fix(cli): disable Claude background sessions in transparent runs

[afourniernv]

Overview

Transparent Claude runs own an ephemeral gateway, plugin, hook configuration,
and settings overlay. Claude Code Agent View can hand a background session to
its persistent supervisor and let the foreground dispatcher exit. Relay then
correctly cleans up its temporary resources while the detached Claude worker
still references them.

This change fails closed instead of silently leaving that worker outside the
transparent run's lifecycle. It uses Claude's documented Agent View controls
at both startup and settings-load time.

• I confirm this contribution is my own work, or I have the right to submit it under this project's license.
• I searched existing issues and open pull requests, and this does not duplicate existing work.

Details

• Set CLAUDE_CODE_DISABLE_AGENT_VIEW=1 in the transparent child environment
  so direct --bg/--background dispatch is rejected before settings load.
• Force both disableAgentView=true and
  env.CLAUDE_CODE_DISABLE_AGENT_VIEW=1 in Relay's existing private settings
  overlay so interactive/background entry points stay disabled after settings
  are applied and child processes inherit the gate.
• Require Claude Code 2.1.169 or newer only for transparent runs. Official
  2.1.121-2.1.168 clients can select an earlier repeated --settings source
  and ignore Relay's final overlay. Persistent and managed integrations keep
  the existing 2.1.121 minimum.
• Centralize that mode-specific floor in a crate-private optional agent
  descriptor field; the shared process layer remains host-agnostic.
• Add focused version-boundary, hostile-settings, uniqueness, materialized
  overlay, compatibility-policy, and architecture-boundary coverage; update
  the Claude guide and support matrix.

Compatibility: this intentionally rejects transparent Claude 2.1.121-2.1.168
and disables Agent View/background session management within accepted
transparent runs. Users needing those sessions should use the persistent
plugin or an administrator-managed Relay deployment. There is no public Rust
API, CLI syntax, config schema, wire/protocol, middleware, cache, daemon,
plugin ABI, worker protocol, async, or cross-crate change. Configured command
wrappers must preserve Relay's injected environment and settings arguments.

Validation:

• uv run pre-commit run passed all staged hooks.
• cargo fmt --all -- --check passed.
• cargo clippy -p nemo-relay-cli --all-targets -- -D warnings passed.
• Claude-filtered tests passed: 80 library tests and 1 CLI test.
• Focused version-policy and architecture-boundary tests passed.
• CARGO_INCREMENTAL=0 cargo build -p nemo-relay-cli --bin nemo-relay passed.
• just docs passed all checks.
• The final canonical workspace artifact recorded 5,057/5,058 passing. The
  sole failure was the unrelated macOS bootstrap process-termination timing
  test, which also reproduced in isolation; no changed file participates in
  that path.
• No-credit real-client smoke with official Claude 2.1.168/2.1.169 binaries:
  2.1.168 was rejected by Relay before spawn, while 2.1.169 reached Claude's
  native disabled gate; neither created job or daemon state.
• A prior current-client Claude 2.1.268 foreground control completed through
  Relay and the live Anthropic provider after the dual-phase settings change.
  OpenAI was not used because this path has no OpenAI request surface.

Where should the reviewer start?

Start with crates/cli/src/agents/claude/launch.rs, then review the
mode-specific policy in crates/cli/src/agents/mod.rs and its boundary test in
crates/cli/tests/coverage/agents/launcher_tests.rs.

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• None yet; keep this PR in draft until a GitHub or Linear issue is linked.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}