Status: verified 2026-09-17 against adapter SOURCES (the /tmp/acp-research copies) + @agentclientprotocol/[email protected] type declarations (npm tarball). Ground truth for W14 (plan / todo display, subagent & permission verification) and W15 (slash commands).

The portal chat currently drops two ACP session-update kinds on the floor: plan and available_commands_update both fall into the daemon's raw fallback (mapAcpUpdate default arm) and the web fold ignores every raw event except hnx/prompt-error. Result: Claude Code's todo list is COMPLETELY INVISIBLE (the wrapper no longer emits TodoWrite as a tool call), and no agent surfaces its slash commands. This doc verifies what each adapter actually emits before W14/W15 build the display.

// session/update — full REPLACE snapshot; the client replaces the whole plan.
{ sessionUpdate: "plan", entries: PlanEntry[] }
PlanEntry = {
  content: string                        // human-readable task description
  priority: "high" | "medium" | "low"
  status: "pending" | "in_progress" | "completed"
  _meta?: ...
}
// Doc note (verbatim): "When updating a plan, the agent must send a complete
// list of all entries with their current status."

// session/update — the agent's slash-command catalog.
{ sessionUpdate: "available_commands_update", availableCommands: AvailableCommand[] }
AvailableCommand = {
  name: string            // NO leading slash (e.g. "init"); MCP ones "mcp:foo"
  description: string
  input?: { hint: string } | null   // unstructured: "text typed after the name"
}

There is NO dedicated command-invocation RPC — a command runs as an ordinary session/prompt whose text is /name rest-of-line (the UnstructuredCommandInput doc: "All text that was typed after the command name is provided as input"). The full 1.4.0 sessionUpdate union (for completeness): user_message_chunk | agent_message_chunk | agent_thought_chunk | tool_call | tool_call_update | plan | plan_update* | plan_removed* | available_commands_update | current_mode_update | config_option_update | session_info_update | usage_update | compaction_update | compaction_summary_chunk (* = UNSTABLE draft kinds — no shipped adapter uses them; they keep falling to raw).

acp-agent.ts streaming switch: chunk.name === "TodoWrite" with Array.isArray(chunk.input?.todos) produces a plan update INSTEAD of a tool_call; isTaskTool(name) (TaskCreate/TaskUpdate/TaskList/TaskGet — the SDK-era task list, NOT the subagent Task tool) is suppressed the same way, with the plan snapshot emitted at tool_result time. Consequences:

1. The W6 TodoCard (registry key TodoWrite) is DEAD for claude-code — the tool call never arrives. The portal's todo display must ride the plan update. (The card stays as a fallback for any adapter that still emits TodoWrite as a tool call.)
2. in_progress entries carry the agent's activeForm text (e.g. "Analyzing dependencies…") instead of the static content — display verbatim.
3. During session/load replay the same mapping runs, so history replay naturally ends at the final plan snapshot — a last-wins panel converges without extra plumbing.

claude-agent-acp carries a draft implementation of ACP PR #1992 (acp-subagents.ts / native-subagents.ts): when the CLIENT advertises capabilities.subagents, it emits subagent_spawned / subagent_state_update and REWRITES child updates to the child's own sessionId. Our daemon does not advertise it (initialize sends clientCapabilities: {}) and W14 will keep it that way: the draft is unstable, and our single-channel fold has no concept of a second session stream — routed child output would corrupt the main transcript. Subagents keep surfacing as Task tool cards (verified path above); W14's subagent work is rig VERIFICATION of that path, not the draft.

• TodoPanel (components/stream/TodoPanel.tsx): a strip ABOVE the input area; empty list renders nothing; collapsed by default (header: icon + title + progress summary + chevron); expanded = per-entry status glyph + single-line ellipsis, capped + internal scroll. Data = last-wins plan snapshot (useAcpConnection case 'plan' → setPlan(entries)); replay converges ("会话打开即恢复面板").
• Commands (commands.ts + ChatInput.tsx): palette opens when the draft starts with / (not while a turn runs); first word after / is the filter (name OR description substring); ↑/↓ cycle, Esc closes, Enter on a BARE command selects + fills /name , Enter WITH args sends as a normal prompt. The portal ALSO ships a hardcoded fallback table — we deliberately DON'T (data-driven honesty: an agent that never pushed commands shows no palette; same rule as dsh's absent selectors).

• plan: ≤128 entries kept, content clamped to 512 chars, unknown status dropped entry-wise, priority passed through when present (display may ignore).
• commands: ≤64 rows, name ≤128, description clamped 512, input.hint clamped 256. Names ride VERBATIM (the web adds the /); MCP mcp:* names keep their prefix.

• W14 renders ONE todo/plan panel fed by the plan snapshot — no per-update transcript rows (full-replace semantics make rows a lie about the model).
• opencode gets NO plan panel (honest absence); its todo tool (if a model uses one) surfaces as an ordinary tool card.
• dsh shows neither panel nor palette — no fallbacks.
• session_info_update (claude goal _meta) and compaction_* stay raw — out of scope for both waves.