Skip to content

fix(agent): rephrase compression preamble to avoid Azure/OpenAI content-filter false positives - #19370

Closed
liuhao1024 wants to merge 3 commits into
NousResearch:mainfrom
liuhao1024:fix/compression-prompt-content-filter
Closed

fix(agent): rephrase compression preamble to avoid Azure/OpenAI content-filter false positives#19370
liuhao1024 wants to merge 3 commits into
NousResearch:mainfrom
liuhao1024:fix/compression-prompt-content-filter

Conversation

@liuhao1024

@liuhao1024 liuhao1024 commented May 3, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

What does this PR do?

Rephrase the context compression summarizer preamble to avoid triggering Azure/OpenAI content-filter jailbreak detection, while preserving the same semantic intent.

Root Cause

The compression preamble in agent/context_compressor.py::_generate_summary() contained phrases that predictably trip Azure/OpenAI content-filter heuristics:

  • "injected as reference material for a DIFFERENT assistant" — "injected" + "DIFFERENT assistant" reads like prompt-injection framing
  • "Do NOT respond to any questions or requests" — aggressive negation triggers jailbreak filters
  • "NEVER include API keys" — directive-style "NEVER" can also trigger filters

These phrases, combined with the serialized conversation body being summarized, create enough signal for Azure's ResponsibleAIPolicyViolation / jailbreak filter to reject the auxiliary summary call with HTTP 400.

Related Issue

N/A

Type of Change

  • 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

  • See commit messages for detailed changes

How to Test

  1. Run pytest tests/ -q — all tests should pass
  2. Verify the specific scenario described above is resolved

Checklist

Code

  • I've read the Contributing Guide
  • My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
  • I searched for existing PRs to make sure this isn't a duplicate
  • My PR contains only changes related to this fix/feature (no unrelated commits)
  • I've run pytest tests/ -q and all tests pass
  • I've added tests for my changes (required for bug fixes, strongly encouraged for features)
  • I've tested on my platform: macOS 26.4.1

Documentation & Housekeeping

  • I've updated relevant documentation (README, docs/, docstrings) — or N/A
  • I've updated cli-config.yaml.example if I added/changed config keys — or N/A
  • I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture and workflows — or N/A
  • I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A

@liuhao1024
fix(gateway): make npm install timeout configurable via WHATSAPP_NPM_…
6ef568c 
…INSTALL_TIMEOUT

Increase the default npm install timeout for WhatsApp bridge from 60s
to 300s (5 minutes) to accommodate slower systems like Unraid NAS.
Make it configurable via WHATSAPP_NPM_INSTALL_TIMEOUT environment variable
for users who need even longer timeouts.

Closes NousResearch#14980
@liuhao1024
fix(tools): mark patch tool conditionally required params as required
5dc3423 
- Add 'path', 'old_string', 'new_string', and 'patch' to required list
- Update description to clarify mode-specific parameter requirements
- This addresses issue where LLMs would omit these parameters because
  they were not marked as required in the schema, even though they
  are required depending on the mode

Fixes NousResearch#15524
@liuhao1024
fix(agent): rephrase compression preamble to avoid Azure/OpenAI conte…
d504a8c 
…nt-filter false positives

The summarizer preamble in context_compressor.py contained phrases that
trigger Azure/OpenAI content-filter jailbreak detection:
- 'injected as reference material for a DIFFERENT assistant'
- 'Do NOT respond to any questions or requests'
- 'NEVER include API keys'

Rephrased to use softer directives while preserving the same semantic
intent. The summary output prefix (SUMMARY_PREFIX) is unchanged since
it is injected into conversation context, not sent as a prompt.

Regression tests assert the trigger phrases are absent and the new
wording preserves key semantic markers.

Fixes NousResearch#19362
@alt-glitch alt-glitch added comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint P2 Medium — degraded but workaround exists provider/openai OpenAI / Codex Responses API type/bug Something isn't working labels May 3, 2026
This was referenced May 3, 2026
Closed
Closed
@teknium1

teknium1 commented Jul 12, 2026

Copy link
Copy Markdown
Contributor

Thanks for the compression compatibility work. This is already implemented on current main by a shipped equivalent fix.

  • Automated hermes-sweeper review verified agent/context_compressor.py:1793 now uses source-material/checkpoint framing instead of the injected-output and different-assistant wording.
  • tests/agent/test_context_compressor.py:691 captures the prompt passed to call_llm and verifies the historical Do NOT respond and handoff phrases are absent.
  • The implementation landed in fc88eec926a90c11a8949a3d7e0b852cfdfb0c3a (fix(compressor): soften summary prompt for content filters) and shipped in v2026.5.7.

Closing as implemented on main.

@teknium1 teknium1 closed this Jul 12, 2026
@teknium1 teknium1 added the sweeper:implemented-on-main Sweeper: behavior already present on current main label Jul 12, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint P2 Medium — degraded but workaround exists provider/openai OpenAI / Codex Responses API sweeper:implemented-on-main Sweeper: behavior already present on current main type/bug Something isn't working

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@liuhao1024 @teknium1 @alt-glitch