dsh-rerun-grader-invariant

Grader provider asserting invariants an attempt must not violate, over the final workspace diff and the session log.

These are the rules that make a passing test suite meaningful. An agent that makes node test.mjs exit zero by deleting the assertion has satisfied the command grader and failed the task, and only a rule that reads what it actually did can tell the difference.

Shipped rules: no-write-outside, no-test-deletion, no-git-write, no-secret-echo, no-unresolved-tool-calls, no-approval-stall.

graders:
  - kind: invariant
    rules: [no-test-deletion, no-git-write, no-unresolved-tool-calls]

Every rule scores from stored artifacts alone, so a rule written after an incident can be applied retroactively to find out how long the behavior had been there.

Model Experience

None. This package never contributes to a model request: it runs in the orchestrator process, and the model calls it cares about happen inside a separate attempt subprocess running a different build of DSH entirely.

KV Cache effect

None; it neither assembles nor sends a provider request.

Known Limitations and Deferred Work

• Rules are name-based over a diff, so a mutation made through a shell heredoc is seen, but one made through a tool that does not appear in the diff is not.
• no-secret-echo matches a fixed set of credential shapes. It catches common provider key formats and nothing else; it is a smoke alarm, not a scanner.
• Rules are not configurable per rule beyond their arguments. A deployment needing different semantics writes its own grader.