Pi: make /plannotator-last message anchoring optional

[vekexasia]

Problem

In the Pi extension, /plannotator-last injects a large amount of redundant context when sending feedback on the most recent assistant response.

With @plannotator/pi-extension 0.27.3, a small annotation produces a message shaped like:

This feedback applies to the earlier assistant response excerpted below:

> [up to 1,000 characters of the original assistant response]

User feedback:
[the actual annotation feedback]

For the normal /plannotator-last workflow, the original response is already in Pi's conversation context. Repeating up to 1,000 characters wastes context and makes short feedback unnecessarily verbose.

Reproduction

1. Use Pi with @plannotator/pi-extension 0.27.3.
2. Generate a normal assistant response.
3. Run /plannotator-last.
4. Add one short annotation and send feedback.
5. Observe that the injected follow-up includes a long excerpt of the original response in addition to the feedback.

Current implementation

apps/pi-extension/index.ts:

• anchorMessageFeedback() prepends the original response excerpt.
• The result is built before getAnnotateMessageFeedbackPrompt() applies prompts.annotate.runtimes.pi.messageFeedback.

Therefore, setting messageFeedback to {{feedback}} removes only the wrapper. It cannot disable the original-message excerpt.

Expected behavior

Provide a way to disable or limit this anchoring for Pi, for example a configuration option that allows feedback to contain only the annotations. The current behavior could remain the default for stale or previously selected messages, while normal last-message feedback could avoid duplicating context.

Related issues

• Context lost after plan approval when using plannotator-annotate in Claude Code #193 requested reducing annotation message overhead.
• feat: Context Anchoring — preserve plan decision reasoning across sessions #431 requested context anchoring and was consolidated into Plan context panel and feedback policy controls #900.
• Plan context panel and feedback policy controls #900 tracks plan context and feedback policy controls, including omitting or summarizing feedback context.

[ashish921998]

I've taken this one — #1371 adds an annotateMessageAnchoring config seam (default true, so current behavior is unchanged) with a PLANNOTATOR_ANNOTATE_MESSAGE_ANCHORING env override. The anchor still fires only for possibly-stale targets via the existing shouldAnchorLastMessageFeedback gate; message-scope feedback from the picker never anchors. Happy to reshape the option (name, placement, or making it a template-level control as discussed in #900) if you'd prefer a different seam.