Stop Claude runs inheriting the host's user settings via --setting-sources

[yasyf]

Summary

ClaudeConfig had no way to reach the claude CLI's --setting-sources, so every spawn with use_host_config: true inherited the host's user-level settings, hooks included, and nothing could turn them off. On a machine with a real hook stack the hooks dominate the call.

Measured on macOS with an identical one-word prompt to claude-haiku-4-5-20251001, MCP already disabled via --strict-mcp-config --mcp-config '{"mcpServers":{}}':

invocation	wall
as spawnllm built it before this change	63.0s, exiting with SessionEnd hook … failed: Hook cancelled
--bare	2.4s, but Not logged in: it skips keychain reads, so subscription auth cannot use it
--setting-sources project	6.4 / 7.5 / 5.7s, correct answer, auth intact
--settings '{"hooks":{}}'	44.7s: settings merge, so this does not remove hooks

The downstream victim was slop-cop, whose per-chunk lint calls each paid that cost and blew their timeouts.

The change

setting_sources: Option<Vec<String>> joins ClaudeConfig, resolved at one site in rust/spawnllm-core/src/plan/claude.rs:

• an explicit list joins with commas and wins over everything, isolation included;
• unset with isolated keeps emitting --setting-sources "", so every isolated invocation is argv-identical to before;
• unset with host config now emits --setting-sources project. This is the new default and the point of the change.

It is a list passthrough rather than a boolean on purpose. CodexConfig's enable_hooks/enable_mcp are booleans because codex's knob is one (-c features.hooks=false); --setting-sources is a list that governs more than hooks, and the nil-vs-empty Option<Vec<String>> shape is the one tools and disallowed_tools already use.

Every layer carries it, by hand (the bindings have no generator): the core wire type and plan, the Rust crate's mirrored spec, Go (SettingSources []string on ClaudeConfig, the wire struct, and coreClaudeOf), and Python (setting_sources on ClaudeConfig plus a new TSettingSource literal alias). The Python SDK backend previously hardcoded ["user", "project", "local"] for non-isolated runs; it now follows the same resolution as the CLI plan.

Conformance: 16 existing claude vectors change, but only two change argv, claude-isolated-false and claude-strict-mcp, the two non-isolated cases, which both gain --setting-sources project. The other 14 are isolated and only gain "setting_sources": null in their input. Three new vectors cover the explicit restore (user,project,local), an empty list resolving to "", and an explicit list beating isolation.

Breaking change

Anyone relying on use_host_config: true to pick up user-level settings loses them. MCP servers declared in ~/.claude/settings.json, user-level permissions, and hooks all stop loading for non-isolated runs.

To restore the previous behavior, name the sources explicitly:

• Python: ClaudeConfig(setting_sources=("user", "project", "local"))
• Go: ClaudeConfig{SettingSources: []string{"user", "project", "local"}}
• Rust: setting_sources: Some(vec!["user".into(), "project".into(), "local".into()])

--settings is not a workaround. Settings merge rather than replace, so --settings '{"hooks":{}}' leaves the inherited hooks in place; that invocation still measured 44.7s above.

The CHANGELOG entry under [Unreleased] / Changed describes the same default change and restore path.

Test plan

Green:

• Rust core: 32 unit tests, conformance replay (3), freshness (1); cargo fmt --check; clippy -D warnings; cargo test -p spawnllm --no-default-features
• Go: go vet, go build, go test -race ./...
• Python: pytest 415 passed with 2 platform skips; ruff check and ruff format; ty at its 21-diagnostic baseline
• Docs build 19/19

Not green, and shown pre-existing by running the same commands against an unmodified HEAD in a separate worktree:

• Two rust/spawnllm/tests/integration.rs cases, call_returns_text_via_auto_selected_backend and run_on_times_out_and_kills_the_child, fail at baseline too. They want real authenticated CLIs and hold 10s wall-clock deadlines.
• TestSelectBackendSpecialtyPromotes flaked once under -race on a loaded machine and passes on an idle one.
• golangci-lint is not installed on this machine, so task go:lint did not run.

https://claude.ai/code/session_014gKfZvjihGgFrEVCr3NJJb