Power user feedback: Agent teams, MCP permissions, context management, and more

Context

I'm a power user running Claude Code on a production app (KW Compliance Tracker — FastAPI + React + Postgres on Railway). Over the past few weeks I've run 12+ agent teams, 135+ tasks, configured 8 custom agents, 25 custom skills, 14+ MCP servers, and 3 hook scripts. This feedback comes from real, daily usage — not hypotheticals.

---

1. Agent Teams — Lifecycle & Visibility

Pain: Running 4-6 agents blind is the #1 friction point. There's no way to see which agents are alive, idle, working, or crashed without reading inbox JSON files manually.

Wishes:

• /team status command — show agent state (alive/idle/working/dead), current task, and last message at a glance
• Persistent agent mode — a first-class way to say "keep this agent alive until I say stop" instead of relying on idle notifications + manual shutdown requests
• Agent auto-recovery — if an agent crashes mid-task, there's no retry or handoff. The team lead just never hears from it again
• Team templates — save a team config (roles, agent defs, task templates) as reusable. Recreating the same 8-agent setup every session is repetitive

2. Task Dependency Management

Pain: I build real dependency chains (blocks/blockedBy) across execution phases, but there's no visualization and the team lead must manually dispatch work when tasks unblock.

Wishes:

• Task graph visualization — even ASCII art of the DAG would help
• Automatic unblocking — when a task completes, auto-notify the assigned agent of newly-unblocked dependents
• Shared team scratchpad — agents working on the same codebase can't share findings except through messages. A team-scoped notepad would reduce duplication

3. MCP Permissions

Pain: Wildcards don't work (mcp__* fails). Every tool must be listed explicitly. With 14+ servers and dozens of tools, this is tedious and error-prone.

Wishes:

• Wildcard/prefix permissions — e.g., mcp__github__* to allow all GitHub tools
• /mcp permissions list — show what's allowed/denied so I don't have to parse settings.json
• Hot-reload MCP servers — adding a server currently requires a session restart

4. Context Window Transparency

Pain: It's unclear how much context is used, what /compact actually does, or how much context it recovers. This causes anxiety and wasted compaction calls.

Wishes:

• Clear real-time "X% used" indicator that updates as you work
• Explain what /compact does — summarize? Truncate? How much is recovered?
• Selective compaction — keep CLAUDE.md, agent defs, and key decisions; compress file reads and tool output
• Auto-compact between team phases — the team lead's context fills up fast tracking multiple agents

5. Cross-Session Continuity

Pain: I maintain a TODO list in MEMORY.md that the next session has to manually re-read. The handoff is fragile.

Wishes:

• "Resume where I left off" mode — auto-detect the TODO list from memory and offer to pick up pending items
• Session handoff notes — auto-generate a summary when ending a session that the next session consumes
• Memory versioning — when updating MEMORY.md across sessions, show what changed

6. Developer Experience

Command discovery:

• The / palette shows only ~6 options at a time. With 17+ plugins and dozens of skills, finding the right command is a hunt
• Wish: Searchable/filterable command palette, or show more results

Hooks (mostly a win!):

• Hooks work great — notify-done.py, notify-question.sh, and auto-format (ruff format on PostToolUse) are genuinely useful
• Wish: Hook debugging mode — when a hook fails silently, there's no log
• Wish: Hook templates/marketplace for common patterns (notify on completion, auto-lint)

Custom agents:

• The .claude/agents/ system is powerful. Model selection (Opus for leads, Sonnet for specialists) works well
• Wish: Agent metrics — tokens used, runtime, tasks completed per agent
• Wish: Agent dry-run — test an agent definition against a sample task before using it in a team

7. Testing Integration

• Wish: Native test result parsing — structure pytest output into actionable categories instead of raw text
• Wish: Test failure auto-triage — group "timestamp str→datetime" failures separately from "auth fixture wiring" failures and suggest fixes per group
• Wish: CI pipeline awareness — "wait for deploy to finish, then health check" without manual polling

8. Safety & Guardrails

• Wish: Project-scoped danger zones — my CLAUDE.md documents "deploying to Postgres service kills the database." It'd be great to enforce these as hard blocks, not just docs
• Wish: Secret scanning in context — flag if a secret appears in tool output before it gets stored in memory

9. Miscellaneous

• Plan-branch linking — auto-link plan files to git branches and update status as tasks complete
• Multi-project dashboard — unified view of branches, test status, and deploys across projects
• Conversation export — export a transcript or curated highlights for sharing
• Team cost tracking — understand token/cost impact of team operations to optimize

---

Top 5 Highest-Impact

| # | Wish | Why |
|---|------|-----|
| 1 | Agent team dashboard (/team status) | Running 4-6 agents blind is the #1 friction |
| 2 | MCP permission wildcards | 14 servers × dozens of tools = tedious explicit listing |
| 3 | Context transparency + smart compaction | Black box context meter causes anxiety |
| 4 | Cross-session resume | TODO-in-MEMORY.md is manual and fragile |
| 5 | Team templates | Recreating 8-agent configs every session is repetitive |

---

Happy to provide more detail, screenshots, or config examples for any of these. Love the product — this feedback comes from wanting it to be even better.