Skip to content

CHANGELOG.md

Latest commit

 

History

History
115 lines (69 loc) · 11.4 KB

CHANGELOG.md

File metadata and controls

115 lines (69 loc) · 11.4 KB
Raw

Changelog

Notable changes per release. The ## X.Y.Z section for each version is published verbatim as that version's GitHub Release notes (.github/workflows/pypi-release.yaml). Newest first. Releases before 0.1.1 predate this changelog; see the compare views at https://github.com/max-sixty/tend/compare for their history.

0.1.7

Improved

  • Generated workflows pin actions/checkout to v7. All generated workflows (and tend's own) move from checkout v6 to v7. The review workflow opts into v7's fork-PR checkout guard (allow-unsafe-pr-checkout: true), which otherwise refuses to check out a fork's refs/pull/N/{merge,head} under pull_request_target (the "pwn request" guard), so fork-PR reviews keep running. (#725)
  • Both Claude harnesses update to claude-code 2.1.185. (#719)
  • The bot surfaces a blocking scope rule instead of silently routing around it. When a running-in-ci scope restriction blocks the right action — e.g. engaging an existing upstream thread in another repo — the bot now surfaces the blocker on the triggering thread and offers either to take the upstream action on approval or to relax the rule via the consuming repo's running-tend overlay, rather than substituting a second-best local workaround without signaling it hit a wall. (#717)

Fixed

  • CI-poll loops fit the Bash tool's 10-minute cap. The bundled running-in-ci poll recipes cap their sleep loops at 9 iterations and call the Bash step with timeout: 600000, so the harness no longer auto-backgrounds a longer loop and strands the gated follow-up (dismissing a stale approval, posting failure analysis). (#695)
  • Nightly workflow-regen bases its worktree on an open PR, not branch-ref existence. The nightly skill's regen step now bases on the tend/update-workflows branch only when an open PR rides it, and otherwise bases on HEAD and drops any leftover remote branch. A PR previously closed without merge no longer leaves a stale branch that inflates the diff, produces an inaccurate PR body, or defeats the no-value skip. (#721)

Documentation

  • The codex effort value list in the README and the install-tend skill is corrected to low | medium | high | xhigh. (#710)

Internal

  • Composite-action step bodies are de-duplicated into scripts under shared/steps/, and each harness action lives under a harness-named path. Generated workflows now invoke max-sixty/tend/claude@X.Y.Z (and claude-interactive) rather than the bare-root default; existing pinned refs keep resolving and the nightly regen stamps the new path automatically. (#712)
  • review-reviewers documents the pull_request_review self-trigger as expected (non-)behavior, and the worker-deploy comment corrects the live-stream count to two. (#707, #711)

0.1.6

Improved

  • The default claude harness runs the official binary headless behind the credential proxy. The root action.yaml was rewritten to run claude -p as a non-sudo tend-sandbox user behind the same credential-injecting mitmproxy the interactive harness uses, replacing the anthropics/claude-code-action@v1 wrapper that handed the bot PAT and the Anthropic credential to the agent directly. Both credentials now live only in the proxy and never enter the agent's environment or disk; completion is the claude -p exit code. The action gains claude_version, timeout_seconds, and mitmproxy_version inputs and drops the unused claude-code-action passthroughs. (#704)

Internal

  • Bundled skills replace guidance duplicated from running-in-ci — triage's recheck-before-posting and review-runs' read-only-mount workaround — with references to the canonical sections. (#703)
  • The claude-smoke workflow that exercises the headless harness end-to-end becomes workflow_dispatch-only, matching interactive-smoke. (#706)

0.1.5

Improved

  • Interactive harness isolates both credentials behind the proxy. Phase 2 extends the credential-injecting proxy to the Anthropic model credential, so the sandboxed agent holds only dummies for both the GitHub PAT and the model token while the runner-owned proxy injects the real values per host. The agent toolchain now installs directly as the non-sudo sandbox user (dropping a ~200 MB per-run copy), and the proxy also injects the PAT for raw.githubusercontent.com. (#686, #683, #684)
  • install-tend isolates each bot's auth in a per-bot GH_CONFIG_DIR. Bot credentials live in a dedicated ~/.config/gh-bots/<bot-name> dir selected per command and stored outside the OS keychain, removing the gh auth switch choreography that could strand a bot as the active account and 403 a maintainer's pushes. (#688)
  • Interactive harness updated to claude-code 2.1.179. The pinned claude binary resolves --model opus to Opus 4.8. (#697)

Internal

  • Bundled skill refinements: nightly skips stamp-only workflow-regen PRs and scopes "Notable changes" to adopter-relevant entries, review-reviewers keeps an audit trail on empty-window cycles, and over-prescriptive guidance is reframed as examples and open frames. (#693, #692, #689, #690, #675)
  • tend-repo maintenance: a weekly task keeps the pinned agent binaries current, integration-fixture secrets reseed outside the sandbox, and the secret env-gating rejection analysis is recorded alongside a CLAUDE.md restructure. (#696, #685, #687)

0.1.4

Improved

  • Claude harnesses run with bypassPermissions. The previous dontAsk mode hard-denies writes to Claude Code's protected paths (.github/, dotfiles), blocking autonomous fixes that touch those files. Everything the bot writes still lands through a reviewed PR. (#677)
  • GitHub Releases publish on tag push. The release workflow extracts the version's section from CHANGELOG.md and creates the release; 0.1.1–0.1.3 are backfilled. Nightly workflow-update PRs now summarize notable upstream changes instead of pasting a file list. (#678)
  • install-tend triages an existing bot PAT before minting a new one. The bot-token step runs the scope audit and routes to reuse, scope refresh, or first-time login. (#680)
  • The review skill checks the PR's check rollup before approving, so visible CI failures aren't rubber-stamped. (#667)

Fixed

  • The interactive harness passes GitHub Actions context env vars (GITHUB_RUN_ID, GITHUB_REPOSITORY, …) into the sandbox; skill recipes for run self-reference and URL construction depend on them. (#664)

Documentation

  • README clarifies that the weekly workflow approves dependency PRs rather than auto-merging them. (#673)

Internal

  • Skill refinements across running-in-ci, triage, and ci-fix: end the turn only when work is shipped, defer test suites to PR CI, split CI monitoring into gated/ungated cases, label transient-tracker issues tend-outage, and carve out bot-authored machine-report issues. (#661, #671, #669, #670, #666)

0.1.3

Improved

  • Interactive harness isolates the GitHub PAT. The agent runs as a non-sudo tend-sandbox user; the PAT lives only in a local credential-injecting proxy that adds it for GitHub hosts, so it never enters the agent's environment. (#652)
  • Prior-run context recall. The bot recalls context from earlier runs on the same issue or PR instead of starting cold each invocation. (#649)
  • tend now dogfoods the claude-interactive harness for its own review/mention/triage/ci-fix workflows. (#622)

Fixed

  • Worker reliability: throw on Search failures instead of caching an empty payload, return 503 (not a 200 all-zero payload) on cold-cache failure, and share the /activity payload across colos via a KV tier. (#648, #650, #653)
  • Site liveData polling self-schedules so ticks can't overlap. (#655)
  • Mention workflow skips no-op sessions for undirected bot comments. (#608)
  • review-reviewers pre-creates the monthly tracking issue to eliminate a matrix race. (#657)

Internal

  • Skill refinements across running-in-ci, notifications, review-runs, ci-fix, and running-tend; dead-input and template cleanups in the actions and generator.

0.1.2

Improved

  • claude-interactive harness. A PTY-supervised alternative to the Agent SDK: runs the official claude binary under a script(1) supervisor with a Stop-hook sentinel. (#609)
  • Per-workflow harness override. Trial a different harness (and matching model) on one workflow at a time. (#612)

Fixed

  • Mention workflow uses comment.updated_at so edit events report accurate queue delay. (#595)
  • Interactive harness: token-usage jq parser double-iterator fix, and the -newer filter dropped from the session-JSONL parser. (#616, #614)

Internal

  • Skill hardening: close the env-filter loophole for ALL_INPUTS secrets, recheck PR state before pushing follow-up commits, raise the bar for repo-overlay PRs, and trigger upstream-bot rebases instead of manual conflict resolution. (#599, #573, #604, #605)

0.1.1

Internal

  • Skill refinements: a weekly integration-test recipe and a release-workflow fix. (#590, #589)