pearl th-7fdfa9: tmux load-buffer for newlines + ban slash commands in driver LLM

[brentrager]

Summary

Fix two bench harness bugs surfaced by a single-task score-tui --debug smoke run against a healthy Big Smooth.

• Bug 1: newlines mangled into j characters. TmuxDriver::send used tmux send-keys -l TEXT, which interprets every embedded \n as the C-j keysym; in literal mode (-l) C-j degrades to the bare letter j. The pearl debug log showed the affine-cipher task prompt rendering as affine-cipher (python).jjWorking directory:…jFiles present:j - INSTRUCTIONS.mdj…. Switched to tmux load-buffer + tmux paste-buffer, which inserts the payload as raw bytes — newlines preserved verbatim.
• Bug 2: driver LLM uses Claude-Code-style slash commands. The default driver model (smooth-summarize) was emitting /open, /read, /help instead of plain English, and the TUI rejected them as unknown commands (and in two cases accidentally fired skills). Hardened the system + user prompts with explicit "no slash commands; you have no shell/file access" directives, and added a slash-command guard in run_human_loop that drops /-prefixed turns, logs them to the pane-debug log, re-asks with a reinforcement prompt, and after 3 consecutive bad turns bails with TASK_STUCK.

Pearl: th-7fdfa9.

Tests

• tmux_driver::send_preserves_newlines_no_j_leakage — 3-line payload through cat > tmpfile asserts exactly 3 lines and no stray j.
• human_driver::run_human_loop_marks_stuck_after_three_slash_commands — fake driver returning /open 3x asserts LoopExit::Stuck after exactly 2 retries.
• human_driver::run_human_loop_accepts_plain_english_message — fake driver returning plain English asserts zero retries fired.
• is_slash_command_detects_leading_slash, build_driver_prompt_warns_against_slash_commands, build_slash_retry_prompt_contains_offending_text_and_attempts.

All 125 smooth-bench tests pass. cargo fmt --check, cargo clippy, cargo build clean.

Verification

Re-ran the smoke that originally caught these bugs:

cargo run --release -p smooai-smooth-bench -- score-tui --pr \
  --th-binary $(which th) --task-limit 1 --debug \
  --output /tmp/bench/smoke2.json --budget-usd 2

New debug log: ~/.smooth/bench-runs/80c092b0/python-affine-cipher.pane.log. Initial prompt now renders with real newlines:

You:
You are solving an Aider Polyglot coding benchmark task: `affine-cipher` (python).

You:
Working directory: the current directory.

You:
Files present:

You:
  - INSTRUCTIONS.md

(Zero j artifacts.) Driver follow-ups are all plain English — sample turns from the log:

Please read INSTRUCTIONS.md and affine_cipher_test.py.
Once you've finished reading, please run the tests.
This is a Python project. Please read INSTRUCTIONS.md and affine_cipher_test.py.
Please run the tests to see what needs fixing.
Please read the contents of affine_cipher.py.

(Zero /-prefixed turns.) The single affine-cipher task itself FAILed — expected, it's a hard task on a single attempt. Harness is healthy.

Test plan

• CI green (cargo fmt, clippy, test, build)
• Reviewer can run a multi-task --pr sweep when ready

🤖 Generated with Claude Code

[changeset-bot]

🦋 Changeset detected

Latest commit: 7ac910c

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@smooai/smooth	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR