Metadata-Version: 2.4
Name: llm-router-ledger
Version: 0.1.3
Summary: Route LLM calls through one send_message() across OpenAI-compatible providers and Anthropic's native Messages API; keep a JSONL ledger of every request and response for offline cost reconciliation.
Project-URL: Homepage, https://github.com/nirmalyaghosh/llm-router-ledger
Project-URL: Issues, https://github.com/nirmalyaghosh/llm-router-ledger/issues
Author-email: Nirmalya Ghosh <1637161+nirmalyaghosh@users.noreply.github.com>
License: MIT License
        
        Copyright (c) 2026 Nirmalya Ghosh
        
        Permission is hereby granted, free of charge, to any person obtaining
        a copy of this software and associated documentation files (the
        "Software"), to deal in the Software without restriction, including
        without limitation the rights to use, copy, modify, merge, publish,
        distribute, sublicense, and/or sell copies of the Software, and to
        permit persons to whom the Software is furnished to do so, subject to
        the following conditions:
        
        The above copyright notice and this permission notice shall be
        included in all copies or substantial portions of the Software.
        
        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
        EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
        MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
        NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
        BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
        ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
        CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
        SOFTWARE.
License-File: LICENSE
Keywords: anthropic,azure,deepseek,ledger,llm,minimax,ollama,openai,openrouter,router,tracking,usage,zhipu
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.12
Classifier: Topic :: Software Development :: Libraries
Requires-Python: >=3.12
Requires-Dist: openai<3,>=1.55.0
Requires-Dist: pydantic>=2.7
Requires-Dist: python-dotenv>=1.0
Requires-Dist: pyyaml>=6.0
Provides-Extra: anthropic
Requires-Dist: anthropic>=0.30; extra == 'anthropic'
Provides-Extra: dev
Requires-Dist: anthropic>=0.30; extra == 'dev'
Requires-Dist: mypy>=1.10; extra == 'dev'
Requires-Dist: pytest>=8.0; extra == 'dev'
Requires-Dist: ruff>=0.5; extra == 'dev'
Requires-Dist: twine>=5.0; extra == 'dev'
Requires-Dist: types-pyyaml>=6.0; extra == 'dev'
Description-Content-Type: text/markdown

# llm-router-ledger

Route any LLM call through one `send_message()` and keep a JSONL ledger
of every request and response for offline cost reconciliation.

## Provider support

| Status | Adapter | Providers |
|---|---|---|
| Supported | direct | Anthropic |
| Supported | OpenAI-compat | Azure OpenAI, DeepSeek, Local Ollama, MiniMax, OpenAI, OpenRouter, Qwen, Zhipu / GLM |
| Supported | via OpenRouter | ByteDance Seed, Xiaomi MiMo |
| Planned | direct | Gemini |

- All "Supported" rows in 0.1.2 are live-smoke-verified end-to-end.
- Anthropic requires the optional `[anthropic]` extra: `uv pip install llm-router-ledger[anthropic]`.
- For ByteDance Seed and Xiaomi MiMo, use `provider: openrouter` with the appropriate model id.

## Install

```bash
uv pip install llm-router-ledger
```

## Quickstart

Set `OPENROUTER_API_KEY` in `.env` and create `llm_endpoints.yaml` in the working directory. The fastest path is to copy `examples/llm_endpoints.example.yaml` to `llm_endpoints.yaml` in your working directory and edit it.

```python
from llm_router_ledger import UsageTracker, send_message

tracker = UsageTracker(
    log_path="logs/usage.jsonl",
    project_id="my-blog",
)
text, usage, gen_id = send_message(
    endpoint_name="openrouter-mimo-v2.5",
    system="You are concise.",
    user="Explain prompt caching in two sentences.",
    tracker=tracker,
)
```

Or against a local Ollama server, with no API costs:

```python
text, usage, gen_id = send_message(
    endpoint_name="local-llama",
    system="You are concise.",
    user="Explain prompt caching in two sentences.",
    tracker=tracker,
)
```

- `send_message()` returns `(response_text, usage_dict, generation_id)`.
- `UsageTracker` appends paired `llm_request` / `llm_response` events to the JSONL log, stamped with `project_id`, `run_tag`, `run_label`, and `purpose` for later grouping.

## Per-endpoint request params

Model-specific knobs belong in config, not in every caller. Give an endpoint an `extra_body` and it is sent on every call to that endpoint:

```yaml
endpoints:
  openrouter-deepseek:
    provider: openrouter
    model: deepseek/deepseek-chat
    api_key_env: OPENROUTER_API_KEY
    base_url: https://openrouter.ai/api/v1
    extra_body:
      reasoning:
        enabled: false
```

- An `extra_body` passed to `send_message()` replaces the endpoint's value outright. The two layers are not merged, so a caller that wants both must combine them itself. An opaque vendor passthrough carries no merge rules to memorise as a result.
- **Known limitation:** `provider: anthropic` ignores `extra_body`, so the field has no effect there. `provider: openrouter` reaches Claude with `extra_body` intact.

## Mirroring usage elsewhere

`UsageTracker.subscribe()` registers a callback that receives every ledger entry, so usage can be mirrored to another store without this library depending on it:

```python
tracker.subscribe(lambda entry: my_container.upsert_item(entry))
```

- Each entry is written to the JSONL ledger before any subscriber runs.
- A callback that raises is logged and skipped. The entry is already in the ledger, the call that produced it is unaffected, and the remaining subscribers still run.
- Each subscriber receives its own copy of the entry.
- Callbacks are synchronous and run on the calling thread, so a slow one delays every call. Queue the work inside the callback if the destination is remote.

## JSONL ledger schema

- `UsageTracker` writes two events per `send_message()` call: an `llm_request` before the call, and an `llm_response` after.
- Both share a `request_id` so they can be paired. Top-level fields on each event include `project_id`, `provider`, `model`, `purpose`, `run_tag`, `run_label`, and `timestamp`.
- The `llm_response` event additionally carries `usage` (with `prompt_tokens`, `completion_tokens`, `total_tokens`) and a response preview.

**Identifying a response for billing reconciliation:** the response id is routed to one of two fields based on prefix:

- `generation_id`: set when the id starts with `"gen-"` (OpenRouter convention). Use this when joining against OpenRouter's CSV export, which calls the column `generation_id`.
- `provider_response_id`: set for everything else. OpenAI, Azure OpenAI, Ollama, and most direct-provider endpoints return ids like `"chatcmpl-..."` that land here. Use this when joining against OpenAI-family billing exports or any provider-native log that exposes a
  chat completion id.

Exactly one of the two fields is populated per `llm_response` event; queries that join the ledger to billing data should `COALESCE` over both or branch on `provider`.

## CLI

```bash
llm-router-ledger list                          # show configured endpoints
llm-router-ledger validate llm_endpoints.yaml   # validate the YAML
llm-router-ledger stale --days 30               # endpoints with stale pricing
llm-router-ledger chat --endpoint openrouter-mimo-v2.5 --system "You are concise." --user "Hello." --log-path logs/usage.jsonl --project-id my-project
```

## Env vars

| Variable | Purpose |
|---|---|
| `LRL_RUN_TAG` | Stamped on every JSONL event. |
| `LRL_RUN_LABEL` | Stamped on every JSONL event. |
| `LRL_CONFIG_PATH` | Default YAML path when `load_config()` is called with no argument. |

## Development

```bash
git clone https://github.com/nirmalyaghosh/llm-router-ledger
cd llm-router-ledger
uv sync --extra dev
pytest tests/unit
```

Verify a local Ollama setup end-to-end with
`python examples/smoke_test_ollama.py` (see prerequisites at the top
of the script).

## License

MIT. See `LICENSE`.
