aidpf-error-triage

aidpf-error-triage - route failures to the right recovery path

This skill is the read-only front door for failures. It turns pasted errors, diagnostic artifacts, and failed-run summaries into a safe next step.

It does not fix anything itself. It identifies the failure class, names the evidence, and routes to the skill or CLI command that owns the recovery.

When to use

• User pastes output containing AIDPF-*.
• User gives .aidp/diagnostics/<run_id>/... or a run id.
• User says validate, bootstrap, seed, incremental, dashboard, OAC, dataset setup, or workbook authoring failed.
• User asks "what should I do with this error?"
• Autopilot hits an error and needs to route without guessing.

When NOT to use

• Clean first-run setup with no failure -> /aidp-fusion-autopilot.
• Known single action with no error, such as "run seed" -> use that skill.
• Authoring new marts -> /mart-author.
• Creating workbooks from a verified dataset -> /workbook-authoring.
• Full error-code reference browsing -> read docs/aidpf-error-codes.md.

Inputs to collect

Collect what is available, but do not block on everything:

• exact pasted error output,
• command that failed,
• active phase: validate, bootstrap, seed, incremental, dashboard, OAC dataset, workbook,
• run id,
• dataset/node id,
• layer,
• diagnostic artifact path or .aidp/diagnostics/<run_id>/ directory,
• whether this is dev/sandbox or production.

Never ask the user to paste passwords, OAuth tokens, private keys, full OCIDs, or full OAC connection payloads.

Workflow

1. Extract the failure signal

Search the pasted text and any artifact paths for codes:

AIDPF-\d{4}

If the user gives a diagnostics directory, list the files:

find .aidp/diagnostics/<run_id> -maxdepth 1 -type f -print

If the user gives a diagnostic JSON file, read it and extract:

• errorCode,
• runId,
• tenant,
• datasetId or node,
• layer,
• variation point name,
• missing columns,
• observed/live schema fields,
• companion logs such as cluster_stdout.log.