Skip to content

Releases

Antigravity Harness v2.1.1

Choose a tag to compare

@hadbilen hadbilen released this 16 Nov 01:04

Cross-platform release & CI hardening for the 2.1.x line (ubuntu-24.04, macos-15, windows-2022), Windows path / file:// URI / guard.log handle & 8.3 short-path fixes, PyInstaller frozen guard.hooks package resolution, and full standalone binary + Linux package asset publication (681 hermetic tests).

Highlights in v2.1.1

  • Cross-platform CI & release asset pipeline (.github/workflows/ci.yml, skills/media-studio/scripts/media_qc.py). Migrated macOS test and build-binaries runners from retired macos-14 to macos-15 (arm64), provisioned ffmpeg/ffprobe across Linux, macOS (ffmpeg-full for libvorbis/libwebp parity under Homebrew FFmpeg 8+), Windows (choco + MSYS2 pacman fallback), and trusted-boundary, and pinned utf-8 (errors="replace") subprocess decoding in media_qc.py for Windows cp1252 environments.
  • Transient 0o600 audit log handler (guard/audit_log.py). Replaced persistent open RotatingFileHandler descriptors with an RLock-guarded _TransientRotatingFileHandler that enforces 0o600 permissions across log rollovers, releases file handles immediately after each emit (preventing Windows WinError 32), and recovers automatically from transient open errors without latching NullHandler.
  • Windows path, file:// URI & Governance Shield parity (guard/hooks/pre_tool_use.py, guard/agy_hooks.py). Added token-anchored Windows drive normalization (C:\...), root-clamped C:/.. collapsing, shared _parse_file_uri (file://C:\..., file://C:/..., file:///C:/..., %5c/%3A decoding, and %00 null-byte rejection across write_file and MCP arguments), escaped-backslash inline script detection, Windows %LOCALAPPDATA%\Programs\antigravity-guard protection, .parent / space-quoted plugins scrubbing guards, backslash out-of-workspace rm/mv and .agents\hooks.json gates, and a PyInstaller sys.frozen PYZ guard in guard/agy_hooks.py.
  • agy-guard init, MCP config migration & doctor fixes (guard/cli.py, guard/environment.py, guard/doctor.py, install.py). Preserved lexical non-8.3/non-symlink-expanded --target paths in cmd_init while rejecting symlinked --target roots and fixing _uninstall_init_profile template cleanup, supported POSIX /-rooted antigravity-harness script repointing on Windows in migrate_harness_mcp_config, kept ensure_skill_script_permissions directly testable on Windows, and included normalized relative .agents/hooks.json paths in doctor collision warnings (681 hermetic tests across Linux, macOS, and Windows on Python 3.10–3.14).

Cumulative 2.1.0 architectural capabilities included

  • In-slot constitutional slimming & 34 -> 20 skill consolidation (ADR-0009). GEMINI.md condensed below 6.5 KB (Rule 4: Simplicity First (YAGNI)), 20 top-level skills with automatic RELOCATED_SKILL_SCRIPTS mcp_config.json migration, and transcript-aware mid-turn PreInvocation silence.
  • Lazy event-scoped guard/hooks/ package (ADR-0008). -98.6% isolated guard/agy_hooks.py import latency and -38% to -52% cold hook CLI startup.
  • Antigravity v2.32.0 systemMessage, Swarm / delegation gates, linked-worktree TestBoundaryGuard, and Rust 2024 / game / media / legacy modernization (ADR-0005–ADR-0007).

Verify all standalone binaries (agy-guard-linux-x86_64, agy-guard-macos-arm64, agy-guard-windows-x86_64.exe) and Linux packages (.deb, .rpm, .pkg.tar.zst, PKGBUILD) against their accompanying .sha256 checksum files before installation.

Antigravity Harness v2.1.0

Choose a tag to compare

@hadbilen hadbilen released this 15 Nov 22:44

Constitutional slimming (GEMINI.md < 6.5 KB, Rule 4: Simplicity First (YAGNI)), 34 -> 20 top-level skill consolidation with automatic mcp_config.json script migration (ADR-0009), lazy event-scoped guard/hooks/ modularization (ADR-0008), native Antigravity v2.32.0 systemMessage and swarm / delegation hook coverage (ADR-0005, ADR-0007), linked-worktree TestBoundaryGuard support, and ecosystem auto-routing + Rust 2024 / game / media / legacy modernization (ADR-0006; 680 hermetic tests).

Breaking & structural changes

  • In-slot constitutional slimming (GEMINI.md, ADR-0009). Condensed GEMINI.md below 6,500 UTF-8 bytes (>= 63% reduction from 17.9 KB) while preserving canonical Rule 1..18 numbering and Rule 15 (a)–(h) collision anchors; elevated Rule 4: Simplicity First (YAGNI) across GEMINI.md and all 9 auditor subagents (agents/*.md).
  • 34-to-20 top-level skill consolidation (skills/, ADR-0009). Folded 14 sub-module and workflow-helper skills into on-demand references/ and scripts/ directories under skills/antislop/, skills/harness/, skills/technical-writing/, and skills/audit/, cutting per-turn <skills> frontmatter load while keeping 100% of scripts, assets, templates, /slash routing (SKILL_HINTS), and Porter suitability slugs (porter/analyzer.py). install.py and agy-guard init automatically repoint relocated mcp_config.json script entries (RELOCATED_SKILL_SCRIPTS) and retire superseded skill directories.
  • Event-scoped guard/hooks/ package (guard/agy_hooks.py, guard/hooks/, ADR-0008). Split the 11,390-line guard/agy_hooks.py monolith into a 347-line lazy CLI dispatcher and PEP 562 compatibility facade backed by guard/hooks/{common,pre_invocation,pre_tool_use,stop_gate}.py, cutting isolated module load latency by 98.6% (236.10 ms -> 3.22 ms) and cold hook CLI startup by 38%–52% while maintaining full symbol parity.

Lifecycle hooks, Guard & Porter

  • Antigravity v2.32.0 contract & top-catalog model calibration (ADR-0005, ADR-0007). Migrated PreInvocation and upstream_watcher.py from deprecated ephemeralMessage to native systemMessage (metadata.user_markdown for operator-only notices), added Gemini 4 Hydrogen (XHigh), Gemini 4 Carbon (High/XHigh), and Orchestrator effort parsing, and added fail-closed MODEL_PLACEHOLDER_* escalation with <USER_SETTINGS_CHANGE> transcript override resolution.
  • Swarm, delegation & workflow PreToolUse gates (ADR-0005, ADR-0007). Extended hooks.json and PreToolUse to cover swarm, post_message, send_message, define_subagent, and schedule (alongside run_workflow AST inspection), blocking multi-step worker/coordinator customization model downgrades and validating daemon cron schedules.
  • Transcript-aware mid-turn PreInvocation silence (ADR-0009). Inspected bounded transcriptPath tails so mid-turn tool continuations stay silent without prematurely clearing "in_flight" recovery records (preserving post-crash recovery even on repeated short prompts) and prevented zero-match grep/rg/diff/cmp/test exit codes 1/2 from escalating to build-error-resolver.
  • Linked-worktree TestBoundaryGuard, Stop & doctor hardening (ADR-0007, ADR-0008). Added resolve_linked_worktree_root (tasks.md authorization fallback), diff + AST assertion relaxation detection (== -> >=, !=, is not None, toBeTruthy), black-listed abnormal stop reasons so normal completions (NO_TOOL_CALL, model_stop, TERMINAL_STEP_TYPE, INJECTED_RESPONSE) always run Stop gates, shielded ~/.gemini/{GEMINI,AGENTS}.md, expanded Bash Guard destructive SQL/FS/container checks, and added plugins/ + .agents/skills/ symlink audits in guard/doctor.py.
  • Porter shared-negation & untracked audit coverage (porter/sanitizer.py). Preserved binary comma-free shared imperative prohibitions across bare or (e.g. "Never weaken assertions or delete tests") and paired git diff with bounded git ls-files --others --exclude-standard inspection across harness and audit.

Skill & ecosystem updates (ADR-0006, ADR-0007)

  • Ecosystem auto-routing: prioritized explicit /slash and backticked slug mentions before regex gating, added tiered hint saturation, bounded workspace/transcript tail signals, Swarm vs. unlazy vs. marathon topology rules, and Tier-3 Pre-Code Plan red-teaming (YAGNI Brake).
  • rustify: updated to Rust 2024 Edition (#[unsafe(no_mangle)], unsafe extern "C", &raw const/&raw mut, core::mem::DropGuard, Cargo min-publish-age cooldowns, and C++ virtual/template/std::move matrices).
  • game-engineering, media-studio & vibecoder: split Profile A (Web/Canvas 16.67 ms) and Profile B (3D/Native), added save schemaVersion quarantine and clean-room rea_rea bridges, and added media_qc.py GAME-1..GAME-3 checks (--game-sfx, --ogg-codec, --check-loop, --trim-loop, --pixel-art, --sprite-atlas).
  • sweetspot & renew: defined --net Controllable vs. Uncontrollable External Path Null-Action thresholds (sweetspot) and [HARDWARE/ENV SHIM] record/replay + Turkish/decimal locale invariants (renew).
  • Upstream sync: synchronized morluto/rea (rea-agents-6.9.0, 255 MCP tools), miqdadbadjuber/anti-slop (40 rules), and mattpocock/skills.

Verification & Architecture Decision Records

  • Recorded ADR-0005 through ADR-0009 under docs/adr/ and expanded the hermetic standard-library test suite from 570 to 680 tests across Linux, macOS, and Windows (Python 3.10–3.14).

Verify downloads with the accompanying .sha256 files. Note: standalone binaries and Linux system packages (.deb, .rpm, .pkg.tar.zst, PKGBUILD) are built and uploaded automatically by the tag CI pipeline once matrix verification completes.

Antigravity Harness v2.0.0

Choose a tag to compare

@github-actions github-actions released this 30 Oct 06:55

Native Antigravity 2.0 lifecycle hooks (PreInvocation, PreToolUse, Stop), full ecosystem auto-routing (skills, auditor subagents, MCP servers, and plugins), AST/dataflow/FTS5 code mapping (scripts/code_map.py), domain-aware UI archetypes (MODE: UTILITY, MODE: EDITORIAL, MODE: GAME & INTERACTIVE), deterministic UI slop scanning (check_ui_slop.py), 13 new engineering skills (34 total), and 2 new auditor subagents (9 total).

Breaking changes

  • Constitution expanded to 18 rules (GEMINI.md). Added Rule 9 C9 (prohibition on test-fitted production code and test-runner environment sniffing), Rule 17 top-catalog model and highest-effort requirement on customization/governance edits, and Rule 18 (Surgical Change & Reuse Discipline); codified built-in platform collision resolutions in Rule 15 (a–e).
  • Native Antigravity 2.0 hooks.json & guard/agy_hooks.py (harness-gate). PreToolUse enforces an explicit "decision" contract ("allow", "force_ask", "deny"), blocks direct agent writes to ~/.gemini/config governance seams (while permitting runtime subdirectories plugins/, sidecars/, projects/, cache/), blocks mutations to customization or governance files unless the active model and effort match the top catalog model (skills/upstream-auditor/upstream_state.seed.json), blocks agent execution of human-only agy-guard administration commands and --no-verify git bypasses, and prompts on destructive shell/git operations.
  • Domain-aware DESIGN.md archetypes. Added MODE: UTILITY, MODE: EDITORIAL, and MODE: GAME & INTERACTIVE alongside the computed WCAG AA contrast table.

Added

  • guard/agy_hooks.py:
    • PreInvocation: bilingual (EN / TR) automatic skill suggestion across the skill catalog (SKILL_HINTS) plus automatic Subagent, MCP Server (rea_rea, antislop-contrast), and Plugin (rea) routing (ECOSYSTEM_ROUTING_HINTS) with per-conversation deduplication.
    • PreToolUse: Governance Shield, Top-Catalog Model & Effort Customization Gate (GEMINI.md Rule 17), and Bash Guard with command-boundary anchoring.
    • Stop: diff integrity gate catching tautological assertions, C9 test-runner sniffing in production code, unapproved test skip markers, malformed .json/.toml/.svg syntax, structural UI/game slop (UI-1–UI-5), and broken workspace-local Markdown links, guarded by a 3-strike circuit breaker.
  • scripts/code_map.py: zero-dependency Python AST profile, callers, flow (backflow/forwardflow), blast-radius (Tier 1/2/3 threshold check), gate-syntax, deterministic runtime trace, SQLite FTS5 BM25 passage search (fts-search), plus brace-aware C/C++ (.c/.h/.cpp/.cc/.cxx/.hpp/.hh/.hxx) function/method indexing and lifetime-safe Rust (.rs) lexer scrubbing (ADR-0003).
  • skills/antislop/scripts/check_ui_slop.py: deterministic structural scanner for UI/HTML/JSX/TSX/CSS files (UI-1 cloned card grids, UI-2 nested card-in-card boxification, UI-3 unearned visual tells, UI-4 raw RGB/CMY game primaries, UI-5 dead interactive stubs).
  • 13 new skills (34 total): test-design, mistake-proofing, refactoring, dependency-upgrade, retro, technical-writing, data-analysis, game-engineering, marathon, sweetspot, media-studio, renew, and rustify.
  • skills/rustify: 5-stage C/C++/Go/Python -> memory-safe Rust migration and FFI bridging package (SKILL.md, references/c-cpp-abi-and-embedded.md, references/managed-go-python.md) with two-stage unsafe purification and a 3-tier rustc/clippy/miri/sanitizer/conditional cargo kani verification ladder (ADR-0003).
  • skills/renew/SKILL.md: 5-stage contract-first modernization pipeline (Silent Autopsy -> Renewal Contract -> Golden Master or Equivalence Ledger -> Layered Strangler Cutover & Security Sanitization -> Parity Gate) for generational legacy software and classic games (ADR-0002).
  • skills/media-studio: container/header recovery (media_qc.py salvage), still-image restoration (JPG/PNG/TIFF/BMP with scanline decode verification), two-tier fidelity-bounded quality enhancement (references/salvage-and-enhancement.md), aspect-drift-locked compare, and side-by-side ab-sheet visual inspection (ADR-0002).
  • 2 new read-only auditor subagents (9 total): test-adequacy-auditor and contract-change-auditor.
  • skills/vibecoder/SKILL.md: added Game Prototype Mode (Game Vibe Mode) bridged with skills/game-engineering/SKILL.md and DESIGN.md (MODE: GAME & INTERACTIVE) for zero-friction single-file rich HTML game prototypes (<canvas> + Diegetic HUD + fixed-timestep loop + seeded PRNG + WebAudio Input Feedback Triad) and a 3-question Game Vibe-Grill.
  • skills/antislop-copywriting: added Turkish synthetic-marker rules (TR-1–TR-7) and monotonous sentence cadence (6.7) in check_synthetic_markers.py.
  • skills/upstream-auditor: added release-tag ("track": "release") and subpath tracking, registering morluto/rea (reverse-engineer-anything, 8 tracked repositories total).
  • Guard CLI, GUI & Packaging synchronization: .zip archive inspection in porter_bridge.py and GUI, tasks.md ## Authorized Test Modifications support in test_boundary.py, hooks.json/install_receipt.json checks in doctor.py, and updated PyInstaller/Linux packaging scripts.
  • Architecture Decision Records: formal ADR process under docs/adr/ (ADR-0001 through ADR-0004) covering v2 remediation, renew/media-studio, rustify/codemap, and tiered MCP hook policy with dual-layer Linux package activation (agy-guard init).
  • Bilingual Architecture: Full parity for English workflows with native, out-of-the-box Turkish intent detection in PreInvocation lifecycle hooks and Turkish AI-marker detection (TR-1–TR-7) in copywriting audits.
  • Terminal & Process Hardening: Sealed terminal and process injection vectors against PTY spawns, multiplexer escapes (tmux, screen), and controlling-terminal prompt bypasses (guard/approval.py).

Verify downloads with the accompanying .sha256 files.

v1.4.0 — Leaner Constitution, Reviewed Leases, Audit Trail & Safer Discovery

Choose a tag to compare

@hadbilen hadbilen released this 29 Sep 20:58

Governance-quality release: fewer always-on rules, more mechanical enforcement, and fixes for every finding of a fresh forensic audit of Guard, Porter and CI. Full details in CHANGELOG.md.

Highlights

  • Leaner constitution. GEMINI.md shrinks from 18.5 KB to 10.4 KB with rule numbers unchanged. Procedures moved into the harness and diagnosing-bugs skills. Handoffs now carry a goal and exit criteria and are triggered by observable events instead of a guessed "context saturation". A single end-of-phase Delivery Block reports progress as criteria met.
  • Leases you can trust. Records are matched by lease id, so a concurrent grant can no longer be erased and leave the tree open. Failed relocks are retried and logged, and a corrupt lease store fails closed. Changes made during an unattended window wait for review ([REVIEW]) instead of being trusted automatically.
  • Audit trail. Approvals, refusals, locks, leases, baselines, restores, boot checks and every WARNING or CRITICAL alert go to ~/.local/state/antigravity-harness/logs/guard.log.
  • Safer discovery. Other agents are discovered through their rule files only (CLAUDE.md, .claude/agents|commands|skills, .cursor/rules, …) and start as monitored. The home directory is never treated as a workspace; a GUI started from a launcher used to register ~/.claude and ~/.gemini wholesale.
  • Integrity and boot sentinel. Backup prefixes are excluded at the top level only, the sentinel locks before it verifies, forensic snapshots are deduplicated, and systemd unit injection through control characters is rejected.
  • Upstream watchdog. The PreInvocation notice is operator-only and shown once per change; silence it with upstream_watcher.py --ack. It no longer tells the agent to recommend /audit-upstream on every turn.
  • Porter. The sanitizer is clause-scoped and flags uncertain lines instead of deleting them. porter.py inspect bundle.zip inspects archives without extracting anything. The fetcher enforces an overall deadline and blocks more IPv6 forms. Exports keep executable bits and templates, and the Claude export writes tools: from the agents' new access: key.
  • New tooling. A deterministic scanner for signs of AI-written prose in antislop-copywriting (English and Turkish), a thin visualizer skill, and skill updates for deep-grill, audit, vibecoder, research and the auditors.
  • CI and contributions. The trusted-boundary job never falls back to the candidate's own verifier, and fork pull requests need maintainer approval before workflows run. See CONTRIBUTING.md.
  • Windows. The CLIs no longer crash on consoles that cannot encode a character, and fetch deadlines are reported correctly despite early timer firing.

Upgrade notes

  • New human approval gates: lowering an environment policy, notify disable, startup enable and snapshot prune (exit code 3 without confirmation).
  • After a lease that expired while you were away, run agy-guard status. If it shows [REVIEW], inspect the changes and accept them with agy-guard rebaseline.
  • Environments registered by 1.3.x with whole .claude, .cursor or .codex seams are migrated to the rule-file whitelist on first load.
  • Snapshots without a readable content manifest can no longer be restored.

Verify

Every asset has a .sha256 file. For example: sha256sum -c antigravity-guard_1.4.0_amd64.deb.sha256.

v1.3.1 — Honest Governance Remediation, Human-Only Administration & Lossless Porter Export

Choose a tag to compare

@hadbilen hadbilen released this 22 Sep 11:22

Antigravity Harness v1.3.1 — Honest Governance Remediation, Human-Only Administration & Lossless Porter Export

Antigravity Harness v1.3.1 is a remediation release for five independent external audits (121 validated findings). It makes Antigravity Guard enforce what it claims: locks never follow symlinks, administration requires a human, baselines and snapshots leave the tree they protect, leases relock on their own, Porter exports every skill file and agent losslessly, and the self-audit checks what the documentation actually says. The README now states an explicit threat model: a user-space lock is advisory against processes running as the same OS user.


Highlights & Features

1. Honest Write Shield & Threat Model (guard/os_adapter.py, guard/environment.py, README.md)

• Governance-Seam Lock Scope: Locks GEMINI.md, AGENTS.md, DESIGN.md, MISTAKES.md, hooks.json, mcp_config.json, skills/, agents/, templates/ and .harness/ only. Antigravity's own runtime state (config.json, projects/, sidecars/, plugins/) and the config root stay writable, because Antigravity saves config.json by atomic replace.
• Symlink-Safe Locking: Never follows symlinks, reports links that leave the tree as unprotected, records original permission modes and restores them exactly, and fails loudly (PARTIAL LOCK) instead of reporting success.
• Windows ACL Status: Lock state is read from the NTFS ACL instead of writing probe files.
• Explicit Threat Model: agy-guard status reports USER-SPACE LOCK (advisory against the same OS user); the README documents what is and is not protected.

2. Human-Only Administration (guard/approval.py, guard/cli.py, GEMINI.md Rule 12)

• --auto-approve Removed: unlock, rebaseline, snapshot restore, porter stage, test-boundary snapshot, startup disable and doctor --fix require a human to type yes in a terminal or confirm in the GUI; headless requests exit with code 3.
• Constitutional Rule: Rule 12 reserves Guard administration for the human operator; agents may only request a maintenance window.

3. Out-of-Tree State, Auto-Relocking Leases & Verified Restore (guard/paths.py, guard/lease.py, guard/integrity.py, guard/snapshot.py)

• Per-User State Directories: Registry, integrity baselines, leases and lock-mode records live in ~/.local/state/antigravity-harness/; snapshots and the runtime in ~/.local/share/antigravity-harness/ (XDG honoured, same layout on every OS).
• Leases That Relock: The lease is persisted before unlocking, a detached watcher relocks on expiry, several environments can hold leases, and closing relocks first, then writes a change report and a new baseline.
• Honest FIM: Symlinks are recorded as links, unreadable files are reported, a missing baseline is shown as BASELINE MISSING, and every rebaseline archives the previous one.
• Verified Journaled Restore: Snapshots exclude config.json (no token copies), are verified against a content manifest, refuse to write through destination symlinks, always take a pre-restore backup and roll back on failure.
• Boot Sentinel Quarantine: Drift at boot captures a forensic snapshot and raises a CRITICAL notification without rebaselining.

4. Porter: Lossless Export, Negation-Aware Sanitizer & SSRF Pinning (porter/, porter.py)

• Lossless Export: All 20 skills with every support file and in-skill symlink, plus all 7 subagents, for Claude Code, Cursor (valid .mdc YAML), Universal (AGENTS.md), Aider and Generic targets; existing files are only overwritten with --force.
• Strict Frontmatter Parser: Standard-library YAML subset parser/dumper verified against every skill and agent in the repository.
• Negation-Aware Sanitizer: Neutralises "skip the failing tests" while keeping "never skip tests"; covers multi-line and Turkish phrasing and applies to export as well as import.
• SSRF Hardening: DNS pinning against rebinding, globally routable destinations only (6to4, NAT64, site-local and IPv4-mapped loopback blocked), proxies ignored.
• Hash-Bound Staging: agy-guard porter stage writes exactly the reviewed content, refuses destination symlinks and always relocks.

5. Six-Pass Self-Audit & Hardened Graders (scripts/meta_audit.py, scripts/verify_invariants.py, Rule 17)

• Six Real Passes: Frontmatter loadability, cross-references, mutual exclusion and tier thresholds, export parity by content, CLI/README sync (every agy-guard example must parse; counts and policies must match the code) and recomputed DESIGN.md contrast ratios.
• --strict Mode: Warnings fail CI.
• Invariant Scanner: Detects bash <(curl …) and source <(…), checks Markdown only inside code fences, supports --base-ref and exits 2 when a check cannot run.

6. Supply-Chain Hardened CI & Packaging (.github/workflows/ci.yml, installers/)

• Pinned Actions & Tools: Every action is pinned to a commit SHA and build tools to exact versions; the matrix covers Linux, macOS and Windows on Python 3.10–3.14.
• Trusted-Base Boundary Job: Pull requests run the candidate code against the base branch's own tests and graders in a clean worktree.
• Verifiable Assets: Every release asset ships with a .sha256 file, assets are never clobbered, and the PKGBUILD checksum is rendered from the real tag tarball.
• License Compliance: Packages ship LICENSE and THIRD_PARTY_NOTICES.md and declare MIT and Apache-2.0 (bundled procoder skill).

7. Copy-Mode Installer & Upgrade Path from 1.3.0 (install.py)

• Real, Lockable Files: Governance files are copied into ~/.gemini/config (--link keeps symlinks for development), hooks.json is merged instead of replaced, and backups go outside the config tree.
• Clean Upgrade: In-tree snapshots and baselines, read-only backup_* folders and old runtime symlinks left by 1.3.0 are moved to ~/.local/state/antigravity-harness/ (mode 0700), and runtime state frozen by the 1.3.0 lock is made writable again. Nothing is deleted.


Breaking Changes

• request-unlock --auto-approve no longer exists; administration commands need a human (exit code 3 otherwise).
• Guard state moved from ~/.gemini/config and the repository to the per-user XDG state/data directories.
• The installer defaults to copy mode; the lock scope is the governance seams only.
• Test boundary bugfix mode now allows new regression test files; existing tests, runner configuration and fixtures must stay byte-identical.

Upgrading from 1.3.0

• Run python3 install.py from a checkout and confirm with yes; then check agy-guard status.
• Review and delete ~/.local/state/antigravity-harness/legacy/ once you no longer need it: 1.3.0 snapshots contained copies of config.json. Rotate any tokens stored there if the snapshots were ever shared.


Verification

• Unit Test Suite: 179/179 tests passing across Linux, macOS, and Windows on Python 3.10, 3.11, 3.12, 3.13, and 3.14.
• Deterministic Invariant Guard: Passed via scripts/verify_invariants.py --all.
• Autonomous Self-Audit: 0 critical, 0 warnings across all 6 passes via scripts/meta_audit.py --all --strict.
• Canonical Manifest: Up to date (porter.py manifest --check).
• CI Matrix & Release Artifacts: Standalone executables for Linux, macOS, and Windows plus .deb, .rpm, and .pkg.tar.zst packages, each with a .sha256 checksum.

v1.3.0 — Native Linux Packaging, Multi-Environment Governance & Autonomous Self-Audit

Choose a tag to compare

@github-actions github-actions released this 20 Sep 07:48

Antigravity Harness v1.3.0 — Native Linux Packaging, Multi-Environment Governance & Autonomous Self-Audit

Antigravity Harness v1.3.0 marks a major architectural milestone, introducing native Linux distribution packaging (.deb, .rpm, .pkg.tar.zst), multi-environment OS write protection across foreign coding agents (Claude Code, GPT Codex, Cursor, Aider), human-in-the-loop lease-based unlock, a low-frequency ergonomic desktop notification engine, and the autonomous 5-pass harness self-audit engine (Rule 17).


Highlights & Features

1. Native Linux Distribution Packaging (installers/packaging/package_linux.py, PKGBUILD)

• Pure Python Debian Package Builder: Zero external dependencies; implements a pure Python ar archive generator producing compliant .deb packages with md5sums, Debian control fields, and FHS directory structures.
• RPM Package Generation: Automated spec file generation and rpmbuild execution for Fedora, RHEL, openSUSE, and AlmaLinux.
• Arch Linux Package: Generates .pkg.tar.zst packages with complete .PKGINFO metadata for Arch Linux and Manjaro.
• FHS Standards Compliance: Installs executable to /usr/bin/agy-guard, library modules to /usr/lib/antigravity-guard/, documentation to /usr/share/doc/, scalable application icon to /usr/share/icons/hicolor/scalable/apps/, and Bash completion script.

2. Multi-Environment Governance Matrix (guard/environment.py, guard/cli.py)

• Co-Located Agent Governance: Discovers and manages foreign agent runtimes (Claude Code, GPT Codex, Cursor, Windsurf, Aider).
• Seam Protection Without Project Lock: Locks and verifies discrete governance seam files (CLAUDE.md, .cursorrules, AGENTS.md, .aider.conf.yml, GEMINI.md) while leaving active project source files completely writable.
• Cross-Environment Drift Matrix: agy-guard drift compares configuration drift across all registered environments.
• CLI Environment Management: agy-guard env (list|detect|add|remove|policy).

3. Human-in-the-Loop Time-Bounded Lease Unlock (guard/lease.py, guard/cli.py)

• Zero Permanent Exposures: Replaces permanent manual unlock with operator-controlled time leases (agy-guard request-unlock --duration 60 --reason "...").
• Auto-Relock & Rebaseline: When the lease duration expires or upon running agy-guard lock-complete, the environment is automatically relocked and rebaselined cryptographically.

4. Low-Frequency Ergonomic Notification Engine (guard/notifier.py, guard/cli.py)

• Native OS Desktop Integration: Dispatches alerts via notify-send (Linux DBus), AppleScript (macOS), PowerShell Toast (Windows), and terminal bell.
• Spam-Resistant Cooldown Cache: Persistent cooldown cache (.notify_cache.json) enforcing 30-minute intervals for INFO, 15-minute intervals for WARNING, and immediate dispatch for CRITICAL alerts.
• CLI & Quiet Modes: Supported --quiet and --disabled modes ensuring zero developer distraction during deep coding workflows.

5. Autonomous Harness Self-Audit Engine (scripts/meta_audit.py, agents/meta-auditor.md, Rule 17)

• Deterministic 5-Pass Self-Auditor:

  1. Schema Validation: Validates YAML frontmatter across all skills and subagents.
  2. Cross-Reference Integrity: Checks dead links, templates, and subagent invocations.
  3. Mutual Exclusion Barriers: Ensures divergent skill modes (vibecoder vs harness, procoder vs standard) do not collide.
  4. Universal Manifest & Porter Parity: Validates lossless transpilation across all target platforms.
  5. CLI & Architecture Sync: Confirms subcommands and agents match documentation.
    • Constitutional Invariant: Enforced as Rule 17 in GEMINI.md.

6. Expanded Subagent Suite & Test Boundary Hardening

• meta-auditor Subagent: Dedicated subagent executing harness meta-consistency checks.
• Total Subagents: 7 specialized autonomous auditor subagents.


Verification

• Unit Test Suite: 77/77 tests passing across Linux, macOS, and Windows on Python 3.10, 3.11, and 3.12 (OK, ~4.1s).
• Deterministic Invariant Guard: 100% verified via scripts/verify_invariants.py --all.
• Autonomous Self-Audit: Verified 0 critical, 0 warnings across all 5 passes via scripts/meta_audit.py --all.
• CI Matrix & Release Artifacts: Complete build and release of .deb, .rpm, .pkg.tar.zst, and standalone executables for Linux, macOS, and Windows.

v1.2.9 — Test & Config Trust Boundary Shield, Run Provenance & Reproducibility Gate

Choose a tag to compare

@github-actions github-actions released this 20 Sep 06:07

Antigravity Harness v1.2.9 — Test & Config Trust Boundary Shield, Run Provenance & Reproducibility Gate

Antigravity Harness v1.2.9 introduces architectural trust boundaries for test suites and configurations, eliminating Goodhart gaming and subtle test weakening through cryptographic trust hashing, run provenance manifests, 2x isolated reproducibility verification, and base-commit CI splits.


Highlights & Features

1. Test & Configuration Trust Boundary Shield (guard/test_boundary.py, guard/cli.py)

• External Trust Anchor: Establishes test suites (tests/, spec/) and test runner configs (pytest.ini, setup.cfg, tsconfig.json, package.json, jest.config.*) as immutable verification anchors.
• Dual Verification Modes:

  • bugfix: Strictly forbids test mutations, additions, or deletions without authorization.
  • tdd: Permits adding new test files or appending test cases while preventing assertion relaxation on existing tests.
    • Anti-Goodhart Gaming: Detects and flags fixture tampering, mock relaxing, timeout extensions, and assertion omissions.
    • CLI Commands: agy-guard test-boundary snapshot and agy-guard test-boundary verify [--mode bugfix|tdd].

2. Run Provenance & Execution Audit Trail Manifest (guard/provenance.py, guard/cli.py)

• Deterministic Audit Trail: Generates .harness/provenance.json capturing base commit SHA, touched files, boundary verification status, and environment overrides (MOCK_*, SKIP_*).
• Markdown Delivery Proofs: Emits formatted execution proofs for delivery gates and compliance tracking.
• CLI Commands: agy-guard provenance generate and agy-guard provenance status.

3. Targeted Reproducibility Gate (guard/test_boundary.py, guard/cli.py)

• Flakiness Elimination: Added agy-guard test-boundary run-reproducible --cmd "<test-command>" executing target tests across consecutive isolated runs (default: 2x) to verify determinism and eliminate timing jitter.

4. Base-Commit CI Trust Boundary Split (.github/workflows/ci.yml)

• Trusted Baseline Testing: CI checks out test suites directly from the base branch (git checkout origin/${{ github.base_ref }} -- tests/), ensuring candidate PRs are judged against trusted, unmodified test assertions.


Verification

• Unit Tests: 52 passing unit tests across test_guard.py, test_test_boundary.py, and test_provenance.py.
• Invariant Guard: 100% verified (scripts/verify_invariants.py --all).
• CI Matrix: Full multi-OS matrix passing across Linux, macOS, and Windows.

v1.2.8 — GUI Upstream Resilience, Safe Key Resolution & Model Drift Alignment

Choose a tag to compare

@github-actions github-actions released this 19 Sep 23:50

Antigravity Harness v1.2.8 — GUI Upstream Resilience, Safe Key Resolution & Model Drift Alignment

Antigravity Harness v1.2.8 delivers critical UI stability fixes for the desktop GUI upstream dashboard, ensuring defensive data resolution for model drift and tracked community repositories under all configuration states.


Highlights & Features

1. GUI Upstream Dashboard Crash Fix (guard/gui.py, guard/upstream.py)

• KeyError Resolution: Fixed fatal KeyError: 'tracked_ecosystems' crash during AntigravityGuardApp initialization.
• Defensive Data Access: UpstreamAuditorBridge.get_model_drift_status() dynamically computes and returns "tracked_ecosystems" count with fallback defaults.
• Tab Construction Safety: _build_tab_upstream() in Tkinter GUI now accesses status dictionaries defensively with safe fallbacks, preventing crash loops when state files are missing or incomplete.

2. Upstream Drift Contract Test (tests/test_guard.py)

• Automated Regression Guard: Added test_upstream_auditor_model_drift_status_keys unit test ensuring drift status model and ecosystem contracts remain satisfied across releases.


Verification

• Unit Tests: 38 passing unit tests across tests/test_guard.py.
• GUI Smoke Test: Zero crash on desktop GUI initialization and tab switching.
• Invariant Guard: 100% passed (scripts/verify_invariants.py --all).

v1.2.7 — Decoupled Control Plane, Strict Baseline FIM & Upstream Resolution

Choose a tag to compare

@hadbilen hadbilen released this 19 Sep 23:11

Antigravity Harness v1.2.7 — Decoupled Control Plane, Strict Baseline FIM & Upstream Resolution

Antigravity Harness v1.2.7 hardens the installation and runtime boundaries by cleanly decoupling the control plane executable from configuration workspaces, enforcing strict baseline trust in File Integrity Monitoring, providing microsecond-precision snapshot tracking, and synchronizing upstream auditor state resolution.


Highlights & Features

1. Decoupled Control Plane & Policy Workspace (install.py, install.sh)

• Standalone Binary Symlinking: Control plane (agy-guard) installs and links cleanly to ~/.local/bin/agy-guard without touching or mutating existing user policy directories unless --full is specified.
• Safe Upgrades: Enables upgrading Guard executable logic independently of user prompt contracts and custom skills.

2. Strict Cryptographic FIM Trust Anchor (guard/integrity.py)

• No Blind Baseline Auto-Creation: If .guard_integrity.json is missing, verify() returns is_intact=False with [BASELINE MISSING] instead of silently establishing an unverified baseline.
• Tamper-Evident Security: Prevents rogue actors or wiped environments from masking configuration drift.

3. Atomic Re-Lock Verification (guard/porter_bridge.py)

• Post-Ingest Gate: Captures the return value of os_adapter.lock(); explicitly flags failure and stops execution if relocking fails, eliminating exposure windows after staging external rules.

4. Microsecond Precision Snapshots (guard/snapshot.py)

• High-Frequency Snapshot IDs: Switched timestamp format to %Y%m%d_%H%M%S_%f, completely eliminating file collision risks during automated or rapid consecutive snapshot creation.

5. Upstream Auditor State Synchronization (guard/upstream.py, scripts/upstream_watcher.py)

• Synchronized XDG Resolution: Unified state file paths across the headless watchdog and Guard CLI, prioritizing UPSTREAM_STATE_FILE and $XDG_STATE_HOME/antigravity-harness/upstream_state.json.

6. Top-Level Dependency & Dynamic Manifest Alignments (porter.py, porter/manifest.py)

• CLI Import Fix: Added top-level import re resolving NameError during porter.py import.
• Dynamic UniversalManifest Binding: Aligned version tracking dynamically across manifest compilation.


Verification

• Unit Tests: 37 passing unit tests across tests/test_guard.py.
• Invariant Guard: 100% passed (scripts/verify_invariants.py --all).
• Cross-Platform Matrix: Linux, macOS, and Windows standalone executables validated.

v1.2.6 — Pre-Session Boot Sentinel, Cross-Platform Autostart & Security Hardening

Choose a tag to compare

@github-actions github-actions released this 19 Sep 22:41

Antigravity Harness v1.2.6 — Pre-Session Boot Sentinel, Cross-Platform Autostart & Security Hardening

Antigravity Harness v1.2.6 introduces pre-session boot protection ensuring developer environments are locked and verified before desktop AI IDEs or agent processes start, robust SSRF immunity with safe redirect validation, strict path traversal guards across snapshot and staging bridges, and seamless symlink tracking in File Integrity Monitoring.


Highlights & Features

1. Pre-Session Boot Sentinel (guard/startup.py, guard/cli.py, guard/gui.py)

• Cross-Platform Early Boot Hook: Automatically locks and verifies configuration directories during user login before graphical IDE sessions initialize.
• Linux systemd Integration: Registers user-level agy-guard-boot.service (Before=graphical-session.target xdg-desktop-autostart.target).
• macOS launchd: Native LaunchAgent (com.antigravity.guard.plist) executing early-boot verification.
• Windows Task Scheduler: Configures high-privilege logon task (schtasks /SC ONLOGON /RL HIGHEST).
• CLI & GUI Controls: New commands agy-guard startup enable|disable|status, headless agy-guard boot-check, and graphical settings toggle.

2. SSRF Network Immunity & Safe Redirects (porter/net.py, guard/porter_bridge.py)

• Centralized Fetch Engine: Implemented SafeRedirectHandler preventing open-redirect SSRF bypasses to link-local (169.254.169.254), loopback, or private RFC-1918 subnets.
• Strict Protocol Filtering: Enforces HTTPS/HTTP scheme validation and safe DNS resolution across remote rule ingestion gates.

3. Path Traversal & Identifier Hardening (guard/snapshot.py, porter.py)

• Regex Snapshot Validation: Enforced strict regex matching (^[a-zA-Z0-9_-]+$) on all snapshot identifiers.
• Strict Path Containment: Enforces Path.relative_to boundary checks preventing directory traversal attacks during snapshot restore or file viewing.
• Target Identifier Sanitization: Sanitized --as-skill and --as-agent names against injection vectors.

4. File Integrity Monitoring Symlink Traversal (guard/integrity.py)

• Symlink Resolution with Cycle Detection: Added followlinks=True with visited directory inode tracking, guaranteeing modular skills installed via symlinks are continuously audited by FIM.

5. Runtime State Separation & Portability (guard/upstream.py, hooks.json)

• State Relocation: Moved mutable upstream_state.json to $XDG_STATE_HOME/antigravity-harness/, preventing write collisions when config is locked.
• Clean Hooks: Removed machine-specific paths from repository hooks.json.


Verification

• Unit Tests: 32 passing unit tests across test_guard.py.
• Deterministic Invariants: 100% verified via scripts/verify_invariants.py --all.
• CI Matrix: Standalone binaries built across Linux, macOS, and Windows.