Agent tools and evidence
Why OpenStar answers should be grounded in tool results.
OpenStar Chat is agent-centered. The model decides what to do, and the harness executes tools.
The answer is only as strong as the evidence
OpenStar should show tool activity and reject unsupported repository claims instead of hiding them behind a polished paragraph.
Find candidates
Search repositories and collect candidates that can be inspected by later steps.
Read sources
Fetch metadata, README content, releases, and activity before forming a judgment.
Act on taste
Save, Watch, hide, or update Memory when the user gives a clear signal.
Tool surface
| Tool category | Used for | User-visible value |
|---|---|---|
| Repository search | Discover candidates by query, topic, or owner/repo. | The model has a real candidate set. |
| Metadata lookup | Read stars, license, language, topics, and ownership. | Claims can be checked. |
| README and releases | Inspect positioning, setup, roadmap, and recent change. | Answers can explain fit and risk. |
| Summary and comparison | Turn multiple repos into a decision. | The user gets a concise trade-off view. |
| Save and Watch | Capture follow-up intent. | Chat can move work into the product loop. |
| Memory updates | Learn explicit interests and negative signals. | Personalization improves over time. |
Harness behavior
The model can answer when referenced repositories and facts are backed by tool output. The answer should cite what was inspected, not imply unseen certainty.
If the model needs more information, the harness should let it call another tool round rather than guessing. Repair is useful for missing README details, alternatives, or activity checks.
If the model references repositories that were not returned by tools, OpenStar should fail clearly and ask for a narrower prompt or another attempt.
Evidence principles
- Do not let the model answer repo questions from memory alone. Repository names, owners, releases, and maintenance signals change constantly.
- Treat tool evidence as the contract between the agent and the user.
- Keep tool calls visible so users can tell whether Chat is searching, reading, comparing, or acting on their behalf.