DBOPS Changeset Generation

Generate Liquibase database changesets from natural language descriptions, anchored to a real schema and instance from the Harness DBOPS service.

When to Use

When the user asks to:

• Create, add, modify, or drop tables, columns, indexes, constraints, or views as Liquibase changeset operations
• Generate a database migration or changeset
• Fix a failed migration or update an existing changeset
• Apply schema changes as Liquibase YAML

Prerequisites

The Skill tool loads this file (SKILL.md) automatically — do not Read it again.

Before generating Liquibase YAML, Read reference.md using the exact path listed in the Skill tool result. It contains all Liquibase YAML rules, change types, rollback patterns, batch migration guidelines, MongoDB JSON string requirements, and the clarifying questions policy. Apply those rules in Step 4 when generating the changeset.

Full Workflow

Step 0: Check for Pre-Seeded Context (ALWAYS do this first)

At conversation start, the initial message may contain a JSON block with pre-selected identifiers injected by the Harness frontend UI:

{
  "dbops_schema_id": "users-schema",
  "dbops_instance_id": "prod-instance"
}

Read this JSON block (if present) before doing any tool calls.

Step 1: Discover Database Schemas (skip if dbops_schema_id is in context)

Call harness_list for the available schemas:

harness_list(resource_type="database_schema")

The response is a plain JSON array of objects. Each object has:

• identifier — used as the schema ID in subsequent calls
• name — human-readable name
• migrationType — "Liquibase" or "Flyway" (only Liquibase supported)
• instanceCount — number of linked instances
• primaryDbInstanceId — the designated instance for LLM authoring (may be empty). If non-empty, Step 2-A uses this directly as the instance_id without further discovery.

Disambiguation:

• If the array has one schema, select it automatically.
• If the array has multiple schemas, use AskUserQuestion (selection mode): Show: [<name> (<identifier>) — <migrationType>, <instanceCount> instances] for each.
• If the array is empty: "No database schemas found in this project. Create one in Harness DBOPS first."
• If migrationType is "Flyway": "This skill supports Liquibase schemas only."

Step 2-A: Resolve Instance from Schema (when only dbops_schema_id is known)

When you have a schema_id but no instance_id (from context OR after Step 1 selection):

1. Call harness_get to fetch the full schema:

harness_get(resource_type="database_schema", dbschema_id="<schema_id>")

2. Check the primaryDbInstanceId field in the response:

• If set: use it as the instance_id. Inform the user: "Using primary instance <primaryDbInstanceId> for schema <name>." Then fetch the instance to obtain the connector field needed in Step 3:

harness_get(resource_type="database_instance", resource_id="<primaryDbInstanceId>",
            dbschema_id="<schema_id>")

Read connector from the response.

• If empty or missing: fall through to Step 2-B.

Step 2-B: Auto-select Best Instance (when primaryDbInstanceId is empty)

Use the database_default_authoring_instance resource to let the server pick the best instance (oldest with a metadata snapshot; falls back to oldest overall):

harness_get(resource_type="database_default_authoring_instance", dbschema_id="<schema_identifier>")

The response is { identifier: string, name: string, hasSnapshot: boolean, connector: string }.

Handling the response:

• If the tool call returns results: Use identifier as the instance_id and connector as the connector identifier for Step 3 (no separate instance detail fetch needed). Inform the user: "Using instance <name> (<identifier>) for schema <schema_name>." If hasSnapshot is true, the instance has metadata available for snapshot lookups — proceed to Step 3-B for any changeset intent (create, modify, or drop). If false, skip Step 3-B entirely.
• Not found (404 OR empty result): Reply: "No instances found for schema <schema_name>. Please add an instance in Harness DBOPS first — open Harness → DBOPS → Schemas → <schema_name> → Add Instance." — STOP. Do NOT prompt the user for a connector identifier. Do NOT proceed to Step 3. Do NOT generate a changeset.

Step 3: Resolve Database Engine Type from Connector

Invariant. By this point you MUST have a non-empty connector value from Step 0 (fast-path instance fetch), Step 2-A, or Step 2-B — whichever applied. If you do not — for any reason — STOP and surface the Step 2 "instance not found" message to the user. Never ask the user to provide a connector identifier directly; connectors are always derived from an existing DBOPS instance.

Use the connector field obtained in Step 0, Step 2-A, or Step 2-B (whichever ran). Fetch the connector to determine the engine type:

harness_get(resource_type="connector", resource_id="<connector>",
            org_id="<org_id>", project_id="<project_id>")

The connector is always a JDBC connector. Read connector.spec.url and map the prefix:

If the URL prefix is unrecognised, proceed with generic ANSI SQL. Do NOT ask the user.

Step 3-B (Optional): Snapshot metadata check

Run this step for any changeset intent (create, modify, or drop) when hasSnapshot is true.

Purpose:

• For CREATE intents: verify the target object does not already exist (conflict check).
• For modify/drop intents: align names with the latest snapshot.
• For foreign key intents: resolve column mappings from snapshot metadata (Step 3-B-4).

3-B-1: List existing object names

Call with dbschema_id, dbinstance_id, and the relevant object_type ("Table", "Index", "View", etc.):

harness_list(
  resource_type="database_snapshot_object",
  filters={
    dbschema_id: "<schema_id>",
    dbinstance_id: "<instance_id>",
    object_type: "Table"
  }
)

Returns { data: string[] } — names of objects of that type in the current snapshot.

If the list API errors or returns an empty data array, treat the snapshot as stale — proceed to Step 4 without blocking the user (stale-snapshot tolerance).

3-B-2: Conflict check for CREATE intents

If the user's intent is to create an object (new table, index, view, etc.) and the target name appears in the data list returned above:

• Do NOT generate a CREATE changeset.
• Inform the user clearly:

  "The <object_type> <name> already exists in the current snapshot for instance <instance_name>. A CREATE changeset would fail. Did you mean to modify it instead? I can generate an ALTER TABLE (add column / modify column / add index / add constraint) changeset if you describe the change you need."

• Stop and wait for the user's response. Resume from Step 4 once the user clarifies.

If the name is not in the list (expected case for a genuine new object), proceed to Step 4.

3-B-3: Fetch full metadata for modify/drop intents

When modifying or dropping an existing object, fetch its full metadata (columns, indexes, etc.) — requires non-empty object_names:

harness_get(
  resource_type="database_snapshot_object",
  resource_id="<instance_id>",
  params={
    dbschema_id: "<schema_id>",
    object_type: "Table",
    object_names: ["<table_name>"]
  }
)

If the target name is not in the snapshot list for a modify/drop intent, proceed anyway — the snapshot may be stale; do not block the user.

3-B-4: Resolve foreign key column mappings

Run when the intent involves adding or modifying a foreign key (addForeignKeyConstraint), including FK constraints added during createTable.

Fetch snapshot metadata in one harness_get call. Which tables to include depends on intent:

• addForeignKeyConstraint on an existing table: fetch both baseTableName and referencedTableName — needed for snapshot name alignment and column scoring on both sides.
• createTable with inline FK constraints: fetch only referencedTableName — the new base table is not in the snapshot yet; base columns come from the changeset definition.

harness_get(
  resource_type="database_snapshot_object",
  resource_id="<instance_id>",
  params={
    dbschema_id: "<schema_id>",
    object_type: "Table",
    object_names: ["<base_table>", "<referenced_table>"]
  }
)

For createTable with inline FK, use object_names: ["<referenced_table>"] only.

Apply column-resolution rules in reference.md → Foreign Key Column Resolution → Column resolution. If snapshot metadata is unavailable, proceed using user-provided names only.

Canonical DBOPS resource_type values (see dbops toolset): database_schema, database_instance, database_default_authoring_instance, database_snapshot_object, database_execute_llm_authoring_pipeline.

Step 4: Generate the Changeset

Produce a valid Liquibase YAML changeset. All YAML generation rules (structure, required fields, rollback, preconditions, timestamps, batch migrations, MongoDB formats, clarifying questions) are in **reference.md** in this directory — follow them exactly.

Changeset author: Derive the author name from .harness_context/user_context.json (if present) or the ## User Context: block in the conversation. Apply these rules in order:

1. Check for username: Look for user.display_name, uid, or username in the user context.
2. Normalize the username: Apply these transformations in order:

• Convert to lowercase
• Remove extra spaces
• Convert internal spaces into - (hyphen)
• Remove all special characters (keep only a-z, 0-9, and -)

3. Fallback to email: If no username is available, extract from the email field:

• Take the part before the @ symbol (e.g., john.doe@harness.io → johndoe)
• Apply the same normalization rules above

4. Final fallback: If neither username nor email is available, use harness-ai.

Examples:

Multi-turn accumulation (follow-up prompts): If the conversation already contains a previously generated databaseChangeLog YAML (from any prior turn, whether accepted, denied, or still pending review), that YAML is the running baseline. On every follow-up turn you MUST produce the full accumulated changelog — not just the delta for this turn.

Classify the user's intent and merge accordingly:

Always present the complete, merged databaseChangeLog (all accumulated changeSet entries) for review — never only the new delta.

Update mode (refining a single changeSet): If the user is explicitly refining or correcting one specific changeSet from the baseline (e.g., "change the column type in the dog table changeset"), apply the modification to that changeSet block in-place. Preserve the id and author from that changeSet and apply the requested modifications. Still present the full accumulated changelog.

Error recovery mode: See Error Recovery Mode section below.

Step 5: Present for Review

The present_for_review tool is not available. Present the change for review in chat:

1. Show the full generated Liquibase YAML in a fenced yaml code block.
2. Clearly ask the user to choose one of these three options:
   • Accept — save the changeset only (do not execute a pipeline)
   • Accept & Commit — save and execute the DBOPS authoring pipeline
   • Deny — discard / request changes and regenerate

Example prompt after the YAML:

Please review the changeset above and reply with one of: Accept, Accept & Commit, or Deny.

CRITICAL:

• Always present all three options: Accept, Accept & Commit, Deny.
• Always present the complete, merged databaseChangeLog — never only the new delta.
• Wait for the user's choice before proceeding to Step 6.
• If the user edits the YAML before choosing, always use their edited YAML, not your original generation.

Step 6: Handle the User's Action

"accept" — Save only

Confirm the changeset is saved. Do NOT execute any pipeline. Reply with a short confirmation message referencing the schema and instance names.

"deny" — Discard

Ask what they want to change and regenerate from Step 4. DO NOT re-run the schema/instance discovery — the schema and instance are already selected.

"accept_commit" — Save + Execute Pipeline

After the user accepts:

IMPORTANT — use already-resolved identifiers: The schema_id, instance_id, and the instance's connector were resolved in Steps 1–2 and are still in context. Do NOT re-list schemas, re-fetch instances, or list connectors — use the values you already have.

IMPORTANT — changeset encoding: Send the changeset field as PLAIN-TEXT YAML, exactly as the user accepted it. Do NOT base64-encode it..

IMPORTANT — conversation_id (REQUIRED before harness_execute): Always pass conversation_id explicitly when calling harness_execute for the database_execute_llm_authoring_pipeline resource. Resolve it from agent context in this order:

1. conversation_id — Harness chat / platform-injected conversation ID (preferred)
2. session_id — use the session ID as the conversation_id value when present in agent or system context
3. Cursor chat / agent transcript ID — when running in Cursor and neither of the above is injected, resolve the active chat UUID and use it as conversation_id without asking the user:
   • Prefer any Cursor-injected session field if present (e.g. session_id, composerId, chatId).
   • Otherwise locate this chat’s agent transcript under the project’s agent-transcripts/<uuid>/ directory (typically ~/.cursor/projects/<project-slug>/agent-transcripts/<uuid>/<uuid>.jsonl). Match the transcript that contains the current conversation’s recent user query (or the most recently updated transcript for this chat), and use that <uuid> as conversation_id.
   • Do not invent a random UUID. The value must come from Cursor’s chat/session identity or transcript metadata for this conversation.

It is a hard requirement before making the execute-pipeline tool call. Prefer platform-injected IDs when both a Harness conversation_id and a Cursor session ID exist. Never fabricate a UUID from scratch — conversation_id is the billing / idempotency key. Only if steps 1–3 all fail, STOP — do not call harness_execute. Tell the user pipeline execution cannot proceed without a conversation/session ID.

IMPORTANT — chat runner gate protocol. The user already consented via Accept & Commit, but the chat runner may render a separate permission prompt before forwarding harness_execute. Handle it deterministically:

Branch-lock. Once Step A picks custom vs default, the decision is locked until harness_execute returns a result or the user cancels — never re-evaluate it after a gate event, even if the NG setting could have changed.

Args lost to context compaction. If a user confirmed event arrives but your context no longer holds the original harness_execute args, reply: "Approval received but I need to regenerate. Can you re-confirm Accept & Commit on the changeset?" — do NOT speculatively reconstruct args.

Post the success message ONLY after harness_execute returns a non-empty executionId.

Step A — Read the NG setting to pick the branch

harness_list(resource_type="setting", filters={ category: "DBOPS", scope: "project" }, compact=false)

Find the entry with identifier == "dbops_llm_authoring_pipeline". Read its value field.

• If value is a non-empty string → custom-pipeline branch (continue to Step B).
• If value is empty or the setting is absent → default-pipeline branch (skip to Step E).

Step B — Fetch the pipeline's runtime-input template (custom-pipeline branch only)

harness_get(
  resource_type="runtime_input_template",
  resource_id="<value from Step A>",
  account_id="<account>", org_id="<org>", project_id="<project>"
)

The response's inputSetTemplateYaml lists every <+input> token the admin's pipeline declares. The skill is responsible for these three reserved tokens (the server expects these EXACT YAML keys — see the dbops database_execute_llm_authoring_pipeline resource schema):

• dbSchema — auto-filled with schema_id.
• dbInstance — auto-filled with instance_id.
• changeset — auto-filled with the PLAIN-TEXT changeset YAML.

The pipeline's stepGroupInfra.spec.connectorRef is also a <+input> token but is filled server-side from the DBOPS instance — the skill never sends a value, never asks the user for it.

If a template field's name does not match any of the canonical names or aliases above, treat it as non-reserved and elicit it in Step C. Do NOT silently auto-fill it from a similarly-named context value (e.g. a template field called schema or db_schema is non-reserved — ask the user).

Step C — Diff and elicit the remaining required inputs (custom-pipeline branch only)

Server-side note (current state): the db-devops-service execute endpoint currently rejects any non-empty runtime_inputs map with the error "custom runtimeInputs are not yet supported (pipeline-service contract pending)". Until that contract lands, if the template declares any required field outside the reserved set, stop and tell the user the configured pipeline is incompatible with chat execution. Skip the elicitation loop below entirely. Once the server lands custom-input support, remove this note and follow the loop as documented.

For every non-reserved template field, ask the user for a value via AskUserQuestion — treat all such fields as required (we cannot reliably infer optionality from the template).

• Use the raw field name as the prompt label.
• One question per field — AskUserQuestion is single-question by design; do not batch.
• Do not attempt to parse default values from the template; always elicit fresh.

Collect responses into a runtime_inputs map.

Cancellation rule. Stop and do NOT call Step D if any of these hold:

• AskUserQuestion returns cancelled (no answer, user dismissed the prompt).
• The response is empty or whitespace-only.
• The response matches cancel, skip, stop, or abort (case-insensitive whole word).

Any other freeform text is a real answer — copy verbatim into the runtime_inputs map. Do not proceed to Step D with a partial map.

Step D — Trigger the pipeline (custom-pipeline branch)

Call the consolidated dbops resource (single tool call — server triggers execution and records the billing event atomically):

harness_execute(
  resource_type="database_execute_llm_authoring_pipeline",
  action="run",
  schema_id="<schema_id>",
  instance_id="<instance_id>",
  conversation_id="<conversation_id or session_id from agent context>",
  changeset="<plain-text accepted YAML>",
  pipeline_identifier="<value from Step A>",
  runtime_inputs={ <field>: <value>, ... }   # OMIT this key entirely when no fields were elicited; do NOT pass {}
)

If the call returns an error or times out, surface the exact error message verbatim and stop. Do NOT retry automatically and do NOT fall back to the default-pipeline branch. Do NOT fabricate navigation paths, execution status claims, or retry invitations. The openInHarness link MUST only be shown when the response provides it — never construct it manually.

Skip to Step F.

Step E — Trigger the pipeline (default-pipeline branch)

harness_execute(
  resource_type="database_execute_llm_authoring_pipeline",
  action="run",
  schema_id="<schema_id>",
  instance_id="<instance_id>",
  conversation_id="<conversation_id or session_id from agent context>",
  changeset="<plain-text accepted YAML>",
  use_default_pipeline=true
)

The server performs get-or-create of dbops_default_pipeline on first use, then triggers the execution. The server resolves the K8s connector for the step-group infra from the DBOPS instance's connector — the skill sends no infra hints.

If the call returns an error or times out, surface the exact error message verbatim and stop. Do NOT fabricate navigation paths, execution status claims, or retry invitations. The openInHarness link MUST only be shown when the response provides it — never construct it manually.

Proceed to Step F.

Step F — Show the live execution link

Both branches return:

{ executionId, pipelineIdentifier, openInHarness }

Post immediately — only if openInHarness is a non-empty string in the response:

Pipeline started — <a href="{openInHarness}" target="_blank">view execution</a>

If openInHarness is absent or empty, post instead:

Pipeline started — execution ID: `<executionId>`.

Do NOT construct the URL manually. The link must come from the response.

Do NOT proactively poll. When the user asks for status, call harness_get(resource_type="execution", execution_id="<executionId>") once — this hits the same pipeline-service execution endpoint that powers the openInHarness link.

Status triage:

• Running / Queued / Paused / Waiting — report: "Running — step <stepName> is <stepStatus>".
• Success — "Changeset applied successfully" — if openInHarness is present, append — <a href="{openInHarness}" target="_blank">view execution</a>. Do NOT construct the URL manually.
• Failed / Aborted / Expired — read the failed step name and failureInfo.responseMessages[0].message.
  • If the failed step name contains changeset/liquibase/migration OR the message contains those keywords or checksum / precondition / validation error — this is a changeset error. Offer error recovery: enter error-recovery mode (see Error Recovery Mode section below) and regenerate a corrected changeset.
  • For any other step failure (infra, connector, K8s, networking, permissions) — report the error verbatim and stop. Do NOT auto-debug. Reply: "Pipeline failed at step <stepName>: <message> — view execution. This is an infrastructure or configuration issue — please check your Harness pipeline setup."

Important Rules

• Present the changeset for review in chat and always offer all three options: Accept, Accept & Commit, Deny.
• Always include the full YAML in the review message.
• Use the YAML the user accepted/edited — not your original generation if they changed it.
• Before any harness_execute for database_execute_llm_authoring_pipeline, resolve conversation_id from agent context (conversation_id, else session_id, else Cursor chat/transcript UUID). Do not execute without it, and do not ask the user for a Cursor chat ID when the transcript UUID can be resolved locally.
• For accept_commit, post the openInHarness link (if present) and stop. Do NOT proactively poll — the server reconciles status asynchronously. If the user later asks for status, call harness_get once as described in Step F.
• For deny, skip re-discovery — schema and instance are already known.
• For error recovery, explain the fix BEFORE presenting the corrected YAML.
• If the DBOPS service returns an error during schema/instance discovery, report the error clearly and suggest the user check their Harness DBOPS project setup.

Error Recovery Mode

Trigger: Enter this mode ONLY for changeset/Liquibase failures:

• A FAILED pipeline status event resumes the conversation (automated polling path), OR
• The user asks to fix a failed migration or reports a Liquibase / checksum / precondition / OPA error

Do NOT enter this mode for infrastructure or configuration failures (broken connector, K8s scheduling failure, missing secret, pipeline misconfiguration, etc.). For those, report the error verbatim and reply: "Pipeline failed at step <stepName>: <message> — this is an infrastructure or configuration issue — please check your Harness pipeline setup." Do NOT attempt to regenerate the changeset.

Do NOT re-run Steps 1–2 (schema/instance discovery). schema_id, instance_id, and connector are already resolved from the original session. Jump directly to the steps below.

Steps

1. Extract the failure details from context:

• From the system event (automated path): failed_step, failure_message, execution_id, open_in_harness link.
• From the user message (manual path): the error text the user pasted or described. If the system event provides an execution_id but no failure_message, call:

harness_diagnose(resource_type="execution", execution_id="<execution_id>")

Read the full step log from the response to get the Liquibase error text before proceeding to root-cause analysis.

1a. Classify the failure before proceeding:

2. Identify the root cause. Read reference.md → "Author, Update Mode, and Error Recovery" section. Common causes:

• Missing rollback block
• Illegal or reserved table/column name for the target engine
• Wrong data type for the target engine (e.g., DATETIME2 vs TIMESTAMP)
• Duplicate changeSet id within the changelog
• Invalid precondition syntax or missing onFail
• OPA policy violation — extract the policy's error text verbatim, correct the changeset to satisfy the policy

3. Patch the offending changeSet:

• Fix only what the error requires. Preserve id, author, and all unaffected changeSets.
• Produce the full accumulated databaseChangeLog (never emit just the delta).

4. Explain the fix in a text message BEFORE presenting the corrected YAML: State: what the error was, which changeSet was affected, and what was changed to fix it.

5. Present the corrected YAML for review (Step 5) with all three options: Accept, Accept & Commit, Deny.