fix(agent): use standard shell code fences as the action format

[larstalian]

What

Shell actions in the agent loop now use the standard markdown code fence a model is trained to emit — ```bash / ```sh / ```shell / ```console / ```zsh — instead of the bespoke ```run_shell token.

Why

parse_action only recognized ```run_shell / ```finish. Any other fence failed the match and fell through to the "no recognized block → finish" path, which ends the episode and discards the action (including a submit). Instruct/cold models reach for ```bash constantly — it has overwhelming pretraining mass, while ```run_shell appears essentially only in our own prompt — so a competent rollout could die at a single step with its training signal lost.

open-range's design is one shell primitive + a finish (no per-verb tool menu), so there is nothing for a custom fence token to disambiguate; it is pure friction against the model's prior. This also matches the CodeAct finding — executable code as the action space outperforms bespoke/JSON tool encodings (Wang et al. 2024).

Changes

• parse_action accepts bash|sh|shell|console|zsh as a shell command and finish as the terminal block.
• The last recognized block wins, so an illustrative snippet earlier in a reply is not executed in place of the action the model actually settled on.
• The internal action identifier stays run_shell; only the wire format the model emits changes. No back-compat alias for the old ```run_shell fence (the only emitters were our own test fixtures).
• finish stays a sentinel — there is no standard markdown fence for "done".
• Tests updated, plus a new case covering the standard fences and last-block selection.

Test

pytest tests/test_agent_harness.py tests/test_rllm_shim.py → 13 passed, 5 skipped (rllm not installed locally). ruff check, ruff format --check, and mypy clean on the changed files.

🤖 Generated with Claude Code