Skip to content

v0.56.1

Latest

Choose a tag to compare

@github-actions github-actions released this 14 Nov 23:37
· 51 commits to main since this release
27c1d6c

Highlights

GitLab runs now send GITLAB_TOKEN only to the run's own GitLab host, and some setups need a change on upgrade. The run stops before setup in these cases: a local fullsend run --forge gitlab with none of FULLSEND_GITLAB_URL, GITLAB_API_URL or CI_SERVER_URL set (set FULLSEND_GITLAB_URL to your instance's https URL, or https://gitlab.com); a GitLab URL that is http:// for a host other than localhost, 127.0.0.1 or ::1 (use the https URL); and a provider profile that uses YAML merge keys (<<), alias keys, duplicate keys or more than one document (write each key out once, one document per file). Agents that call GitLab under another host name no longer receive the token, and the run log lists the hosts removed. Scaffolded GitLab CI jobs set the URL for you and need no change, but they now fail on an unpinned CI_SERVER origin. See the Upgrading table in the fullsend run docs.

Jira agents can run on a service account's OAuth 2.0 client credentials. Set FULLSEND_JIRA_CLIENT_ID and FULLSEND_JIRA_CLIENT_SECRET instead of FULLSEND_JIRA_TOKEN, and fullsend run mints the sandbox's Jira token itself through a new built-in jira provider, refreshing it before it expires. The client secret never enters the sandbox, and the minted token is read-only and limited to the run's own Jira site. Minting needs tracker.jira.base_url and project, and cannot run with --offline or a custom tracker.jira.api_url. The built-in triage and code agents in agents v0.56.1 use the jira provider and accept either credential; static-token setups keep working unchanged. The provider name jira and profile id fullsend-jira are now reserved: an existing local definition under those names gets a warning and still takes precedence. The agents' jira-ro provider and profile are deprecated, so custom harnesses that list them should switch to jira before the next release.

Repositories can tighten command scanning by committing .tirith/policy.yaml. fullsend reads it from the trusted base commit, and tirith merges it as tightening only: block rules, approval rules and network denies, never allowlists or a bypass. It is enforced under Claude Code and Codex; under pi it is best effort. No repository policy applies yet to GitLab merge request runs or to GitHub review checkouts without the base commit; those runs use tirith's baseline and log a note. The sandbox also moves to tirith 0.5.0, and the hook now runs the scanner offline in an environment it builds.

Also in this release: issue transitions work the same way across trackers; github setup, github set and the token-scope checks follow GITHUB_API_URL for GitHub hosts other than github.com (no change on github.com); triage marks an issue as a duplicate only of an open issue; GitLab CI jobs skip the CLI download when the runner image already has the pinned release; and pi sub-agents keep, and are capped at, the parent's MCP and tool permissions.


Changelog

Features

  • 7fbb9ce: feat(#9220): mint the sandbox Jira token from client credentials (@jwm4)
  • 81769d5: feat(#9486): let a repository tighten tirith scanning with .tirith/policy.yaml (@fullsend-ai-coder[bot])
  • a5bfb6a: feat(issues): add tracker-neutral issue transitions (@ederign)

Bug Fixes

  • 8ee4a79: fix(#9039): honour GITHUB_API_URL in github setup and github set (@fullsend-ai-coder[bot])
  • efdade2: fix(#9041): honour GITHUB_API_URL in token-scope, reconcile, eval (@fullsend-ai-coder[bot])
  • a2a6fcf: fix(#9187): export COMMENT_BODY for GitLab Jira agent runs (@fullsend-ai-coder[bot])
  • 1aa500a: fix(#9220): harden the minted Jira provider after review (@waynesun09)
  • 9eba775: fix(#9461): bump tirith to 0.5.0 and update real-binary tests and docs (@waynesun09)
  • 70b109f: fix(#9461): make the offline and approval tests conclusive, correct the docs (@waynesun09)
  • 7ad4a0d: fix(#9461): tirith hook runs offline and blocks approval verdicts (@fullsend-ai-coder[bot])
  • d3d8ae5: fix(#9465): fail GitLab agent job on unpinned CI_SERVER origin (@fullsend-ai-coder[bot])
  • 4d1872e: fix(#9469): bound total grace wait time and retry only lost dispatches (@waynesun09)
  • 5f09b30: fix(#9469): harden dispatch diagnostics and skip teardown summary (@waynesun09)
  • 6f1039a: fix(#9469): retry any driver's lost dispatch and refund cut grace (@waynesun09)
  • 887dd3e: fix(#9469): share grace window and mark partial waits inconclusive (@waynesun09)
  • 4d02080: fix(#9472): classify GitHub server errors as infrastructure (@fullsend-ai-coder[bot])
  • 8916a6f: fix(#9480): quote removed endpoints that hold control characters (@waynesun09)
  • 1464551: fix(#9481): build tirith's environment and working directory in the hook (@waynesun09)
  • c19c9bd: fix(#9481): run the tirith scanner in an environment the hook builds (@fullsend-ai-coder[bot])
  • 73750e8: fix(#9481): stop tirith's policy search at the per-call directory (@waynesun09)
  • 8841a57: fix(#9487): mask the header credential in place (@waynesun09)
  • 4bd7fbe: fix(#9487): redact scheme-prefixed credentials in both redactors (@fullsend-ai-coder[bot])
  • 7de9174: fix(cli): harden the per-site Jira profile renderer and docs (@waynesun09)
  • c5c13c7: fix(forge): drop the tracker key from the allowlist doc comment (@tpouyer)
  • b85ed7c: fix(forge): narrow agent remove's deny-all fallback to used prefixes (@tpouyer)
  • 0f1bb2f: fix(forge): resolve the config allowlist the same way in every loader (@tpouyer)
  • 0699a6b: fix(forge): validate merged managed agents against the effective list (@tpouyer)
  • 20de06d: fix(forge): widen the agent-editing commands to the effective allowlist (@tpouyer)
  • 8fa21fb: fix(forge): widen the agent-editing commands to the effective allowlist (@tpouyer)
  • a955f13: fix(forge): write only the overlay allowlist in agent add and remove (@tpouyer)
  • d0f94cf: fix(gitlab): bind the GitLab credential profile to the GitLab host only (@waynesun09)
  • 770a8b5: fix(gitlab): compare CI_SERVER origin parts and run the check on bash 3.2 (@waynesun09)
  • e8386cc: fix(gitlab): limit GITLAB_TOKEN profiles to the run's GitLab host (@fullsend-ai-coder[bot])
  • 35ec01c: fix(issues): address Jira transition review findings (@ederign)
  • 0e52563: fix(issues): enforce transition boundaries and sync docs (@ederign)
  • 48f3db8: fix(jira): limit the minted Jira token to the run's own site (@fullsend-ai-coder[bot])
  • bdd9b80: fix(runtime): cap pi children at active parent tools (@shanemcd)
  • dea4c51: fix(runtime): deny unselected MCP tools in pi children (@shanemcd)
  • 25cb102: fix(runtime): fail closed on pi child activation errors (@shanemcd)
  • 4240d94: fix(runtime): filter MCP resources in pi parent and child runs (@shanemcd)
  • acb8627: fix(runtime): map Claude MCP resources and remove dead child mode (@shanemcd)
  • a55448d: fix(runtime): pass mcp__ tool names through the pi tool mapping (@shanemcd)
  • fc1c01b: fix(runtime): preserve MCP permissions in pi sub-agents (@shanemcd)
  • daf0e60: fix(runtime): preserve callable MCP tools in pi children (@shanemcd)
  • 584653d: fix(runtime): preserve inactive pi child permissions (@shanemcd)
  • 5bf1eba: fix(runtime): retain implicit pi extension permissions (@shanemcd)
  • 3a554cd: fix(runtime): return structured pi manifest rejections (@shanemcd)
  • e6c345c: fix(runtime): tolerate delayed pi child tool registration (@shanemcd)
  • e4ea267: fix(runtime): translate pi MCP server selectors and hook names (@shanemcd)
  • f56e82a: fix(runtime): validate pi MCP selectors before rendering (@shanemcd)

Refactoring

Others

  • d277dbb: perf(#9251): skip the GitLab CLI download when the image matches the pin (@fullsend-ai-coder[bot])

Agents changes

The fleet agents ship in fullsend-ai/agents v0.56.1; its release notes list the agents changes in this release.