From an FMP request
to a maintainable pipeline.

“Build batch pipelines for all symbols.”

That request starts a clarification process. “All symbols” needs a declared universe, dataset eligibility, and a feasible freshness plan for the user’s entitlements and API budget.

1. Agree on the scope.

   The agent asks which datasets, symbol universe, history, destination, freshness target, API limits, and alert destination matter. Credentials are supplied through the chosen secret mechanism, not embedded in project files.

2. Review a versioned project.

   The intended plugin guides the agent to supported pipeline configuration, data contracts, custom-code references, and runner configuration. Humans can inspect the same files. Proposed changes need review before deployment.

3. Verify a bounded load and rerun.

   Check a small declared dataset against the actual destination, including rejected records, publication rules, and a safe repeat. A passing record fixture alone cannot prove these properties.

4. Deploy recurring batch jobs.

   The first planned reference uses Prefect for scheduling and compute running PyCharter as a library. The project declares its deployment, secrets, durable state, and ownership. Closing the coding-agent session should leave the deployed jobs operational.

5. Monitor coverage independently.

   Compare expected partitions and freshness with committed destination data. An independent monitor must detect missed work even when no job ran; a green process exit is insufficient evidence of complete ingestion.

6. Review and verify maintenance.

   Use diagnostics, provenance, and progress to propose a repair. Review the configuration change, verify it, and perform a bounded rerun or backfill. Retries, checkpoint recovery, and lineage are advertised only after their runtime evidence passes.

Declare who operates each part.

In the default intended batch mode, you operate the scheduler, compute, destination database, durable state, and monitoring. PyCharter runs inside jobs that can exit between runs. An always-on PyCharter API, worker, or MCP server is not required by this mode; selected shared-service features may add their own requirements.

Agent guidance helps create and maintain the project. Authorization and database permissions belong to the declared runtime and infrastructure. The plugin alone cannot prevent alternate code or SQL from bypassing a check.