Method runSandboxed

src/agent/loop.ts:649–871

* Sandboxed wrapper around run(). When `config.sandbox.enabled` and we're in a * git repo, this runs the whole task on a hidden `qodex/sandbox- ` branch: * the agent experiments freely, and we squash-merge onto the user's branch only * if the task completed (a `final` event, not an error

(
    messages: Message[],
    sessionId: string,
    options: AgentOptions,
  )

        `${providerPromptCfg.override}\n\n` +
        `Working directory: ${this.cwd}\n` +
        `Git branch: ${gitBranch ?? '(none)'}\n` +
        `Available tools (the only ones you can call this turn): ${effectiveTools.join(', ')}\n` +
        `If a task needs web data, use \`web_search\` / \`web_fetch\` (if listed above). Do not claim you lack internet access — those tools ARE your internet access.`;
    } else {
      // Classify the user's intent so the prompt can inject task-shaped reasoning.
      const taskClass = this.classifyForPrompt([{ role: 'user', content: userPrompt }]);
      // Stack-specialist expertise: detect from the user's words + what's on disk, then
      // inject the deep how-an-expert-builds-THIS block(s). Orthogonal to task class.
      const stacks = detectStacks(userPrompt, projectSignals);
      const stackAddendum = buildStackAddendum(stacks);
      if (stacks.length > 0) logger.info('Stack specialist profiles active', { stacks });
      sysPrompt = buildSystemPrompt({
        cwd: this.cwd,
        mode,
        modelFamily: detectModelFamily(modelId),
        modelId: resolvedModelId,
        providerName,
        projectInfo: { ...projectInfo },
        projectRules: mergedProjectRules,
        knowledgeFacts,
        directoryTree,
        gitBranch,
        availableToolNames: effectiveTools,
        taskClass,
        stackAddendum,
        skillsBlock: buildSkillsSystemBlock({ prompt: userPrompt }),
        identityBlock: identity.block,
      });
    }

    // Provider-specific guidance, appended on top of whatever base prompt was built
    // (default, role-override, or provider-override). Additive behavioral steering —
    // does NOT replace the read-before-write rule or tool list above it.
    if (providerPromptCfg?.append) {
      sysPrompt = sysPrompt +
        `\n\n# Provider-specific guidance (${providerName})\n${providerPromptCfg.append}`;
    }

    // Static/volatile split: injections are routed into two buffers so the prompt-cache
    // boundary lands between them. `stableTail` holds session-stable guidance (code style,
    // failure lessons) that is byte-identical across turns → folded into the CACHED core.
    // `volatileTail` holds genuinely per-turn context (retrieval, dep-graph, episodic recall)
    // that changes with the query → kept AFTER the boundary so it never invalidates the core.
    let stableTail = '';
    let volatileTail = '';

    // ── Auto-retrieval pre-pass (best-effort) ──
    // Embed the request and inject the most semantically-relevant files so the model
    // starts at the right place in a large codebase instead of grepping blind. Only for
    // top-level normal turns (sub-agents/vision are already scoped). NEVER blocks: bounded
    // by a short timeout and any failure is swallowed to null.
    if (mode === 'normal' && !customSysPromptOverride && !isTrivialMessage(userPrompt)) {
      const ctxCfg = (this.config as any).context ?? {};
      if (ctxCfg.autoRetrieve !== false) {
        try {
          const buildInline = ctxCfg.buildIndexIfMissing === true;
          const files = await retrieveRelevantFiles(this.cwd, userPrompt, {
            embeddingModel: ctxCfg.embeddingModel,