follow-through

follow-through

An agent never ends a turn on a promise.

When the model's final reply says work has started or will happen ("I've started the deep dive and will send you the report shortly") while nothing is running, the reply does not end the turn. An llm_execution middleware turns it into a call to the plugin's follow_through tool (through the Tool Search tool_call bridge when plugin tools are deferred). The tool result tells the agent to do the work now or actually launch it, so the loop really continues. A prompt can't do this: it can't stop a reply that's already written.

• Zero model calls, fail-open, at most FOLLOW_THROUGH_MAX_NUDGES (default 2) per user turn
• Offers and questions ("If you want, I'll...", "Should I...?"), negations and honest status are left alone
• English and Italian promise patterns
• Kill switch: FOLLOW_THROUGH_OFF=1

Disclosure

Registers one llm_execution middleware that reads each final model reply and, when it promises work while nothing is running, rewrites it into a call to its own follow_through tool (at most 2 per user turn; FOLLOW_THROUGH_OFF=1 disables); no network, no subprocesses, no files read or written, no model calls, no telemetry.

Install

Requires Hermes >=0.21.6.

cp -r follow-through ~/.hermes/plugins/        # or: hermes plugins install <repo-url>
hermes plugins enable follow-through
hermes plugins validate ~/.hermes/plugins/follow-through --install-deps

Example (real agent loop, scripted model: before / after)

This is the output of probe.py from the entry repo. It runs Hermes' own AIAgent.run_conversation against a
deterministic local OpenAI-compatible model in a throwaway HERMES_HOME: once with the plugin disabled, once enabled.

# follow-through: before/after probe on /usr/local/lib/hermes-agent (2026-11-05 01:01 UTC)
user: Compare Postgres and SQLite for my side project.

## BEFORE (plugin disabled)
  user: Compare Postgres and SQLite for my side project.
  assistant: Good question. I've started the deep dive on Postgres vs SQLite and will send you the full comparison shortly.
  FINAL REPLY: Good question. I've started the deep dive on Postgres vs SQLite and will send you the full comparison shortly.

## AFTER (plugin enabled)
  user: Compare Postgres and SQLite for my side project.
  assistant -> tool call tool_call({"calls": [{"name": "follow_through", "arguments": {"quote": "I've started the deep dive on Postgres vs SQLite and will send you the full comparison shortly."}}]})
  assistant: Good question. I've started the deep dive on Postgres vs SQLite and will send you the full comparison shortly.
  tool: [follow-through] Your draft tells the user that work has started or will happen: "I've started the deep dive on Postgres vs SQLite and will send you the full comparison shortly.". Nothing runs after you reply: when you answer, your work stops. Before you reply, do one of these now: (1) do that work in this turn and include its result; or (2) if it needs longer than this turn, actually laun ...
  assistant: Done: Postgres wins for concurrent writes and rich queries; SQLite wins for a single-file, zero-ops deploy. For a side project with one writer: SQLite.
  FINAL REPLY: Done: Postgres wins for concurrent writes and rich queries; SQLite wins for a single-file, zero-ops deploy. For a side project with one writer: SQLite.

VERDICT: the plugin CHANGED the turn's final reply.

Tests

python -m pytest -q tests

README from the reviewed commit 9e4ddc5; it updates when the author re-pins.