Talend ANS Agent Skill

Use this skill when modifying the Talend Alert Noise Suppression source code in this repository.

First Read

1. Read AGENTS.md.
2. Read function_app.py.
3. Check whether alert_recipients.json or any edited file contains unresolved merge markers.

Code Map

• Settings: all environment-driven runtime configuration.
• TalendClient: Talend API calls.
• SqliteAlertStore and AzureSqlAlertStore: duplicate alert persistence contracts; keep schemas aligned.
• EmailNotifier: daily digest SMTP delivery.
• OpenAIErrorSummarizer: optional JSON-mode summary fallback.
• Decision helpers: retry/noise classification, rule-based summaries, master-job dependency parsing, plan dependency parsing.
• Timer triggers: poll_talend_alerts and send_daily_digest.

Endpoint Flow To Preserve

Shared polling starts with GET /processing/executables/tasks/executions. Read response items, group by taskId, and keep the latest run using finishTimestamp, startTimestamp, then triggerTimestamp.

For every latest failed execution, call GET /monitoring/observability/executions/{executionId}/component.

For MANUAL task failures:

1. Use taskId, executionId, executionStatus, executionType, errorMessage, and component metrics to classify the failure.
2. Retry only transient network/engine/platform failures with POST /processing/executions and payload {"executable": "<taskId>", "logLevel": "WARN"}.
3. Treat data, component, permission, syntax, and ambiguous failures as valid alerts.

For MANUAL Master Jobs:

1. Detect child jobs from component metrics where connector_type == "tRunJob".
2. Order children by numeric suffix in connector_id, such as tRunJob_1 before tRunJob_2.
3. Use the first tRunJob with a stack trace as the failed child.
4. Report later tRunJob components as downstream jobs that should not be expected to run.

For PLAN failures, enrich context in this exact order:

1. GET /monitoring/observability/executions/{executionId}/component
2. GET /processing/executables/plans/executions, matching by planId to find planExecutionId
3. GET /processing/executions/plans/{planExecutionId}/steps, extracting failed step IDs/status
4. GET /orchestration/executables/plans/{planId}, extracting plan name, plan executable, step names, task names, chart, nextStep, and flows

For retryable plan failures, call POST /processing/executions/plans with payload {"executable": "<plan executable>"} where the executable comes from /orchestration/executables/plans/{planId}.

Change Checklist

• Keep new settings in Settings with safe defaults.
• Keep Talend parsing tolerant of missing keys and alternate sample response shapes.
• Update both storage backends when alert row fields change.
• Classify unknown failures as valid failures.
• Retry only transient platform/network failures.
• Preserve digest filtering through include_in_digest.
• Do not expose secrets from local.settings.json.
• Run python -m compileall function_app.py after Python changes.

Common Extension Points

• New transient failure: add a precise pattern to TRANSIENT_PATTERNS.
• New non-retryable failure: add a precise pattern to NON_RETRIABLE_PATTERNS.
• New human-readable summary: update summarize_error_rule_based.
• New Talend component relationship: extend parse_master_job_dependency or add a similarly tolerant helper.
• New plan response shape: adjust flatten_plan_steps or enrich_plan_context without breaking the linked chart.nextStep shape.
• New alert column: update SQLite DDL/migration, SQL Server DDL/migration, insert field lists, digest HTML if user-facing, and any test/sample data.

Verification

Minimum syntax check:

python -m compileall function_app.py

For runtime verification, install dependencies and start the Functions host:

.\.venv\Scripts\python -m pip install -r requirements.txt
func host start

Use real Talend/OpenAI/SMTP calls only when the required environment variables are present and the user expects networked verification.