[Catalog](https://skills.polygres.com/index.md) / [polygres-troubleshooting](https://skills.polygres.com/polygres-troubleshooting/SKILL.md)

Published skills release: 0.9.0. [Compatibility and source](https://skills.polygres.com/manifest.json).

Scripts and templates mentioned here require the installed package; reading this page does not install them.

# MCP operation recovery

Read [`mcp-tool-contract.md`](https://skills.polygres.com/polygres-troubleshooting/references/mcp-tool-contract.md) first. Preserve the operation kind, operation ID,
project ID, request ID, last status, progress, error code, retryability, and
timestamp.

Use `get_operation` or the domain status tool to re-read current state. Treat
context, import, graph, and sync as distinct operation kinds. Use
`wait_for_operation` in bounded intervals and report the latest observed state
when its watch period ends.

For managed embeddings, use `get_embedding_configuration` to inspect progress
and `last_error_code`; generation does not supply a generic operation ID. A
linked Context operation retains its own ID and status. Follow
[embedding diagnostics](https://skills.polygres.com/polygres-troubleshooting/references/embeddings.md) for allowance, provider, and model
binding failures before proposing `process_embeddings` recovery.

Classify the failure as connection or OAuth, installation scope, current role
or project access, project mode or state, Central API, Runtime API, source
database, or durable operation execution. A tool missing from discovery can
indicate feature selection, read-only mode, scope, project compatibility, or
catalog version.

Recommend the smallest documented correction. When the user's authorization
covers recovery, route supported actions as follows:

- `retry_operation` for a retryable Context operation
- `cancel_operation` for an eligible Context or import operation
- `retry_synchronization` for eligible synchronization state
- an exact Context reconciliation, graph build, or import restart through its
  normal workflow

Each action uses its own server proposal and returned digest. An ambiguous
row-write response remains a write-outcome investigation, not a failed durable
operation. Reuse its exact idempotency key only after public evidence supports
replay.

For an uncertain query embedding outcome, retain the original input and key.
Inspect the reported state and await administrator-verified provider usage
reconciliation. A fresh key can cause another charge; a read timeout alone is
not evidence that a new query attempt is needed.