[heydryft]

Summary

• Adds Runcrate as a named OpenAI-compatible LLM provider, in the same shape as the Cheaper Inference and OrcaRouter routes (feat(llm): add Cheaper Inference as a named OpenAI-compatible gateway #223, feat(llm): add OrcaRouter as a named OpenAI-compatible gateway #205).
• funclip/llm/openai_api.py: a runcrate/ prefix that is stripped before the request (model ids keep their vendor, e.g. deepseek-ai/DeepSeek-V4.1-Flash), base URL https://api.runcrate.ai/v1 (overridable with RUNCRATE_API_BASE), key from the APIKEY box or RUNCRATE_API_KEY, never OPENAI_API_KEY.
• funclip/launch.py: runcrate in SUPPORT_LLM_PREFIX, the runcrate/ route to openai_call, and two dropdown entries: runcrate/deepseek-ai/DeepSeek-V4.1-Flash and runcrate/openai/gpt-oss-120b.
• README and README_zh: a short "Using Runcrate" section next to the other providers.
• tests/test_runcrate_launch_integration.py: launcher dispatch, prompt roles and vendor-qualified dropdown ids (added in 37daeda after review).
• tests/test_runcrate_api.py: base URL, prefix stripping, env key, RUNCRATE_API_BASE override, empty model, and no fallback to the OpenAI key.

User impact

Users can run LLM-assisted clipping on open models through Runcrate. Runcrate has a free allowance on some open models, with the limits in its free inference terms; usage beyond it is paid.

Validation

• I ran the relevant tests or commands: pytest tests/test_runcrate_api.py tests/test_cheaperinference_api.py tests/test_orcarouter_api.py tests/test_minimax_api.py tests/test_openai_api.py tests/test_*_launch_integration.py tests/test_model_selection.py: 39 passed.
• I checked the affected README/docs links, if changed.
• I verified the affected UI, API, ASR, LLM, or clipping path, if changed: live openai_call("", "runcrate/<model>", <prompt + SRT>, demo_prompt) with RUNCRATE_API_KEY set and no OpenAI variables. Both deepseek-ai/DeepSeek-V4.1-Flash and openai/gpt-oss-120b returned a segment in the 1. [start-end] text format that the AI Clip step parses.

Notes for reviewers

Disclosure: I work on Runcrate. This PR was prepared with AI assistance (an AI coding agent wrote the change and ran the tests and the live check above).

[LauraGPT]

Reviewed the five-file diff at 4c3073a and the actual shared API/launcher implementation.

The five added Runcrate tests and 26 existing related provider/launcher regressions passed independently on Python 3.12 with OpenAI SDK 2.44.0 and external socket connections denied. These are offline mocked-client/AST checks, not a live provider or full Gradio/AI Clip acceptance. A separate four-case real-SDK/mock-transport check produced passing JUnit/log output, but its wrapper failed to capture the child exit code; it is not counted as a complete process receipt.

Before merging, please address:

1. The English and Chinese README statements that open models are free with a GitHub account omit material conditions. The official free-inference terms specify account eligibility, shared session/week allowances and capacity limits, with paid fallback when a workspace has credit. Please use neutral wording linking to these conditions, rather than implying unrestricted free clipping. A canonical link without campaign tracking is preferable.
2. Please attach the already-run live-check evidence described in the PR: exact head/dependency versions, date, both model IDs, a small synthetic SRT/prompt, and redacted response/parsed segment results. Do not include credentials or private recordings. Existing evidence is sufficient; no need to repeat successful calls just for this review. I have not made provider calls or incurred charges.
3. Please retain a Runcrate launcher regression in the repository, covering its dispatch and prompt roles as well as the vendor-qualified model IDs. The new tests cover openai_call directly but would not catch loss of the new launcher route. Existing providers already have corresponding launcher tests.

The prefix/key isolation and vendor preservation worked in the checked offline cases. This review is about the documented cost boundary, durable launcher coverage, and evidence for the claimed live path; it is not a request for another administrator to approve.

[heydryft]

Thanks for the careful review. All three points are addressed in 37daeda.

1. README wording. Both READMEs now link the canonical pages with no campaign parameters (https://www.runcrate.ai/inference) and say: "Runcrate has a free allowance on some open models, with eligibility, usage and capacity limits described in its free inference terms; usage beyond the allowance is paid." README_zh.md says the same in Chinese.

2. Live-check evidence. I didn't keep a record of the original check, so I reran it once at 4c3073a (the Runcrate code is unchanged in 37daeda; only README and tests changed), 2026-11-01 10:12 UTC, Python 3.12.13, openai 2.54.0. openai_call("", "runcrate/<model>", user + "\n" + srt, system) with RUNCRATE_API_KEY set and no OPENAI_* variables, using the launcher's default Chinese system and user prompts and this synthetic SRT:

1
00:00:00,000 --> 00:00:04,000
Welcome to the weekly product update.

2
00:00:04,000 --> 00:00:09,500
The biggest news is that exports are now three times faster.

3
00:00:09,500 --> 00:00:13,000
Next, a quick note about the office move.

Responses, then extract_timestamps() on each:

• runcrate/deepseek-ai/DeepSeek-V4.1-Flash: 1. [00:00:00-00:00:09,500] Welcome to the weekly product update. The biggest news is that exports are now three times faster. gives [], because the model left the milliseconds off the start time. A run a few minutes earlier with a short English prompt gave 2. [00:00:04,000-00:00:09,500] The biggest news is that exports are now three times faster., which matches the parser.
• runcrate/openai/gpt-oss-120b: 1. [00:00:00,000-00:00:13,000] Welcome to the weekly product update. The biggest news is that exports are now three times faster. Next, a quick note about the office move. gives [[0, 13000]].

So the route, prefix stripping and key isolation work for both models. Whether a reply matches the strict HH:MM:SS,mmm timestamp format depends on the model, the same as for the other OpenAI-compatible providers. I kept the two dropdown entries as they are, but I can drop the DeepSeek one if you would rather list only gpt-oss.

3. Launcher regression. New tests/test_runcrate_launch_integration.py, in the same shape as the Cheaper Inference, MiniMax and OrcaRouter launcher tests. It checks that the branch returning openai_call(...) covers model.startswith('runcrate/') and passes apikey, model, user_content + '\n' + srt_text and system_content in that order; that runcrate is in SUPPORT_LLM_PREFIX; and that the dropdown keeps the vendor-qualified ids runcrate/deepseek-ai/DeepSeek-V4.1-Flash and runcrate/openai/gpt-oss-120b. Removing the runcrate/ condition from launch.py makes the dispatch test fail (checked locally, then restored).

Tests on 37daeda, Python 3.12:
python -m pytest -q tests/test_runcrate_api.py tests/test_runcrate_launch_integration.py tests/test_cheaperinference_api.py tests/test_orcarouter_api.py tests/test_minimax_api.py tests/test_openai_api.py tests/test_cheaperinference_launch_integration.py tests/test_minimax_launch_integration.py tests/test_orcarouter_launch_integration.py tests/test_model_selection.py tests/test_installation_docs.py: 47 passed.

I wrote this follow-up and the new test with AI assistance (omp with Claude). The test runs and the live calls above were run while preparing it.