Tabs
What the figures cover
Figures cover agent sessions that reported this repository’s Git remote. The Atlan CLI reports the remote for Claude Code and Codex sessions run inside a Git checkout. Older sessions without Git metadata, sessions run outside a checkout, and harnesses that do not report one are left out. A date range and filters for Branch, Harness, Model, and Person narrow every card on the tab at once, and a link to the page carries the range and filters chosen. When no session in the range reported the repository’s remote, the profile says so instead of drawing zeros and suggests a longer range. When the repository has no valid HTTPS Git remote, it asks you to add one, because there is nothing to match runs against. Where Atlan is available in your organization, cards offer Explain: select it and Atlan accounts for the card’s figures in a note under the card.Overview
Repository summary shows Sessions, People, Tokens, Recorded cost, and Error rate, each with its change from the previous period. Recorded cost is lower than real spend where a model reports no price, and the error rate is the share of traces with a failed step in this repository. Activity charts the period against the one before it; switch it between Tokens, Traces, and Sessions. Who works here lists the credential owners with Git-attributed sessions in the repository and the harness each used most, beside the share of traces by harness and the traces by model. Select a person, a harness, or a model to filter the page. Context in this repository counts the skills, agents, and scorers sourced from it, and Next steps names the loops Atlan can run from what it has seen here.Insights
- Flagged names what stands out: skill and tool error rates at least twice their trace-weighted peer average, and a branch’s tokens per trace at least twice the repository’s. Each needs enough runs to mean something. A flagged skill’s action opens Improve, and a flagged branch’s action filters the page to it.
- Branches, Skills, Tools and MCP servers, and Harnesses and models rank traces, loads, calls, recorded cost, and error rate. A trace can count for several tools.
- Error rate and Recorded cost chart the range against the previous period.
- Latency shows median, p90, and p99 seconds per trace. It cannot be split by branch, skill, or tool yet.
- Cache and tokens shows the cache hit rate, output tokens, and tool calls.
Context
Skills from this repository lists the skills published from it with Loads here, Error rate, Cost, and an Improve action for each. Usage is for this repository only; the skill page shows every repository. Borrowed skills lists the skills agents loaded in this repository in the selected range that come from another source. Nothing borrowed in the range leaves the card empty. Instruction files are listed as coming soon.AGENTS.md and CLAUDE.md are not in the registry yet, so Atlan cannot show which ones agents read here.
Agents and scorers lists the agents and scorers sourced from this repository.
Improve
Improve skills lists the repository’s skills, each with the state of its check: select Get suggestions to start one, follow its progress, and review what it proposes on the skill. A check without enough runs to judge from says so, and so does a row whose skill you cannot edit. Applied suggestions become a new version in Atlan. The next publish from GitHub replaces it unless the change also lands in the repository. Learnings from traces offers Ask Atlan for learnings: Atlan reads the repository’s runs and names up to three lessons, each with the figure behind it and where it belongs. Fix AGENTS.md and Fix agents are listed as coming soon; improvement runs cover skills only for now. Improvement runs lists the recent runs across the repository’s skills, newest first, with each run’s status and its changes. Select a run to open it on the skill and review its suggestions.See also
- Compare agent usage: The Repositories report and the other Reports pages.
- Review skill improvement suggestions: Run a check and review a skill’s suggestions.
- How skill source works: Where a published skill’s repository link comes from.