Fix Pipeline Dbt Task

Fix → validate on a branch → confirm → merge. Narrate each step briefly, like a data engineer. Never print secrets/tokens.

0. Access

MCP is preferred but optional. If the Orchestra MCP tools aren't present, use the public REST API directly with $ORCHESTRA_API_KEY (check the env — it's often already set). For git, $GITHUB_TOKEN is usually set too. If neither MCP nor a key is available, read ../../../references/orchestra/mcp/setup.md and stop.

Base: https://app.getorchestra.io/api/engine/public — header Authorization: Bearer $ORCHESTRA_API_KEY.

1. Inputs from identify-pipeline-error

identify-pipeline-error hands you: the pipeline run ID (RID), pipeline ID (PID), the failed dbt task run ID (TRID), integration=DBT_CORE, the run branch/commit, the task's taskParameters, and the cause classification (a dbt code/config issue). Use them directly — don't re-run the identification.

If invoked standalone (no handoff): run identify-pipeline-error first, or do its minimum yourself — GET /pipeline_runs?pipeline_run_ids= then GET /task_runs?pipeline_ids=&page_size=50 (filter to this RID client-side; the pipeline_run_ids filter does not work on /task_runs, 7-day window) to find the FAILED dbt task.

Capture taskParameters exactly — especially branch, commands, project_dir, use_state_orchestration. These are run inputs you must preserve on retry. The pipeline run's branch is where the Orchestra YAML lives; the dbt clone branch is a separate task param (commonly an input like dbt_branch).

Pull the logs you'll fix against:

GET /pipeline_runs//task_runs//logs                                  # list filenames
GET /pipeline_runs//task_runs//logs/download?filename=1/debug_task   # URL-encode spaces, e.g. 1/dbt%20run

debug_task = dbt debug (config/YAML validation); dbt run / dbt build = compile + execution.

2. Scope check — you're here to fix dbt code

You were routed here because the cause is a dbt code/config issue, so fix it in the repo. If while reproducing you find the real cause is not a code fix — a data-quality / source-freshness failure (don't edit data), an infra/timeout, an Orchestra-pipeline misconfig, or an upstream ingestion problem — stop and hand back to identify-pipeline-error rather than editing code. (Source-freshness ERROR STALE lines are a data issue, not code; they rarely fail a dbt run — don't chase them unless they're the terminal error.)

3. The reported error is usually just the FIRST bug

A failed dbt run reports one error; broken branches often hide several layered behind it. Don't fix-one-and-retry blindly — reproduce locally and clear them in one branch:

git clone --depth 1 -b  https://x-access-token:$GITHUB_TOKEN@github.com//.git
cd /
export DBT_PROFILES_DIR=$PWD            # write a dummy snowflake profiles.yml; parse never connects
pip install -q dbt-
dbt parse --no-version-check            # catches YAML + ref/source + compilation errors offline

dbt parse finds YAML and ref/source errors with no warehouse. Runtime SQL errors (e.g. invalid identifier) only surface on an actual dbt run — so expect to iterate steps 3–5.

Fixing strategy: typos (ref('customersdfdfd_clean')→customers_clean, site_nadfdfdme→site_name) fix in place. For damage beyond a typo (missing *, passthrough, dropped commas, mangled blocks), diff against a known-good branch (usually main, which builds green) and restore that file rather than guessing the intended logic.

4. Fix on a branch, validate before prod

1. New branch off the failed dbt branch; commit each fix with a clear message. Open a PR (git-backed pipelines can only be fixed via repo; check storageProvider). Never label a code fix "user action" — open the PR.
2. Validate without merging if the pipeline exposes branch/command inputs — trigger against the fix branch:

POST /pipelines//start
{ "branch": "",
  "continueDownstreamRun": true,
  "runInputs": { "dbt_branch": "", "dbt_command": "run" } }

If branch isn't an input, tell the user it can't be hotfixed until they parameterise it (add inputs.dbt_branch) — don't guess.

1. Run the whole pipeline, not just the dbt task. A real ETL flow is python → dbt → reverse-ETL/BI; dbt is rarely the last step. A full /start (no taskIds) runs the entire DAG top-to-bottom — prefer it. If you instead target the failed task ("taskIds": [""]) or use "retryFromFailed": true to skip an expensive upstream, you must also send "continueDownstreamRun": true, or every task after the fixed one is skipped and the pipeline only looks fixed. Targeted retries also assume the fixed task's upstream inputs already exist — if not, do a full run.
2. dbt run / build materialises tables. If the env is Production, get the user's OK first unless they've said to proceed. Then it's their call.

5. Poll, loop, confirm

Poll GET /pipeline_runs?pipeline_run_ids= ~every 20s, one line per check:

▶ Run e2049b86 — RUNNING (0:42)
✅ Run e2049b86 — SUCCEEDED (1:44)

A very-fast SUCCEEDED can be a stale first read — re-check completedAt/message before trusting it. On failure, immediately diagnose the new error (it's progress if the run got further) and loop to step 3.

Confirm the whole pipeline, not just the dbt task. "Fixed" means the pipeline run is SUCCEEDED with every task green — not just that the dbt task passed. Check the run status, then list its task runs (GET /task_runs?pipeline_ids= filtered to the RID) and confirm none are FAILED. A downstream task (reverse-ETL, BI refresh, notification) can fail on the dbt output even when dbt itself is fine. When green end-to-end, merge the PR. As a last check, re-run with the original failing params to prove the real config is fixed.

6. Summary

## Fixed: 
- Final run:  — SUCCEEDED
- Root cause: 
- Fixes: 
- PR: 

Notes

• Preserve task parameters on every retry.
• Validate end-to-end: full /start by default; for targeted/retryFromFailed retries always add "continueDownstreamRun": true, and confirm every task in the run is green.
• Confirm before retrying data-writing pipelines in Production unless told to proceed.
• Only Orchestra-backed pipelines use update_pipeline; GitHub-backed need repo commits.
• Batch list_* calls; mind the 7-day metadata window.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: orchestra-hq
• Source: orchestra-hq/orchestra-skills
• License: MIT

Install and usage instructions live in the source repository linked above.