feat(e2e): in-network mock provider (OpenAI + Anthropic + Gemini)

e2e/.env.example

@@ -33,3 +33,13 @@ MODEL_OPENAI=
 # ─── misc ─────────────────────────────────────────────────────────────
 # Override default proxy port (4011) if you have a conflict.
 # E2E_PROXY_PORT=4011
+
+# ─── mock provider tuning (only used with `proxy start --with-mock`) ──
+# In-network mock that serves OpenAI + Anthropic without real provider
+# calls. Useful for memory / retry / timeout / callback testing.
+# See e2e/_config/mock_provider.py + e2e/cases/23_mock_memory_pressure.md
+#   MOCK_TTFT_MS=0           # ms before first streamed chunk
+#   MOCK_TPS=100             # streamed chunks per second
+#   MOCK_CHUNKS=100          # total chunks per streamed response
+#   MOCK_CALLBACK_DELAY=0    # seconds callback sinks sleep
+#   MOCK_FAIL_RATE=0         # 0..1 fraction of provider calls returning 503

e2e/README.md

@@ -94,6 +94,53 @@ e2e/tools/proxy start
 e2e/tools/proxy stop
 ```
 
+## Mock provider (zero provider cost)
+
+For tests that don't need real LLM behavior — memory pressure, retry
+amplification, slow streaming, callback queue retention, error paths —
+start the proxy with the in-network mock instead of a real provider:
+
+```bash
+e2e/tools/proxy start --with-mock
+```
+
+This also brings up a second container (`litellm-e2e-mock`, only when
+this flag is set) that serves both OpenAI- and Anthropic-shape endpoints
+plus Langfuse / generic-webhook callback sinks. Two extra `model_list`
+entries get added automatically to the rendered config:
+
+- `mock-openai`     → `openai/mock-model`        → `http://mock:8080/v1`
+- `mock-anthropic`  → `anthropic/mock-claude`    → `http://mock:8080`
+
+Behavior is controlled per-process by env vars (set in the shell before
+`proxy start`):
+
+| Env var | Default | What it does |
+|---|---|---|
+| `MOCK_TTFT_MS` | `0` | delay (ms) before first streamed chunk |
+| `MOCK_TPS` | `100` | streamed chunks per second |
+| `MOCK_CHUNKS` | `100` | total chunks per streamed response |
+| `MOCK_CALLBACK_DELAY` | `0` | seconds the callback sinks sleep before responding |
+| `MOCK_FAIL_RATE` | `0` | 0..1 fraction of provider calls returning 503 |
+
+Per-request overrides go in the JSON body (e.g. `mock_chunks`,
+`mock_ttft_ms`, `mock_full_chars`). See the docstring at the top of
+`e2e/_config/mock_provider.py` for the full contract.
+
+Example use cases (canonical reproducer: `cases/23_mock_memory_pressure.md`):
+
+```bash
+# Match a production "slow streaming" profile (TTFT=15s, TPS=30):
+MOCK_TTFT_MS=15000 MOCK_TPS=30 MOCK_CHUNKS=1500 \
+  e2e/tools/proxy start --with-mock
+
+# Retry-amplification: 30% of provider calls return 503
+MOCK_FAIL_RATE=0.3 e2e/tools/proxy start --with-mock
+
+# Slow Langfuse / webhook consumer (drains 8s per batch)
+MOCK_CALLBACK_DELAY=8 e2e/tools/proxy start --with-mock
+```
+
 ## How Claude uses this
 
 Tell Claude:

e2e/_config/docker-compose.yml

@@ -30,6 +30,41 @@ services:
       start_period: 5s
     restart: "no"
 
+  # Optional zero-cost provider — does NOT start unless `--profile mock` is set.
+  # Mounts mock_provider.py (stdlib HTTP server) and exposes:
+  #   POST /v1/chat/completions    OpenAI-shape (stream + non-stream)
+  #   POST /v1/messages            Anthropic-shape (stream + non-stream)
+  #   POST /api/public/ingestion   Langfuse-style callback sink
+  #   POST /api/hooks/*            GenericAPILogger callback sink
+  # Behavior controlled via env vars in the compose project's .env or shell:
+  #   MOCK_TTFT_MS    delay before first chunk (default 0)
+  #   MOCK_TPS        chunks per second when streaming (default 100)
+  #   MOCK_CHUNKS     total chunks per streamed response (default 100)
+  #   MOCK_CALLBACK_DELAY  seconds the callback sinks sleep before responding (0)
+  #   MOCK_FAIL_RATE  0..1 fraction of provider calls returning 503 (default 0)
+  # See e2e/_config/mock_provider.py header for the full request-time override
+  # protocol (each request can also override these per-call via JSON fields).
+  mock:
+    image: python:3.13-slim
+    container_name: litellm-e2e-mock
+    command: ["python", "/mock/mock_provider.py"]
+    volumes:
+      - ./mock_provider.py:/mock/mock_provider.py:ro
+    environment:
+      MOCK_TTFT_MS: "${MOCK_TTFT_MS:-0}"
+      MOCK_TPS: "${MOCK_TPS:-100}"
+      MOCK_CHUNKS: "${MOCK_CHUNKS:-100}"
+      CALLBACK_DELAY: "${MOCK_CALLBACK_DELAY:-0}"
+      FAIL_RATE: "${MOCK_FAIL_RATE:-0}"
+    healthcheck:
+      test: ["CMD", "python", "-c", "import urllib.request,sys; sys.exit(0 if urllib.request.urlopen('http://localhost:8080/healthz',timeout=1).status==200 else 1)"]
+      interval: 2s
+      timeout: 2s
+      retries: 10
+      start_period: 2s
+    profiles: ["mock"]
+    restart: "no"
+
   litellm:
     build:
       context: ../..              # repo root — picks up local source changes
@@ -40,7 +75,7 @@ services:
       - "127.0.0.1:4011:4000"
     environment:
       LITELLM_MASTER_KEY: "sk-e2e-test"
-      LITELLM_LOG: "INFO"
+      LITELLM_LOG: "WARNING"
       # Match the user's production env: force using the locally-bundled
       # model_prices JSON instead of fetching the latest from GitHub.
       LITELLM_LOCAL_MODEL_COST_MAP: "True"
@@ -70,11 +105,6 @@ services:
       # Rendered at proxy-start time by `e2e/tools/proxy` from .env values.
       # See ../tools/proxy `render_config()`. The file is gitignored.
       - ./.litellm.rendered.yaml:/app/config.yaml:ro
-    # NOTE: no `entrypoint:` override here. The fork's
-    # docker/prod_entrypoint.sh already routes through
-    # `python -m litellm_extras.entrypoint`, so PublicReqMiddleware loads
-    # automatically. This keeps e2e behavior aligned with production —
-    # if the wrapper breaks in either path, both surfaces catch it.
     command:
       - --config=/app/config.yaml
       - --port=4000