Verdict

A verdict is the binary output of a grade: either approve or request_changes (with reasons). It’s computed deterministically from a report’s findings against a configured set of thresholds.

Shape

{ "kind": "approve" }

{
  "kind": "request_changes",
  "reasons": [
    "1 error finding (max allowed: 0)",
    "63 warn findings (max allowed: 50)"
  ]
}

The verdict is part of every Report.

Thresholds

Verdicts are derived from [verdict] in .sextant/config.toml:

[verdict]
max_errors = 0     # any errors → request_changes
max_warns = 50     # > 50 warns → request_changes (default: u32::MAX)

Field	Default	Effect
`max_errors`	`0`	Findings with `severity: error` over this count flip the verdict.
`max_warns`	`u32::MAX`	Same, for warns. The default is “never block on warns”.

info findings never affect the verdict.

Modes

A verdict has two modes — controlled by which command produced the report:

Absolute mode

Used by:

sextant grade (default whole-file mode)
sextant grade --diff
The grade_diff and grade_files MCP tools

Counts every finding against thresholds. Whatever’s in the report is what’s measured.

Regression mode

Used by:

sextant grade --pr
The GitHub Action

Counts only findings that are new relative to a baseline-graded base SHA. A PR that exposes pre-existing problems but doesn’t introduce new ones approves.

This is why the Action runs sextant grade --pr: it gives “this PR makes the codebase worse” semantics rather than “this codebase has debt” semantics. Pre-existing debt is its own problem.

Reasons

When the verdict is request_changes, reasons lists the specific threshold checks that failed. Both PR mode and absolute mode populate this — agents and review comments can quote it directly.

Default behaviour

Out of the box (max_errors = 0, no max_warns), Sextant blocks only on errors. Warns surface in the report and the agent’s context but don’t gate.

To make warns blocking, set max_warns in your config:

[verdict]
max_errors = 0
max_warns = 0   # any new warn blocks the verdict