Codex supports thread automations with /goal, /btw, and heartbeat wake-ups

• jxnlco's thread-automation explainer describes thread automations as recurring wake-up calls attached to the current thread, which keeps scheduled work inside the same context instead of restarting from a fresh prompt.
• Users are treating jerryjliu0's /goal post as a higher-level control surface, while reach_vb's workflow note says long Codex runs survive compaction and can recover relevant context from the session when needed.
• Mid-run steering is now part of the pattern: dejavucoder's /btw example shows instructions queued into an active /goal loop, and dejavucoder's CLI note frames /side or /btw as an ephemeral fork for side conversations.
• The more interesting workflow hack is adversarial review on repeat, where PerceptualPeak's loop prompt has Codex spawn three high-reasoning sub-agents, fix issues, and rerun until all three return clean.
• OpenAI is shipping the surrounding product surface quickly: OpenAIDevs on mobile GA added goals, side chat, notifications, file previews, and inline review comments, while gdb's internal-adoption post says Codex usage has spread far beyond engineering inside OpenAI.

According to OpenAIDevs on mobile GA, the product surface now includes goals, side chat, notifications, and inline review comments. jxnlco's definition of thread automations gives the clearest public explanation of heartbeat wake-ups, dejavucoder's example shows how people interrupt and steer those loops in practice, and reach_vb's adoption screenshot suggests the bigger story is delegation becoming the default interface for work.

The cleanest public definition came from jxnlco on thread automations, who described them as heartbeat-style recurring wake-up calls tied to the current thread. That makes them distinct from a scheduled job that starts from scratch, because the same conversation state stays live.

The thread around jxnlco's /loop question suggests the UX is still in flux. In a follow-up, jxnlco's sleep joke reduced the distinction to one line: loop is scheduled, goal would have to learn to sleep.

Two different users describe the same shift from opposite ends. jerryjliu0's post frames /goal as a move from prompt engineering toward goal and eval engineering, while reach_vb's workflow note says single-project threads, thread spin-offs, and compaction made them stop thinking much about context at all.

That workflow shows up in shorter operator recipes too. mattlam_ summarizing loop usage boiled a day-to-day loop down to four parts:

• define the goal
• give the agent tools to self-test
• keep looping until improvement or a block
• use /notify when the run needs a human

The steering interface is smaller than the orchestration interface. dejavucoder's example says that while a /goal or other loop is running, you can either queue an instruction into Codex or use /btw to ask what it is doing and what it should try next.

A second post from dejavucoder's CLI note describes /side or /btw as an ephemeral fork. The missing piece, according to that post, is a way to keep the branch and jump back to older forks, which dejavucoder's follow-up compares to Claude Code's rewind behavior.

The most concrete loop pattern in the evidence is not scheduling, it is review. PerceptualPeak's prompt has Codex write an implementation spec, define a rubric, spawn three GPT 5.5 adversarial sub-agents with x-high reasoning, apply their fixes, and repeat until a full batch returns zero issues.

That recipe matters because it turns evaluation into the stopping condition instead of the prompt length. tokengobbler on /goal describes the feeling more bluntly: a small loop prompts a bigger loop, and average session length jumps by an order of magnitude.

OpenAI is shipping the interface around these workflows as fast as users are inventing them. According to OpenAIDevs on mobile GA, Codex in the ChatGPT mobile app is now generally available with one-to-one device pairing, notifications, goals, side chat, file previews, and inline review comments.

The internal adoption screenshots are the stronger tell. gdb's post shows Codex's share of work climbing across departments including finance, recruiting, and legal, while reach_vb's screenshot claims 28.6% of users at OpenAI run five or more agents concurrently and that skill usage has risen from 5.4% to 26.6%, reaching 96.2% internally.

That makes the /goal, /btw, and heartbeat discussion look less like a bag of power-user tricks and more like the control grammar for long-running delegated work.