hermes - 💡(How to fix) Fix Add self-improvement guardrail for repeated equivalent tool timeouts

Feature Description

Add an agent/runtime reliability guardrail that detects repeated equivalent tool failures (especially terminal timeouts) and prompts or automatically creates a structured self-improvement issue instead of continuing to spin.

This was prompted by a Hermes Telegram session where the agent spent many minutes retrying external-volume directory enumeration despite another local coding agent completing the same checks in milliseconds.

Related incident issue: #22111

Motivation

Hermes markets itself as self-improving through skills/memory, but failures in basic computer operations need a stronger escalation path:

• detect repeated same-class tool failures
• stop wasting user time
• classify likely runtime/tool/session issue vs target-system issue
• capture reproducible evidence
• open a source-code ticket or local diagnostic artifact automatically when appropriate

This is especially important for gateway sessions where the user cannot easily inspect the underlying shell state and the agent may misattribute failures.

Proposed Solution

Add a guardrail in the agent/tool loop, terminal tool wrapper, or both:

1. Track repeated failures by class/signature:

   • same tool
   • same path/host/resource
   • same timeout/no-output pattern
   • equivalent commands attempted repeatedly

2. After N bounded failures, require a strategy change:

   • do not run another equivalent command
   • classify failure mode explicitly
   • suggest runtime/session refresh or diagnostic route
   • if repo/auth available, draft/open a GitHub issue with commands, outputs, environment, and impact

3. Add source-code hooks for self-improvement issue capture:

   • include exact commands
   • include timeout durations
   • include host/platform/gateway context
   • include comparison evidence if another tool/agent succeeded
   • include relevant logs if safe

4. Provide user-visible behavior:

   • “I have hit repeated equivalent terminal timeouts. I’m stopping retries and opening a bug with the evidence.”

Acceptance Criteria

• Repeated terminal timeouts on equivalent commands trigger a stop/escalate path rather than continued retries.
• The guardrail distinguishes user-target failure from Hermes runtime/tool failure when metadata/control checks contradict the timeout behavior.
• A generated issue body includes enough data for maintainers to reproduce/debug.
• The behavior works from gateway sessions, not just interactive CLI.

Impact

Reduces user frustration, prevents long wheel-spinning loops, and turns failed tasks into actionable Hermes improvements.