tests_e2e Wire E2E Guide

Goals and Scope

• Test only kimi --wire JSON-RPC + wire messages; no Shell UI/Print/ACP/Term/shortcuts.
• Do not test --agent okabe.
• Do not test: W-23, W-26, W-29, W-27 (env overrides).

Execution Rules

• Tests run via uv run kimi by default; set KIMI_E2E_WIRE_CMD to override the base command (e.g. ../kimi-agent-rs/target/debug/kimi-agent or kimi-agent on PATH). --wire is appended if missing.
• Always isolate HOME, USERPROFILE, and KIMI_SHARE_DIR, and use a temporary --work-dir to avoid touching real ~/.kimi.
• Use inline_snapshot for snapshot testing; snapshots can start empty and be updated later.
• Wire traffic is line-delimited JSON; event/request/responses may interleave.

Test Matrix

Startup and Protocol

• W-01 Handshake: initialize returns protocol_version=1.1, server, and non-empty slash_commands.
• W-02 External tool registration and call: register external_tools, trigger a ToolCallRequest, return ToolResult, turn finishes.
• W-03 External tool conflict: registering a tool name that conflicts with a built-in is rejected with a reason.
• W-04 Prompt without handshake: send prompt without initialize, turn still completes.
• W-05 LLM not set: missing config yields -32001 (LLM is not set).
• W-06 Max steps: set a tiny --max-steps-per-turn, response is status=max_steps_reached.

Prompt and Event Stream

• W-07 Basic turn: TurnBegin/StepBegin/ContentPart(text)/StatusUpdate flow.
• W-08 Multiline input: TurnBegin.user_input preserves newlines.
• W-09 Content parts input: user_input as ContentPart[] (text/image/audio/video) with capabilities enabled.
• W-10 Thinking toggle: --thinking emits ContentPart(type=think), --no-thinking does not (real LLM).
• W-11 Concurrent prompt: second prompt returns -32000 (turn already in progress).
• W-12 Cancel turn: cancel returns {}, original prompt becomes status=cancelled (may emit StepInterrupted, real LLM).

Tools and Approvals

• W-13 Shell approval: Shell triggers ApprovalRequest, approve yields ToolResult.
• W-14 Approval reject: reject ends turn; tool does not run.
• W-15 Session approval: approve_for_session removes subsequent blocking approvals.
• W-16 YOLO: --yolo skips blocking approvals.
• W-17 DisplayBlock coverage: Shell/WriteFile/StrReplaceFile/SetTodoList emit expected display types.
• W-18 Tool args streaming: ToolCallPart.arguments_part can be stitched into full args.

Sessions and Context

• W-19 Session files: --session <id> writes context.jsonl and wire.jsonl.
• W-20 Continue session: --continue appends to the same session files.
• W-21 Clear context: /clear ensures the next prompt does not rely on prior context.
• W-22 Manual compaction: /compact triggers CompactionBegin/CompactionEnd.
• W-24 Status stats: StatusUpdate.context_usage/token_usage types are valid and change over time.

Config and Runtime Flags

• W-25 --config inline: JSON/TOML string config works.
• W-28 CLI override: --model overrides default_model.
• W-30 Work dir: prompt asks "where is the current directory", verify --work-dir (real LLM).

Extensions (Agents/Skills/MCP)

• W-31 Built-in agent boundary: default calling SendDMail fails or is rejected.
• W-32 Custom agent: --agent-file excludes tool and calls are rejected.
• W-33 Subagents: prompt "parallel call two task tools, each runs shell sleep 0.5 and 1 second"; verify Task/SubagentEvent streaming/ToolResult/multiple approvals (real LLM).
• W-34 Skill call: create a test skill, /skill:test injects SKILL.md (wire has no /help).
• W-35 Flow skill: /flow:<name> runs the flow until END.
• W-36 MCP: use a Python fastmcp test server, load via --mcp-config-file, verify tool works.

Resilience and Errors

• W-37 Invalid JSON: malformed JSON line returns -32700.
• W-38 Invalid request: missing fields returns -32600.
• W-39 Unknown method: returns -32601.
• W-40 Invalid params: bad structure returns -32602.
• W-41 Cancel without active turn: returns -32000.
• W-42 LLM errors: unsupported model -32002 and service error -32003.

Real LLM Placeholders

• W-10/W-12/W-30/W-33 require a real provider; everything else uses _scripted_echo.