version: "1.0.0" name: debug-failed-job description: Find out why a Keboola job failed and propose a fix. Use when the user says a Keboola job, extractor, transformation, writer or flow failed or errored, or asks why their pipeline is red.

Debug a failed job

Find the failed job, read its error and logs, explain the cause, and fix it only after the user agrees.

Input

A job ID, a configuration or flow name, or just "my last job failed". Ask for nothing more until step 4 shows whether it is needed.

Steps

1. Pick the project. Call get_accessible_projects (with with_llm_instruction=true on the first call in the

conversation) and ask the user which project the job ran in. Do not guess. Do this even when the session looks scoped already or a data tool would work without it: this plugin always narrows the session to the projects the user confirms. If get_accessible_projects fails because no programmatic token is present, the login is bound to one project: confirm it with the user and skip step 2.

1. Scope the session to that project. Call set_project_scope with project_ids=[<the confirmed id>]. This

must happen before any write in step 7. If the result carries a non-null scope_token, pass it on every later tool call. Every write in step 7 also passes the confirmed id as project_id: other conversations on the same login share the scope, and a scope they moved makes the write fail instead of landing in another project.

1. Find the job.

• Known job ID: go to step 4.
• Known configuration or flow name: find its ID with search (`item_types=["configuration", "transformation",

"flow"]), then call get_jobs with job_ids=[], status="error", component_id=... and config_id=...`.

• Nothing known: call get_jobs with job_ids=[], status="error" and limit=10, and ask the user which job

they mean if more than one is recent.

1. Read the failure. Call get_jobs with job_ids=["<id>"], include_logs=true and

log_event_types=["error", "warn"]. If the error event is not enough, call again with a larger log_tail_lines=... and no event type filter.

1. Read what ran. Call get_configs with configs=[...] for the job's configuration. For a flow job, call

get_flows with flow_ids=[...] and find the failed task, then repeat steps 3 and 4 for that task's job.

1. Explain. Name the cause in one or two sentences and quote the error line. Typical causes: an SQL error in

a transformation (check columns with get_tables and reproduce the failing SELECT with query_data), a missing input table, expired credentials of an extractor or writer, or a schema change in the output table. Use docs_query for component-specific errors. Propose a fix.

1. Fix, only on request. Ask before changing anything.

• SQL transformation: update_sql_transformation with project_id=<the confirmed id>, its

configuration_id=..., the parameter_updates=[...] and a change_description=....

• Flow: modify_flow with project_id=<the confirmed id>, its configuration_id=..., the flow_type=... it

already has and a change_description=....

• Another component: get_components for its schema. When the fix belongs in a configuration row, call

update_config_row with project_id=<the confirmed id>, component_id=..., configuration_id=..., configuration_row_id=... and a change_description=.... Otherwise call update_config with project_id=<the confirmed id>, component_id=..., configuration_id=... and a change_description=....

• Expired OAuth credentials need the user: create a link with create_oauth_url,

project_id=<the confirmed id>, component_id=... and config_id=..., and ask them to reauthorize.

Then, with the user's consent, call run_job with project_id=<the confirmed id>, component_id=... and configuration_id=..., and check the new job with get_jobs. If the server refuses a write because the user's role is read-only, give the user the proposed fix to pass to a project admin.

Rules

• No write before step 2, and no write at all without the user's go-ahead.
• Never rerun a job just to see whether it fails again; read the logs first.
• Report what the logs say. If they do not show a cause, say so and list what you checked.