Problem

gstack skills (especially /plan-eng-review) warn about context running low even when context usage is only ~27%. On a 1M context window (Claude Opus 4.6), 27% means ~730K tokens remaining — far from any real constraint.

Root Cause

The prompt in plan-eng-review/SKILL.md:307 contains:

"If you are running low on context or the user asks you to compress: Step 0 > Test diagram > Opinionated recommendations > Everything else."

This instruction makes the model context-anxious regardless of actual available space. The model doesn't have precise token-counting ability — it relies on heuristics like conversation length. When the prompt explicitly mentions context management, the model becomes overly cautious and starts warning/compressing prematurely.

This was likely calibrated for 200K context windows. On 1M models, the thresholds are completely wrong:

Suggested Fix

Make context management context-window-aware. Options:

1. Detect model context size — Check if the model has extended context (1M) and adjust thresholds accordingly. Could use a gstack-config setting like context_window: 1m.

2. Remove the proactive instruction — Instead of telling the model to watch for low context, only react when Claude Code's own compaction system triggers. The system already handles this automatically.

3. Add a user config — gstack-config set context_warning off to let users opt out of context management hints in the prompt.

Environment

• gstack v0.11.19.0
• Claude Opus 4.6 (1M context)
• Skills affected: at least plan-eng-review, likely others with similar instructions