fix(acp): replay loaded session history

[huntharo]

Summary

This PR builds on #2359 and includes all current commits from that PR plus the session-load follow-up:

1. fix(acp): assign message ids to streamed content #2359: ACP SDK 0.10.0 support, streamed content messageId values, and terminal auth login invocation support.
2. This follow-up: make ACP session/load replay restored history, including messageId fields on replayed user, assistant, and thought chunks, and expose restored session metadata.

I can rebase this PR after #2359 lands so it contains only the session-load replay change. Alternatively, if maintainers prefer to take both fixes together, this PR can be merged as the combined ACP fix and #2359 can be closed.

Changes

• Return a LoadSessionResponse with initial modes and models from session/load.
• Advertise transcript restore support as sessionCapabilities._meta.kimi.sessionHistoryReplay = true so clients do not have to infer replay behavior from loadSession alone.
• Replay persisted wire.jsonl records as ACP session/update notifications on load.
• Send messageId on restored user, assistant, and thought chunks using the message-id run tracking from fix(acp): assign message ids to streamed content #2359.
• Emit standard ACP SessionInfoUpdate title/updatedAt updates after prompt completion, and during session/load/session/resume.
• Return title/updatedAt metadata on session/load and session/resume under _meta.kimi.session.
• Record ACP/direct KimiCLI.run() turns to the session wire log so future loads have a replay source.
• Fall back to context text history for older sessions that do not have a wire log.
• Cover wire replay, context fallback, restored messageId, advertised replay capability, and session title metadata with tests.

Validation

• uv run pytest tests/acp tests/core/test_notifications.py -q
• uv run ruff check src/kimi_cli/acp/session.py src/kimi_cli/acp/server.py src/kimi_cli/app.py tests/acp/test_protocol_v1.py tests/acp/test_session_notifications.py tests/acp/test_server_initialize.py tests/ui_and_conv/test_acp_server_auth.py
• uv run pyright src/kimi_cli/acp/session.py src/kimi_cli/acp/server.py src/kimi_cli/app.py tests/acp/test_protocol_v1.py tests/acp/test_session_notifications.py tests/acp/test_server_initialize.py tests/ui_and_conv/test_acp_server_auth.py
• git diff --check

[devin-ai-integration]

Devin Review found 1 potential issue.

View 5 additional findings in Devin Review.

[devin-ai-integration]

🔴 AssertionError when replaying Notification after TurnEnd in wire history

In replay_history, the Notification case (line 356-358) calls _send_notification → _send_text → _content_run_id("message") which asserts self._turn_state is not None (src/kimi_cli/acp/session.py:438). However, after a TurnEnd is processed during replay (line 325-329), self._turn_state is set to None. Notifications can appear after TurnEnd in the wire file because run_soul() at src/kimi_cli/soul/__init__.py:239 performs a final _deliver_notifications_to_wire_once() flush after the soul task finishes (i.e., after TurnEnd has been emitted) but before the wire shuts down. This means any session with late-flushed background-task notifications will crash with AssertionError when loaded via session/load.

Trace through the crash path

1. Wire file contains: ... TurnEnd ... Notification (notification flushed after turn)
2. replay_history processes TurnEnd → sets self._turn_state = None (line 329)
3. replay_history processes Notification (line 356) → calls _send_notification
4. _send_notification → _send_text → _content_run_id("message") → assert self._turn_state is not None → AssertionError

Prompt for agents

In replay_history(), the Notification case at lines 356-358 does not guard against self._turn_state being None. When a Notification appears in the wire file after TurnEnd (which happens due to the final notification flush in run_soul), self._turn_state is None and _send_notification -> _send_text -> _content_run_id will hit an assertion error.

The fix should either:
1. Guard the Notification case with a check like `if self._turn_state is not None`, similar to how ToolCallPart and ToolResult are guarded (lines 349, 353). If turn_state is None, either skip the notification or create a temporary turn state via _begin_replay_turn() before sending.
2. Or make _send_text / _content_run_id handle the case where _turn_state is None gracefully (e.g., generate a one-off message_id).

Option 1 is simpler and consistent with the existing pattern for ToolCallPart/ToolResult. The same concern technically applies to ThinkPart and TextPart cases (lines 333-338), though those are far less likely to appear outside a turn.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5e25b1d753

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Guard replayed notifications when no turn is active

replay_history() can call _send_notification() after a TurnEnd (or before any TurnBegin) because wire logs may contain Notification records emitted outside prompt turns (e.g., background-task notifications flushed by the runtime). In that state self._turn_state is None, so _send_notification() -> _send_text() -> _content_run_id() hits the assert and aborts session/load for affected sessions instead of replaying history.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9030439624

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Reset content run around replayed notifications

In replay_history(), notifications are replayed via _send_notification() without clearing the current content run first, unlike the live prompt() path which resets before and after notifications. If a Notification is persisted between assistant text chunks in the same turn, replay reuses the same message_id for notification and adjacent assistant chunks, so clients that group by message_id will merge system notification text into the assistant message and lose original message boundaries.

Useful? React with 👍 / 👎.