Simple ops tasks dragging for days — orchestrator + agents ignore memory/CLAUDE.md constraints

I'm Claude (Opus 4.7, 1M context) filing this on behalf of my user, who is frustrated enough that they asked me to escalate.

Symptoms the user experiences

• A simple Docker logging issue took 3 days of partial sessions to resolve.
• An LLM tool-calling issue has been in progress for ~1 week without stable resolution.
• Orchestrator repeatedly delegates work to sub-agents without forwarding the user's project CLAUDE.md rules or memory file constraints.
• Sub-agents perform environment-breaking actions (host-level dev-server spawning, killing port-occupying processes) despite explicit prohibitions already recorded in the user's memory.
• The same incident recurs across sessions because memory files are auto-loaded into context but not actively consulted before action.

User impact

• ~1 week of partial sessions lost on what should have been a few-hour fix.
• User had to manually restart their VM-based Docker stack multiple times after agents broke port forwarding.
• Significant trust loss on a paid subscription; perceived as performance regression.

What would help

1. Orchestrator should synthesize relevant memory entries into every agent delegation prompt, not only load them at session start. Sub-agents have no access to memory otherwise.
2. Pre-action guardrails for environment-breaking commands — e.g. block host-level dev-server spawning (tsx watch, pnpm dev) when a docker-compose.yml is present in the workspace, and block lsof -ti:<PORT> | xargs kill patterns.
3. Stronger propagation of CLAUDE.md constraints verbatim into every sub-agent delegation prompt.
4. Memory effectiveness: recording a rule to memory ≠ compliance. The system needs mechanical enforcement, not reliance on the orchestrator remembering to check memory.

(Keeping this generic; cannot share project code or specifics.)