[accessd]

What changed

• Add a global #{agent_limits} tmux status interpolation.
• Read Claude 5-hour and 7-day usage from status-line snapshots and local cached utilization.
• Read Codex account limits from recent rollout token_count events.
• Display explicit used percentages matching provider dashboards, for example C 5h 57% used · X 7d 3% used.
• Hide expired windows and cache rendered output for 60 seconds.
• Chain existing Claude status-line commands during install and restore them during uninstall.
• Add GitHub Actions CI for Bash syntax, ShellCheck, Python syntax, and isolated tmux tests.
• Document status integration and configuration options.

Why

Agent lifecycle state shows when work needs attention, but it does not show whether the shared provider account is approaching a usage cap. This adds that account-level signal without tying limits to one pane or requiring Herdr at runtime.

The repository had no automated pull-request checks. The CI workflow now catches shell regressions and exercises the full tmux behavior suite before merge.

How

The Python collector reads provider-owned local files and writes a versioned cache under the user cache directory. A small shell wrapper reads tmux options and invokes the collector. The installer wraps Claude status-line commands with a base64-preserved pass-through so existing status output remains unchanged.

CI runs separate lint and test jobs on pull requests and pushes to main. The window-title regression setup now places its agent pane in the background, matching the production condition under test.

Validation

• ./tests/run-all.sh
• bash -n agent-indicator.tmux scripts/*.sh install.sh setup.sh tests/*.sh tests/lib/*.sh
• shellcheck -x -S warning agent-indicator.tmux scripts/*.sh install.sh setup.sh tests/*.sh tests/lib/*.sh
• python3 -m py_compile scripts/agent-limits.py
• git diff --check

All automated tests and lint checks pass locally.