Gravitas is an open-source reliability runtime for AI coding agents. It combines portable Agent Skills with optional Antigravity and OpenCode integrations, a shared policy engine, validators, a hash-linked evidence ledger, and a completion gate. The project changes the engineering process around a model, not the model itself: important actions and completion claims are expected to have inspectable evidence.
Coding agents can edit files they have not read, write outside a task's intended scope, retry a failed command without changing the diagnosis, or report success from tests run before the final edit. Gravitas addresses these process failures with six explicit invariants, declared write scope, post-change validators, and a gate that refuses completion when current evidence is missing or stale.
I contributed to PR #83, which was merged after maintainer review and follow-up hardening. The work strengthened the evidence path used by the runtime: failure resolutions are bound to anchored records, affected files must be re-read, and completion requires a fresh validator pass for linked criteria. It also records successful OpenCode mutations by session so earlier validation becomes stale without treating a tool completion as a passing test. Benchmark worktrees are private on POSIX where supported, fixture failures clean up temporary worktrees, and exported episode reports remove hidden-validator output while retaining its digest.
PR #88 was merged into Gravitas main. It fixes four workflow-integrity issues: benchmark episodes are validated before reports are generated, read-before-write freshness is tied to file content hashes, reproducer runs are linked to the same criterion and command, and shell writes with unknown paths invalidate verification. The PR also adds a deterministic bundle-manifest generator with normalized line endings, stable cross-platform ordering, and coverage for nested checksum files.
I contributed cross-platform impact-graph path normalization, switched validator and test fixtures to the running Python interpreter, and added regression coverage for Windows path and test-environment behavior. PR #49 was closed as superseded, not merged: the maintainer confirmed the relevant fixes and tests were incorporated upstream through PR #53, and closed the related Windows portability issue #50.
Gravitas is a research preview. Its hooks and guards are workflow controls, not an operating-system sandbox, and the project README does not publish comparative performance results. PR #83 recorded 258 tests passed and 2 skipped; PR #88 recorded 327 tests passed and 8 skipped, with 15 manifest tests passing and the bundle-drift check clean.
Repository · Merged PR #83 · Merged PR #88 · Superseded PR #49 · Issue #50