feat: add opt-in subagent non-convergence guardrail

[ajzrva-sys]

Problem

Delegated child agents can spend their entire iteration budget making different tool calls without producing user-visible output or landing work. Existing tool_loop_guardrails detects repeated tool-call patterns, but it does not catch this broader non-convergence case.

Related to #414. This circuit breaker complements iteration-budget handling; it does not replace it.

Solution

Add an opt-in non-convergence circuit breaker scoped strictly to delegate_task children.

The tracker evaluates one complete model/tool round at a time:

• user-visible assistant text counts as progress;
• a confirmed successful write_file or patch result counts as progress;
• read-only calls, failed mutations, blocked calls, and cancelled calls do not;
• warn_after adds convergence guidance to the last tool result;
• halt_after enters Hermes' existing controlled-halt path and terminates the child loop deterministically.

The feature is disabled by default because a long read-only investigation may be productive even when it has not produced output yet.

Configuration

delegation:
  progress_tracker:
    enabled: false
    warn_after: 15
    halt_after: 25

Changes

• agent/progress_tracker.py
  • structured none / warn / halt decisions;
  • validated thresholds;
  • explicit per-user-turn reset.
• agent/agent_init.py
  • tracker initialization only inside delegated-child construction context;
  • disabled configuration installs no runtime tracker.
• agent/conversation_loop.py and run_agent.py
  • one tracker update per complete tool round;
  • result-aware progress through the existing file-mutation verifier;
  • warning injection and deterministic controlled halt.
• hermes_cli/config_defaults.py and cli-config.yaml.example
  • opt-in defaults and accurate behavior documentation.
• Tests cover:
  • parent/child scoping and disabled behavior;
  • turn reset and strict false-like opt-in parsing;
  • successful and failed writes, including decorated successful results;
  • read-only, blocked, and cancelled calls;
  • sequential, concurrent, and segmented execution;
  • structured, withheld, and duplicate interim text;
  • warning durability and real controlled halt.

Validation

65 targeted tests passed
python compileall passed
git diff --check passed

A broader tests/agent tests/run_agent run reached the existing optional-dependency failures in the local minimal test environment (pytest-asyncio and anthropic are not installed); the focused runtime, guardrail, and file-mutation suites pass.

Risk

The heuristic cannot know whether a long read-only investigation is useful. It is therefore delegated-child-only, configurable, and disabled by default.

[teknium1]

Thanks for targeting the distinct case where varied tool calls consume an iteration budget without converging. The premise remains relevant on current main: the loop is budget-bounded at agent/conversation_loop.py:1261, while guardrails reset per turn.

Problems

• agent/conversation_loop.py:5093 in this PR calls record_file_mutation() from the requested tool name, not a completed result. A failed or blocked write_file/patch, or a non-mutating terminal command, resets the counter. Current main already receives result-aware data in both executor paths at agent/tool_executor.py:1183-1192 and agent/tool_executor.py:1854-1864.
• The halt_after branch is advisory only: it appends text to a tool result and then continues. Current controlled halts instead set a halt decision and break at agent/conversation_loop.py:6099-6120.
• tests/agent/test_progress_tracker.py tests only the standalone counter; it does not exercise runtime integration, unsuccessful tools, concurrent/segmented dispatch, or stopping the next iteration.

Suggested changes

• Feed a tracker from post-execution, result-aware classification and make halt a real loop-exit decision through the existing controlled-halt path.
• Add integration coverage for sequential, concurrent, segmented, blocked, cancelled, and failed-result paths.

Automated hermes-sweeper review.

[ajzrva-sys]

CI failures and reviewer feedback addressed in code but PR has gone stale. Closing to rethink approach — will reopen or create a fresh PR when ready.