{"slug": "thumbgate-1-28-0-a-safer-path-from-agent-feedback-to-enforcement", "title": "ThumbGate 1.28.0: A Safer Path from Agent Feedback to Enforcement", "summary": "ThumbGate 1.28.0 introduces a safer path from agent feedback to enforcement, focusing on the boundary between detecting risk and implementing prevention rules. The update includes an interactive quickstart, five explicit commands for inspecting operations, and a replay mechanism that tests candidate rules against previously allowed actions before promotion. The release also adds a Hermes Agent adapter, strengthens self-protection, and hardens package-integrity checks.", "body_md": "Most AI-agent safety systems are comfortable detecting risk. The harder problem is deciding when human feedback is specific enough to become enforcement, then proving that the new rule will not block legitimate work.\n\nThumbGate 1.28.0 focuses on that boundary.\n\nThe new interactive quickstart walks an operator through installing a first prevention rule:\n\n```\nnpx thumbgate@1.28.0 quickstart\n```\n\nThe flow is TTY-aware. In non-interactive environments it prints guidance rather than pretending an operator approved a rule.\n\nFive explicit commands make the operational surface easier to inspect:\n\n`/thumbgate-guard`\n\nturns a concrete mistake into a prevention rule.`/thumbgate-rules`\n\nshows active rules and learned lessons.`/thumbgate-blocked`\n\nreports what enforcement actually stopped.`/thumbgate-protect`\n\nshows branch and release governance.`/thumbgate-doctor`\n\nchecks hook, MCP, and agent-readiness wiring.A bare thumbs-down is evidence, not a permanent policy decision. In 1.28.0, a thumbs-down followed by a concrete `never ...`\n\ncorrection can take the immediate force-gate path. It is surfaced to the operator rather than silently installed.\n\nThumbs-up guidance remains guidance. An `always ...`\n\npreference does not automatically become a hard block because positive preferences rarely define a safe denial boundary on their own.\n\nThat distinction is deliberate: fast enforcement for explicit negative constraints, without turning vague sentiment into a lockout.\n\nBefore a candidate prevention rule hardens, ThumbGate replays it against previously allowed actions. A candidate that would catch known-good work is quarantined to warning mode instead of promoted.\n\nThis makes rule promotion testable. The question is not only \"does this match the failure?\" but also \"what legitimate behavior would this rule break?\"\n\nFor autonomous loops, setting `THUMBGATE_AUTONOMOUS=1`\n\nmakes approval gates fail closed when no human is present to answer. The release also strengthens self-protection around environment overrides and process-kill attempts.\n\nThe goal is not to claim that every risky action is universally blocked. It is to make the active enforcement posture visible, preserve explicit owner escape paths, and prevent an unattended agent from treating silence as approval.\n\nVersion 1.28.0 adds a Hermes Agent adapter and hardens package-integrity checks so dynamically loaded runtime files are verified before publication. Release checks now validate the package users actually install, not only the source tree that produced it.\n\nThumbGate is MIT-licensed and published on npm:\n\n```\nnpx thumbgate@1.28.0 quickstart\n```\n\nDisclosure: I built ThumbGate. Technical criticism of the feedback, promotion, and lockout tradeoffs is welcome.", "url": "https://wpnews.pro/news/thumbgate-1-28-0-a-safer-path-from-agent-feedback-to-enforcement", "canonical_source": "https://dev.to/igorganapolsky/thumbgate-1280-a-safer-path-from-agent-feedback-to-enforcement-38en", "published_at": "2026-07-13 10:01:02+00:00", "updated_at": "2026-07-13 10:15:57.859603+00:00", "lang": "en", "topics": ["ai-safety", "ai-agents", "developer-tools"], "entities": ["ThumbGate", "Hermes Agent"], "alternates": {"html": "https://wpnews.pro/news/thumbgate-1-28-0-a-safer-path-from-agent-feedback-to-enforcement", "markdown": "https://wpnews.pro/news/thumbgate-1-28-0-a-safer-path-from-agent-feedback-to-enforcement.md", "text": "https://wpnews.pro/news/thumbgate-1-28-0-a-safer-path-from-agent-feedback-to-enforcement.txt", "jsonld": "https://wpnews.pro/news/thumbgate-1-28-0-a-safer-path-from-agent-feedback-to-enforcement.jsonld"}}