Laguna-S-2.1-Agentic-Chat-Template-Jinja

Laguna S 2.1 chat template

A drop-in Jinja chat template for Poolside Laguna S 2.1, rendering the model's native conversation format. It carries the same v24.0-loopguard agentic hardening as the Qwen fork (https://huggingface.co/sanjxz/Qwen-Fixed-Chat-Templates/tree/main), re-expressed in Laguna's <system> / <user> / <assistant> token scheme.

• Internal version: laguna-s21-froggeric-v24.0-loopguard
• Engines: llama.cpp / LM Studio / MLX / vLLM and any engine with a HuggingFace-Jinja (minja) runtime.

---

Files

---

Native format at a glance

Laguna does not use <|im_start|> ChatML. The template emits:

Example assistant turn with a tool call:

<assistant><think>I should check the weather.</think><tool_call>get_weather<arg_key>location</arg_key><arg_value>NYC</arg_value></tool_call></assistant>

---

Feature set

The agentic internals are identical to the Qwen fork — see the "What the fork adds over the original" section of README.md for the full write-up. In brief, this template adds over chat_template_laguna-orig.jinja:

• Escalating loop-guard — consecutive-failure escalation (nudge → warn → hard halt), plus tool suspension: on hard halt a pre-scan replaces the <available_tools> block with a suspension notice so the model cannot emit a tool call.
• Repeat / ping-pong detection — fingerprints each turn by tool name + arguments over a two-turn window (catches A-B-A-B alternation); mutating tools exempted from result-repeat, caught instead by the ping-pong nudge.
• Four-tier failure heuristics — hard / harness / generic / permission errors, kept in sync between the pre-scan and the main loop, with $ (shell) and took (timing) false-positive escapes. Read tools are exempt from error-text matching; results are attributed via tool_call_id → name.
• Security hardening — only system/developer messages may toggle thinking; user content has the tokens stripped but never applied. sanitize_tool_tokens strips literal <|think_off|> / <|think_on|> from untrusted tool output.
• Harness integration — structural skill-tool triggering, vanished-tool warnings, strftime_now date injection, OpenAI {"type":"function",…} wrapper unwrapping, middle-out tool-response truncation, <__media__> marker stripping, unclosed-reasoning recovery.

---

Developer role

OpenAI-style APIs send a developer role. Laguna has no native developer tag, so the template folds developer into <system>:

• A leading developer message becomes the system message.
• A mid-conversation developer message renders as <system>…</system> (not the generic <user>[developer]: … fallback).
• Developer messages inherit trusted-role behavior: they may toggle thinking and their content is treated as system content (images rejected).

If your client sends developer-role prompts, set its supportsDeveloperRole flag to true.

---

Template kwargs

Override via your engine's chat_template_kwargs (llama.cpp: --chat-template-kwargs '{"…":…}').

---

Tool call formats

History tool calls and the format reminder are rendered in one of two formats, selected with tool_call_format.

XML (default) — arguments as <arg_key>/<arg_value> pairs:

<tool_call>get_weather<arg_key>location</arg_key><arg_value>NYC</arg_value></tool_call>

JSON (tool_call_format='json') — a single JSON object inside the <tool_call> envelope:

<tool_call>
{"name": "get_weather", "arguments": {"location": "NYC"}}
</tool_call>

max_tool_arg_chars truncates per-argument values in XML mode and the whole serialized arguments object in JSON mode.

---

Running with llama.cpp

Point --chat-template-file at the template and enable --jinja. Example for Laguna S 2.1 APEX:

import os
import subprocess
import sys

env = os.environ.copy()
env["GGML_CUDA_NO_PINNED"] = "1"
env["GGML_CUDA_DISABLE_GRAPHS"] = "1"

cmd = [
    r".\llama-server.exe",
    "-m", r"D:\Laguna-S-2.1-APEX-i-compact.gguf",
    "--reasoning-preserve",
    "--reasoning-budget", "16000",
    "--reasoning-budget-message", "I've thought enough. Answering now with what I have.",
    "-ngl", "999",
    "--n-cpu-moe", "47",
    "--rope-scaling", "yarn",
    "--rope-scale", "20",
    "--yarn-orig-ctx", "8192",
    "--yarn-attn-factor", "1.0", # experimental fix -- https://www.reddit.com/r/LocalLLaMA/s/tZmaypD6zp
    "--load-mode", "mlock", #recent llama.cpp only ; --mlock & --no-mmap alternative
    "-c", "160000",
    "--cache-type-k", "q8_0",
    "--cache-type-v", "q8_0",
    "-np", "1",
    "-fa", "on",
    "-t", "8",
    "-tb", "8",
    "-b", "3840",
    "-ub", "3840",
    "--jinja",
    "-kvu",
    "--temp", "0.6",
    "--top-p", "0.95",
    "--top-k", "20",
    "--min-p", "0.02",
    "--dry-multiplier", "0.6",
    "--dry-base", "1.75",
    "--dry-allowed-length", "4",
    "--dry-penalty-last-n", "-1",
    "--dry-sequence-breaker", r'\n,:,\",*,;,{,}',
    "--samplers", "top_k;top_p;min_p;temperature;dry",
    "--alias", "laguna-s-2.1",
    "--cache-reuse", "256",
    "--cache-ram", "1024",
    "--host", "127.0.0.1",
    "--port", "8080",
    "--verbosity", "4",
    "--chat-template-file", r"G:\xlam3\laguna\chat_template_laguna.jinja",
]

if __name__ == "__main__":
    try:
        print("Starting llama-server...")
        subprocess.run(cmd, env=env, check=True)
    except KeyboardInterrupt:
        print("\nServer stopped by user.")
    except Exception as e:
        print(f"\nError running server: {e}")

---

Authorship & license

MIT.