Schema Sync Fully Guided

Schema Sync — Fully Guided

Read dev_tools/README.md and dev_tools/skill_examples/03_fully_guided/references/schema-layers.md before starting. They contain field-level detail about what each layer holds and what to change.

One-command fix

If the migration matches the standard merged → merged_at rename pattern, the full procedure is automated:

# Preview what will change (nothing is written)
uv run python dev_tools/schema_sync.py --table github_events --dry-run

# Apply all four patches
uv run python dev_tools/schema_sync.py --table github_events

# Confirm no drift
uv run python scripts/validate_schema.py

A SchemaSyncReport is printed listing every change made across all four layers.

Step-by-step procedure (for non-standard migrations)

Use this when the one-command fix does not cover the migration, or when you need to apply patches selectively.

Pre-check — confirm drift exists

uv run python scripts/validate_schema.py

If "No drift" is reported, the system is already in sync. Stop here.

Step 1 — Introspect live ClickHouse schema

uv run python dev_tools/scripts/clickhouse_introspect.py --table github_events

Note every column that is new, removed, or type-changed compared to the YAML contract. See references/schema-layers.md for the expected post-migration schema.

Step 2 — Patch YAML contract

# Preview
uv run python dev_tools/scripts/yaml_patch.py --table github_events --dry-run

# Apply
uv run python dev_tools/scripts/yaml_patch.py --table github_events

If patching manually: edit agentic_system/schema/github_events.yaml.

• Add the new column entry (name, type, description).
• Update the description of any renamed or type-changed column to document post-migration semantics.
• Do NOT remove columns that still exist in the DB (even if zeroed out).

Step 3 — Patch ChromaDB RAG chunks

# Preview
uv run python dev_tools/scripts/chroma_patch.py --dry-run

# Apply
uv run python dev_tools/scripts/chroma_patch.py

If patching manually: use the ChromaDB HTTP API.

• GET /api/v2/tenants/default_tenant/databases/default_database/collections — list collections.
• For each collection, POST .../get with {"include": ["documents", "metadatas"]}.
• For items where metadata.stale == True, apply text replacements to the document.
• Re-embed with the 64-dimension deterministic hash (see references/schema-layers.md#embedding).
• POST .../upsert with updated ids, documents, embeddings, and metadatas.

Step 4 — Patch agent prompts

# Preview
uv run python dev_tools/scripts/prompt_patch.py --dry-run

# Apply
uv run python dev_tools/scripts/prompt_patch.py

If patching manually: edit the two system prompt files.

• agentic_system/agents_core/nl2sql/prompts/system.md
• agentic_system/agents_core/rag/prompts/system.md

Replace every occurrence of:

• Old field name → new field name (in the schema table, type annotations, prose)
• Old SQL predicate → new predicate (e.g. merged = 1 → merged_at IS NOT NULL)
• State labels (pre-migration → post-migration)

See references/schema-layers.md#prompt-replacements for the complete replacement list.

Post-check — confirm drift resolved

uv run python scripts/validate_schema.py
# Expected: No drift — live schema matches YAML contract exactly.

# Confirm prompts no longer reference old predicate
grep -r "merged = 1" agentic_system/agents_core/
# Expected: no output

Rollback

# Restore YAML and prompts from the pre-sync snapshot
uv run python dev_tools/schema_sync.py --table github_events --rollback

# Restore ChromaDB chunks from source files
make seed-vectors

The rollback snapshot is saved to dev_tools/.schema_sync_rollback.json by the one-command fix. If you used step-by-step patching, save file snapshots before making changes.

References

• references/schema-layers.md — field mapping, embedding algorithm, prompt replacement list
• dev_tools/README.md — full file map and quick-start
• scripts/validate_schema.py — drift detection tool