All plugins

proofgate

No proof, no done. Verifies an agent's completion claims (file created, message sent, command ran, tests passed, figures quoted) against the tool evidence recorded in the same session, before the final reply is delivered. Unproven claims are flagged as unverified or withheld, fail closed, and every verdict is written to an append-only ledger stored locally in $HERMES_HOME/proofgate/. Disclosure — observes every tool call and final reply in profiles where it is enabled and stores tool arguments (sensitive-named keys redacted) and the first 800 characters of every tool result, unredacted, in append-only files under $HERMES_HOME/proofgate/ with no rotation; makes no network calls; in the default enforce mode it holds reply streaming and may append warnings to or entirely withhold a reply judged by keyword heuristics to contain unproven claims.

Whydiriki700Toolsv0.1.0Updated 53 minutes ago

About

No proof, no done. Verifies an agent's completion claims (file created, message sent, command ran, tests passed, figures quoted) against the tool evidence recorded in the same session, before the final reply is delivered. Unproven claims are flagged as unverified or withheld, fail closed, and every verdict is written to an append-only ledger stored locally in $HERMES_HOME/proofgate/. Disclosure — observes every tool call and final reply in profiles where it is enabled and stores tool arguments (sensitive-named keys redacted) and the first 800 characters of every tool result, unredacted, in append-only files under $HERMES_HOME/proofgate/ with no rotation; makes no network calls; in the default enforce mode it holds reply streaming and may append warnings to or entirely withhold a reply judged by keyword heuristics to contain unproven claims.

ProofGate

No proof, no "done".

ProofGate is a Hermes Agent plugin that verifies an agent's completion claims against the tool evidence recorded in the same session, before the final reply reaches you.

Most plugins give agents new powers. ProofGate does the opposite: it holds the agent accountable for the powers it already has.

Why

The most dangerous failure mode of an autonomous agent is not a wrong answer. It is a false report that work got done: a file "created" that does not exist, a message "sent" that actually failed, a price "checked" that was never looked up. You find out later, usually after acting on the report.

ProofGate was born from a real incident: the author fired a previous AI agent for repeatedly inventing completed work. This plugin turns the rule that replaced it, never claim without receipts, into a mechanism every Hermes user can install.

How it works

  1. Evidence collector. A post_tool_call hook silently records every tool call in the session: tool name, arguments (sensitive values redacted), result, status, and duration.
  2. Claim checker. A transform_llm_output hook holds the finished reply and extracts factual completion claims in five families: a file was created or updated, a message was sent, a command ran successfully, tests passed, a specific figure is quoted. Each claim is matched against the recorded evidence from the same session. Checking is deterministic, no LLM, no network, milliseconds.
  3. Fail-closed verdicts.
    • PASS: every claim is backed by evidence. The reply ships with a small footer stating how many claims were verified.
    • DOWNGRADE: some claims check out, some do not. The unproven claims are listed in the reply as explicitly unverified, never silently deleted.
    • BLOCK: no claim is backed by evidence, or a claim is contradicted by a tool attempt that actually failed. The reply is withheld and replaced with an explanation of exactly which proof is missing.
    • If the checker itself errors, the reply is downgraded, never passed. A broken gate fails closed.
  4. Ledger. Every checked reply is written to an append-only ledger on your own machine at $HERMES_HOME/proofgate/ledger.jsonl, with the raw evidence trail in evidence.jsonl. Nothing leaves your machine.

Surfaces

  • /proof in chat: recent verdicts and counts.
  • proof_status tool: the agent can inspect its own verification record mid-work and fix itself before the gate has to.
  • hermes proofgate ledger on the CLI: read the ledger from a terminal.
  • A first-turn briefing: when ProofGate is active, the agent is told once per session that its claims will be checked against tool evidence.

Install

hermes plugins install <owner>/proofgate --enable

Or copy this directory into ~/.hermes/plugins/proofgate/ and run hermes plugins enable proofgate, then restart Hermes.

Verify any time with:

hermes plugins validate ~/.hermes/plugins/proofgate
hermes plugins doctor ~/.hermes/plugins/proofgate

Note: plugins validate uses a permissive probe. plugins doctor exercises the real registration contracts. ProofGate passes both.

Configuration

In config.yaml:

plugins:
  entries:
    proofgate:
      settings:
        mode: enforce   # enforce | warn | off
  • enforce (default): replies are footered, flagged, or withheld per verdict.
  • warn: verdicts are recorded in the ledger, replies are never modified.
  • off: ProofGate observes nothing and changes nothing.

Honest limitations

  • Five claim families in v0.1. Claims outside those families (plans, opinions, summaries of reasoning) are not checked and pass through untouched.
  • enforce mode trades streaming for certainty: the finished reply is held until the check completes. Use warn mode if streaming matters more to you than gating.
  • In v0.1, a DOWNGRADE appends an explicit unverified-claims section to the reply instead of rewriting the claim inline.
  • Evidence is matched within the same session. A claim about work done in an earlier session cannot be proven from this session's trail and will be flagged.
  • Statuses count as proof only when Hermes reports the tool call as successful. An unknown status is treated as no proof, by design.

Development

python3 tests/test_checker.py   # 8 scenario tests, no dependencies

The checker (checker.py) is pure Python with no Hermes imports, so the verification logic can be tested and audited in isolation.

License

MIT

README from the reviewed commit b07f2fe; it updates when the author re-pins.

More by Whydiriki700

3 plugins by Whydiriki700 in the catalog.

Whydiriki700money-magnetFinance tracker and money magnet for Hermes Agent: record income and expenses, set a monthly budget and watch it fill, value a crypto portfolio against live prices, and scan for price gaps between exchanges. Transactions and budget live in local JSON under $HERMES_HOME/money-magnet/. Disclosure: the portfolio and arbitrage tools query the public CoinGecko and Binance price APIs (no account or API key; only coin symbols are sent). The tracker tools work fully offline.CommunitytoolsWhydiriki700token-dashboardToken cost and usage dashboard for Hermes Desktop: a right pane with the focused session's live input and output tokens, session cost (backend-reported cost_usd when available, otherwise a labelled estimate from a built-in per-model price table), API calls and context usage bar, plus a recent-sessions overview with message counts and a status bar cost chip. Desktop-only (desktop/plugin.js inside the plugin SDK surface); no agent tools, hooks, middleware, or env vars. Disclosure: reads the focused session usage state and the session list through the Desktop host API and renders them locally; no network calls, no storage.CommunityDesktop app