4itworks/Livvie Code Review

AI code review with multi-perspective reviewers, native suggestion blocks, and REQUEST_CHANGES support.

View on GitHub

Trust Signals

Scorecard Score
not yet scored
Maintenance Recency
Activelast commit Jul 13, 2026
License
MIT

Pinned Snippet

workflow.ymlSHA-pinned
uses: 4itworks/livvie_code_review@e1024d2a5623dda7b0f0934dfc35e2e270a8a147 # v2

tags can be moved; commit SHAs can't. why a SHA?

namedescriptionrequireddefault
github-tokenGitHub token for posting PR reviews.yes${{ github.token }}
llm-api-keyAPI key for your LLM provider. This is the only sensitive input — store it as a GitHub Secret.yes
llm-base-urlBase URL for OpenAI-compatible API. Not sensitive — set as a plain string.nohttps://openrouter.ai/api/v1
modelModel name (e.g. z-ai/glm-5.2, anthropic/claude-sonnet-4). Not sensitive — set as a plain string.yes
review-instructions-fileRepository file with extra review instructions.no.github/code-reviewer.md
max-diff-sizeMaximum diff size in characters per file.no50000
max-output-tokensMaximum tokens for the LLM response (includes reasoning tokens if reasoning is enabled).no16000
reasoning-effortReasoning effort level for models that support it. Options: none, low, medium, high, max. Ignored by models that don't support reasoning.nonone
fallback-modelFallback model if the primary model fails after all retries. Set to empty to disable.no""
request-changes-on-highPost REQUEST_CHANGES when high-severity findings exist.notrue
always-request-changes(Deprecated) Use request-changes-on instead.nofalse
max-commentsMaximum inline comments to post.no25
include-severitiesComma-separated list of severities to include in the review. Options: low, medium, high.nolow,medium,high
include-confidencesComma-separated list of confidences to include in the review. Options: low, medium, high.nolow,medium,high
request-changes-onComma-separated list of severities that trigger REQUEST_CHANGES. Use 'low,medium,high' to always request changes.nohigh
respect-author-repliesWhen true, the bot reads replies to its inline comments and dismisses findings whose author's justification is reasonable.nofalse
ignore-patternsComma-separated glob patterns for files to skip (e.g. *.g.dart,build/). Defaults to generated file patterns.nobuild/**,dist/**,node_modules/**
max-batchesMaximum number of file batches. Each batch is reviewed by every perspective, so total LLM calls = batches × perspectives. 0 = no limit.no0
context-windowContext window size in tokens for the model. Used for token budget calculation. Override if your model has a different context window.no128000
cross-file-budget-maxMaximum tokens to allocate for cross-file context in each batch.no2000
cross-file-budget-ratioPercentage of context window (0-100) to allocate for cross-file context.no5
safety-marginAdditional tokens reserved for message wrappers and response overhead.no500
circuit-breaker-thresholdConsecutive failures before opening the circuit breaker for a model.no3
agents-dirDirectory containing agent .md files that define review perspectives. Each .md file = one reviewer. See README for file format and examples.no.github/livvie_code_review_agents
verboseLog LLM reasoning traces and detailed debug info to the Actions log.nofalse
namedescription
review-idThe ID of the posted GitHub review.
finding-countTotal number of findings in the review.