| preset | Repo preset. Currently `bun-ub` is the production/default preset. | no | bun-ub |
| config | Optional ub-review TOML config path. When set, this overrides preset selection. | no | "" |
| profile | Box profile. `gh-runner` is the zero-config GitHub-hosted default. | no | gh-runner |
| runtime-profile | Runtime profile for proof budgets and local work limits. Set `gh-runner-full` only when leasing broader proof. | no | "" |
| root | Repository root to inspect. | no | . |
| base | Base ref, usually an origin remote branch for the PR target. | no | origin/main |
| head | Head ref. | no | HEAD |
| out | Output directory for the packet. | no | target/ub-review |
| install-tools | Best-effort install of advisory sensors. | no | true |
| tool-bundle | Sensor bundle to install: none, core, bun-fast, or full. Unknown values fail the install step. | no | core |
| setup-rust | Install/select Rust 1.95.0 with rustup when available. | no | true |
| install-mode | Runner install mode for ub-review itself: auto, release, source, or path. | no | auto |
| binary-path | Existing ub-review binary path when install-mode=path. | no | "" |
| release-version | Release tag to download when install-mode=release. Defaults to the action ref. | no | "" |
| release-asset | GitHub release asset name for the Linux x64 ub-review archive. | no | ub-review-x86_64-unknown-linux-gnu.tar.gz |
| allow-heavy | Allow heavy witnesses such as tests/builds/Miri/mutation. Disabled by default. | no | false |
| posting | Posting mode: artifact-only or review. | no | review |
| mode | Execution mode: review-byok or intelligent-ci. Legacy review-direct is accepted as an alias. | no | review-byok |
| review-mode | User-facing review posture. When set, overrides mode/fail-on-gate/[gate].review_forward (with a per-knob warning). advisory = comment only, never blocks; gate = deterministic-floor required check (recommended); strict = + reporter verdict can block. Unset (default) uses the legacy knobs below unchanged. | no | "" |
| run-pass | Run pass identity: auto, opened, reopened, ready_for_review, synchronize, pull_request_other, or manual. Auto maps pull_request opened, reopened, ready_for_review, and synchronize to the matching pass; when posting=review, a pass posts the grouped review only if the profile lists its event action in [gate].post_review_on. | no | auto |
| github-token | GitHub token used to seed PR-thread context and, when posting=review, submit the grouped PR review. | no | "" |
| minimax-api-key | MiniMax provider API key for MiniMax M3 lanes. | no | "" |
| minimax-api-url | Optional MiniMax API URL override. | no | "" |
| minimax-provider-kind | MiniMax request/response family: anthropic or openai. | no | anthropic |
| minimax-model | MiniMax model name. | no | MiniMax-M3 |
| opencode-api-key | OpenCode Go API key for optional direct provider lanes. | no | "" |
| opencode-api-url | Optional OpenCode Go API URL override. | no | "" |
| opencode-model | OpenCode Go model for the canary lane. | no | minimax-m3 |
| opencode-endpoint-kind | OpenCode Go endpoint family: auto, openai-chat, or anthropic-messages. | no | auto |
| model-mode | Model mode: auto or off. | no | auto |
| depth | Review depth selector: quick, standard, or deep. | no | standard |
| lanes | Comma-separated lane IDs to run. Empty uses the selected depth default. | no | "" |
| except-lanes | Comma-separated lane IDs to skip after applying lanes. | no | "" |
| tools | Comma-separated sensor/tool IDs to plan. Empty uses the profile default. | no | "" |
| except-tools | Comma-separated sensor/tool IDs to skip after applying tools. | no | "" |
| provider-policy | Provider policy: auto (defers to [providers].policy in the repo config, else minimax-primary), primary-with-fallback, minimax-primary, minimax-only, opencode-go-canary, or opencode-go-wide. An explicit value overrides config. | no | auto |
| lane-width | Bun model lane width: 6, 10, or 20. | no | 10 |
| model-timeout-sec | Per-model-call timeout in seconds. | no | 300 |
| max-inline-comments | Maximum validated inline comments in the grouped PR review. | no | 8 |
| model-concurrency | Planned model lane concurrency. | no | 8 |
| max-model-calls | Maximum model calls for review generation. | no | 14 |
| review-body-max-bytes | Maximum bytes for the grouped PR review body. | no | 60000 |
| ledger-path | Optional read-only UB ledger path to include in shared context. | no | "" |
| ledger-max-bytes | Maximum UB ledger context bytes. | no | 65536 |
| pr-thread-context | Optional file containing bounded PR thread context, prior replies, or proof receipts to seed into shared context. | no | "" |
| pr-thread-context-max-bytes | Maximum PR thread context bytes. | no | 65536 |
| prior-resolved-candidates | Optional prior review/resolved_candidates.json receipt to carry resolved candidates across passes. | no | "" |
| prior-resolved-candidates-artifact | Artifact name to scan for review/resolved_candidates.json in the previous completed PR run when prior-resolved-candidates is empty. | no | ub-review-gate |
| fail-on-post-error | Fail the action when posting the grouped PR review fails. | no | false |
| fail-on-gate | Gate enforcement: auto, true, or false. The final Enforce gate outcome step runs `ub-review gate-check`, which exits non-zero when review/gate_outcome.json records a fail conclusion and this resolves to true; artifacts, the job summary, and PR review posting always complete first. auto resolves to true for mode=intelligent-ci and false otherwise. | no | auto |
| github-summary | Append running-summary.md to GITHUB_STEP_SUMMARY. | no | true |