feat: response-viewer (eye) support for Codex sessions

[pirronewantlux529-coder]

Problem

The response-viewer currently only reads Claude transcripts from ~/.claude/projects. For Codex panes the eye falls back to a raw terminal-buffer dump (TUI repaint garbage), and there is no way to see the last response or the conversation thread cleanly.

A naive fix (match rollouts in ~/.codex/sessions/** by cwd + newest mtime) breaks as soon as two Codex panes share a working directory — each pane's eye shows whatever pane wrote last. We ran that variant privately for two weeks; this PR is the design that survived real-world usage.

Design: locating THIS pane's rollout

In order of confidence:

1. history match — Session now tracks the pane's last Enter (codexLastSubmitAt, both write() and writeViaMux() paths). Correlating it against ~/.codex/history.jsonl {session_id, ts} entries identifies the thread the pane is actually on — the only signal that survives /resume, /new and /fork typed inside the codex TUI itself. An entry is credited to the pane whose Enter is closest among all codex panes, so a menu keystroke elsewhere can't steal attribution. Resolved pins are cached per (session, submitAt).
2. originator match — codex panes are spawned with CODEX_INTERNAL_ORIGINATOR_OVERRIDE=codeman_<sessionId>; codex (verified on 0.144.1) writes the value into session_meta.originator of every rollout it creates, including new files after /new. Newest match wins.
3. resume-id match — resumed rollouts keep their original session_meta (codex appends without rewriting it), so originator matching can't see them; but the rollout uuid is in the filename and codexConfig.resumeSessionId is known (validated as UUID first).
4. cwd+mtime fallback — for pre-existing panes. Hardened: case-blind path compare (codex records the launch-time case; /mnt paths vary on WSL), and rollouts claimed by other panes' originators are excluded, so an idle pane can't display a sibling's thread.

Reader quality

• User turns come from event_msg/user_message — codex emits one per real user input, and injected context (AGENTS.md, environment_context, …) never appears there, so no fragile filtering heuristics. Legacy response_item user rows are deduped per-text against the event set, so mixed-version rollouts (old turns without event_msg) keep full history.
• Image inputs render an [image ×N] placeholder instead of silently vanishing.
• session_meta identity is parsed once per rollout and cached (it is write-once — verified that resume appends without touching it), so repeat eye clicks cost stats + head reads of new files only.
• ?context=full returns the whole user/assistant thread, same shape as the Claude path.

Frontend

• Conversation role label follows the session mode (Codex / Gemini / OpenCode / Claude) instead of hardcoding "Claude".
• The terminal-buffer fallback is now Claude-only — _cleanTerminalBuffer knows Claude CLI output, but for TUI modes it produced repaint garbage; those now show a clear placeholder.

Testing

Verified on codex-cli 0.144.1, tmux backend, two machines (WSL2):

• Two codex panes created in the same directory within the same second: each eye shows its own response.
• codex resume <id> panes resolve via filename uuid.
• /new inside a resumed pane: originator match takes over (newest wins).
• /resume typed inside the TUI, then a new message: history match follows the pane to the resumed thread.
• Idle pane in a shared directory: empty placeholder, no bleed-over.
• Claude sessions: unchanged behavior (regression-checked).
• tsc, prettier, npm run build all clean.

Note: CODEX_INTERNAL_ORIGINATOR_OVERRIDE is an internal codex interface; if a future codex stops honoring it, layers 1/3/4 still function — the design degrades gracefully.

🤖 Generated with Claude Code

[Ark0N]

Merged — thank you @pirronewantlux529-coder, and welcome as a first-time contributor! 🎉 The 4-layer Codex rollout resolution (history pin → originator → resume-UUID → cwd fallback) is genuinely thoughtful work, and the reviewers found the parser logic sound. Review added the missing test coverage (10 route tests over fixture rollouts covering pin/originator/uuid precedence, dedup, injected-context filtering, and a Claude-mode regression guard) plus small hardening (LRU caches, multi-block text joins). Hope to see more from you! 🙏