feat(team): team enhancements, subagentModel, parallel tool exec

[TheArchitectit]

Summary

Ports the agent-team enhancement layer onto main so the model can spin up
coordinated sub-agent teams for parallel work. 2,095 lines added across 11 files
— zero deletions of existing upstream functionality.

Single clean commit built from upstream/main. No TUI, LSP, or setup-wizard
contamination from other in-flight work.

---

Motivation

The TeamCreate, TeamDelete, and Agent tools already existed, but they
didn't work end-to-end because:

1. subagentModel was silently dropped — /setup saved the sub-agent model
   into settings.json but RuntimeConfig never stored or exposed it, so
   agents always fell back to a hard-coded default.
2. Sub-agents couldn't reach the user's provider — the main CLI calls
   inject_config_as_env_fallbacks() to set CLAWCUSTOMOPENAI_API_KEY /
   CLAWCUSTOMOPENAI_BASE_URL from /setup-saved config, but the sub-agent
   spawn path in tools never did, so custom providers were invisible.
3. providerFallbacks.primary silently overrode the agent's model — the
   configured fallback primary replaced whatever model resolve_agent_model
   picked, routing agents to a dead model instead of the session's provider.
4. 404 model-not-found killed the chain — a dead primary was not retryable,
   so the fallback chain never advanced to working models.

This PR fixes all four and adds the team coordination tools the model needs
to orchestrate multi-agent work.

---

What's included

1. subagentModel config wiring (bug fix)

The subagentModel field was validated by config_validate.rs but never
stored or read — the setting from /setup was silently dropped.

Change	File
RuntimeFeatureConfig.subagent_model field	runtime/src/config.rs
RuntimeConfig::subagent_model() accessor	runtime/src/config.rs
parse_optional_subagent_model() (reads subagentModel / subagent_model, blank → None)	runtime/src/config.rs
runtime::inject_config_as_env_fallbacks() — moved from private main.rs fn to pub fn in runtime so sub-agents can call it	runtime/src/config.rs + runtime/src/lib.rs
resolve_agent_model fallback chain: explicit model → subagentModel → session provider model → DEFAULT_AGENT_MODEL	tools/src/lib.rs

2. Parallel tool execution (perf)

When the model emits multiple tool_use blocks in one response, read-only
tools now run concurrently via std::thread::scope instead of
sequentially.

Parallel-safe tools: read_file, glob_search, grep_search,
WebFetch, WebSearch, ToolSearch, Skill, LSP, GitStatus,
GitDiff, GitLog, GitShow, Agent, AgentMessage, AgentSuggestion,
TeamStatus, TeamInfo, TaskGet, TaskList, TaskOutput

Sequential tools: everything else (writes, side-effects)

Change	File
ToolCall, ToolResult, TurnProgressReporter re-exported from runtime	runtime/src/conversation.rs + runtime/src/lib.rs
ToolExecutor::execute_batch default impl on trait	runtime/src/conversation.rs
CliToolExecutor::execute_batch override with parallel-safe classification	rusty-claude-cli/src/main.rs

3. Team coordination layer (feature)

New tools and capabilities for multi-agent work:

Tool	Purpose
TeamCreate	Spawn N agents (1–6× presets: tiny/small/medium/large/xlarge/mega) with a shared prompt
TeamDelete	Mark a team deleted, clean inbox, release task claims
TeamStatus	Live snapshot, final summary, or event inbox for a team
AgentMessage	Send/receive messages between agents (shared mailbox)
TaskClaim	Claim/list/complete tasks to prevent duplicate work across agents
AgentSuggestion	Agents suggest AGENTS.md patterns/pitfalls

Supporting infrastructure:

• Shared mailbox directory at ~/.clawd-agents/mailbox/team/{team_id}/
• Background team watcher thread — polls agent .json files every 2s,
  prints [team] progress to stderr, writes events to -events.jsonl
• Mode presets — tiny/1x through mega/6x controlling agent count
  and role distribution (Explore, Plan, Verification, Reviewer)
• /team slash command — on|off|status to toggle CLAWD_AGENT_TEAMS
  (required by TeamCreate)
• Enriched tool descriptions on TeamCreate and Agent steering the
  model toward correct usage

4. Model-resolution robustness fixes (bug fix)

Fix	Detail
qualify_for_provider()	Bare model names (e.g. claude-haiku-4-5-20251001) get custom/ prefix when the active provider is custom-openai, so sub-agents route through the same endpoint as the main session
providerFallbacks.primary precedence	The caller's resolved model is now the primary; providerFallbacks.primary + fallbacks are appended as recovery entries, deduped. Previously the config primary silently replaced the caller's model
404 model-not-found chain fallthrough	Added fallback_chain_eligible() that treats 404/400 "not found" as chain-eligible, so a dead configured primary advances to kimi/qwen/etc. Added ApiError::status_code() and response_body() accessors for the detection
Sub-agents inject /setup credentials	run_agent_job() calls runtime::inject_config_as_env_fallbacks() before building the agent runtime, so custom providers set up via /setup work for spawned agents

5. Also fixed

Fix	File
lane_completion.rs test helper set stale team_id/task_id fields	tools/src/lane_completion.rs
/team slash command had no validate_slash_command_input match arm (parsed as Unknown)	commands/src/lib.rs
ApiError gains status_code() and response_body() public accessors	api/src/error.rs
model_token_limit catch-all: _ => Some(...) → _ => None — unknown models skip preflight, let API enforce limits (caught by @Offwhite-Del)	api/src/providers/mod.rs
Lsp variant added to match arms	commands/src/lib.rs + rusty-claude-cli/src/main.rs
DEFAULT_CUSTOM_OPENAI_BASE_URL constant added	api/src/providers/openai_compat.rs
Provider chain precedence test updated to match new behavior	tools/src/lib.rs

---

Verification

Build

cargo check --workspace              ✅
cargo build -p rusty-claude-cli --release  ✅

Tests

tools/src/lib.rs           112 passed  (incl. 3 fallback_chain_tests, 4 provider_runtime_client chain tests)
commands/src/lib.rs         42 passed  (incl. /team parse regression)
runtime team_cron            17 passed
rusty-claude-cli bin        207 passed

Live testing (IdleDefense repo, Godot 4.6 game)

• /team on → TeamCreate with 4 agents → all agents spawned, claimed tasks, read files
• Agents used custom/openclaw provider (via qualify_for_provider)
• TeamStatus showed live progress, inbox messages, task claims
• 2/4 agents completed full reports (Explore + Verification), 2/4 hit iteration limits on complex analysis prompts (expected — increase max_iterations or simplify prompts for large repos)

---

Diff stats

 rust/crates/api/src/error.rs             |   43 +
 rust/crates/api/src/lib.rs               |    3 +-
 rust/crates/api/src/providers/mod.rs     |   39 +-
 rust/crates/api/src/providers/openai_compat.rs | 1 +
 rust/crates/commands/src/lib.rs          |   28 +-
 rust/crates/runtime/src/config.rs        |   70 +
 rust/crates/runtime/src/conversation.rs  |  203 +-
 rust/crates/runtime/src/lib.rs           |    5 +-
 rust/crates/rusty-claude-cli/src/main.rs |    3 +
 rust/crates/tools/src/lane_completion.rs |    2 +
 rust/crates/tools/src/lib.rs             | 1699 ++++++++++++++++++++++++
 11 files changed, 2095 insertions(+), 113 deletions(-)

Key files

File	Lines	What it does
tools/src/lib.rs	+1,699	All team tool logic: TeamCreate, TeamDelete, TeamStatus, AgentMessage, TaskClaim, AgentSuggestion, mode presets, mailbox, watcher, qualify_for_provider, fallback_chain_eligible
runtime/src/conversation.rs	+203	ToolCall/ToolResult types, TurnProgressReporter trait, execute_batch default impl
runtime/src/config.rs	+70	subagent_model field/accessor/parser, inject_config_as_env_fallbacks
api/src/error.rs	+43	status_code()/response_body() accessors for chain-fallback detection
api/src/providers/mod.rs	+39	model_token_limit generic fallback, custom-openai provider entry

---

What this does NOT include

This PR is purely additive on top of upstream/main. It does not include:

• TUI / split-pane dashboard changes (separate PR)
• LSP module restructuring (already merged)
• Setup wizard / provider config changes (already merged)
• Custom-openai namespace routing (separate PR)
• Hooks refactoring (separate PR)

🤖 Generated with Claude Code

[Offwhite-Del]

发现并修复: preflight_skips_unknown_models 测试失败

根因: model_token_limit() catch-all arm (_ =>) 为未知模型分配了默认 context window (131072)，导致 preflight_message_request 对未知模型返回 ContextWindowExceeded。

修复: catch-all 应返回 None，preflight_message_request 已正确处理 None → 跳过预检。

修改: crates/api/src/providers/mod.rs, change _ => Some(ModelTokenLimit {...}) to _ => None (line ~694)

验证: ✅ cargo test preflight — 4/4 passed
修复分支: Offwhite-Del/sun:fix/preflight-unknown-models

建议直接 rebase 到最新 main (main 上已经是 _ => None)，或者 Cherry-pick 这个 commit: 6e28598

[TheArchitectit]

感谢发现这个问题！你说得对，catch-all 应该返回 None 让 preflight 跳过未知模型。已经修复并推送上来了。谢谢！🙏

[1716775457damn]

Nice work on the subagentModel fix — that was a real gap where /setup saved the model but RuntimeConfig never surfaced it. The team coordination layer looks clean. One thing: cargo fmt is failing, you'll want to run cargo fmt --all and push before this can land.