All plugins

escalation-governor

Deterministic stuck-detection for Hermes sessions. Watches tool/API failure signals (consecutive tool errors, repeated identical calls, API error bursts, model calls without progress) and, when a threshold trips, injects a structured escalation packet (goal, what was tried, what failed) directing the agent to delegate the remaining work to a stronger model via delegate_task instead of retrying blindly; writes a per-session JSONL report with escalation and token-usage accounting. Disclosure — local code only: injects an escalation packet (the session's first user message and recent tool targets/errors) into tool results and into each following turn's context once a stuck threshold trips, directing the agent to call delegate_task (which spends model usage on a subagent); reads an optional escalation-governor.json and appends one line per session to a shared escalation-governor-reports.jsonl in the Hermes home; adds a /governor status command; opens no network connections, starts no processes and never blocks a tool call.

jusstAutomationv0.1.0Updated 1 hour ago

About

Deterministic stuck-detection for Hermes sessions. Watches tool/API failure signals (consecutive tool errors, repeated identical calls, API error bursts, model calls without progress) and, when a threshold trips, injects a structured escalation packet (goal, what was tried, what failed) directing the agent to delegate the remaining work to a stronger model via delegate_task instead of retrying blindly; writes a per-session JSONL report with escalation and token-usage accounting. Disclosure — local code only: injects an escalation packet (the session's first user message and recent tool targets/errors) into tool results and into each following turn's context once a stuck threshold trips, directing the agent to call delegate_task (which spends model usage on a subagent); reads an optional escalation-governor.json and appends one line per session to a shared escalation-governor-reports.jsonl in the Hermes home; adds a /governor status command; opens no network connections, starts no processes and never blocks a tool call.

Escalation Governor

A weak model does not know when it is out of its depth. Left alone, an
agent on a cheap model retries the same failing approach until the session
dies — Hermes' own advanced-delegation write-up recorded 0 delegations in
2,900 calls over ~89 hours, because the decision to escalate was left to
the model that was already failing.

Escalation Governor takes that decision away from the model. It watches
deterministic signals only — no LLM judging, no prompt tricks — and when
a session is measurably stuck, it injects a structured escalation packet
into the model's context and directs it to hand the remaining work to a
stronger model.

Signals

| Signal | Default threshold | Source hook |
| --- | --- | --- |
| Consecutive tool errors | 3 | post_tool_call (status != ok) |
| Same tool call repeated (same tool + same args) | 3 | post_tool_call |
| Provider API errors since last successful tool result | 3 | api_request_error |
| Model calls since last successful tool result | 12 | post_api_request |

Any successful tool result resets the error counters — the governor only
fires on sustained failure, never on a single unlucky call.

What happens on a trip

The governor builds a packet — the session goal, the approaches already
tried and how each one failed, and the trip reason — and injects it once per
turn (via transform_tool_result mid-turn, and pre_llm_call at the next
turn's start). The packet directs the agent to:

  1. call delegate_task with the packet as the brief, so a stronger model
    takes over the remaining work without repeating the failed approaches, or
  2. if delegation is unavailable, stop and present the packet to the user as
    the blocker report.

A successful tool call after injection clears the trip (the session
recovered on its own); a subagent_start also clears it (the escalation
was acted on).

At on_session_end, a one-line JSON report is appended to
~/.hermes/escalation-governor-reports.jsonl with turn/API/tool counts,
token usage, escalation count, and whether delegation followed — so you can
measure what the governor saved. /governor shows live session status;
/governor report prints the last session report.

How this differs from what exists

  • The bundled delegation plugin decides which model a subagent uses.
    The Governor decides when escalation must happen at all. They compose:
    the Governor trips, the delegation plugin routes.
  • A community blog hook ("escalade") popularised failure-count escalation.
    The Governor generalises it to four independent signal families, adds
    no-progress detection (calls without a successful tool result), packet
    injection at two points in the turn lifecycle, recovery semantics, and a
    persistent per-session savings report.

Configuration

Defaults work out of the box. Override with a JSON file at
~/.hermes/escalation-governor.json (keys: consecutive_error_threshold,
repeat_call_threshold, api_error_threshold, no_progress_api_calls,
max_packet_actions) or environment variables GOVERNOR_CONSECUTIVE_ERRORS,
GOVERNOR_REPEAT_CALLS, GOVERNOR_API_ERRORS, GOVERNOR_NO_PROGRESS_CALLS.
GOVERNOR_DISABLE=1 is a kill switch.

Safety

  • Fail-open: every callback swallows its own exceptions; a Governor bug
    can never break the agent loop.
  • Non-blocking: the Governor never blocks or modifies a tool call. It
    only adds context and writes its own report file.
  • No credentials, no network: it reads no secrets and calls nothing.
  • The packet contains only the session's own goal text (truncated) and
    metadata-style summaries of its own tool calls (tool name + target hint +
    outcome) — the same visibility the agent already has.

Install

Drop this directory into ~/.hermes/plugins/escalation-governor/, enable
it (hermes plugins enable escalation-governor), and restart Hermes.

Tests

python3 tests/test_governor.py — plain-python tests of the signal engine
(trip conditions, once-per-turn injection, progress reset, recovery,
delegation consumption, usage accounting).

README from the reviewed commit 50fcf38; it updates when the author re-pins.

More by jusst

2 plugins by jusst in the catalog.

jusstconstraint-vaultPins the user exact constraints (numbers, addresses, hard rules) as structured local data and re-injects them verbatim into the model context on every turn and immediately after every context compaction, with a SHA-256 hash-logged injection record, because memory providers and compression silently paraphrase or drop exact values. Not a memory provider and not a compressor: a guard layer above both. Disclosure — local code only, no network or processes; stores constraints in constraint-vault.json in the Hermes home and appends a hash record to constraint-vault-injections.jsonl on every turn; the model can pin or drop constraints through its own tools without asking, and pinned text is injected into the model context on every turn of every session until dropped.Communitygeneral