Roadmap

This page merges the former Chinese-only MVP, tools-parity, roadmap-stubs, and MCP post-MVP notes into one bilingual information architecture. Source of truth for code remains crates/tools/src/catalog.rs, crates/tools/src/agent_tools.rs (Agent / Task), crates/cli/src/bootstrap/mod.rs, and crates/agent/src/agents.rs.

Maintainer execution backlog (now / next / later, decisions): repository docs/roadmap.md — edit that file instead of duplicating task lists here.

MVP scope (frozen)

In MVP

• Read/write/edit files, Glob / Grep, Bash (plus built-in tools documented as P1/P2/P7/P8 in the matrix below).
• Approval / sandbox (SecurityLayer and config).
• At least one practical tool-calling path for z.ai (OpenAI-compatible) and Anthropic.
• Execution logs under ~/.anycode/tasks/<id>/output.log and summary when not TUI-only output.
• CLI: run, repl, tui, channel bridges, scheduler, etc. (HTTP daemon was removed — see ADR 003 and HTTP daemon (removed).)

Out of MVP (separate milestones; does not block an MVP release)

• Full MCP product story (SSE/HTTP, rich OAuth UI, lazy-loaded tools) beyond current stdio v1 when enabled.
• Full LSP subprocess story beyond experimental tools-lsp.
• Sub-agent: full upstream parity for isolation/orchestration (fork, durable cross-process background agents, etc.) remains a separate milestone; worktree-level isolation, run_in_background v1 (in-process registry + TaskOutput / TaskStop), and Claude Agent field parity are described under P5 below.
• Skill plugin market / OpenClaw full parity beyond SKILL.md + Skill tool (see Agent skills).
• Swarm / coordinator, plugin market, telemetry, voice, browser tools, etc.

When MVP boundaries change, update this section and the acceptance list below.

MVP acceptance (suggested)

After each scenario, inspect ~/.anycode/tasks/<task_id>/output.log:

• [task_start], [turn_start] (or equivalent).
• If tools should run: [tool_call_start] (or project convention — see CLI sessions).
• Non-TUI paths: == summary == or documented reason for no summary.

A — Read-only: Ask for *.rs glob, Grep a symbol, FileRead one file — expect matching tool calls and correct answer.

B — Small write: In a temp repo, add/edit a file with approval — expect FileWrite or Edit and disk matches instruction.

C — Bash: Harmless read-only command — Bash succeeds; sandbox still respected if enabled.

D — Web: WebFetch or WebSearch for a stable query — answer grounded in tool output.

E — z.ai first-turn tools (recommended): Set zai_tool_choice_first_turn or ANYCODE_ZAI_TOOL_CHOICE_FIRST_TURN=1 — first request should include tool_calls when tools are required.

F — Memory: With memory.backend=file, seed a .md memory and ask a prompt that should inject Relevant Memories in the system prompt.

Tools parity (P0–P8)

Registry: crates/tools/src/catalog.rs — TOOL_*, DEFAULT_TOOL_IDS, build_registry_with_services, validate_default_registry.
Bootstrap: crates/cli/src/bootstrap/mod.rs — Arc<ToolServices> + registry build; SecurityLayer::set_tool_policy for sensitive tools.
Agent subsets: crates/agent/src/agents.rs — general-purpose = full DEFAULT_TOOL_IDS; explore / plan = EXPLORE_PLAN_TOOL_IDS (still FileRead / Glob / Grep / Bash oriented).

Naming (examples)

Reference name	API tool name	Notes
FileEditTool	Edit	README may say FileEdit
SyntheticOutputTool	StructuredOutput
BriefTool	SendUserMessage (alias Brief)
MCPTool	mcp	lowercase
AgentTool	Agent; legacy Task

Stage	Tools (API)	Status (high level)
P0	Module split, ToolServices, registry build	Done
P1	Edit, NotebookEdit, TodoWrite	Done
P2	WebFetch, WebSearch	Done
P3	mcp, ListMcpResourcesTool, ReadMcpResourceTool, McpAuth	v1 with tools-mcp, ANYCODE_MCP_COMMAND / ANYCODE_MCP_SERVERS, deny rules, dynamic mcp__<server>__authenticate
P4	LSP	Partial: tools-lsp + ANYCODE_LSP_COMMAND JSON-RPC forward; stub when off
P5	Agent, Skill, SendMessage, Task (legacy)	Skill v1 shipped; Agent / legacy Task run nested AgentRuntime (agent_type / subagent_type, nesting depth capped); Claude-style model, isolation: worktree (temp git worktree), cwd resolved to an absolute path; run_in_background: true returns status: started and runs nested work in a background task — poll TaskOutput, cancel with TaskStop on nested_task_id (cooperative flag at nested turn/tool boundaries + AbortHandle fallback; in-process only); SendMessage stored in orchestration snapshot
P6	TaskCreate/Update/List/Get/Stop/Output, team/cron/trigger	v1 orchestration file ~/.anycode/tasks/orchestration.json
P7	Plan/worktree modes, ToolSearch, Sleep, StructuredOutput	Done
P8	PowerShell, Config, Brief, AskUserQuestion, REPL	Done (PowerShell Windows-only)

Features: crates/tools/Cargo.toml — tools-mcp, tools-lsp forwarded from the anycode package; tools-http reserved.

Post-MCP / stub tracking

MCP (beyond stdio v1)

• Transports: stdio ANYCODE_MCP_READ_TIMEOUT_SECS (JSON-RPC line read), optional ANYCODE_MCP_CALL_TIMEOUT_SECS (whole tools/call wall clock), clearer timeout/EOF errors, McpStdioSession::stdio_child_is_running; SSE/HTTP later.
• Protocol: initialize, tools/list, tools/call, timeouts and errors in output.log.
• Registration: single path initialize_runtime → build_registry_with_services.
• Security: allow/deny for mcp__<server>__<tool>; sensitive calls through approval.
• OAuth / McpAuth: align with dynamic tool names; clear errors without a GUI when servers require stdio OAuth.
• Resources: real List / Read MCP resource tools.
• Optional lazy load / ToolSearch integration later.

Code entrypoints: mcp_normalization.rs, mcp_tools.rs, mcp_stdio.rs, bootstrap/mcp_env.rs.

LSP

• Configurable subprocess: lsp_stdio.rs + tools-lsp feature.
• ANYCODE_LSP_COMMAND enables JSON-RPC forwarding when built.

P5 Agent / Skill

• Skill (shipped): multi-root SKILL.md discovery, ToolServices.skill_catalog, system prompt Available skills, path-safe Skill execution (timeout, output cap, optional minimal env), config skills.*, CLI anycode skills list|path|init. Optional skills.expose_on_explore_plan registers Skill for explore / plan.
• Agent / legacy Task: nested runs use the same AgentRuntime (SubAgentExecutor). Claude Code Agent tool parity (current subset): subagent_type (alias agent_type; Explore / Plan / general-purpose normalized), optional description, optional cwd (relative paths resolve against the tool-call working directory, then canonicalized to an absolute path), optional model (sonnet / opus / haiku or a raw model id, mapped for the session provider), optional isolation: "worktree" (git worktree add under the system temp dir, removed after the run). run_in_background: true returns immediately with status: started, the same nested_task_id / output_file hints as sync runs, and a process-local registry; use TaskOutput for background_status / log tail and TaskStop on the UUID: a shared flag is polled at nested turn and tool boundaries and during blocking chat / streaming (tokio::select!, short-interval poll—no hard guarantee the TCP stream stops immediately). AbortHandle::abort remains fallback; syscall-blocking tools may still need abort. Sync success JSON still includes content. Still missing vs upstream: fork-self, durable cross-process background agents, teammate-swarm SendMessage semantics — later milestones.

OpenAI official client

• cargo build -p anycode --features openai — when provider is exactly openai, Chat Completions may use OpenAIClient; gateways often still use ZaiClient.

Recently shipped (reference)

• Nested-task cooperative cancel (v2 + v2.1): TaskStop sets a flag wired to nested TaskContext / TUI coop_cancel; tokio::select! competes with chat, chat_stream, and stream recv (~20ms poll). TaskResult::Failure cancelled maps to background_status: cancelled. Optional later: tokio_util::sync::CancellationToken; syscall-blocked tools stay best-effort (AbortHandle).

Suggested next focus (maintainers)

Pick one primary thread for the next milestone-sized effort (avoid two large refactors in parallel). Open scoped GitHub issues per thread.

Thread	Goal (issue-sized starters)
P5 Agent / Task	Optional: tighter alignment between orchestration records and task ~/.anycode/tasks/<id>/ layouts; fork-self; durable cross-process background agents (beyond in-process registry).
MCP beyond stdio v1	Partial: ANYCODE_MCP_READ_TIMEOUT_SECS (per-line read), ANYCODE_MCP_CALL_TIMEOUT_SECS (whole tools/call); clearer timeout / unexpected-EOF errors (incl. child exit); McpStdioSession::stdio_child_is_running; McpAuth without a GUI documented under Config & security. Still open: richer stdio lifecycle / reconnect; polish List / Read MCP resource tools in UX.
Channel AskUserQuestion	Card / inline-keyboard flows for WeChat / Telegram / Discord (new host implementations).

Docs note: explicit model-instructions file path is only via ANYCODE_MODEL_INSTRUCTIONS_FILE; JSON model_instructions controls discovery only — see Config & security.

Related

• Architecture
• Troubleshooting — MCP / OAuth

Chinese: 路线图.