Add Codex Desktop app monitoring support

[crosswyb]

Summary

Adds a new CodexDesktopCollector that monitors Codex Desktop app sessions by reading token usage data from ~/.codex/logs_2.sqlite. This is complementary to the existing CodexCollector which only covers Codex CLI.

How it works

• Detects the Codex Desktop process (Codex.app/Contents/MacOS/Codex) via ps
• Reads ~/.codex/logs_2.sqlite using the sqlite3 CLI (no new Rust dependencies)
• Parses structured log entries containing input_token_count, output_token_count, cached_token_count per conversation
• Groups by conversation.id to form per-session views
• Registered as "codex-desktop" agent in MultiCollector, respects hidden_agents config
• Sessions auto-mark as Done when the desktop app process exits

Design decisions

• No new dependencies — uses std::process::Command to call sqlite3 CLI with -json output, consistent with existing patterns (lsof, ps, git)
• Separate collector — not merged into existing CodexCollector because the data source (SQLite vs JSONL) and discovery strategy (process name vs codex binary) are fundamentally different
• High-water mark — tracks last_ts to only re-read new log entries each tick

Limitations

• No cwd/project_name/chat_messages/tool_calls (SQLite logs don't expose these at the same granularity as JSONL)
• Context window percentage is estimated (no per-turn breakdown in aggregated SQLite data)
• Requires sqlite3 CLI to be installed (pre-installed on macOS, available everywhere)

Testing

• 3 new unit tests for field extraction logic
• All 139 existing tests pass (updated with_hidden test assertions for the new 4th collector)

Closes: N/A — feature request from user who needed desktop app monitoring alongside CLI support

[graykode]

Thanks for adding Codex Desktop support. The feature direction makes sense for abtop, but I don't think this implementation is safe to merge yet.

Main blockers:

• src/collector/codex_desktop.rs:97 treats logs.ts as milliseconds, but the Codex Desktop DB stores epoch seconds plus ts_nanos. With real DB rows this can make sessions fail the active-window check, and the same unit mismatch affects started_at values later in the file.
• src/collector/codex_desktop.rs:265-345 rereads the latest 500 rows on every collect pass and adds token counts again each time. last_ts is updated but not used as a cursor. This will inflate token totals on every refresh. Rows are also processed newest-to-oldest, so an older event can overwrite last_event_kind after a newer event.
• src/collector/codex_desktop.rs:310-318 skips tool-related events before they can drive SessionStatus::Executing, because conversation.id is only extracted for completed/failed/created/in_progress events.
• src/collector/codex_desktop.rs:140-149 calculates context_percent by subtracting other conversations' total input from the current conversation. A conversation's context should not depend on unrelated sessions; this should use the latest per-conversation context/input value instead.

Verification:

• cargo test passes locally: 141 tests.
• cargo clippy -- -D warnings fails on this branch with derivable_impls, unnecessary_cast, and useless_format in src/collector/codex_desktop.rs.

Suggested direction: use a stable processed-row cursor such as (ts, ts_nanos, id), process rows in ascending order, avoid re-adding already-processed token events, and store latest per-conversation context/input separately from lifetime token totals.