SKILL: Idempotent Tool Result Persistence

Domain: persistence Trigger: Use this when tool result blocks exceed their configured thresholds and you want to persist them once while still returning a preview. Source Pattern: Distilled from reviewed tool execution, streaming, persistence, and output-budget implementations.

Core Method

Persist oversized tool outputs to a deterministic artifact path derived from the tool call ID and content type. Create that artifact with exclusive-write semantics so retries and concurrent writers cannot clobber or duplicate the same payload. After the first successful write, return a compact preview plus metadata pointing at the persisted artifact; later retries should reuse the existing file and regenerate the same preview without rewriting the payload. This keeps large outputs accessible while making persistence safe and repeatable.

Key Rules

• Create the parent directory once per storage root, but write each artifact with exclusive creation semantics.
• If the target artifact already exists, treat that as a successful prior write and move directly to preview generation.
• Keep preview metadata explicit: content type, truncation status, persisted path, and original size should all be recoverable downstream.
• Log the persisted location and size so operators can trace where a large result went.

Example Application

When a log-reading tool returns 60 KB of text, write it once to a stable artifact file such as tool-results/<tool-call-id>.txt, then send only a short preview plus the saved path. If the same result is replayed on retry, skip the write and reuse the same artifact reference.

Anti-Patterns (What NOT to do)

• Don’t re-write the same file on retries; failing to check EEXIST causes EOS errors and clobbers the previous log.
• Don’t drop the preview header; consumers depend on <persisted-output> tags to know the persisted location.