feat: terse persona prompt, bash+ipython default tools, actionable compaction handoff

[samsja]

Summary

The rlm harness, rebuilt on evidence from a k3-focused ablation campaign (9 models, up to 128 tasks/arm, SWE-rebench-V2):

1. Terse coding-agent persona replaces the verbose default system prompt (~3.2kB -> ~1.4kB)
2. RLM_TOOLING preset — dual (default) | tools | skills — one knob for how shell/edit are exposed
3. Kernel-isolation prompt line (stops project imports into the kernel venv)
4. Actionable compaction handoff (fenced command blocks + numbered next steps)
5. bash tool/skill share one guarded runner (bash -c, git-history guard, harness timeout); output-clipping removed

The starting point: something was wrong

kimi-k3 on stock rlm vs a plain bash+edit agent, same tasks:

setup	solve (easy 32)	reasoning tok/turn	solve (SWE-bench Pro 32)
stock rlm (ipython-only, stock prompt)	0.53	327	0.75
bash+edit agent (verifiers harness)	0.69	268	0.94

Ablations isolated the cause: the verbose prompt (not the ipython interface, not %%bash syntax). Same tools, persona prompt: 209-232 tok/turn, solve 0.62-0.75. Frontier models are insensitive either way; kimi-k3 and opus-5 are prompt-sensitive (~2x reasoning inflation under the stock prompt).

Final ablation: the three RLM_TOOLING presets (128 tasks, kimi-k3, same seeded set)

preset	n	solve	turns	reason/turn	error turns
tools	128	0.711	36.1	216	0.8%
skills (with triple-quote guidance)	125	0.704	33.5	256	2.5%
dual	128	0.695	35.4	239	2.1%

Converged on solve and total reasoning (~8k tok). tools is cleanest, skills fewest turns (REPL cells batch multiple actions), dual within noise of both.

The skills preset needed two fixes, each isolated on the same 32 tasks:

skills variant	turns	error turns
buggy skill (ran /bin/sh) + no guidance	53.5	4.2%
+ shell-parity fix (bash -c)	35.9	5.0%
+ triple-quote prompt line	33.5	2.5%

The bugfix bought the turns; the prompt line bought the errors. The skills preset auto-renders the validated line: "Run shell with out = await bash(\'\'\'command here\'\'\') — always triple-quote the command so shell quotes and multi-line scripts never need escaping..."

Cross-model check (GLM-5.3 @32): tools 0.61/1.3% err, skills 0.63 but 2.4x the reasoning tokens and 3.7% err, dual 0.59/0.6% err — skills-only is kimi-validated, not universal; hence dual as default (never the outlier on any model; both alternatives one env var away).

Kernel-isolation line

kimi imported project modules into the rlm kernel (~25 attempts/32 rollouts, 1 in 5 hitting ModuleNotFoundError). The line "The ipython kernel is an isolated venv without the project's packages... everything that executes project code goes through bash" drops kernel project-imports 5 -> 0, ImportErrors -> 0, at lower reasoning (161 vs 182 tok/turn on the probe set) — informational phrasing, no prohibition cost.

Compaction

CHECKPOINT_COMPACTION_PROMPT now demands runnable handoffs (exact repro/test commands + pending edits as fenced code blocks, numbered next steps): validated 0 -> ~4 command blocks per handoff, compacted-rollout solve 0.75-0.88 vs 0.53 under the stock prompt. Also checked: advance compaction notice in the system prompt is unnecessary (models stash state naturally; the at-compaction REPL note suffices; an explicit stash line doubles state-keeping but buys no solve).

Notes

• Merged with main (absorbs feat: built-in bash skill (bash -c, shell parity with a native tool) #145 bash skill and chore: remove RLM_MAX_OUTPUT and RLM_MAX_TOOL_OUTPUT_CHARS #146 knob removal; keeps feat: supervise isolated recursive RLM sessions #134 supervisor semantics — presets respected via _selected() in the tool registry, RLM_SKILLS default now preset-aware in RuntimeConfig.from_env).
• Tests: 125 passed; the 6 test_acp.py failures are pre-existing on clean main in this environment. Ruff clean.

---

Note

Medium Risk
Changes default tooling, prompts, and compaction handoffs that directly steer agent behavior; bash/edit run on the host cwd with file writes, though git guards and timeouts are centralized.

Overview
Default agent surface moves from IPython-only to native bash, edit, and ipython tools, with RLM_TOOLING (dual | tools | skills) and RLM_BUILTIN_TOOLS to override. When RLM_SKILLS is unset, skills now follow the preset via preset_skills() instead of an empty list.

rlm.tools.bash adds shared run_bash (guarded bash -c, cwd, allow_git) used by BashTool, EditTool, and the bash skill; the IPython kernel gets RLM_EXEC_TIMEOUT / RLM_ALLOW_GIT at startup from engine policy.

System prompt is shorter and tool-aware: coding-agent role lines per active tool, kernel-isolation guidance (PROJECT_ENV_PROMPT), conditional bash-skill triple-quote hints, and trimmed IPython/edit copy. Compaction CHECKPOINT_COMPACTION_PROMPT now requires fenced repro commands, pending edit calls, and numbered next steps.

^{Reviewed by Cursor Bugbot for commit 41e2698. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 9b607b0. Configure here.}

[samsja]

Both fixed in the latest commit — run_bash threads allow_git from ToolContext into find_blocked_command (matching the ipython tool's policy-over-env behavior), and EditTool now resolves relative paths against context.cwd like BashTool. Tests added for both.