flow-post-task-sync

Flow Post-Task Sync

Purpose

Leave the project in a synchronized, documented, committed state. This skill runs after Serena memory sync, not instead of it.

Workflow

1. Confirm Serena memories are current via python3 "$(git rev-parse --show-toplevel)"/plugins/rldyour-serena-mcp/scripts/serena_memory_state.py (use git rev-parse --show-toplevel rather than ${CLAUDE_PLUGIN_ROOT} because the script lives in a different plugin; ${CLAUDE_PROJECT_DIR} is documented only for hook commands and stdio MCP env, not skill execution context). If is_current=false, invoke the rldyour-serena-mcp:flow-memory-sync subagent through the Agent tool - pass HEAD SHA, newest synced SHA, and the changed-files list as prompt context. The subagent has narrow tool access (Serena memory tools + read-only Bash/Read/Grep/Glob; Edit/Write/NotebookEdit are disallowed in its frontmatter), enforces fact-only updates, runs commit_serena_knowledge.sh itself, and emits a JSON report. Do not edit .serena/memories/*.md from the main session - keep all memory mutations in the dedicated subagent so anti-hallucination guards apply.
2. Read .serena/.flow_post_task_state.json if present and run python3 ${CLAUDE_PLUGIN_ROOT}/scripts/flow_post_task_state.py. Inspect branch_cleanup_state and run bash ${CLAUDE_PLUGIN_ROOT}/scripts/git_sync_audit.sh when branch/worktree cleanup is not obviously complete.
3. Inspect uncommitted changes deeply. Separate source changes, docs, Serena knowledge, generated junk, runtime markers, and secrets.
4. Run instruction-docs-sync when durable project instructions may have changed. Keep AGENTS.md as the concise root project-instruction file and .claude/CLAUDE.md as the Claude Code-native deep project memory in fullrepo-managed projects.
5. Run applicable quality checks from project scripts and bash ${CLAUDE_PLUGIN_ROOT}/scripts/detect_project_checks.sh.
6. Commit atomically with Conventional Commits. Use separate commits for implementation, tests, docs/instructions, and Serena knowledge when that improves history.
7. Push to upstream when configured. If no upstream exists, ask before creating one.
8. Keep normal branch history clean from agent-only files. Ensure .git/info/exclude contains the rldyour fullrepo block and move tracked agent-only files out of the current branch with python3 ${CLAUDE_PLUGIN_ROOT}/scripts/fullrepo_sync.py --migrate-main only when the project is ready for that migration.
9. Publish the complete project snapshot to fullrepo through python3 ${CLAUDE_PLUGIN_ROOT}/scripts/fullrepo_sync.py --publish. This uses safe --force-with-lease, not a blind force push.
10. Remove merged local and remote branches/worktrees only after verifying they are merged into main and no open PR depends on them. Leave protected branches such as main and fullrepo; report any ambiguous branch ownership instead of deleting silently.
11. Remove .serena/.flow_sync_marker and .serena/.flow_post_task_state.json after successful sync.

Loop Guard

Do not edit runtime marker files except to remove them after sync. If the Stop hook repeats for the same fingerprint (stop_hook_active=true + same SYNC_MARKER content), allow stop instead of forcing new commits.

Fullrepo Branch

fullrepo is the portable AI-context branch. It contains the normal branch tree plus agent-only files such as project AGENTS.md, .claude/CLAUDE.md, REVIEW.md, .serena knowledge, .claude, .cursor/rules, .agents/skills, and similar agent workflow files. The main branch should not track those files in normal projects; they should be restored locally from fullrepo and ignored through .git/info/exclude.

Anti-patterns

• Duplicate Serena memory sync (Serena owns its own freshness; this skill waits for it).
• Force-push main (only --force-with-lease on fullrepo).
• Commit secrets / runtime markers / browser artifacts / accidental junk.
• Auto-edit runtime marker files except to remove them after success.
• Delete branches without verifying merged state (use branch_cleanup_state from flow_post_task_state.py).