feat(v1): bump rlm pin to nano-rlm main, expose guardrail and role-append knobs

[hallerite]

Companion to the recently merged nano-rlm work. Two parts:

Pin bump: 4ef3438 → 240090d

The new default picks up everything merged to nano-rlm main since the compaction pin:

• feat: execution guardrails (turn, context, tree-token, and output caps) nano-rlm#158 — execution guardrails: tree-total max_total_turns / max_total_tokens budgets (tokens counted as new work: completion + uncached prompt; every engine stops gracefully once spent), and max_tool_output_bytes overriding the built-in 20KB per-tool-result truncation.
• feat: role-aware sub-agent prompting (built-in contract + per-tier appends) nano-rlm#151 + fix: prompt clarity for weaker models nano-rlm#166 — role-aware sub-agent prompting: depth-gated built-in sub-agent contract, per-tier system-prompt appends on the contract (subagent_…/leaf_… with leaf → subagent → root fallback), plus root-prompt clarity fixes. Validated on a 3-model, 32-task SWE-bench Pro before/after (never worse, slightly cheaper on every model).
• fix: recognize vLLM's context-length 400 as request-too-large nano-rlm#162 (vLLM overflow phrasing), fix: bash tool survives binary output nano-rlm#156 (bash tool binary-output fix), refactor: publish semantic request edges nano-rlm#153 (semantic lineage edges — the session snapshot already tolerated new fields via extra="ignore").

New harness config knobs

knob	maps to	default
max_total_turns	policy.max_total_turns	None (uncapped)
max_total_tokens	policy.max_total_tokens	None (unbounded)
max_tool_output_bytes	policy.max_tool_output_bytes	None (rlm's 20KB)
max_concurrent_subagents	policy.max_concurrent_subagents	None → max(4, max_depth)
append_to_system_prompt	contract root append (joined after the taskset system prompt)	None
subagent_append_to_system_prompt	contract node append	None
leaf_append_to_system_prompt	contract leaf append	None
builtin_skills	now also accepts "bash", "fetch"	unchanged

Compatibility: the payload always carries the full field set of the runtime-v1 contract at the default pin (unset knobs go as explicit None), validated against nano-rlm's contract models. The harness was already unconditionally sending post-compaction fields, so the effective rule stays what it was: version must be at least the default ref — pinning older refs than the harness's contract era is unsupported.

Related: #2487 (builtin tools passthrough, RLM_BUILTIN_TOOLS) stays separate — this PR only widens the skills literal.

Note

Bump RLMHarnessConfig rlm pin to 240090d and add guardrail and prompt-append knobs

• Bumps the default version from 4ef3438 to 240090d and expands BuiltinSkill to include "bash" and "fetch".
• Adds optional config fields max_concurrent_subagents, max_total_turns, max_total_tokens, max_tool_output_bytes, and three append_to_system_prompt variants (global, subagent, leaf) with precedence/fallback semantics.
• max_depth type changes from int (default 0) to Optional[NonNegativeInt] (default None): None defers to nano-rlm's default of 1, 0 disables recursion.
• _runtime_metadata now conditionally includes policy keys only when configured and no longer forces max_concurrent_subagents to max(4, max_depth).
• Risk: a new cross-field validator rejects configs where max_concurrent_subagents is set lower than max_depth; previous consumers relying on the hardcoded concurrent-subagent floor will now get nano-rlm defaults or None instead.

^{Macroscope summarized a201007.}

---

Note

Medium Risk
Changes default recursion/concurrency behavior when configs omit fields, and wires new execution budgets into the nano-rlm runtime contract.

Overview
Bumps the default nano-rlm install ref from 4ef3438 to 240090d and widens builtin_skills to allow bash and fetch in addition to edit and search.

RLMHarnessConfig gains optional knobs that map into the runtime policy and prompt contract: tree-wide max_total_turns / max_total_tokens, per-tool max_tool_output_bytes, max_concurrent_subagents, plus root/sub-agent/leaf append_to_system_prompt fields. max_depth is now NonNegativeInt | None (unset defers to nano-rlm’s default instead of hard-coding 0), with validation that max_concurrent_subagents ≥ max_depth when both are set.

_runtime_metadata merges the taskset system prompt with append_to_system_prompt, forwards the new sub-agent/leaf append fields, and omits policy keys whose values are None so nano-rlm defaults apply—replacing the prior always-sent max_depth and computed max(4, max_depth) for concurrency.

^{Reviewed by Cursor Bugbot for commit a201007. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This change alters the default RLM execution model by enabling nano-rlm’s recursive default and adds tree-wide budgets, tool limits, role-specific prompts, and new skills through an updated runtime dependency. An unresolved review finding also identifies configurations that can pass local validation but fail when the RLM session starts.

^{You can add or adjust custom eligibility rules. Learn more.}

[mikasenghaas]

should we myb not spell out the actual default value so that this cannot diverge?

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

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

^{Reviewed by Cursor Bugbot for commit a201007. Configure here.}

[cursor]

Concurrent default undershoots max depth

Medium Severity

When max_concurrent_subagents is left unset, the policy omits it and nano-rlm applies a fixed default of 4. That can be below max_depth, which nano-rlm rejects. The new validator only compares the two knobs when both are set, so this config is accepted and then fails at session start.

Additional Locations (1)

• verifiers/v1/harnesses/rlm/harness.py#L170-L176

 

^{Reviewed by Cursor Bugbot for commit a201007. Configure here.}