Skip to main content
Version: 3.0.0

Configuration

Deepseek for Jira is configured once per site. In version 1.1.0, configuration is performed outside Jira during service setup. There is no administration page inside Jira.

Model provider​

The service uses an OpenAI-compatible configuration:

  • Endpoint base URL.
  • Model.
  • API key.

The same configuration feeds the plan and the implementation. The default documented endpoint for regulated customers is self-hosted. Using the public DeepSeek API is a configuration change.

If the public DeepSeek API is used, the currently published models are:

ModelUse
deepseek-v4-proMain model in the V4 family.
deepseek-flashCost-effective variant in the V4 family.

Exact names and availability are documented by DeepSeek. Do not infer them from previous versions.

Data sent to the provider​

The agent sends the following to the model provider:

  • The work item summary and description.
  • The work item comments.
  • Text attachments.
  • The title and URL of the work item's links (repository, branch, pull request, and pages).
  • The repository files it reads to plan, answer, or review.
  • The relevant repository diff.

No secrets, credentials, or Jira token are sent.

Binary attachments

Binary attachments are skipped. The agent says so in the conversation.

Bitbucket OAuth consumer​

Bitbucket access uses a workspace OAuth consumer:

  • Scopes: repository:write and pullrequest:write.
  • No personal API token is used.
  • The app does not request Bitbucket scopes.
  • The service works with a single repository per task. If there is no repository or there are several, the agent asks. GitHub URLs are ignored.

Agent environment network​

The environment where the agent reads and changes code has a closed network. It can only reach two destinations:

DestinationPurpose
Model endpointPlanning, implementation, and review.
developer.atlassian.comReading the public, current Forge documentation.

Any other destination is blocked and there is no web search. If the network cannot be closed, the task does not start and there is no push. The Jira token does not enter this environment.

Usage limits​

The administrator defines three limits per task:

LimitWhat it controls
TokensTotal consumption of the plan and implementation.
StepsNumber of agent steps during implementation. By default, 80 steps per task.
TimeWall-clock time of the task.

When a limit is reached, the task stops. The message states which limit was reached and what was completed. Partial work is not published, and the branch remains as it was before that task.

Change in 1.1.0

The default step limit goes up from 40 to 80, because every change now goes through exploration and review before the push.

The token and time limits have configurable defaults that are not published in this guide.

Isolation and credentials​

  • Each installation is registered and isolated. The agent only accepts work from known installations.
  • A reinstall does not mix tasks from the previous cycle.
  • Uninstalling deletes the site's credentials and tasks.
  • The model key and Bitbucket consumer secret are stored encrypted.