feat(agent): Better Harness, coordinated subagents, and harness tooling

[iptoux]

Summary

This PR delivers the Better Harness agent stack and coordinated subagents on top of the existing BLXCode workbench agent, plus supporting UI, i18n, documentation, and CI.

Better Harness (core skills + slim system prompt)

• Embeds 11 core harness skills (file-access, memory, plans, tasks, rules-skills, harness, environment, shell, git, web, subagents) via CORE_SKILLS in skills_rules/store.rs.
• Adds SkillSourceKind::Core and Skills panel Core / User tabs with core badge and remove guard.
• Reduces system_prompt.rs to a compact checklist + tool name index; full tool docs via skills_read on core skills.
• Web core skill exposes availability: disabled_no_key when no Tavily/Brave key is configured.

Coordinated subagents

• New server tool subagents.run — parallel runs (max 5 agents, default concurrency 3) sharing coordinator provider/model/key/thinking.
• Roles: scout, review, security_analyst with role-specific default allowedToolGroups.
• Per-subagent HTTP loops in subagent_runner.rs for OpenRouter, OpenAI-compatible, and Anthropic (non-streaming).
• Forced submit_result tool with structured JSON; truncation for coordinator payload size.
• Live timeline: SubagentGroup cards, localized labels, 50 ms event debounce (subagent_debounce.rs).
• Hard caps: 8 tool rounds / ~20k output tokens per subagent; no nesting (subagents.run / shell_write stripped from subagent catalogs).

Harness server tools

• environment_detect with session cache; agent_environment_invalidate + frontend call on workspace switch.
• shell_exec with read-only allowlist, child registry, cancel kill.
• Git and workspace tools (git_*, workspace_search, workspace_diff, …).
• Web tools (web_search / web_fetch) with settings envelope + keyring (agent:web:tavily / agent:web:brave) and Harness Web Tools UI.
• Unified tool_dispatch.rs and tool_groups.rs for coordinator + subagent filtered catalogs.

UI & i18n

• Harness Agent pane: web provider + API key management.
• Subagent timeline rendering with AgSubagent*, AgRole*, AgTool*, AgWeb* keys across all 13 locales.
• Rules/skills panel improvements (expandable cards, rule creation, styling) included on this branch.

Docs

• New user guides: docs/user/agent-harness.md, docs/user/subagents.md.
• New developer guides: docs/developer/agent-harness.md, docs/developer/subagents.md.
• Updated doc index, architecture, IPC, troubleshooting, and README links.

Other (already on branch)

• Leptos 0.8 upgrade and related dependency bumps.
• PR cargo check workflow (.github/workflows/pr-check.yml).
• Platform setup script path fixes; plans marked done in .agents/plans/.

Architecture (high level)

flowchart TB
  UI[Leptos Agent panel]
  Bridge[tauri_bridge]
  Orch[session_orchestrator]
  Dispatch[tool_dispatch]
  Sub[subagents.run]
  Skills[CORE_SKILLS + skills_read]

  UI --> Bridge --> Orch --> Dispatch
  Dispatch --> Sub
  Orch --> Skills
  Sub --> Runner[subagent_runner]
  Runner --> API[OpenRouter / OpenAI / Anthropic]

Loading

Test plan

• cargo test -p blxcode — backend unit tests (tool groups, environment cache, skills store, …)
• cargo check -p blxcode and cargo check -p blxcode-ui --target wasm32-unknown-unknown
• cargo tauri dev — open workspace, configure provider + optional web keys
• Agent turn: verify skills_list shows 11 core skills; Skills panel Core/User tabs
• Explicit subagent prompt (e.g. “use subagents to review …”) — timeline Subagents group, debounced updates, final coordinator reply
• Switch workspace — shell/git after environment_detect works; cache invalidation on switch
• Web Tools: save Tavily/Brave key, confirm web_search appears in tool catalog when enabled
• Anthropic + OpenRouter provider paths for main agent and subagents
• Read updated docs under docs/user/ and docs/developer/

Breaking / behavior notes

• System prompt is significantly shorter; agents must call skills_read for detailed tool guidance (by design).
• Subagents only run on explicit user request (documented in system prompt).
• shell_write and subagents.run are never exposed to subagent tool catalogs.

Plans

• .agents/plans/better-harness.md — done
• .agents/plans/coordinated-subagents.md — done

Made with Cursor