Shannon Sands 06e9422324 Keep full trajectory; truncate prompt on per-turn copy ...

Previously _truncate_context() mutated the shared messages list, which could drop older turns and break reward computation/debugging.

Now we keep messages as the full trajectory and apply truncation to a copy (prompt_messages) for each model call.

2026-02-14 09:33:36 +10:00

..

configs

Add support for Atropos Agentic RL environments (requires branch tool_call_support in Atropos atm)

2026-02-07 09:17:16 +00:00

tool_call_parsers

Fix RL training pipeline: context truncation, double-encoding, shaped rewards

2026-02-13 22:21:32 +00:00

__init__.py

Add support for Atropos Agentic RL environments (requires branch tool_call_support in Atropos atm)

2026-02-07 09:17:16 +00:00

agent_loop.py

Keep full trajectory; truncate prompt on per-turn copy

2026-02-14 09:33:36 +10:00

gsm8k_agent_env.py

refactor on SlotPoolEnvironment

2026-02-10 08:30:37 +00:00

hermes_base_env.py

Fix RL training pipeline: context truncation, double-encoding, shaped rewards

2026-02-13 22:21:32 +00:00

hermes_swe_env.py

Add support for Atropos Agentic RL environments (requires branch tool_call_support in Atropos atm)

2026-02-07 09:17:16 +00:00

patches.py

testing training

2026-02-11 22:13:05 +00:00

swe_smith_oracle_env.py

SWE env: keep reward shaping env-defined; log tool-call metrics only

2026-02-14 09:19:22 +10:00

terminal_test_env.py

Enhance async tool execution and error handling in Hermes agent for Atropos integration

2026-02-08 05:00:47 +00:00

tool_context.py

Enhance async tool execution and error handling in Hermes agent for Atropos integration

2026-02-08 05:00:47 +00:00