fix: truncate large hook payloads before forwarding to NeMo Relay

[nanookclaw]

Summary

The NeMo Relay Axum server applies a default ~2 MiB body limit on /hooks/hermes. Large but valid LLM turns silently return HTTP 413, so the interaction completes but Phoenix/ATIF telemetry is missing for that turn.

Root Cause

_forward() sends the full serialized hook payload without checking size. When a Hermes turn produces a large payload (multi-message conversations, verbose tool outputs, long reasoning chains), the Axum Json<Value> extractor rejects it before hermes::adapt(...) runs.

Change Made

File: examples/personal-community-sentiment-triage/agents/hermes/plugins/nemo-relay/__init__.py

Added a _truncate_payload() step in _forward() that progressively shrinks oversized payloads while preserving correlation metadata (task_id, session_id, model, etc.) and hook event structure:

• Phase 1: Truncate message content strings with a proportional per-message budget
• Phase 2: Truncate response.raw_response and assistant_message content
• Phase 3: Truncate tool args and result strings
• Phase 4: Aggressive — replace all content with [truncated] marker, set _truncated: true

Threshold is 1.5 MiB, safely under Axum's ~2 MiB limit. Small payloads pass through unchanged (identity return, zero overhead).

This is the client-side complement to any future server-side DefaultBodyLimit increase in NeMo Relay itself.

.../plugins/nemo-relay/__init__.py   | 146 ++++++++++++++++++++
 1 file changed, 146 insertions(+)

Testing

• Verified small payloads pass through unchanged (identity return)
• Verified 2 MiB message content is truncated below 1.5 MiB
• Verified 10-message conversations get proportional truncation
• Verified correlation metadata (task_id, session_id, model) is preserved through all phases
• Verified response and tool call payloads are handled
• Python syntax validation passes

Fixes #22

[abhisawa-Nvidia]

@nanookclaw would you file a GH issue on NeMo Relay repo please? We can handle the case in Relay directly.

[nanookclaw]

Filed NVIDIA/NeMo-Relay#255 with the Relay-side oversized hook payload handling context.

[WilliamK112]

The Relay-side follow-up requested above looks done now: NVIDIA/NeMo-Relay#261 closed NVIDIA/NeMo-Relay#255 and shipped in NeMo Relay 0.4.0.

That release raises the default hook payload limit to 20 MiB, adds NEMO_RELAY_MAX_HOOK_PAYLOAD_BYTES, and includes regression coverage for hook payloads above Axum's default limit plus structured 413 responses. This example currently pins nemo-relay-cli@0.3.0, so the smallest follow-up in this repo may be to bump the pinned CLI to 0.4.0 instead of maintaining plugin-side truncation here.

I prepared and locally validated a minimal two-file patch for that path (agents/hermes/Dockerfile version/comment update plus one stale 0.3.0 comment cleanup in nemo-relay/finalize-hook). Validation passed:

• python scripts/check_license_headers.py --check
• git diff --check
• bash -n examples/personal-community-sentiment-triage/scripts/*.sh
• python -m py_compile over the example Python files

Happy to send that minimal bump if maintainers prefer this route.