feat: rlm renderer for the RLM chat format

[mikasenghaas]

Adds the rlm renderer for the minimal RLM chat format (PrimeIntellect/RLM-Chat-Template, private), designed for the rlm harness.

Format

<system>...</system><user>...</user><assistant>{content}<ipython>{code}</ipython></assistant><output>{result}</output>

• One message, one tag block. Ten single-token delimiters (reserved Nemotron-3 <SPECIAL_18>–<SPECIAL_27> slots renamed — no vocab resize for Nemotron Super base). </assistant> (id 23) is eos/stop.
• Thinking is never dropped: retention is structurally "all" (config validator rejects anything else); assistant content renders verbatim on every turn and the bridge extends without re-rendering. A reasoning_content field is restored as <think>{reasoning}</think> before content rather than dropped (the only deliberate divergence from the jinja template, which cannot know the field exists).
• ipython is the only tool, rendered inline as <ipython>{code}</ipython>; results are <output> turns (role: "tool"). Exactly one call per assistant turn (mirrors the harness's parallel_tool_calls=False + one-call error in rlm/engine.py).

tools= semantics, end to end

The HF chat template raises on any tools= (deliberate: no schema text ever enters the prompt). This is safe because apply_chat_template is not in the serving path:

• Serving: rlm-harness sends tools=[IPYTHON_SCHEMA] on every chat-completion (rlm/tools/registry.py — "always just ipython") → verifiers interception dialect.parse_request extracts tools → renderers client.generate(..., tools=tools) → this renderer's render(messages, tools=...). The renderer therefore accepts exactly [ipython] as validation-only (renders nothing); any other name or extra tool raises.
• Template raise only fires on direct apply_chat_template(tools=...) use (e.g. naive SFT preprocessing), where failing loudly is the correct signal that this format renders no schemas.

Wiring

• renderers/rlm.py — RlmRenderer (render/render_ids/parse_response/get_stop_token_ids/bridge_to_next_turn, per-token attribution matching the nemotron-3 conventions: assistant body+close sampled/content, scaffold tags neither).
• RlmRendererConfig (name="rlm") in configs + discriminated union; lazy registry, _populate_registry, MODEL_RENDERER_MAP["PrimeIntellect/RLM-Chat-Template"].
• Real-data validation: 200 samples of PrimeIntellect/SyntheticData-GLM5.2-scaleswe-reward0 / scaleswe_rlm_glm51 render through the renderer (197 clean; 3 raise on genuine data noise — a 'ipy' tool-name typo). Tool calls in these traces arrive as flat JSON strings (verifiers trace shape) and compaction produces mid-list system messages; both are accepted, matching the jinja template's behavior.
• tests/test_rlm.py — 16 tests: byte parity with the jinja template at every prefix (with/without generation prompt), thinking retention, reasoning_content restoration, tools validation, tool-call rejection paths, parse round-trip (OK span + unclosed block), stop ids, bridge exact-extension + assistant rejection, config validation, content-mask invariants. The tokenizer fixture reconstructs the template repo from the public Nemotron-3 Super tokenizer, so CI needs no private-hub token. All pass locally (15 passed).

🤖 Generated with Claude Code

Note

Add RlmRenderer for the RLM chat format with inline ipython tool support

• Adds renderers/rlm.py implementing RlmRenderer, a deterministic renderer for the RLM chat format using single-token role tags (<system>, <user>, <assistant>, <ipython>, etc.).
• Validates at startup that the tokenizer encodes all required role tags as single tokens; fails fast if any are missing.
• Supports at most one ipython tool call per assistant turn encoded inline as <ipython>code</ipython>; rejects other tool types or multiple calls.
• Assistant reasoning content is always retained inline as <think>...</think> (thinking retention is forced to "all").
• parse_response extracts content and ipython code blocks into ParsedToolCall entries, handling both closed and unclosed blocks; bridge_to_next_turn trims to a closed assistant turn before appending new messages.
• Registers the renderer under the key "rlm" and routes the PrimeIntellect/RLM-Chat-Template model to it automatically.

^{Macroscope summarized b62d804. (Automatic summaries will resume when PR exits draft mode or review begins).}