feat(commands): fork-agent infra + /btw /simplify /advisor (#33 #37 #62)

[yaohaowei0914]

Summary

Delivers the fork-agent infrastructure and the three commands that depend on it:

• Implement /btw via a forked side-question agent #37 /btw — tool-free, single-turn side-question fork (doesn't touch the main transcript)
• Align /simplify with a public slash entrypoint and multi-agent workflow #62 /simplify — multi-agent review via three parallel reviewer forks (reuse / quality / efficiency)
• Implement /advisor command and advisor-model request plumbing #33 /advisor — show/set/unset + end-to-end advisor_model plumbing into the Messages request

Fork infrastructure (issue #37)

• New engine::agent::fork::run_fork primitive — spawns an ephemeral child QueryEngine with no persistence, configurable tool set, and single/multi-turn caps. Optionally seeds with parent messages for cache-safe reuse.
• SkillContext::Fork in tools/skill.rs now dispatches through run_fork instead of the prior "inline fallback", honoring each skill's allowed_tools.

/btw (issue #37)

• /btw <question> runs a tool-free, single-turn fork.
• Answer returned via CommandResult::Output, so the main transcript is not polluted.
• Side-question role pinned by a short system-prompt fragment to keep cache hits high.

/simplify (issue #62)

• Bridges the bundled simplify skill to a user-facing slash command.
• Multi-agent mode (default): fans out three parallel reviewer forks covering reuse / quality / efficiency via futures::future::join_all, then joins the outputs into a single action-oriented summary.
• --single / -1 runs a single-agent pass; optional scope argument (/simplify src/foo.rs) narrows the review.

/advisor + advisor_model plumbing (issue #33)

• advisor_model: Option<String> added to:
  • AppState (types/app_state.rs)
  • SettingsJson runtime projection (cc-config/runtime_settings.rs)
  • RawSettings / EffectiveSettings + JSON schema (camelCase key advisorModel)
  • ModelCallParams → MessagesRequest (serialized with skip_serializing_if)
• provider_supports_advisor() gates Anthropic / Azure / Bedrock / Vertex; other providers are stripped with a debug! log (engine/lifecycle/deps.rs).
• /advisor supports show / set / unset (aliases: none, off, clear), persists atomically via write_user_settings, and surfaces a clear inactive-provider note in its output.

Test plan

• cargo test -p claude-code-rs --bin claude-code-rs -- commands::btw commands::simplify commands::advisor engine::agent::fork api::client::tests::test_messages_request api::client::tests::test_provider_supports_advisor — 27/27 pass
• cargo test -p claude-code-rs --bin claude-code-rs -- commands:: --test-threads=1 — 299/299 pass (4 pre-existing env-var races surface only under parallelism; none of them are new tests in this PR — verified by running tests against base before stashing)
• cargo build -p claude-code-rs — clean, only the 2 pre-existing session_id warnings in web/handlers.rs

Review notes

• The run_fork primitive is small on purpose: it reuses the existing collect_stream_result from the AgentTool path and skips the global agent-tree + IPC event registration, so /btw stays lightweight.
• Advisor tests use a noop_persist closure so the command tests are hermetic (no env-var mutation, no touching of the user's real settings.json). Disk round-trip is covered separately by persist_to_path_round_trips_via_disk.
• Provider matrix for advisor support is intentionally conservative (Anthropic wire format only). OpenAI-compatible + Google are stripped; when users set advisor on such a provider, /advisor still persists the value but the status line flags it as inactive.

🤖 Generated with Claude Code