feat(coding-agent): add template variables to custom system prompts

[ogulcancelik]

adds {{tools}}, {{context}}, {{skills}} template variables for SYSTEM.md.

when present, replaced with dynamic content. no templates = full replacement (current behavior preserved).

also tracks injection metadata through sdk → session → ui, so the welcome screen only shows "loaded context/skills" when actually injected into the system prompt.

[badlogic]

Thanks! This touches a lot of stuff that's been changed in the great refactor, I will integrate this manually once the refactor is complete.

[badlogic]

Also need to integrate this #896

[ogulcancelik]

Also need to integrate this #896

let me know if you'd like me to tackle that

[scutifer]

The .replace is simple and robust, but I wonder if a minor extension would allow a lot more functionality without bringing in Moustache or Handlebars or NJK.

Something like this for example:

function render(template: string, context: Record<string, any>): string {
  // 1. Resolve nested paths (e.g., "user.name")
  const resolve = (path: string, ctx: any) => 
    path.trim().split('.').reduce((obj, key) => obj?.[key], ctx);

  // 2. Handle Logic Blocks: {% if ... %} and {% for ... %}
  // This regex matches {% tag condition %} content {% endtag %}
  let output = template.replace(
    /{%\s+(if|for)\s+(.*?)\s+%}([\s\S]*?){%\s+end\1\s+%}/g,
    (_, type, condition, inner) => {
      if (type === "if") {
        return resolve(condition, context) ? render(inner, context) : "";
      } else {
        // For loops: expects syntax "item in list"
        const [itemKey, , listKey] = condition.split(" ");
        const list = resolve(listKey, context);
        if (!Array.isArray(list)) return "";
        return list
          .map((item) => render(inner, { ...context, [itemKey]: item }))
          .join("");
      }
    }
  );

  // 3. Handle Variable Interpolation: {{ variable }}
  return output.replace(/{{(.*?)}}/g, (_, path) => {
    const value = resolve(path, context);
    return value !== undefined && value !== null ? String(value) : "";
  });
}

You would use it like this

const context = {
    cwd: resolvedCwd,
    tools: tools.map(t => ({ name: t, desc: toolDescriptions[t] })),
    hasSkills: skills.length > 0,
    skills: skills,
    contextFiles: contextFiles
};

const template = `
You are an expert assistant working in {{ cwd }}.

Available Tools:
{% for tool in tools %}
- {{ tool.name }}: {{ tool.desc }}
{% endfor %}

{% if hasSkills %}
# Skills
{{ skills }}
{% endif %}

# Project Context
{% for file in contextFiles %}
## {{ file.path }}
{{ file.content }}
{% endfor %}
`;

const systemPrompt = render(template, context);

Then you move the actual string to a file. Users can even bring custom templates in the future via extensions.

[ogulcancelik]

@badlogic rebased/refactored implementation after your resourceloader rework.

regarding #896 - i can add that to this pr too, but wanted to check first since it feels like a design choice. currently system prompts aren't stored in sessions at all (no SystemPromptEntry type, nothing in SessionContext). adding persistence would mean storing the prompt when --system-prompt is used and restoring on resume. but then what happens if you want to resume a session with a different prompt? e.g. you updated SYSTEM.md and want existing sessions to pick up the new version. happy to implement either way, just wanted to confirm intended behavior.

@scutifer yes that approach feels a lot comprehensive, though again design choice. i just want my custom system prompt in there for now, either way works for me.

[airtonix]

i'd prefer to have use of nunjucks tbh, means less core churn and more flexibility for end users to extend in their own way