huggingface/Serge

Review pull requests with any OpenAI-compatible LLM and post inline comments on the diff.

Trust Signals

Scorecard Score: not yet scored
Maintenance Recency: Stale
License: None

Inputs

name	description	required	default
llm_api_key	Bearer token for the chat-completion endpoint.	yes	—
llm_api_base	Base URL of an OpenAI-compatible service. Accepts either https://host or https://host/v1.	no	https://api.openai.com/v1
llm_model	Model identifier to pass in the chat-completion request. If omitted, the action auto-discovers the first model from {llm_api_base}/models.	no	—
llm_bill_to	Optional org slug for routing inference billing (sent as the X-HF-Bill-To header). Required when using a Hugging Face token whose Inference Providers permission is scoped to an organization rather than the user.	no	—
llm_max_tokens	Maximum completion tokens requested from the LLM. Bump for reasoning models (e.g. Kimi-K2) that spend tokens on reasoning before emitting JSON.	no	4096
llm_stream	Set to "true" to consume the chat-completion response as a streaming SSE feed. Off by default; useful for endpoints/proxies that drop long-running non-streaming connections.	no	false
mention_trigger	Phrase that triggers reviews. Must be the first word of the comment.	no	@askserge
review_event	One of COMMENT, REQUEST_CHANGES, APPROVE (fallback if the LLM omits one).	no	COMMENT
max_diff_chars	Maximum diff size (characters) sent to the LLM.	no	200000
review_rules_path	Path to repo-specific review rules, read from the default branch.	no	.ai/review-rules.md
helper_tools_path	Path to a repo-specific helper-tools JSON file, read from the default branch. Helpers can expose tightly scoped CLI commands to the reviewer when a local PR checkout is available.	no	.ai/review-tools.json
default_review_rules	Fallback rules used when the repo has no review-rules file.	no	Apply general Python correctness and security standards.
context_script_path	Path to an executable script in the target repo that, given the PR title/body and modified-file list as JSON on stdin, prints additional context to inject into the reviewer prompt. Skipped silently if the file is missing or not executable.	no	.ai/context-script
context_script_timeout	Seconds to wait for the context script before giving up.	no	30
repo_checkout_path	Path to a local checkout of the PR head ref. When set, the reviewer exposes read-only browse tools (read_file/list_dir/grep) rooted here so the LLM can fetch context beyond the diff. Defaults to GITHUB_WORKSPACE; set to an empty string to disable tools entirely. The caller workflow is responsible for checking out the PR head ref (e.g. with `actions/checkout@v4` and `ref: refs/pull/<number>/head`) before invoking this action.	no	${{ github.workspace }}
tool_max_iterations	Maximum number of tool-calling rounds before forcing the model to produce a final review. Each iteration is a separate LLM call.	no	8
headroom_compress	Set to "true" to compress context with headroom-ai before each LLM call (cuts tokens on tool outputs and older turns). Requires the headroom-ai package; install reviewbot with the [headroom] extra. No-op if the package is missing.	no	false
headroom_target_ratio	Optional headroom keep-ratio for text compression (e.g. 0.5 keeps ~50%). Leave empty to let headroom decide.	no	—
headroom_compress_user_messages	Set to "true" to also compress user messages (the annotated diff). Off by default so cited line numbers stay intact.	no	false
headroom_compress_system_messages	Set to "false" to leave system messages uncompressed. On by default.	no	true
headroom_protect_recent	Number of most-recent messages never compressed.	no	4
headroom_min_tokens	Skip compressing messages shorter than this many tokens.	no	250
headroom_kompress_model	Kompress model id for ML compression, or "disabled" to skip it. Leave empty for the headroom default.	no	—
headroom_model_limit	Model context window (tokens) used for sizing.	no	200000
staging	Set to "true" to mark this as a staging deployment. Published reviews then carry a note saying they were posted from staging.	no	false
github_token	Token with pull-requests:write. Defaults to the job token.	no	${{ github.token }}
python_version	Python version to set up for running the reviewer.	no	3.12

Outputs

no outputs