.
 		ParserID: "",
 		// Pricing per Moonshot's platform rates at K3 launch (July 2026):
-		// K3 $3/$15 per MTok with $0.30 cached input, flat across the
-		// 1M-token window; K2 Thinking $0.60/$2.50 with $0.15 cache hits.
-		// Kimi K3 has a single always-on-reasoning SKU — no mini/turbo
-		// variants at launch. The consumer app's "K3 Swarm Max" mode is
-		// not an API SKU, so it doesn't appear here.
+		// $3/$15 per MTok with $0.30 cached input, flat across the 1M-token
+		// window. kimi-k3 is the ONLY model the platform serves newer
+		// accounts — K2-era ids (kimi-k2-thinking) and even the kimi-latest
+		// alias return resource_not_found_error, verified live 2026-07-21 —
+		// so it's the only catalog entry. Grandfathered accounts with K2
+		// access can still type those ids on the provider's model rows.
+		// The consumer app's "K3 Swarm Max" mode is not an API SKU, so it
+		// doesn't appear here.
 		Models: []Model{
 			{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
-			{ID: "kimi-k2-thinking", Label: "Kimi K2 Thinking", InputPer1k: 0.0006, OutputPer1k: 0.0025, ContextWindow: 262144},
 		},
 	},
 	{

proxy/internal/llm/pricing/defaults_pricing.yaml

@@ -163,15 +163,13 @@ openai:
 
   # Kimi / Moonshot AI (kimi_api) — OpenAI-compatible /v1 endpoint. Moonshot
   # reports cache hits OpenAI-style when present; cached input is 10% of
-  # input for K3 ($0.30 vs $3.00 per MTok) and $0.15/MTok for K2 Thinking.
+  # input ($0.30 vs $3.00 per MTok). kimi-k3 is the only model the platform
+  # serves newer accounts (K2-era ids and kimi-latest 404), matching the
+  # management catalog.
   kimi-k3:
     input_per_1k: 0.003
     output_per_1k: 0.015
     cached_input_per_1k: 0.0003
-  kimi-k2-thinking:
-    input_per_1k: 0.0006
-    output_per_1k: 0.0025
-    cached_input_per_1k: 0.00015
 
 anthropic:
   # Claude 4.x family — cache reads ≈10% of input, cache writes ≈125% of input.
@@ -220,7 +218,7 @@ anthropic:
 
   # Kimi / Moonshot AI (kimi_api) via the Anthropic-compatible endpoint
   # (/anthropic/v1/messages — the official Claude Code setup). Same rates
-  # as the OpenAI-shape entries above. "kimi-k3[1m]" is the model id some
+  # as the OpenAI-shape entry above. "kimi-k3[1m]" is the model id some
   # Claude Code guides set for the 1M-context alias; priced identically so
   # cost metering doesn't silently skip those requests.
   kimi-k3:
@@ -231,10 +229,6 @@ anthropic:
     input_per_1k: 0.003
     output_per_1k: 0.015
     cache_read_per_1k: 0.0003
-  kimi-k2-thinking:
-    input_per_1k: 0.0006
-    output_per_1k: 0.0025
-    cache_read_per_1k: 0.00015
 
 bedrock:
   # AWS Bedrock model ids, normalised by the request parser (cross-region