The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Hermes Bug Report: computer_use Screenshot Lifecycle / Context Blowup
Audience: Hermes / AURA development team
Severity: High — causes session degradation, compression loops, and unbounded payload growth in production AURA missions
Date: 2026-04-29
Repo: hermes-agent @ b07791db0508f92625dfc9e75f20c331cc7bb528
Executive Summary
Long-running AURA desktop-control missions degrade into compression/rotation loops because Hermes treats computer_use screenshots as durable conversation history rather than ephemeral observations.
Each computer_use capture returns an inline base64 PNG (176K–253K chars per screenshot). run_agent.py appends the raw multimodal dict directly into the active messages list, bypassing the normal large-tool-result persistence path. The screenshot is then resent on every subsequent provider call, saved in raw session JSON, and protected by compression's recent-tail logic.
This is not a compression configuration problem. Compression is enabled, resolves correctly, and runs when triggered. The problem is architectural: screenshots bypass the storage/truncation safety net, are undercounted by the token estimator, and are reintroduced after every verification loop.
Documents
| Document | Description |
|---|---|
📋 hermes-screenshot-lifecycle-bug.md |
Full bug report — root cause chain, production evidence, five bottlenecks, acceptance criteria |
🔧 hermes-screenshot-fix-optimal.md |
Optimal fix — 4 targeted code changes grounded in actual source code + research (ACON, MemAct, GUI-KV) |
Quick Links — Bug Report
- The Feedback Loop
- Root Cause Chain
- Evidence from Production Sessions
- Five Bottlenecks
- Acceptance Criteria
Quick Links — Optimal Fix
- Why This Specific Fix — ACON vs MemAct vs GUI-KV analysis
- Architecture-Grounded Fix — actual code paths with line numbers
- The Fix: 4 Targeted Changes — ready-to-implement patches
- Tests — 5 test cases covering all change surfaces
- Impact Analysis — projected 60–78% payload reduction
TL;DR — The 4 Changes
| # | File | What | Why |
|---|---|---|---|
| 1 | agent/context_compressor.py |
Make _prune_old_tool_results handle multimodal content instead of skipping it |
isinstance(content, list): continue means screenshots are never pruned |
| 2 | tools/tool_result_storage.py |
Make enforce_turn_budget measure serialized size for dict/list content |
len(dict) returns 3 (key count), not the 180K+ serialized size |
| 3 | run_agent.py |
Add _downgrade_stale_screenshots() before every provider call |
Core fix: enforce one-model-call lifetime for screenshot images |
| 4 | run_agent.py |
Strip base64 from tool results in _save_session_log() |
Prevent 176K–253K base64 per screenshot in session JSON |
- Downloads last month
- 25