Hermes plugin · v0.1.1

Hermes Sentinel start card — Audit · Block · Truncate

Use case

Hermes agents can loop on identical tool calls, overfill context with oversized tool results, and surface credential-shaped strings into model prompts. Operators need deterministic guards and a local record of interventions without sending telemetry.

How it works

Four Hermes lifecycle hooks: pre_tool_call blocks the third consecutive identical tool call in a session; transform_tool_result redacts credential patterns and truncates oversized results before the model sees them; pre_llm_call blocks a new user message that contains a credential-shaped string; post_tool_call records dispatch outcomes in a local SQLite ledger.

Inspect the implementation

The public repository documents hook behavior, install/doctor commands, CLI/slash commands, compatibility floor (Hermes >= 0.22.6), and verification notes. The start-card image comes from the repository assets.

Scope

Local-only: zero external network requests and no telemetry. Credential detection uses patterns and is not exhaustive secret detection. The audit ledger is best-effort (in-memory fallback if storage fails); loop and credential guards remain active. Linux/commercial-model interactive TUI paths remain incompletely verified per the public README.

Execution and evidence

Audit events store tool names, timestamps, sizes, outcomes, and redacted argument previews. Raw prompt and tool-result bodies are not stored. Ordinary unchanged tool results do not create a separate transform event.

The documented run envelope is one search, at most five source pages, one model call, depth 4, and a 24-step budget.

Review procedure

hermes plugins validate .; hermes plugins doctor . --ci; hermes sentinel status; hermes sentinel audit

Content reviewed against repository documentation on November 5, 2026.