feat: add /codex:attach command for live log streaming

[mreza0100]

Problem

Background Codex jobs (/codex:rescue --background) are a black box. Once fired, there's no way to observe what Codex is doing until the job completes. The only options are polling /codex:status (which shows a one-line phase summary) or waiting for /codex:result (which shows the final output only).

Meanwhile, the background worker already writes every event — thread start, turn start, command executions, assistant messages, completion — to a structured log file on disk. That data just goes unobserved.

Solution

Add /codex:attach [job-id] — a new command that tails the job's log file in real time and exits cleanly when the job reaches a terminal status.

/codex:attach                    # attaches to the most recent active job
/codex:attach task-abc123        # attaches to a specific job by ID
/codex:attach task-abc123 --poll-interval-ms 250   # faster polling

Example output:

[attach] Job task-moozar3f-3ly2x1 · status: running
[attach] Task: Security audit of freudche-be/ and freudche-cortex/
[attach] Log: ~/.claude/plugins/data/.../jobs/task-moozar3f-3ly2x1.log
---
[2026-05-03T00:12:01Z] Starting Codex Task.
[2026-05-03T00:12:01Z] Queued for background execution.
[2026-05-03T00:12:02Z] Starting Codex task thread.
[2026-05-03T00:12:04Z] Turn started (019dead8-...).
[2026-05-03T00:12:04Z] Running command: rg -n "api_key|secret" freudche-be/
...

--- Job task-moozar3f-3ly2x1 completed ---

Implementation

plugins/codex/scripts/codex-companion.mjs — new handleAttach() function + attach case in the dispatch switch:

• Resolves the job by ID or falls back to the most recent active job via listJobs() + sortJobsNewestFirst()
• Reads log file path from job.logFile (falls back to resolveJobLogFile())
• Polls for new log content every 500ms (configurable via --poll-interval-ms)
• Simultaneously polls job state for terminal status (completed, failed, cancelled)
• Final log flush + closing line on exit — no content is missed

plugins/codex/commands/attach.md — command definition following the same pattern as status.md and cancel.md.

What this does NOT change

• No existing commands or behaviours are modified
• No new dependencies
• Background mode, foreground mode, review commands — all untouched
• Works entirely with the existing state.mjs / job-control.mjs infrastructure

Context: why not MCP streaming?

We investigated whether codex mcp-server streaming notifications could solve this. The Codex MCP server does emit events as MCP notifications during execution, but Claude Code silently drops them (the MCP client does not forward notifications/progress or notifications/message to the model — tracked in anthropics/claude-code#4157, closed "not planned"). The log file approach is the reliable path and requires no protocol changes on either side.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c310f05758

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Select only active jobs when no ID is provided

When attach is run without a job ID, this fallback picks jobs[0] whenever no queued/running job exists, so a previously completed/failed job is treated as attachable instead of reporting that there is no active job. In repositories with old job history, users will unexpectedly stream stale logs and immediately exit rather than getting the intended guidance to start /codex:rescue, which contradicts the command contract in plugins/codex/commands/attach.md.

Useful? React with 👍 / 👎.