mirrors/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-05-04 02:21:47 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 5

Updated terminal_tool with SlotPoolEnvironment

Browse source
This commit is contained in:
Shannon Sands 2026-02-10 07:23:08 +00:00
parent f82c3081f2
commit c8b30e9efa
8 changed files with 373 additions and 367 deletions
Download patch file Download diff file Expand all files Collapse all files

16
memory-bank/progress.md
View file

@ -21,11 +21,19 @@ Test results:
- [x] Test Phase 1 (OpenAI server type) with Nous API — WORKS
- [x] Test Phase 2 (ManagedServer) with RunPod SGLang — WORKS
- [x] Port SWE env to `HermesAgentBaseEnv` with multiplexed sandboxing
- [ ] End-to-end test with Modal sandbox (needs live Modal)
- [ ] Remove redundant `atropos/agent/` and `atropos/envs/agent_env.py`
- [ ] Clean up redundant `atropos/tools/`
- [x] End-to-end test: Qwen 3 8B + Modal sandbox + tool calls in sandbox + pytest verification
- [x] Add `_SlotPoolEnvironment` to `tools/terminal_tool.py` (TERMINAL_ENV=slot_pool)
- [x] Remove redundant `atropos/tools/` files (4 of 8)
- [ ] Remove redundant `atropos/agent/` and `atropos/envs/agent_env.py` (deferred)
- [ ] Test end-to-end with Tinker trainer (blocked on billing)
- [ ] Test with actual tool calls (model producing tool_calls, not just text)
### ✅ End-to-End SWE + Modal Sandbox Verified (Feb 10, 2026)
- Qwen 3 8B on RunPod SGLang (endpoint `0tx0ruuuo4f10c`)
- Phase 2 ManagedServer with hermes tool call parser
- 5 terminal commands executed in Modal sandbox: ls, git status, git log, cat parse.py, cat tests/
- In-sandbox verification: install deps + pytest → score 0.0 (model inspected but didn't fix)
- Full token tracking with logprobs via /generate endpoint
- Key finding: Llama-3-8B template drops tools= silently; Qwen 3 has full Hermes tool format
## Completed Features