[Waishnav]

Until now a running subagent was a black box. LocalAgentRuntime.run() only returned when the turn ended, the host had no way to see what the worker was doing, and nothing identified an individual turn, so there was nothing to tie a subagent's work to later review changes. On top of that, agents show hid why a turn failed: hosts saw Claude agent turn failed. while the provider's real error ("the selected model does not exist", "usage limit reached") stayed inside the error cause.

This PR gives every turn an app-owned trn_ id and a small, normalized event stream defined in DevSpace's own terms. The types are turn_started, message, tool_started, tool_completed (tool, short summary, status) and turn_completed (completed/failed/stopped). The manager allocates the turn id and emits the two lifecycle events itself. Each driver maps its native frames into activity through a new onEvent callback, so no provider frame enters the domain model. Before persisting, the manager redacts and bounds each event. Tool summaries keep only the identifying argument (a command, path, or query), never full inputs or outputs. A turn keeps at most 500 events and 256 KiB, with the last slot reserved for turn_completed. Only an agent's 20 most recent turns keep their event logs. The log is exposed through a new agent.events daemon method (protocol v8) and devspace agents events <id> [--turn <id>] [--follow], written as the existing XML fragments. The daemon protocol is request/response, so --follow polls every 500 ms for events after the last seq it has seen. It also pins the turn it started on, so it stays on that turn even if the agent is continued. show, ls, run, and wait now include turn="trn_…". Failed turns also persist the provider's own message, redacted and capped at 1,000 characters, shown as <cause> in XML and error.detail in JSON.

Streaming support is reported by capabilities.turns.streaming. It sits next to interrupt to match the existing nested shape. Claude maps complete SDK assistant/tool_result messages. Pi maps tool_execution_* and message_end RPC frames. Codex maps app-server item/started/item/completed for commands, file changes, MCP tools, and web search. ACP maps tool_call/tool_call_update and joins agent_message_chunk deltas into one message per stretch of text. OpenCode v1 and v2 report streaming: false: both run the turn as one blocking prompt call, and their SSE event stream would be its own change. They still get turn_started/turn_completed from the manager. Claude and Pi were verified live. Codex and OpenCode were not, because both accounts are out of quota. Codex and ACP are covered by fake-transport tests: a fake codex app-server process and the ACP notification path.

I did not change the open_workspace card. The MCP server has no connection to the agent daemon today, since subagents run only through the CLI and daemon. The card is also a one-time snapshot of open_workspace with no refresh path. Showing live subagents there would need two things: the server opening a read-only client to the daemon (or the store) during open_workspace, and a way for the card to refresh, which today means a new app-callable tool. That is a product decision, not a drive-by. Once both this PR and #407 land, WorkspaceActivityJournal.currentTurn(workspaceId) can be backed directly by agent.events. It would list the workspace's agents, take each one's latestTurnId, and page events by afterSeq. The trn_ id is the stable key a review checkpoint can record to link changes back to the turn that made them.

Live smoke used the built CLI (node dist/cli.js) with a temp DEVSPACE_CONFIG_DIR, a throwaway git repo, and Pi 1.4.0 with an isolated PI_CODING_AGENT_DIR (xai/grok-4.5). The state dir had to live under a short /tmp path, because a long scratchpad path exceeds the Unix-socket length limit; that limit predates this PR. Trimmed transcripts are below. The Claude turn shows its sandbox rejecting a read and edit outside the workspace (status="failed"). The Pi turn shows parallel tool calls being correlated. A Pi turn stopped during sleep 30 ended the follower with turn_completed status="stopped" code="PROVIDER_CANCELLED"; Pi sends no tool_execution_end for the cancelled tool, so that tool_started stays unpaired. A Claude turn with a bogus model showed the new cause:

<agent id="agt_82131682" status="failed" turn="trn_2dcbaf62" code="PROVIDER_EXECUTION_ERROR" retryable="false">Claude agent turn failed.<cause>Claude returned an error result: There's an issue with the selected model (claude-nonexistent-9). It may not exist or you may not have access to it.</cause></agent>

agents events --follow during a real Claude turn (events 9–12 elided)

<event seq="1" type="turn_started" turn="trn_cd9eb384" at="2026-10-25T21:49:09.229Z"/>
<event seq="2" type="tool_started" turn="trn_cd9eb384" tool="Read">/home/waishnav/math.js</event>
<event seq="3" type="tool_completed" turn="trn_cd9eb384" tool="Read" status="failed">/home/waishnav/math.js</event>
<event seq="4" type="tool_started" turn="trn_cd9eb384" tool="Bash">cat ./math.js</event>
<event seq="5" type="tool_completed" turn="trn_cd9eb384" tool="Bash" status="completed">cat ./math.js</event>
<event seq="6" type="tool_started" turn="trn_cd9eb384" tool="Bash">node -e "import(\"./math.js\").then(m =&gt; console.log(m.add(2, 3)))"</event>
<event seq="7" type="tool_completed" turn="trn_cd9eb384" tool="Bash" status="completed">node -e "import(\"./math.js\").then(m =&gt; console.log(m.add(2, 3)))"</event>
<event seq="8" type="message" turn="trn_cd9eb384">Now I'll fix the bug by changing the subtraction to addition:</event>
<event seq="13" type="tool_started" turn="trn_cd9eb384" tool="Edit"><repo>/math.js</event>
<event seq="14" type="tool_completed" turn="trn_cd9eb384" tool="Edit" status="completed"><repo>/math.js</event>
<event seq="15" type="tool_started" turn="trn_cd9eb384" tool="Bash">node -e "import(\"./math.js\").then(m =&gt; console.log(m.add(2, 3)))"</event>
<event seq="16" type="tool_completed" turn="trn_cd9eb384" tool="Bash" status="completed">node -e "import(\"./math.js\").then(m =&gt; console.log(m.add(2, 3)))"</event>
<event seq="17" type="message" turn="trn_cd9eb384">Fixed the `add` function by changing `a - b` to `a + b`, and the test now correctly returns 5.</event>
<event seq="18" type="turn_completed" turn="trn_cd9eb384" status="completed" at="2026-10-25T21:49:27.040Z"/>

agents events --follow during a real Pi turn

<event seq="1" type="turn_started" turn="trn_34b8f3b4" at="2026-10-25T21:49:39.459Z"/>
<event seq="2" type="message" turn="trn_34b8f3b4">I'll read `math.js` and run the add check next.</event>
<event seq="3" type="tool_started" turn="trn_34b8f3b4" tool="read"><repo>/math.js</event>
<event seq="4" type="tool_started" turn="trn_34b8f3b4" tool="bash">node -e 'import("./math.js").then(m =&gt; console.log(m.add(2, 3)))'</event>
<event seq="5" type="tool_completed" turn="trn_34b8f3b4" tool="read" status="completed"><repo>/math.js</event>
<event seq="6" type="tool_completed" turn="trn_34b8f3b4" tool="bash" status="completed">node -e 'import("./math.js").then(m =&gt; console.log(m.add(2, 3)))'</event>
<event seq="7" type="message" turn="trn_34b8f3b4">`add` is subtracting; fixing that and re-running the check.</event>
<event seq="8" type="tool_started" turn="trn_34b8f3b4" tool="edit"><repo>/math.js</event>
<event seq="9" type="tool_started" turn="trn_34b8f3b4" tool="bash">node -e 'import("./math.js").then(m =&gt; console.log(m.add(2, 3)))'</event>
<event seq="10" type="tool_completed" turn="trn_34b8f3b4" tool="edit" status="completed"><repo>/math.js</event>
<event seq="11" type="tool_completed" turn="trn_34b8f3b4" tool="bash" status="completed">node -e 'import("./math.js").then(m =&gt; console.log(m.add(2, 3)))'</event>
<event seq="12" type="message" turn="trn_34b8f3b4">Fixed `add` so it returns `a + b` instead of `a - b`, and the check now prints `5`.</event>
<event seq="13" type="turn_completed" turn="trn_34b8f3b4" status="completed" at="2026-10-25T21:49:46.724Z"/>

Migration 11 adds turn_id and error_detail to local_agent_turns, latest_turn_id and error_detail to local_agent_sessions, and a local_agent_turn_events table, backfilling trn_<rowid> ids for existing turns. The migration runner rejects unknown versions, so any other open PR that claims version 11 will need renumbering when the two meet.

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features
  • Added devspace agents events <id> to view an agent turn’s activity, with --follow to watch until it ends and --turn to select an earlier turn.
  • Agent activity can include messages and tool updates, with turn IDs and sequence numbers. Providers without streaming report turn start and completion.
• Improvements
  • Failed-agent results can include a redacted, truncated explanation from the provider.
  • Activity history is limited and sensitive information is redacted.

[Waishnav]

Re the outside-diff finding (local-agent-daemon-protocol.ts:279): fixed in d3734ee. LocalAgentWaitResult now has an optional turnId. waitResultFromTurn sets it from the selected turn, waitResultFromAgent sets it from the agent's latestTurnId, and decodeAgentWaitResults keeps it, so agents wait shows turn="…". The manager test now asserts the ids match each agent's latest turn (f5a3fa8), and a protocol test covers decoding.