[github-actions]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

github-agent-workflow@1.10.1

Patch Changes

• f92015c: diffWorlds(before, after) in scripts/scenario.mjs turns the change a run left behind into the effects the scenario checker takes, so a real run can be judged by the same scope, invariant and oracle checks as a golden scenario. A change that no effect can express (a title or body edit, an edited or removed comment, a reopened item) is reported as an unmodelled effect, which checkScenario rejects unless the contract lists a new unmodelled write scope. Tests round-trip every golden scenario and 600 generated worlds.
• 302d78c: A gh simulator for tests (tests/sim/bin/gh) answers gh calls from a world file in the scenarios' format and logs every call. This part covers the read paths: issue list|view, pr list|view|checks, label list, repo view, auth status and run view, with --json fields checked against the ones gh really has. A command, flag or field it does not cover exits with unsupported: <what>, and a test fails when a skill calls a gh command that is not in the manifest. The write paths, api and fault injection follow.
• 862e759: The gh simulator for tests now writes. issue edit|close|create|comment|reopen, pr edit|create|comment|merge and label create change the world file the way gh would and refuse what gh refuses (an unknown label, a merge pinned to the wrong head, a duplicate PR), leaving the file untouched when a call fails. A run's changes are read back by diffWorlds. For that, a create_issue or create_pr effect can carry the new item's title, body, creation time, author and branch as data, and a close effect its reason.
• b0f6069: grill-via-github-comments has a ## Contract (kind: utility, pure, used_by: [unattended-triage]) and three scenarios: two independent decisions asked as one round, a dependent decision left for a later round, and an open question the code answers, stated as a finding. The contract says what the text already said, that the skill returns the round and posts nothing, and its description now says it drafts the questions for the caller to post. The scenario checker accepts a list as a document value and has an oracle that checks a round against the subject's decision tree: the whole frontier is asked, nothing whose prerequisite is open, nothing that can be looked up, and every question has a recommended answer. See ADR 0013.
• 3342d51: The five group skills have scenarios (three each). A group scenario lists expect.dispatches, the group's results and its one run-log entry, and for triage-github-issues the reconcile call that comes first. The checker picks the batch from the real item skill's eligibility oracle, oldest first, so a dispatch of an ineligible item, one past the batch, or in the wrong order fails. It reads what a dispatch must carry from the real item contract, write_log=false and the run's inputs that the item lists in params, so a dispatch that drops trusted_authors, or forwards something the item does not list, fails. explore-github-repo is checked as its own kind of group: findings ranked best first, one call to file-github-issues with limit and the shared inputs, only filed issues counting toward the cap, and the count line and deferred list in the log. The coverage gate wants three scenarios for a group skill, one that dispatches and one that does not. See ADR 0012.
• f25834a: merge-safety has a ## Contract (kind: utility, pure, used_by: [review-github-pr]) and five scenarios: a breaking change and a migration with no rollback plan (BLOCK), a clean small PR (PASS), and a moved head and a pending check (HOLD). The scenario checker has a mode for a skill that returns a document: the report must have the contract's lines and a verdict it allows, a pure skill must leave no effects, and the world settles what it can (a draft, a pending check or a head that is not the pinned one must give HOLD). The coverage gate counts a utility skill's verdicts. One behaviour is now written down: when head is given and is not the PR's current head, merge-safety assesses nothing and reports HOLD. See ADR 0013.
• 6ae692f: claim and renew now edit this account's own lease comment for the step in place instead of posting a new one each time, so an item taken through three claims of a step has one lease comment, not three. Each step keeps its own comment, a comment from another account is never edited, and if the edit fails the helper posts a new one. Because a comment's position in the thread no longer says which lease is newest, lease_live and leaseState now take the lease written last (updated_at, else created_at, else thread order); the expiry is still the expires= in the marker. New helper ownLeaseComment and CLI command own-lease in helpers.mjs, shared fixtures for the ordering, and shell tests for one comment per step, other accounts, the failed-edit fallback and the ordering (One lease comment per claim costs a comment on every item on every run #69).
• e5fad13: Scenario checks now have an eligibility oracle for every item skill: triage-github-issue, implement-github-issue and review-github-pr join shepherd-github-pr. Each mirrors its skill's requires, models the lease as lease: live | stale | none on the item, and honours lease_held. A test compares each skill's number of requires entries with the oracle's, so a changed precondition prompts a look at the oracle. A skipped item no longer has to be free of lease labels, because the label may belong to the run that holds it. Test tooling only, with no change to any skill. See docs/adr/0012-scenarios.md.
• b563fd8: prs_for_issue and issue_for_pr work on a real gh. They asked gh pr list and gh pr view for authorAssociation, which gh rejects as a PR field (Unknown JSON field), so neither could find the PR that belongs to an issue. Both now read the REST pull request, which has the author's association, and apply the same trust rule as before. issue_for_pr reads GitHub's own issue links with gh 2.72 or newer, and without that it still reads the description. It now prints the issue numbers in numeric order.
• 8eb937a: A scenario can say what a second run finds, so that idempotent in a contract is checked. expect.rerun is ineligible (a terminal outcome), eligible (a failed run, which leaves the item for a retry) or {skip: <reason>}, and the checker asks the skill's own eligibility oracle about the same item in the world the run leaves behind. A push puts the PR's checks back to pending, so a pushed fix is not eligible again until CI reports. A scenario with interrupted: true returns no result and must be eligible again, with a stale lease, and its rerun_effects may not repeat what the interrupted pass already did. All 30 scenarios that complete a run now carry a rerun, two scenarios cover an interrupted triage and an interrupted implementation, and the coverage report lists any scenario that has none, which npm test fails on. A test with a deliberately broken oracle shows a terminal state that is still eligible being rejected.
• dcfca22: retrospective has a ## Contract (kind: utility, pure, used_by: []) and five scenarios over a world of run-log issues: an entry older than days is not used, step and only_runner filter, ranked suggestions never exceed max_suggestions, and a log with only idle entries gives an empty report. The scenario world can now describe run-log entries (run, step, runner, at, outcome, problems) and a scenario carries a now date. The skill now says what the report contains when no entry is left: Entries considered: 0, a summary that says nothing was found, and no suggestions. The validator accepts used_by: [] for a skill that is not user-invoked only when the skill's text says it is run by hand, which retrospective now does. See ADR 0013.
• cd2e13c: docs/RUNNERS.md now covers the sweeper and the newer inputs. It says when to schedule reconcile-github-issues (and that Triaging also runs it), what a runner that dispatches items itself must pass (the item's URL, write_log=false and the group's inputs), and where write_log, lease_held and limit are defined. The README's schedule table lists the sweeper.
• 14b9c41: setup-github-agent-workflow has a ## Contract (kind: utility, user-invoked, used_by: []) and five scenarios: an empty repository gets every label, an existing agent: label is reset and an existing triage label is kept, a second run changes nothing, the settings go into CLAUDE.md when it exists, and only the ## GitHub workflow section is updated. Every skill now has a contract, and CONTRACT_PENDING is empty. The skill lists its nine settings in an Inputs table, which the protocol's Defaults rows now name it as a user of, and the protocol says that repository writes include repository settings such as the label set. The scenario world can hold the repository's labels and files, with label_create, label_update and write_file effects that diffWorlds also produces. The label set a scenario expects is read from references/labels.md. See ADR 0013.
• 87feef1: tests/lib.test.sh now checks for jq, git and node first and exits with one line naming the missing tool, instead of printing dozens of jq: command not found errors and ending with a long list of failures. The hooks check in npm test now says how to install the hooks (npm ci) and how to skip the check (HUSKY=0). AGENTS.md lists what the checks need.
• 85dd01f: unattended-triage has a ## Contract (kind: utility, pure, used_by: [triage-github-issue]) and four scenarios: a complete bug (ready-for-agent), a feature with no use case and an issue with an open question (needs-info), and a wontfix recommendation (ready-for-human, never a close). The skill's description said it writes the outcome comment, while its body and its caller say the caller posts it. The description now says it drafts the comment and the caller posts it, as the body always did. The skill's text now says that an open question or a missing use case is needs-info and that wontfix is ready-for-human, which triage-github-issue already applied after the fact. The scenario checker gains an oracle for the state, the comment structure and these rules. See ADR 0013.