fix(agent): optimize dflash stall recovery telemetry

[OmarB97]

What changed

• Enables the existing tool-use enforcement prompt for the local dflash model alias.
• Adds bounded stall-retry telemetry with per-turn summaries and local NDJSON logging.
• Adds a retry-only continuation nudge so the fallback lane is told to emit the promised tool call instead of repeating the preamble.
• Keeps truly empty responses on the existing empty-response recovery path.
• Makes low-information terminal redirects recoverable by default, with hard halt reserved for hard_stop_enabled.

Why

Recent dflash MeshBoard sessions mostly recovered, but still wasted turns on action-preamble stalls and repeated low-information terminal probes. One latest-run trace showed the new guardrail ending the turn too bluntly after repeated terminal probes; this patch turns that into recoverable tool feedback by default.

How to Review

Review agent/stall_retry.py and agent/conversation_loop.py for the retry and telemetry path, then review agent/tool_guardrails.py for the low-information terminal redirect change. The tests in tests/agent/test_stall_retry.py and tests/run_agent/test_tool_call_guardrail_runtime.py cover the new behavior directly.

Evidence

• scripts/run_tests.sh tests/run_agent/test_run_agent.py tests/run_agent/test_tool_call_guardrail_runtime.py tests/agent/test_stall_retry.py tests/agent/test_tool_guardrails.py tests/agent/test_prompt_builder.py -- -q passed: 529 tests, 0 failed.
• git diff --check passed with rc=0.
• python -m py_compile agent/stall_retry.py agent/conversation_loop.py agent/prompt_builder.py agent/tool_guardrails.py hermes_cli/config.py passed with rc=0.

Verification

scripts/run_tests.sh tests/run_agent/test_run_agent.py tests/run_agent/test_tool_call_guardrail_runtime.py tests/agent/test_stall_retry.py tests/agent/test_tool_guardrails.py tests/agent/test_prompt_builder.py -- -q
git diff --check
python -m py_compile agent/stall_retry.py agent/conversation_loop.py agent/prompt_builder.py agent/tool_guardrails.py hermes_cli/config.py

Risks / Gaps

• Telemetry is local bounded JSONL and omits raw request messages; low risk, no follow-up needed.
• Upstream PR will be stacked on existing dflash retry PRs until Nous review lands; already tracked by the upstream mirror.

Collaborators

Participants: Omar, Codex.

Process: Reproduced from live Hermes/dflash MeshBoard traces, added focused regression coverage, then ran the broader main-loop tests.

Task context: MeshBoard task hermes-dflash-stall-optimization-pass.

Related work: Upstream mirror PR NousResearch#35694. Builds on fork PRs #28-#36 and the open upstream dflash retry stack.

[github-actions]

🔎 Lint report: fix/dflash-stall-telemetry-optimization vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 9551 on HEAD, 9545 on base (🆕 +6)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 4959 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.