GRAVITAS

An open-source reliability runtime for coding agents — measured, not claimed.

GRAVITAS is an open-source Agent Skill and optional Google Antigravity runtime for coding work. It asks an agent to inspect before editing, plan in proportion to risk, verify against real output, recover from failures, and report only what the evidence supports.

GRAVITAS changes the engineering process, not model weights.

npx skills add udaysharmadev/Gravitas --skill gravitas --agent antigravity --yes

Use Gravitas to fix this bug and verify the result.

npx skills add udaysharmadev/Gravitas --skill gravitas --agent antigravity --yes

python3 .agents/skills/gravitas/scripts/doctor.py

git clone --depth 1 https://github.com/udaysharmadev/Gravitas.git
cd Gravitas
bash scripts/build-native-bundle.sh
agy plugin install "$PWD/dist/gravitas-antigravity"

agy plugin validate "$PWD/dist/gravitas-antigravity"
agy plugin list

Validation reports gravitas-native components; the list confirms it is staged. In the agy TUI, /hooks shows the active hook registrations. See the complete, version-aware installation guide.

Or just ask Antigravity

Copy this into an Antigravity conversation:

Install GRAVITAS from https://github.com/udaysharmadev/Gravitas using its documented installation guide. Prefer Gravitas Core in this workspace unless I explicitly request Native. Verify the installed Core skill with its doctor script, tell me which mode is active, and do not alter unrelated Antigravity configuration.

Use it

You do not need a ceremonial prompt. GRAVITAS is designed to activate for implementation, debugging, refactoring, review, and migration work. Naming it explicitly is useful when you want to be certain of the workflow.

Use Gravitas to find and fix why pagination occasionally skips records. Verify the fix.

Use Gravitas to add rate limiting to this API without breaking existing clients.

Use Gravitas to review this authentication change for correctness and regression risk. Do not modify files.

Use Gravitas in plan mode. Investigate migrating this project to PostgreSQL and give me an implementation plan. Do not change anything.

What GRAVITAS changes

Without: task → edits → “done”
With:    task → contract → reconnaissance → scoped change
        → external validation → recovery when needed
        → evidence-backed completion

It does not assume a model is careless. It makes the engineering loop explicit: strong model + a better harness can produce a more auditable process.

Core vs Native vs OpenCode

| Capability | Core | Native (Antigravity) | OpenCode | |---|:---:|:---:|:---:| | Agent Skills compatible | ✓ | ✓ | ✓ | | Planning, read-before-write, verification policy | Instruction | Instruction + hook checks | Instruction + permissions | | Adaptive effort and plan-mode guidance | Instruction | Instruction + action lock | Instruction + permission profiles | | Evidence ledger | — | Enforced runtime record | Manual via CLI | | Pre-tool checks | — | Enforced where host hooks apply | Shim + permissions | | Completion gate | — | Enforced where host hooks apply | Manual (summarize + evidence) |

Core is the practical default. Native is optional, and its hooks are workflow controls—not an operating-system sandbox. OpenCode setup is one command: gravitas init --host opencode (see quickstart and hosts).

Evidence

There are 30 registered task specifications across 10 planned lanes in GravitasBench. They are benchmark design and infrastructure, not completed model results. No public FSR, requirement-coverage, false-completion, token, time, Claude-gap, or model-ranking number belongs in this README yet.

The repository does verify implementation behavior: the test suite covers action locking, scope checks, session isolation, evidence-chain tampering, completion gating, recovery state, schemas, and benchmark plumbing. Those are runtime tests, not evidence of model performance.

When comparative results are eligible, they will disclose the task corpus, unique-task and run counts, model and effort, Antigravity and GRAVITAS versions, date, raw trajectories, methodology, uncertainty, and overhead. Results will apply only to that listed configuration—not as universal model rankings.

Read the benchmark protocol, release gates, and install verification record.

Why harness engineering?

The research case is modest and concrete:

• ReAct supports grounding reasoning in actions and observations.
• Plan-and-Solve motivates explicit decomposition for multi-step reasoning; applying that to coding is a GRAVITAS hypothesis.
• Huang et al. found that unaided self-correction can fail on studied reasoning tasks, which is why GRAVITAS privileges external validators.
• SWE-bench motivates testing repository state instead of trusting final prose.
• Recent Harness-Bench and Harness-IF preprints motivate evaluating model-plus-harness configurations as systems.

How it works

flowchart LR
  U[User task] --> S[GRAVITAS skill]
  S --> C[Task contract]
  C --> P[Adaptive execution]
  P --> T[Agent tools]
  T --> V[Validators and evidence]
  V --> G{Completion gate}
  G -->|missing proof| P
  G -->|evidence complete| R[Verified result]
  N[Native hooks] -. guard .-> T
  N -. gate .-> G

Core supplies the behavioral policy. Native wraps selected Antigravity tool events with local checks. For implementation detail, see architecture.

Adaptive execution

GRAVITAS scales effort with risk, uncertainty, ambiguity, blast radius, and failed attempts. A read-only question should not trigger a committee; an auth, schema, or destructive change should receive deeper reconnaissance and stronger verification.

The built-in profiles are eco, balanced (the design default), deep, and team. They describe intended process cost; no profile performance advantage is claimed before comparative data exists.

Verification and trust

The Core doctor reports the installed skill path and content hash:

python3 .agents/skills/gravitas/scripts/doctor.py

Update Core with:

npx skills update gravitas --project --yes

Update Native by pulling the repository, rebuilding the bundle, and re-running the same agy plugin install command. Full uninstall instructions are in docs/install.md.

Native hooks are defense in depth. Shell classification is tokenizer-based with capability levels, not a regex denylist — but hooks still do not replace host permissions, sandboxing, code review, or CI. See SECURITY.md.

Documentation

• Quickstart (Core, Native, OpenCode)
• Concepts (contract, context, evidence, policy, recovery)
• Hosts (capability differences, honestly)
• Configuration (profiles, overrides, contracts)
• Benchmark methodology
• Install and troubleshooting
• Contributing (architecture + test expectations)
• Security policy (what is and is not enforced)
• Research foundations
• Architecture decisions

Contributing

Contributions that improve evidence are especially useful: licensed benchmark tasks, hidden deterministic validators, live Antigravity compatibility checks, privacy review, and independent replications. Start with CONTRIBUTING.md.

Citation

Use CITATION.cff and pin a commit until a versioned release and archived benchmark dataset exist.

License

MIT © Uday Sharma.