@@ -233,6 +236,52 @@ AI-powered calendar and time-blocking assistant that automatically schedules foc
**Common uses:** Protecting focus time on your calendar, auto-scheduling recurring habits (lunch, exercise), optimizing meeting-heavy weeks.
## Support agentic harnesses
They typically connect [Duo CLI](https://docs.gitlab.com/cli/duo/cli/) or `[OpenCode](https://opencode.ai/)` to GitLab and other systems through the [Model Context Protocol (MCP)](https://docs.gitlab.com/user/gitlab_duo/model_context_protocol/), and are tailored to ticket work, investigation, and workflow automation.
A governed containerized Support toolkit that packages `glab` + Duo plus Glean, GitLab MCP, Semantic Search, and Ticket Search over sanitized past Zendesk tickets.
**Common uses:**
1. Running `glab` and Duo in a ready-to-use, governed environment.
1. Searching sanitized historical tickets with semantic and ticket search.
### Request for Help (RFH) Agent - Search and Create issues
**Approval:** Approved (Red)
**Reference:**[Request for Help](https://gitlab.com/gitlab-com/request-for-help/)
An AI agent that lives in the RFH project, used to search for and create issues. As a Duo custom agent, it operates under Red data approval.
**Common uses:**
1. Finding existing RFH issues.
1. Creating new RFH issues directly from the RFH project.
An agentic investigation harness for SOS bundles, logs, configs, screenshots, and ticket conversations. It provides Glean access and 27+ purpose-built MCP tools.
> [!note]
> The harness itself handles Red data locally, but its Glean integration is restricted to **Orange data only**.
**Common uses:**
1. Investigating SOS bundles and correlating logs and configs.
1. Reviewing screenshots and ticket conversations during deep-dive troubleshooting.
## Non-Approved Tools
The following AI tools are **not approved** for internal use in normal workflows. Using them with GitLab or customer data violates internal policy.
@@ -118,22 +118,29 @@ When Red data is involved (full ticket content, logs, configuration snippets), G
### Model selection
You can select which large language model (LLM) to use with DAP. The choice affects response quality, speed and reasoning depth.
You can select which large language model (LLM) to use with DAP. The choice affects response quality, speed and reasoning depth, and the right choice depends on whether you are using Duo Chat directly or an agentic harness.
| Model | Speed | Reasoning depth | Best for |
|-------|-------|-----------------|----------|
| Claude Haiku 4.5 | Fastest | Basic | Quick lookups, simple summaries, reformatting text |
| Claude Sonnet 4.6 | Fast | Strong | Most day-to-day support work: diagnosis, drafting, research |
| Claude Opus 4.6 | Slow | Strongest | Deep investigation, ambiguous problems, long multi-step reasoning |
| Claude Haiku | Fastest | Basic | Quick lookups, simple summaries, reformatting text |
| Claude Sonnet | Fast | Strong | Most day-to-day support work: diagnosis, drafting, research |
| Claude Opus | Slow | Strongest | Deep investigation, ambiguous problems, long multi-step reasoning |
**Recommended approach:**
#### Using Duo Chat directly
1. Start with **Sonnet 4.6** for most tickets. It handles diagnosis, solution research and drafting well.
1. If the response lacks depth or misses key details, retry the same prompt with **Opus 4.6**.
1. Use **Haiku 4.5** for high-volume, low-complexity tasks where speed matters, such as batch-summarising tickets or converting bullet points into prose.
1. For deep-dive investigations that cross-reference code, issues, MRs and documentation, **Opus 4.6** produces the most thorough results.
1. Start with **Sonnet** for most tickets. It handles diagnosis, solution research and drafting well.
1. If the response lacks depth or misses key details, retry the same prompt with **Opus**.
1. Use **Haiku** for high-volume, low-complexity tasks where speed matters, such as batch-summarising tickets or converting bullet points into prose.
1. For deep-dive investigations that cross-reference code, issues, MRs and documentation, **Opus** produces the most thorough results.
The model you select does not affect what data the tool can access. All models within DAP operate under the same Red data approval.
#### Using an agentic harness
When using an agentic harness such as `internal-soslab` or Agent Capabilities for GitLab Support, prefer a larger model. The bigger context window and stronger reasoning handle the log-heavy, multi-source investigation these tools are built for.
1. Start with a large model such as **Opus** when working a ticket end to end.
1. Drop down to a smaller model such as **Sonnet** for well-scoped, individual tasks.
The model selected with Duo does not affect what data the tool can access.
