AI coding agents run for hours without pausing to reflect. Humans take breaks. This post describes wiring the two together: when Stretchly fires a break, the agent stops, reviews what it did, and answers five standup questions before continuing.

Built as an OMP extension. No external dependencies. See also: Adding timestamps to AI coding agent conversations.

The problem

Long coding sessions with AI agents have no natural checkpoints. The agent keeps going — sometimes in circles — without stepping back. You come back from a break and have no idea what happened, what worked, or what was wasted effort.

The design

stretchly-sync already pauses tool execution during Stretchly breaks. The addition: snapshot the activity buffer when a break fires, then inject a configurable prompt as a follow-up message after the break ends.

tool call → tool call → tool call → [BREAK] → snapshot → wait → resume → retro prompt → continue

The retro prompt is configurable via stretchly-sync.json:

{
  "onBreak": "~/.claude/commands/retro.md"
}

Point it at any file — a retro template, a checklist, a standup format. The extension reads it, strips YAML frontmatter, prepends an activity summary, and sends it as a follow-up user message via pi.sendUserMessage().

The retro prompt

~/.claude/commands/retro.md:

---
description: Review work done since last break
---
You are running a break-time retrospection. Answer these questions concisely:

1. **Done**: What was accomplished in this segment?
2. **Efficiency**: What could have been done faster or more directly?
3. **Stuck points**: Any repeated tools, failures, or backtracking?
4. **Web search**: Would a search unblock the current task? If yes, what?
5. **What we can do better next**

and do it

Also works as /retro from the command palette for manual invocation.

Activity tracking

Every tool_call event logs { timestamp, toolName } into an in-memory buffer. On break, the buffer is snapshotted to ~/.omp/agent/retro-pending.json and reset. The retro prompt gets a one-line summary:

Segment: 8min, 19 tool calls (11 bash, 3 write, 2 web_search, 2 read, 1 grep)

Break detection

Breaks are detected at multiple points to avoid missing them:

tool_call — before each tool execution
context — before each LLM call
turn_start / turn_end — at turn boundaries
Idle poll (setInterval) — catches breaks when no events fire

A synchronous checking flag prevents the same break from triggering multiple retros.

What we learned building it

OMP caches compiled extensions. /reload doesn’t bust the module cache. Full process restart required after code changes.
pi.sendMessage() silently fails from event handlers. Known issue (#2860). pi.sendUserMessage() with deliverAs: "followUp" works. input event transform is the only reliable injection point.
Always compile-check before committing. A missing } silently prevented the extension from loading for 5 reload cycles. esbuild --bundle catches this instantly.
Race conditions with multiple hooks. Four event hooks + an idle poll all detecting the same break window = triple-firing. Fixed with a synchronous guard flag.

Setup

# Clone
git clone https://github.com/ankitg12/stretchly-sync ~/tools/stretchly-sync

# Config (~/.omp/agent/config.yml)
extensions:
  - ~/tools/stretchly-sync

# Config (~/.omp/agent/stretchly-sync.json)
{
  "debug": true,
  "onBreak": "~/.claude/commands/retro.md"
}

Requires Stretchly running and the fixed-time breaks scheduled task for clock-aligned breaks.

Source

stretchly-sync — break detection, activity tracking, configurable onBreak prompt

This post was written during a session where the retro extension was being built and tested — the retro fired multiple times during writing, catching its own bugs.