Fullsend

Appearance

156. Separate the Jira API URL from the site URL ​

Date: 2026-11-02

Status ​

Accepted

Context ​

Fullsend's Jira integration (the jira-poll input driver of ADR 0063, status notifications, and fullsend issues) authenticates as a single non-human Jira identity. Atlassian Cloud offers two kinds of identity for that purpose, and Atlassian accepts their credentials at different hosts. The Jira integration guide covers setting up either kind.

A bot account is an ordinary Atlassian user account that a team creates for automation. It has an email address, can log in like a person, holds project roles like a person, and authenticates with a classic API token. Atlassian documents that token for use at the site URL (https://<site>.atlassian.net). Its strengths are that any Jira Cloud site supports it and no organization administrator is involved beyond the usual user invitation. Its weaknesses are that it occupies a user seat, it depends on a password and two-factor setup that someone must own, and its token is an unscoped user credential that organizations may limit in lifetime or forbid for automation.

A person's own Atlassian account authenticates the same way and is indistinguishable from a bot account to Fullsend. It is not a supported choice: the poller treats the account it signs in with as itself, so that person's own comments and label changes never dispatch agents (see the Jira guide's Use a dedicated account for the poller). This ADR uses "bot account" for a regular account created for Fullsend alone.

A service account is an Atlassian-managed identity created by an organization administrator (Atlassian Administration > Directory > Service accounts). It has no password login. Its credential is a scoped API token (or OAuth 2.0 client credentials), and Atlassian accepts it only through its API gateway (https://api.atlassian.com/ex/jira/<cloudId>/...), never at the site URL. Its strengths are that it is designed for integrations: no human login to protect, least-privilege scopes, and administration separate from people. Its weaknesses are that it requires organization-administrator provisioning and that a site must have the feature enabled.

The two tokens look alike. A bot account's token is sent with the account's email, and a service account's token is sent on its own as a Bearer token. The bigger difference is the host. The gateway accepts both kinds: Atlassian documents it for scoped and service-account tokens, and a classic bot token has been observed to work there as well, though Atlassian's documentation does not promise it. The site URL accepts only the bot token.

Before this decision (verified at 6f45fa15b), Fullsend had one URL setting, the site URL, and used it for three jobs: every API call, the links to issues that it writes for humans, and the sandbox egress rule for Jira. It therefore supported only bot accounts (#9007 recorded the gap, with the file-level evidence in this ADR's pull request). A team that had a service account could use it only by pointing the site URL setting at the gateway, which worked for API calls but produced wrong links and left the sandbox without Jira access.

Fullsend supports Jira Cloud only; Jira Data Center is tracked as future work. Atlassian also operates cloud variants with their own hosts (a government cloud with a different gateway, and an isolated cloud with a per-customer domain), which Fullsend has not been tested against.

Options ​

Explicit account kind. Keep the site URL as the only URL and add a setting that names the kind of account, from which Fullsend picks the host. Simple for the user, who knows what they were given. Costs a second authentication code path, requires the user to answer a question that the credential itself could answer, and hard-codes the gateway host for the service-account case.

API URL defaulting to the site URL. Add an API URL setting beside the site URL, used for API calls, defaulting to the site URL. Existing deployments are untouched. A service-account deployment sets the API URL to the gateway address, which Atlassian's own service-account instructions hand them. Fullsend never needs to know the account kind. Costs the user one copied URL, including the cloud ID, and leaves the default on the only host that rejects service accounts.

API URL defaulting to the gateway. The same two settings, but the default API URL is the gateway address for the site, built from the site's cloud ID, which Fullsend discovers from the site URL. Both account kinds work with no extra setting; the override remains for cloud variants with other hosts. Costs a change of host for every existing deployment on upgrade, resting on observed rather than documented behaviour for bot tokens, and a discovery call at startup.

Decision ​

We separate the two URLs. The site URL stays the one required setting and keeps its human-facing jobs: links to issues and discovery of the site's cloud ID. API calls go through the Atlassian API gateway. A separate API URL setting names the gateway host; it defaults to https://api.atlassian.com and can be overridden. Fullsend always calls {API URL}/ex/jira/{cloudId}/..., so an override must serve the same path layout. Fullsend does not ask which kind of account it holds: a token with the account's email is sent with Basic auth, as a bot account's is, and a token without an email is sent as a Bearer token, as a service account's is.

Fullsend always discovers the cloud ID from the site URL. If discovery fails, the run stops and reports the failure rather than falling back to the site URL, because a silent fallback would turn a service-account rejection into an unexplained failure. Fullsend resolves the gateway address once and hands only the resolved {API URL}/ex/jira/{cloudId} to scripts it runs on behalf of agents, so that nothing downstream repeats the discovery.

Every deployment, bot accounts included, talks to the gateway, and there is no setting that sends API calls to the site URL: one would keep unscoped classic tokens on the site host. Upgrading therefore requires a credential that works through the gateway. A deployment pinned to an earlier Fullsend version keeps its current behaviour until it switches. A classic bot token working through the gateway rests on observed behaviour that Atlassian does not promise, and we accept that risk.

A service account can also authenticate with OAuth 2.0 client credentials instead of a token. Fullsend exchanges them at Atlassian's token endpoint for a short-lived Bearer access token, refreshes it before it expires, and sends it to the same API URL. This is another way of authenticating, not another host, so it needs no further URL or account-kind setting. A deployment supplies either a token or client credentials, never both (#9176). How a deployment obtains and rotates the credential, including the longer-term goal of short-lived credentials minted centrally (#2269, #4885), is a separate question.

Forge scope ​

Forge-neutral. The Jira identity and URLs are the same whether the poller runs from GitHub Actions (ADR 0063) or GitLab CI (ADR 0067); only the mechanism that delivers the secret to the job differs per forge, and this ADR does not change it. Forgejo has no Jira poll driver today, so nothing changes there.

Consequences ​

  • Teams can adopt Fullsend with whichever Jira identity their organization provisions, bot account or service account, with the same settings.
  • Existing deployments switch to the gateway host on upgrade and need a credential that works there. A bot token there rests on observed behaviour, and there is no supported way back to the site URL if that ever changes.
  • The sandbox egress profile for Jira must admit the gateway host as well as the site, within the credential-isolation rules of ADR 0017.
  • Fullsend passes the resolved gateway address to agent runs (#9218). Every component that calls Jira, including built-in agents' scripts in fullsend-ai/agents, must call it rather than the site URL before service accounts work end to end.
  • Atlassian's government and isolated cloud variants may be reachable through the API URL override, if their gateways serve the same path layout, but remain untested and unsupported until someone can try them.