[BUG] /model and /effort mutate global settings.json — breaks agents/fleet view; no way to configure per-agent model or effort

Summary

/model and /effort write to ~/.claude/settings.json immediately and globally. In the agents/fleet view (claude agents) this means there is no supported way to run a fleet of agents with intentional, independent model and effort configurations.

Environment

• Claude Code: 2.1.169
• Platform: Windows 11

The problem

~/.claude/settings.json is a single shared file. Every Claude Code session reads from it on start. /model and /effort both write to it immediately on invocation.

The slash commands look and feel like session controls, not "save my settings" commands. In a single session this is a confusing default. In the agents/fleet view it becomes a correctness bug:

1. You have three agents running: one on Opus doing deep analysis, one on Sonnet doing fast search, one on Haiku doing summarisation.
2. The Sonnet agent completes a phase and a skill invokes /effort high.
3. Any agent started after that point inherits high effort on whatever model that last write left behind.
4. There is no warning. There is no way to know which agent last mutated the shared state.

Launch ergonomics compound this: when you spawn a new agent from the fleet view, it inherits whatever settings.json currently contains — not the model or effort you intended when you hit "new agent." There is no way to launch an agent with an explicit, intentional model/effort configuration other than CLI flags, which must be known before the session starts and cannot be changed once running.

Note: v2.1.169 fixed background agents ignoring ANTHROPIC_MODEL env values — confirming that even the env var workaround for model configuration was broken for background agents until yesterday.

Steps to reproduce

1. Open claude agents
2. Launch agent A on Opus — note current model/effort
3. Launch agent B — in that session run /model and select Sonnet
4. Launch agent C from the fleet view
5. Agent C starts on Sonnet, not Opus

Expected behaviour

• /model and /effort should be session-local by default and should not write to settings.json.
• A distinct, explicit action — e.g. /model save, --save flag, or a dedicated /settings set model command — should be required to persist a change.
• When launching a new agent from the fleet view, the model and effort should be specifiable at launch, independently of any global default.
• Each agent in the fleet view should maintain its own model/effort state for its lifetime.

Related issues (all closed as stale or duplicate without a fix)

• #20745 — Model setting changes globally across all sessions (regression)
• #27608 — Status bar shows stale model after settings.json change from another session (closed as duplicate)
• #27627 — /model command is global, not per-session (closed as duplicate of #27608)
• #36288 — /model affects all concurrent sessions instead of being per-session (closed as duplicate)
• #37303 — effortLevel not persisted for Max mode; model/effort have no per-session isolation
• #47820 — /effort session command does not override effortLevel in settings.json
• #49076 — /model silently persists effortLevel to settings.json, overriding user's preferred default (closed as duplicate)
• #49166 — /effort is global across concurrent sessions instead of session-scoped
• #53416 — /effort setting is global across sessions instead of per-session (closed as duplicate of #49166)
• #65598 — feat: support effort level in agent definition frontmatter (open — feature request for the same gap)
• #66346 — effortLevel in settings.json not honoured on session start (open)

This issue is not a duplicate of any of the above. Those issues cover single sessions and concurrent terminal sessions. This issue covers the agents/fleet view, where the lack of per-agent launch configuration means there is no supported way to run agents with intentional, independent model and effort settings.