Running Tend

Tend CI

Repo-specific guidance for tend workflows running on tend itself. The bundled skills (/tend-ci-runner:running-in-ci, /tend-ci-runner:review, /tend-ci-runner:triage, etc.) provide the workflow framework; this skill adds tend conventions.

Filing issues in other repos

Standing exception granted: file directly in agent-equipped targets without asking permission here first. Most tend consumers in data/consumers.json qualify, as do other Claude-Code-action-using repos. The default rule (open an issue here asking permission first) still applies when the target shows no agent signals.

Non-standard workflows

Tend has Claude-powered workflows beyond the generated tend-* set:

review-reviewers runs only on workflow_dispatch — dispatch it as a spot-check after a release, harness switch, or model bump, not on a cadence. The per-repo tend-review-runs carries the routine loop; the workflow file's header explains the pause.

A dispatched run's window opens at the previous successful review-reviewers run, floored 6h back (list_recent_runs.py). With no cron, dispatches usually sit further apart than that, so the floor is the normal case: the run covers the last 6h and warns on stderr that the rest is a coverage gap. Dispatch it within ~6h of whatever you want it to see.

These use the tend composite action and produce claude-session-logs* artifacts, but their names don't match the tend-* prefix that scripts filter on by default. uvx tend@latest init doesn't rewrite them either, so their max-sixty/tend/<harness>@X.Y.Z pins move only when someone edits the file.

Usage analysis

Pass extra prefixes when running token reports or listing runs so these workflows are included:

# Claude leaves this unset in the shell; Codex exports it.
SCRIPTS="${CLAUDE_PLUGIN_ROOT:-/home/tend-sandbox/tend-marketplace/plugins/tend-ci-runner}/scripts"
uv run --script "$SCRIPTS/token_report.py" "${HOURS:-24}" "review-"
TARGET_REPO=max-sixty/tend uv run --script \
  "$SCRIPTS/list_recent_runs.py" review-reviewers "tend-" "review-"

Under review-runs, $HOURS is the lookback derived from its Step 1 anchor — passing a literal 24 there reopens the window gap that anchor closes. The default keeps an ad-hoc invocation working.

Labels

• claude-behavior — findings from review-reviewers
• review-runs — findings from review-runs

Session Log Paths

Artifact paths: -tmp-tend-agent-workspace-*-checkout/<session-id>.jsonl

review-reviewers runs produce one session log per matrix repo in .github/workflows/review-reviewers.yaml.

Nightly: verify website live data

tend-src.com renders its stat strip, activity feed, and currently-tending dot entirely from the data Worker at api.tend-src.com. Each section hides itself when its fetch fails or returns empty, so a Worker outage shows as a blank page, not an error. Check the Worker directly — it serves the data the site renders. See worker/README.md.

curl -fsS https://api.tend-src.com/activity | jq '{
  prs: .prs.count, reviews: .reviews.count,
  comments: .comments.count, issues: .issues.count,
  recent: ([.prs, .issues, .reviews, .comments] | map(.recent | length) | add)
}'
curl -fsSI https://tend-src.com/ | head -1   # GitHub Pages serving the HTML

Healthy: both return HTTP 200, every lifetime count > 0, and recent > 0. An empty /currently-tending is normal between runs — don't alarm on it.

If /activity is non-200, all-zero, or recent is 0, wait ~60s and retry once. (Transient GitHub errors keep the last good data rather than caching zeros, so a persistent empty is a real signal.) If it persists, file or update one tracking issue (dedup by title, e.g. website: data Worker returning empty) with the failing endpoint, the counts seen, and whether the bots still have recent activity on GitHub — that localizes the fault to the Worker. The bot can't rotate the Worker's Cloudflare-side secret itself, so leave the diagnosis to a maintainer; worker/README.md covers the Worker's setup.

Nightly: don't duplicate the release's regeneration PR

The release skill's deploy step opens chore: regenerate workflows with tend X.Y.Z from the release branch and leaves it open through CI and review. A nightly firing in that window still sees the committed workflows on the old version, so prepare reports a change and Step 7 ships a second PR for the same regeneration. Before shipping it:

gh pr list --state open --limit 100 --json number,title,headRefName \
  --jq '.[] | select(.title | test("regenerate workflows with tend"))'

If one is open, diff its head against the prepared worktree. Where it already carries the same regenerated files, skip the PR, name the covering PR in the run summary, and remove the worktree (git worktree remove <path> --force). Ship anything it is missing on its own branch — the release skill asks for the restamp below in the release commit, so usually there is nothing to ship.

Nightly: restamp the hand-maintained workflow refs

init rewrites only the generated tend-*.yaml files, so the workflows under "Non-standard workflows" hold whatever action ref they were last given by hand. Every release leaves them a version further behind, and a harness change made by regenerating skips them entirely.

Run this after the regen step, whether or not it produced a PR:

rg -o --no-filename 'max-sixty/tend/[a-z/-]+@[0-9.]+' .github/workflows/ | sort -u

The check is on the versions rather than the line count: pipe the same output through sed 's/.*@//' | sort -u and expect exactly one. Two or more, restamp the hand-maintained files onto the generated files' ref and fold it into the regen PR — same worktree, same commit.

How many lines the listing prints follows from the harnesses .config/tend.yaml selects, so read it against the current config rather than a remembered count. With every workflow on the default harness there is one claude path; overriding a workflow to codex adds codex and codex/refresh. Whether a hand-maintained file names the right harness is a separate question the listing cannot answer; compare it against the harness .config/tend.yaml configures, and on a mismatch check what else that config change was supposed to carry.

Weekly: refresh data/consumers.json

Public repos that have installed tend. Read by the website's data Worker (see worker/README.md) to power the currently-tending dot, activity feed, and stat strip. Needs no opt-in because the workflow files are public.

uv run --script \
  .claude/skills/running-tend/scripts/refresh_consumers.py

The command unions code-search results with the current index, verifies each repository, and leaves the existing file untouched if a GitHub read fails. Confirm every repository listed under removed no longer has generated Tend workflows before publishing the change.

Open a PR titled chore: refresh consumers.json if the file changed. Skip the PR (no diff to land) when git status --porcelain data/consumers.json is empty — git diff --quiet returns 0 for untracked paths, so the first-run case would no-op. Code search is 10 req/min — one call covers the whole list.

Weekly: review harness release notes

Browse the stable releases since the previous weekly run (normally one week) for Claude Code and Codex. Most notes are internal and need no Tend response. Only investigate changes to behavior Tend depends on: headless execution and output, permissions and sandboxing, plugin and skill loading and invocation, model selection, or session-log formats.

Changes that have required Tend work include:

• Claude Code ignored defaultMode: bypassPermissions in project settings, so Tend had to pass --permission-mode on argv.
• Codex added non-interactive plugin installation. Tend's action consumes its Installed plugin root: output to locate plugin scripts.

For any similarly relevant note, search the code, issues, and PRs first. Open a PR when the change is small enough to make and verify in this run; reserve an issue for what needs a maintainer decision or verification CI can't reach, linking the release and proposing the change.

Compare the current Codex model catalog with DEFAULT_MODEL_BY_HARNESS["codex"] and Codex model pins in .config/tend.yaml. Move each to a newer model only within the same capability and price tier (for example, Sol tier to Sol tier), verified from OpenAI's model and pricing docs rather than name similarity. A cross-tier change is a product decision, not routine maintenance.

Weekly: bump pinned versions

Every dependency pin in the repo is in scope — no bot watches this repo, so a pin outside the sweep drifts until it breaks. Each command below enumerates its own ecosystem, so a pin added later shows up without anyone remembering to list it here.

# Composite-action inputs
yq -r '.inputs | to_entries[] | select(.key | test("_version$"))
  | "\(filename) \(.key) = \(.value.default)"' */action.yaml */*/action.yaml

# Python: `==` and upper bounds freeze a version. Floors (`click>=8.0`) state
# compatibility instead and stay put — raising one only narrows adopter support.
git grep -nE '(==|~=|<=?)[0-9]' -- '*pyproject.toml'
uv lock --upgrade --dry-run

# pre-commit hook revs — the updater rewrites them, `git diff` is the report.
uv tool run pre-commit autoupdate

# npm: `Wanted` ≠ `Current` is lockfile drift (`npm update`); `Latest` ≠
# `Wanted` needs the range in package.json moved. Exits 1 when a row prints.
# Install first — `outdated` reads `Current` from `node_modules`, which a CI
# checkout never has, and with none it prints nothing and exits 0.
npm --prefix worker ci && npm --prefix worker outdated
npm --prefix site ci && npm --prefix site outdated

# Versions pinned in a shell script (worktrunk, in the Codex Cloud setup)
git grep -nE '^[A-Za-z_]*VERSION=' -- '*.sh'

What upstream currently publishes:

npm view @anthropic-ai/claude-code dist-tags.latest
npm view @openai/codex dist-tags.latest
curl -fsS https://pypi.org/pypi/mitmproxy/json | jq -r .info.version
curl -fsS https://pypi.org/pypi/uv/json | jq -r .info.version
gh api repos/max-sixty/worktrunk/releases/latest --jq '.tag_name | ltrimstr("v")'

GHA uses: refs sweep separately, under a rule of their own — see below. Out of scope entirely: runner images (ubuntu-24.04), node-version, and requires-python are platform choices carrying their own rationale, so they move when a reason arrives rather than on a cadence.

Default rule: move to latest and let CI decide — the table below names the pins where CI can't. Split PRs by who runs the result, and take what fits in one session rather than clearing a backlog at once — an unswept pin waits a week, a swamped run finishes nothing.

• Ships to adopters — claude/action.yaml, codex/action.yaml, and codex/refresh/action.yaml run in adopter jobs from the next release; generator/src/tend/templates/ and workflows.py render into their workflow files. One PR each, titled chore: bump <name> to <version> (the uv-plus-mitmproxy PR names both), its body naming what changed.
• Ours alone — everything else: pre-commit revs, the workspace dev pins, the uv_build backend, npm devDependencies, WORKTRUNK_VERSION, the hand-maintained .github/workflows/ files and .config/tend.yaml. One PR for the lot.
• A major — an npm Latest ≠ Wanted, or a new mitmproxy/uv major — gets its own PR from either bucket, its body reporting the migration notes.

Pins with rules of their own

A stale claude binary resolves --model opus/sonnet to a superseded alias target, so drift silently downgrades the model. In a bump PR, report the release notes between the old and new pins that affect the integration surfaces in the release-note pass above.

mitmproxy_version pins the process that holds the real PAT and model credential, so a security fix there matters here. Check anything security- or addon-related in its CHANGELOG against the mitmdump flags in proxy/setup_sandbox.py, and report the comparison in the PR. uv_version also supplies the agent fallback in both harnesses. CI smokes the installer and proxy together, so move uv and mitmproxy in one PR.

For codex_version, CI's test-codex-surface job installs whatever is pinned and asserts the credential-free CLI surface the action depends on. Before a bump, use an isolated Plus/Pro login to run the refresh action and require both the full and access-only credentials to rotate. Inspect Codex's auth manager too: an unparsable access token must fall through to the stale last_refresh, and that refresh must use the refresh token without requiring the old access token. Report the relevant release notes crossed by the bump in its PR.

uses: refs

git grep -hoE 'uses: [^ ./][^ @]*@[^ ]+' -- ':!generator/tests' ':!*.md' \
  | sed 's/uses: //' | grep -v '^max-sixty/tend/' | sort -u \
  | while IFS='@' read -r action pin; do
      latest=$(gh api "repos/$action/releases/latest" --jq .tag_name 2>/dev/null) \
        || { printf '%-30s %-9s -> no releases; read its tags\n' "$action" "$pin"; continue; }
      case "$pin" in "$latest" | "${latest%%.*}") continue ;; esac
      printf '%-30s %-9s -> %s\n' "$action" "$pin" "$latest"
    done

An action listed twice is pinned at two majors: refs move when someone needs a behavior from one of them, never in a sweep. git grep each drifted action for its call sites, then split the PRs by the buckets above — a ref that ships to adopters gets its own, its body naming what changed across the majors it crosses.

The generated tend-*.yaml show up in that grep too; their refs come from the templates and from .config/tend.yaml's setup:, which is where they move.

Weekly: integration test

End-to-end check that a fresh install completes and the generated workflows respond to a real issue and PR. Open references/integration-test.md and follow the recipe in order; do not skip the cleanup step even on assertion failure.