[Bug]: Copilot Autopilot final answers stay hidden in tool details #26697

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Searched open and closed issues and PRs for copilot, task_complete, end_turn, copilot autopilot, copilot final and copilot summary. Nothing covers this. #17714 and #24086 deal with Copilot subagent output. #26651 (session recovery) is unrelated.

Area

apps/server

Steps to reproduce

1. Enable the ACP Registry agent GitHub Copilot CLI (github-copilot-cli, Copilot CLI 1.0.105) and sign in.
2. Start a thread with Copilot and set Mode to Autopilot. Copilot advertises it as https://agentclientprotocol.com/protocol/session-modes#autopilot.
3. Ask for a small multi-step task, for example "Read the build scripts in this repo and tell me what build does."
4. Wait for the turn to finish.

Expected behavior

The turn ends with Copilot's final answer shown as an assistant message, the same as in Agent mode.

Actual behavior

The turn completes, but the final answer often never becomes an assistant message. In Autopilot, Copilot 1.0.104/1.0.105 ends its work by calling its native task_complete tool. The summary argument of that call is the answer, and Copilot writes no assistant text after the call. Upstream doesn't promote that summary into a reply, so when the turn's earlier assistant text is only progress notes, those notes are all the thread shows.

On web, the last progress note is treated as the turn's reply. The task_complete call folds into the "Worked for …" row with the rest of the turn's work. Its row is titled Tool: the completing tool_call_update carries no title, so the row's label falls back from "Task complete" to "Tool". To read the answer, the user has to expand "Worked for …", then expand that Tool row. The answer then shows as a plain-text summary argument line, and again in the output as raw {"content": …, "detailedContent": …} JSON (packages/client-runtime/src/work-log/itemDetail.ts, apps/web/src/components/chat/V2ItemInspector.tsx). Nothing in the web, mobile or client-runtime code handles task_complete specially.

Upstream has no step that turns that summary into a reply (main @ 9826dbb):

• packages/provider-acp-registry/src/server/copilotAcp.ts:15-25: the Copilot client capabilities subscribe only to subagent.* and assistant.*_delta session events. Nothing observes tool.execution_start for task_complete or session.task_complete.
• packages/provider-acp/src/server/adapter.ts:7613-7656: when session/prompt resolves, the turn is finalized straight from result.stopReason. The flavor has no hook to append a final answer reported outside agent_message_chunk.
• packages/provider-acp/src/server/adapter.ts:3620-3621: the task_complete ACP tool_call is projected like any other tool, with the summary kept only in its rawInput/rawOutput.

Copilot's own session logs show the pattern. Seven Copilot sessions ran in Autopilot from T3 over ACP on this machine (1.0.104 ×3, 1.0.105 ×4), and every one ended a turn through task_complete:

• 6 of 7: the turn's own assistant text was only short progress notes (18–296 chars each). The task_complete summary held the actual answer (384–2,548 chars), and nothing came after it.
• The 7th: Copilot had already written an 883-char answer, and an autopilot continuation then called task_complete with a restated summary.

No session outside Autopilot called task_complete.

session/prompt itself behaves correctly. In T3's provider log it resolves about 20 ms after task_complete completes and the native turn ends, so the turn is not cut short. The answer is simply delivered somewhere T3 doesn't read it.

Impact

Major degradation or frequent failure

In six of the seven Autopilot tasks I inspected, the thread looked finished but showed no answer; all seven ended through task_complete, whose summary upstream doesn't promote into a reply. In my case the threads were delegated child tasks, where that summary is the whole result.

Version or commit

main @ 9826dbb (replay test and code reading); behaviour observed with GitHub Copilot CLI 1.0.104 and 1.0.105

Environment

macOS (Darwin 27), GitHub Copilot CLI 1.0.104 and 1.0.105 via the ACP Registry agent github-copilot-cli, Autopilot mode

Logs or stack traces

# Copilot CLI 1.0.105, Autopilot. Copilot's native ~/.copilot/session-state/<session>/events.jsonl
# for the last turn of one interaction. Text replaced by lengths; ids shortened.
21:07:29.393Z assistant.turn_start     turnId=157
21:07:34.832Z assistant.message        content=""  toolRequests=[task_complete]
21:07:34.832Z tool.execution_start     toolName=task_complete  arguments.summary=<1,565 chars: the final answer>
21:07:34.834Z tool.execution_complete  success=true
21:07:34.835Z session.task_complete    summary=<1,565 chars>  success=true
21:07:34.837Z assistant.turn_end       turnId=157
# T3 provider log, same thread:
21:07:34.857Z session/prompt succeeded
# The earlier assistant messages in this interaction were short progress notes (27-84 chars).

A synthetic reconstruction of Copilot's session updates, exercised through upstream's real adapter on main @ 9826dbb. It is not a captured Autopilot wire trace or a browser reproduction. The fields follow Copilot's ACP frames for its other tools and the task_complete entries in its session logs. It is a throwaway copy of the existing routes interleaved Copilot subagent output into subagent threads case in apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.test.ts, with the same Copilot flavor wiring (makeCopilotSubagentRouting, extractCopilotSubagentUpdate):

1. agent_message_chunk "Checking the build scripts.", preceded by its assistant.message_delta session event.
2. tool_call {toolCallId, title: "Task complete", kind: "other", status: "pending", rawInput: {summary: "FINAL-ANSWER-I1: …"}}.
3. tool_call_update {status: "completed", content: [text summary], rawOutput: {content, detailedContent}}.
4. session/prompt → {stopReason: "end_turn"}.

# vp test run apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.test.ts -t SCRATCH-I1
turn.terminal: status=completed
assistant_message items: ["Checking the build scripts."]
summary in any assistant_message: false
items containing the summary: ["dynamic_tool"]
dynamic_tool title over time: "pending:Task complete" -> "completed:Tool"
dynamic_tool.input  = {"summary": "FINAL-ANSWER-I1: `build` runs `turbo run build`, ..."}
dynamic_tool.output = {"content": "FINAL-ANSWER-I1: ...", "detailedContent": "FINAL-ANSWER-I1: ..."}

Screenshots, recordings, or supporting files

None. The replay under Logs shows what upstream produces for this turn shape.

Workaround

Use Copilot's Agent mode instead of Autopilot. Copilot then writes its answer as normal assistant text.

Additional context

I'm happy to open a focused PR once the expected behavior is agreed.

Investigated by Opus 5.5 with Claude Code. Reviewed by GPT 6.1 Sol.