Skip to content

claude.md

Latest commit

 

History

History
76 lines (66 loc) · 5.04 KB

claude.md

File metadata and controls

76 lines (66 loc) · 5.04 KB
summary Claude Code support in CodexBar: PTY probing, parsing, and UX.
read_when
Debugging Claude usage/status parsing
Adjusting Claude provider UI/menu behavior
Updating Claude CLI detection paths or prompts

Claude Code support (CodexBar)

Claude Code support is implemented: CodexBar can show Claude alongside Codex (one status item per provider) and keeps provider identity fields siloed (no Claude org/plan leaking into Codex, and vice versa).

UX

  • On launch we detect CLIs:
    • Codex: codex --version
    • Claude Code: claude --version
  • Settings → General: toggles for “Show Codex usage” and “Show Claude Code usage” (Claude defaults on when detected).
  • Menu: each enabled provider gets its own status item/menu card.

Claude menu-bar icon (crab notch homage)

  • Same two-bar metaphor; template switches to the Claude “crab” style while keeping the same bar mapping.

Data path (Claude)

How we fetch usage (no tmux)

  • We launch a single Claude CLI session inside a pseudo-TTY and keep it alive between refreshes to avoid warm-up churn.
  • Driver steps:
    1. Boot loop waits for the TUI header and handles first-run prompts:
      • “Do you trust the files in this folder” → send 1 + Enter
      • “Select a workspace” → send Enter
      • Telemetry (y/n) → send n + Enter
      • Login prompts → abort with a nice error (“claude login”).
    2. Send the /usage slash command directly (type /usage, press Enter once) so we land on the Usage tab.
    3. Re-press Enter every ~1.5s (Claude sometimes drops the first one under load).
    4. If still no usage after a few seconds, re-send /usage + Enter up to 3 times.
    5. Stop as soon as the buffer contains both “Current session” and “Current week (all models)”.
    6. Keep reading ~2s more so percent lines are captured cleanly, then exit.
  • Parsing:
    • We strip ANSI codes, then look for percent lines within 4 lines of these headers:
      • Current session
      • Current week (all models)
      • Current week (Sonnet only) (optional)
    • X% used is converted to % left = 100 - X; X% left is used as-is.
    • If the CLI surfaces Failed to load usage data with a JSON blob (e.g. authentication_error + token_expired), we surface that message directly ("Claude CLI token expired. Run claude login"), rather than the generic "Missing Current session" parse failure.
    • We also extract Account: and Org: lines when present.
  • Strictness: if Session or Weekly blocks are missing, parsing fails loudly (no silent “100% left” defaults).
  • Resilience: ClaudeStatusProbe retries once with a slightly longer timeout (20s + 6s) to ride out slow redraws or ignored Enter presses.

What we display

  • Session and weekly usage bars; Sonnet-only weekly limit if present.
  • Account line uses Claude CLI data (email + org + login method). Provider identity fields stay siloed.

Notes

  • Reset parsing: Claude reset lines can be ambiguous; the parser keys off known “Current session / Current week …” section headers so “Resets …” cannot be attributed to the wrong window.
  • Debug: the Debug tab can copy the latest raw CLI scrape to help diagnose upstream CLI formatting changes.

Open items / decisions

  • Which template asset to use for the Claude icon (color vs monochrome template); default to a monochrome template PDF sized 20×18.
  • Whether to auto-enable Claude when detected the first time; proposal: keep default off, show “Detected Claude 2.0.44 (enable in Settings)”.
  • Weekly vs session reset text: display the string parsed from the CLI; do not attempt to compute it locally.

Debugging tips

  • Quick live probe: LIVE_CLAUDE_FETCH=1 swift test --filter liveClaudeFetchPTY (prints raw PTY output on failure).
  • Manually drive the runner: swift run claude-probe (if you add a temporary target) or reuse the TTYCommandRunner from a Swift REPL.
  • Check the raw text: log the buffer before ANSI stripping if parsing fails—look for stuck autocomplete lists instead of the Usage pane.
  • Things that commonly break:
    • Claude CLI not logged in (claude login needed).
    • CLI auth token expired: the Usage pane shows Error: Failed to load usage data: {"error_code":"token_expired", …}; rerun claude login to refresh tokens. CodexBar now surfaces this message directly.
    • Enter ignored because the CLI is “Thinking” or busy; rerun with longer timeout or more Enter retries.
    • Running inside tmux/screen: our PTY driver is standalone, so disable tmux for this path.
    • Settings > General now shows the last Claude fetch error inline under the toggle to make it clear why usage is stale.
  • Codex parity: when credits are missing because the Codex CLI shows an update prompt, our PTY driver now auto-sends Down+Enter, re-runs /status, and retries once with a longer timeout; if it still fails, run LIVE_CODEX_STATUS=1 swift test --filter liveCodexStatus to dump the raw screen.
  • To rebuild and reload the menubar app after code changes: ./scripts/compile_and_run.sh. Ensure the packaged app is restarted so the new PTY driver is in use.