fix(claude-code): filter injected harness turns from retention

[runbgp]

Fixes #3023.

Claude Code records an invoked skill's full SKILL.md as an ordinary user message, and injects <system-reminder> / <task-notification> blocks into user turns. Both were retained verbatim. On my transcripts that was ~50% of everything sent to extraction, and one skill invocation alone was a single 566 KB user message.

Same shape as the Codex fix in #1346:

• is_synthetic_user_message() — skips user turns starting with Base directory for this skill: , applied once in prepare_retention_transcript() so it covers both the text and JSON formats.
• strip_harness_blocks() — removes the two block types, alongside the existing strip_memory_tags() call on the retain path.

Recall query composition is left alone; this only touches retention.

Tests added for both functions and for the retention path. uv run pytest tests passes (213). No new ruff findings — the two pre-existing F541s in setup_hooks.py and the import-sort in test_content.py are untouched.

[nicoloboschi]

Closing: superseded by Coding Agents, which handles this class structurally — and your report is part of why it now handles it properly.

Measured against 400 real transcripts there: skill bodies arrive with isMeta: true (8/8) and are dropped with every other meta line, and <system-reminder> rides inside tool_result blocks, which the reader drops entirely. What was leaking were two cases neither of us had listed — <task-notification> (39 records) and Claude Code's compaction summary (29 records, 474,016 chars, retained as if the user typed them). Both are fixed: #3378 and #3383.

npx @vectorize-io/hindsight-coding-agents install claude-code (docs).