GSD pause work very annoying

[Munkhbileg0711]

GSD Version

1.22

Runtime

Claude Code

What happened?

it is keep stopping itself in middle and saving its work, instead lets it keep doing and auto get its continue here, i dont want that manual process, it can just save its progress and compact itself and keep going never stops, what it is doing right now is very annoying, i like the idea but very annoying.

What did you expect?

use /compact automaticallly instead stops and keep saying use clear and continue, also it is keep saving its session.

Steps to reproduce

just exhaust the context

Relevant logs or error messages

[skellorok]

This is related to #884 — the context monitor hook (gsd-context-monitor.js) uses imperative language ("STOP new work immediately. Save state NOW") that causes the agent to autonomously save state and halt, which is the behavior you're experiencing.

However, I want to explain why GSD uses /clear + resume instead of /compact — it's an intentional architectural decision, not a limitation:

Context rot is the core problem. As a context window fills up, the agent becomes progressively less reliable — it starts missing instructions, contradicting earlier decisions, and making mistakes. This happens even before you hit the context limit.

/compact does not solve context rot — it compounds it. Compaction summarizes the current conversation and continues in the same session. But:

• The summary loses details (tool outputs, exact error messages, nuanced decisions)
• The compacted context still carries accumulated noise, outdated reasoning, and stale assumptions
• Each subsequent compaction summarizes a summary, degrading signal further
• The agent keeps operating on an increasingly lossy representation of what actually happened

/clear + resume is a clean slate. When GSD pauses work and you start a fresh session:

• The agent gets a full, fresh context window with zero accumulated noise
• State is captured explicitly in structured files (STATE.md, PLAN.md) — not as a lossy summary
• The fresh agent reads the state files and picks up exactly where work left off
• You get the agent's full capability on every phase, not a degraded version

Think of it like this: compaction is like photocopying a photocopy — each generation loses quality. /clear + resume is like going back to the original document each time.

That said, the real UX problem is that the hook shouldn't be forcing autonomous pauses and state saves. Even inside GSD, the agent should warn you so you can choose when to pause — not halt on its own mid-work. #884 proposes making the messages advisory so you stay in control of when /gsd:pause-work → /clear → /gsd:resume-work happens.