GRAVITAS skill
Skills and hooks that make Gemini in Antigravity and OpenCode close a task only after a passing check
- 390 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The task contract runs verification commands defined in the project
- With hooks on it blocks writes outside the declared file list and calls the classifier flags as dangerous
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/gravitasDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The task contract runs verification commands defined in the project
- With hooks on it blocks writes outside the declared file list and calls the classifier flags as dangerous
This is third-party code. Review the repository files before installing.
What it does
GRAVITAS is a set of skills with optional hooks that teaches a coding agent engineering discipline: read the file, its tests and callers first, then edit. The main gravitas skill describes six rules, and separate skills cover high-stakes work, security and performance. There are three install levels. Core is skills only with no enforcement, and the author treats it as the default. Native adds an Antigravity plugin with hooks before and after tool calls and a check when the session stops. For OpenCode there are permission profiles and a shim that checks every tool call. With hooks on, the agent declares which files may change up front, and writes outside that list are denied. A task counts as done when a check run by GRAVITAS itself after the last edit returned success. Repeating an identical failed call is blocked until the diagnosis or strategy changes. Every step goes into a hash-chained ledger.
Who it is for. For developers who use Gemini in Antigravity or OpenCode and want the agent to back up finished work with real checks.
Good fit when
- Gemini in Antigravity edits files blind or declares "done" without running tests
- You need to limit repository writes to the list of files declared at the start of a task
- The agent loops and repeats the same failed call, burning quota
- A team wants a decision ledger showing which check confirmed each requirement
- A multi-file change or migration needs a plan whose depth scales with the effort
Not a fit when
- You need OS-level isolation: a container or a separate OS user
- You need a proven quality gain: the author publishes no comparative measurements yet
- You need hard blocking without hooks: in Core the rules stay an instruction to the model
- A one-line fix: the contract and checks do not pay off
Example request
Use Gravitas to fix the failing test in the payments module and verify the result by running the testsLimitations
The project is a research preview. The author has not measured its effect yet: the repository holds 30 benchmark tasks with no results, and every claim of benefit is labelled a hypothesis. OpenCode needs python3 and the gravitas command on PATH, and session completion there is checked manually. Antigravity hooks run with the agent's own privileges and are not a sandbox. The classifier only estimates how dangerous a shell command is, so it treats any unknown command as mutating.
How to disable. Remove Core with `npx skills remove gravitas`. Uninstall the Antigravity plugin with the agy plugin command; for OpenCode delete the block from opencode.json and the task contract file. Detailed steps are in docs/install.md.
FAQ
How do I install GRAVITAS for Gemini in Antigravity?
The base setup is `npx skills add udaysharmadev/Gravitas --skill gravitas --agent antigravity --yes`. For hook mode, clone the repository, run `bash scripts/build-native-bundle.sh` and install the bundle with `agy plugin install`.
How does Core differ from Native and OpenCode?
Core is skill text only. Native adds the Antigravity plugin: writes outside declared files are blocked, the ledger is automatic, and the session cannot stop without evidence. In OpenCode a shim applies the same policy to tool calls, while the ledger and final check are run manually by command.
What does "closed with evidence" mean?
A criterion counts as met only after a check launched by GRAVITAS itself exits with code 0 later than the last file write. A green run from before the final edit, the agent's self-review and a chat summary do not count.
How does GRAVITAS size a plan?
The `gravitas decide --signals` command emits a record with a planning level and reason codes. There are four levels: direct, compact, deep and replan. A one-line fix in one file gets direct, an auth migration across twelve files gets deep with a reviewer.
Does GRAVITAS protect against a hostile agent?
No. Hooks run at the agent's own privilege level, so they catch accidents and make tampering visible, but they do not replace a container, a separate OS user or read-only access.
Discussion
No questions or reviews yet.
Related
SkillsAn adversarial code-audit skill: Hunter finds bugs, Skeptic challenges them, Referee rules, and no source edit happens without your permission
Unit tests for skills: checks that an agent finds and uses your skill. Node 20 and Docker are needed, plus a model key
Reviews the git diff and the conversation with the agent: catches bugs and gaps between your request and the agent's result. CLI, skill, GitHub Action
Antithesis Skills
antithesis-skills
Official Antithesis skills: the agent analyzes your system, picks reliability properties, builds a harness, launches tests and triages the results
OpenQodex reviews your change before a push: scanners with no model first, then a separate Claude Code or Codex process triages findings on your login
A Claude Code and OpenCode plugin that will not let a session end until 1C edits pass standards, architecture and XML checks
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| GRAVITAS (this entry) | no | 390 | |
| RigorPilot Skills | no | 486 | runs research repositories and keeps verifiable run evidence |
| Trail of Bits Claude Code Config | no | 2,143 | a roundup of Claude Code settings: sandbox, permissions and hooks, with no evidence ledger |
README in short
The README opens with typical coding-agent failures: editing unread files, overly broad changes, running tests before the last edit. Six rules and three install options follow. The end covers the ledger hash chain and the test suite.
README badge
Original SKILL.md
---
name: gravitas
description: >
Use for software implementation, debugging, refactoring, review, migration,
and other repository-changing engineering work where correctness must be
verified. Enforces scoped reconnaissance, adaptive planning, evidence-backed
verification, requirement tracking, failure recovery, and budget-aware
delegation. Model-agnostic reliability runtime for coding agents --
measured, not claimed.
license: MIT
---
# GRAVITAS -- Engineering Reliability Runtime
**One rule above all:** Report what actually happened, not what you intended.
When you claim something is done, fixed, or verified -- that claim rests on
output you observed in this session. If you did not check, say you did not
check.
---
## THE 6 INVARIANTS
These hold across models and hosts. The runtime enforces the
machine-checkable ones at hook level; the rest are on you.
1. **Ground before mutating.** Read the target, its tests, and its callers
before editing. Never act on assumed file state.
2. **Stay in declared scope.** Writes outside `allowed_write_scope` are
denied. If the scope is wrong, amend the contract -- don't route around it.
3. **Completion needs evidence.** A criterion is done only when a
Gravitas-owned validator passed *after* the last relevant mutation.
Stale evidence and self-review don't count.
4. **Never exact-retry a failure.** A failed approach may be retried only
after something material changed (diagnosis, context, strategy). The
runtime blocks byte-identical retries.
5. **Satisfy the contract.** Acceptance criteria and required validators in
`contract.json` gate completion. Generic "tests passed" doesn't cover
unrelated criteria.
6. **Say what you can't verify.** Mark uncertain or unverifiable claims
explicitly. Downgrade confidence instead of fabricating certainty.
---
## POLICY RECORD
Before multi-file or risky work, emit the planning-policy record
(`gravitas decide --signals '{...}'` or the equivalent judgment):
```json