Why is Plannotator Ask AI not using the current session? Version and async-review checks

First identify which Ask AI mode you need

If you want review questions answered by the Agent writing the code, look for Ask this session and the Claude Code, Pi or OpenCode name. This mode uses the context of the session that initiated review. If a separate provider or model selector remains visible, check connection requirements first. An open review page does not establish connection to the original session.

0.28.0 was officially released at 00:50 UTC on October 5, 2026, followed that day by 0.28.1 at 04:11 UTC, 0.28.2 at 07:19 UTC and official patch 0.28.3 at 09:41:38 UTC. Chinese-language discussion posts appeared later; that does not make the version in a post the latest. This article explains conditions from that day’s official notes, not proof that your machine is updated.

After installation, check the command and plugin separately

New users should choose their operating system in the official installation documentation below. Full installation configures detected Agents; minimal installs only the binary, so it does not establish that plugins, hooks and commands are ready. Open a new terminal after installation and run plannotator --version and plannotator --help to confirm the version actually executed and available commands.

Claude Code also requires /plugin marketplace add backnotprop/plannotator followed by /plugin install plannotator@plannotator in the session, then a restart. Existing plugin users should run claude plugin marketplace update plannotator followed by claude plugin update plannotator@plannotator in an external terminal. Update the binary using the official installation documentation as well, then restart Claude Code. Refreshing the marketplace alone does not update the plugin.

Why Claude Code may still be waiting for review

Plannotator’s Claude mod path requires Claude Code 2.1.287 or later and an interactive terminal. Run claude --version first, then check whether the plugin is enabled and whether PLANNOTATOR_CLAUDE_MOD=0 or claudeCodeMod=false was previously set. Organizational policy restrictions belong with administrators; disabling policies is not a general troubleshooting step.

Native Windows, older Claude Code versions, claude -p and SDK sessions retain the classic flow. Remote mode, --tailscale, OpenCode 1 and running plannotator directly outside an Agent session also cannot assume current-session connectivity. An old binary can return plan review to blocking mode with an update notice. These environmental differences do not necessarily mean plugin installation failed.

Run a minimal check that the question reaches the original session

In a supported interactive Claude Code session, run /plannotator-last to review the previous response, or /plannotator-review to open the current code diff. Select a small diff in code review and choose Ask AI from the annotation toolbar; general questions can use the AI tab. Verify the current-session label first, then ask a small question about the visible content. In the official connection flow, the question should appear in Agent chat and the answer should stream back into the review panel.

When the Agent is busy, either wait for the current turn to finish or intentionally interrupt and ask. If a new message you type in the original chat takes over the same turn answering Ask AI, 0.28.3 protects that turn: Stop then only closes the panel question, and Interrupt and ask now does not interrupt your own work. This protection has a specific trigger; it does not mean every Stop in every version leaves the Agent running. If the original session is closed or unreachable, check the initiating session and reopen the corresponding review rather than expecting an automatic switch to another AI provider.

Async return does not mean the plan is approved

Enter plan mode first and let Claude draft a plan. Its ExitPlanMode call opens review automatically; no separate plan-review command is needed. The mod returns control, but plan mode stays active and the Agent is told approval is pending. Once a decision is made, feedback returns as a plugin message. Do not manually leave plan mode during review: official guidance says Claude may then begin editing. An unblocked terminal is not confirmation of approval.

When the Agent revises the plan during review, the same page updates and retains comments. Decisions for an old version are rejected; review the new version. After approval, Claude calls ExitPlanMode again and proceeds according to the approved text and selected permission mode. To restore the classic flow, set PLANNOTATOR_CLAUDE_MOD=0 or set claudeCodeMod to false in existing configuration; it takes effect on the next Claude Code start.

Distinguish Pi, OpenCode and the scope of the 0.28 series

Pi plan decisions also return later as session messages, with planning restrictions retained until approval. OpenCode 2 code review, annotate and last can initiate full turns in the session. During plan review it is still waiting for submit_plan, so Ask AI supplies only a quick context-based answer, without tools or chat-history writes. Use that label as a quick-explanation mode.

0.28.1 fixes duplicate export of folder-annotation feedback and adds Ask AI to diagram comments. The button becomes available after a question is entered, while Enter still submits the comment. It also fixes some OpenCode link-display, duplicate-reply and notification-order issues. These patches have specific scopes and do not establish that every host, remote connection or review problem is fixed.

Are your own chat replies appearing in Ask AI? How 0.28.3 handles this

This patch addresses a specific overlap between review and chat: Ask this session is streaming an answer when the user enters a new message in the same Agent chat. Older versions could also stream the reply to that new message into Ask AI, and clicking Stop could cancel the user’s own work. When 0.28.3 detects that a new message not sent by Plannotator has taken over the same turn, it stops the panel receiving replies, retains the partial answer already received and explains that subsequent replies went to the new message.

The officially listed paths are Claude Code with the Plannotator mod enabled, Pi and OpenCode 2. After takeover, Stop closes only the panel question, and interrupt-and-ask-now does not stop the taken-over turn. Let your own task finish before asking another review question. If Ask AI’s answer completed before the new message, the full answer remains. Agent-internal events such as background-task completion, system notifications or compaction messages are not this user-takeover scenario; messages in other Claude sessions are not either.

Update the binary and corresponding host plugin separately, restarting Claude Code afterward. For Pi, follow the official pi update --extensions steps; for OpenCode, rerun the official installer and update the plugin cache. After checking versions, use a small non-sensitive question to observe the partial answer, notice and actual result in the original chat. Some pull-host paths combining an old server and updated plugin end the question with a partial answer and explanation; this does not prove all hosts are updated. This describes the official fix and checks, not a test of cancelling or restoring user tasks.

Does empty Done still send feedback? Check the 0.28.2 update path

0.28.2 fixes clicking Done in an annotate session with no annotations and the equivalent /plannotator-last closing path. Official notes say Claude Code mod, Pi, OpenCode, Amp and Droid no longer send an extra prompt asking the Agent to process feedback; instead they record that the session closed without feedback. This is not plan approval or the behavior of an empty Approve in code review.

Claude Code mod, OpenCode CLI bridge, Amp and Droid require the Plannotator binary update. Binaries 0.28.0 and 0.28.1 do not emit the new signal, so updating only the host plugin can retain the old behavior. Pi and OpenCode embedded runtimes bundle the server and receive this fix through plugin updates. Pi uses the official pi update --extensions steps; OpenCode reruns the official installer, which also clears the OpenCode 2 plugin cache. Claude Code’s binary and marketplace plugin still require separate updates.

Codex and Gemini skill-based paths, and Claude Code with the mod disabled, still read the original plain-text no-feedback sentence. PR1701 also explains that the OpenCode 1 embedded command itself still runs an empty turn; only the second feedback turn is removed. Do not interpret the patch as no turns on every host or a guarantee of no usage charges. Identify your own runtime path first.

What to preserve before seeking help if it still does not work

Record, in order, binary version, plugin version, Claude Code version, interactive or non-interactive launch, local or remote mode and panel label. Retry with a non-sensitive sample response or diff and save the exact error text. Do not immediately clear all Agent configuration; establish whether the issue concerns connection conditions, a version combination or a particular document review.

Ask this session places questions and answers into the original Agent conversation. Independent-provider mode processes the relevant context through the selected provider. Use examples without keys or customer data and check calls under existing service usage rules; plugin connectivity is not an additional free allowance. Review Agents, Code Tour and Guided Review manage model settings separately.