Repository navigation
0.2.11
·
332 commits
to main
since this release
Immutable
release. Only release title and notes can be modified.
Improved
- The twelve cross-cutting
references/files are skills of their own, andrunning-in-ciis nowrun-tend.references/posting.mdis/tend-ci-runner:post-to-github,references/pushing.mdis/tend-ci-runner:push-commits, and so on forrespond-on-thread,open-pr,monitor-ci,ground-claims,check-requester-access,dismiss-approval,act-in-other-repos,read-session-logs,propose-instructions, andfix-a-bug; the skill listing every session already holds is the index, andnightly,weekly,notifications, andci-fixcite each skill at the step that takes the action.review's four references stay references — every skill whose steps reach them can be named,weeklyandreview-runsamong them. (#1305, #1306) model:accepts any name the harness CLI accepts, so a consumer can pin an exact id (claude-opus-5) rather than an alias that moves when a new model is promoted behind it. Neither harness is enumerated in the generator now;claudeandcodex execjudge the name, and a typo fails the job with the bad value named rather than failinginit. (#1294)install-tendhands the Claude OAuth mint to the user — two commands in their own terminal, withgh secret setrun by them so the token never leaves it, or pasted back for the agent to set. The pty wrapper that droveclaude setup-tokenand its tests are deleted. (#1295)- The sandbox boundary's Ubuntu and npm packages resolve as of one recorded instant. apt reads
snapshot.ubuntu.comand installs bubblewrap, socat, and ripgrep by exact version; npm gets--beforethe same instant, which freezes the Sandbox Runtime dependency tree that builds the bwrap argv and terminates the agent's TLS. The weekly refresh resolves every architecture and stops on a divergence rather than writing a pin that installs on one. (#1289, #1291) tend-mentionno longer wakes on a third-party review bot's approval. An approval with no inline comments is terminal when its author is aBot, whatever its body says; aCOMMENTEDreview from the same bot still runs, since only a reader can tell a real finding from a template. (#1292)- The
tend-outagetracker names a stale action pin as a remedy when the workflow that failed pins a release behind tend's newest, and says to runuvx tend@latest init. A boot-path failure takes out the regeneration that would pick up its fix, so the tracker is the only artifact a consumer gets in that window. (#1288) - Shipped instructions carry mechanics rather than tend's own calls, after a survey of every file that reaches a consumer's repo:
reviewno longer objects to compatibility layers on a published library's behalf,nightly's checklist no longer applies tend's placement rules to a consumer's instruction files, and the escalation ladder offers a CI job rather than adding one. The autonomy and skill-loading directives move from Python constants intoshared/system-prompt.md, which reaches both harnesses — so Codex sessions get the autonomy directive for the first time. (#1284, #1280, #1287)
Fixed
tend checkverifies release immutability without repository admin.GET /repos/{repo}/immutable-releasesis admin-only, so the check reportedSKIPin every scheduled run whether the setting was on or off — which also made the nightly's "every line PASS, close the drift issue" branch unreachable in every consumer. It now reads theimmutableflag on the most recently published release when the settings endpoint 404s. (#1290)review-runsreads outage trackers a maintainer closed, not just open ones, since the live scan cannot distinguish a PR whose review died in the outage from one the maintainer merged without waiting. Trackers still open at the end of the drain are closed; one someone else closed stays closed. (#1299)token_report.pyadmits a run on when it finished, not when it was created, matching the census window. A run that started before the window's anchor and finished inside it was priced by neither sweep, and those runs are by construction the longest ones — so the published spend was short by exactly the sessions the figure exists to surface. (#1301)list_recent_runs.pyasks for 200 workflows and warns on hitting the limit.gh workflow listfetches 50 and says nothing when it truncates, so on a large repo atend-*workflow could fall outside the listing and its runs read as a quiet window rather than an incomplete one. (#1293)- The worktree recipes in
review-runsandpropose-instructionskeep theircds inside subshells, so removing the worktree no longer leaves the session standing in a deleted directory with no working directory for the steps after it. (#1297) triageends the run on a bot-maintained evidence tracker without commenting — an issue whose body says Do not close manually and which later runs append to carries no report. (#1303)
Internal
- The CI suite runs every six hours as well as on push and PR.
test-sandboxinstalls bubblewrap from the live Ubuntu archive, so it is the check that catches an archive update changing the sandbox boundary — and every other watch that would notice boots an agent inside that same sandbox. (#1282)