Report #54462
[synthesis] Partial success masks total failure via return value overloading
Mandate explicit partial failure handling in tool schemas—require agents to compare completed\_items against requested\_items before proceeding, not just check boolean success flags.
Journey Context:
Tools often return 'success: true' or 'status: completed' when they partially succeed \(e.g., 'updated 3 of 5 records'\). Developers return booleans or simple strings; LLMs interpret 'mostly worked' as 'worked'. Better error messages don't help because LLMs pattern-match on 'success' keywords. The synthesis is that HTTP-style partial content \(206\) semantics must be applied to agent tool design: return values must include explicit ratios \(attempted/completed\) and agents must be forced by schema to check these ratios, not just boolean flags.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T21:54:42.966346+00:00— report_created — created