BAN-148: Fix Sonnet continuing conversation instead of summarising

[RemoteCTO]

Sonnet treated USER:/ASSISTANT: transcript entries as a
conversation to continue, producing ~700 char in-character
responses instead of structured summaries. Both retry attempts
failed validation, leaving sessions unresumable.

Root cause: The system prompt and input framing were too
weak — Sonnet couldn't distinguish between "here is a
conversation to summarise" and "here is a conversation to
continue".

Changes

• System prompt reframed as "SUMMARISATION TOOL" with
  explicit anti-roleplay instructions. Conditional: uses
  generic format instructions when SEAMLESS_PROMPT_FILE
  is active to avoid heading conflicts with custom prompts.
• Transcript wrapping in <transcript> XML tags creates
  a clear data boundary between instructions and input.
• Closing instruction appended after the transcript:
  "Produce the structured summary now."
• User prompt restructured: numbered section prefixes
  removed (### 1. → ## ), explicit "include ALL six
  sections" rule, 2000 char minimum target.
• Plugin manifest version synced with package.json.

Test plan

• Manual compaction of failed session db080aef —
  produced 15,735 chars valid output (was 676/716 chars)
• All 158 existing tests passing
• Biome lint clean
• Validated against real 75K char infrastructure
  transcript (326 entries)

[Copilot]

Pull request overview

This PR fixes a critical bug where Claude Sonnet was treating the conversation transcript as an ongoing conversation to continue rather than data to analyze, resulting in ~700 character in-character responses instead of comprehensive structured summaries. The fix implements three defensive measures: reframing the system prompt as a "SUMMARISATION TOOL" with explicit anti-roleplay instructions, wrapping the transcript in XML tags to create a clear data boundary, and adding a closing instruction to produce the summary.

Changes:

• Updated system and user prompts with anti-roleplay instructions and explicit section requirements
• Wrapped transcript input in <transcript> XML tags instead of using bare text after a separator
• Version bumped to 0.2.1 with corresponding CHANGELOG entries

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
scripts/compactor.mjs	Updated PROMPT and SYSTEM_PROMPT with anti-roleplay instructions; wrapped transcript in XML tags; added closing instruction; removed section numbering
package.json	Version bump from 0.2.0 to 0.2.1
.claude-plugin/plugin.json	Version sync from 0.1.1 to 0.2.1 to match package.json
CHANGELOG.md	Added comprehensive 0.2.1 release notes documenting the bug fix and changes

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.