Can Bilmez

Reliability · Testing · Open Source · Execution Integrity

I work on systems where “valid-looking” is not the same as “safe to execute.”

AtomicBot-ai/atomic-agent#144 prevents truncated or malformed native tool-call arguments from silently becoming executable input.

The final patch covers:

• malformed non-empty function.arguments → parse failure, never silent {} fallback;
• bare EOF with pending tool calls → fail closed unless a real terminal signal was observed;
• Qwen tagged calls → same termination-safety decision as native calls;
• final SSE events without a trailing blank line → flushed and parsed correctly at EOF;
• parallel tool calls and stream-read failures → zero dispatch;
• zero-argument calls and clean provider termination → preserved.

The maintainer re-ran the original probes plus additional EOF/UTF-8/abort cases before merging the patch to main.

#19063 report → #19066 v7 → #19120 v6 → #19121 v5 → ai@7.0.70

A tool call should not execute just because its arguments parse. The terminal model state is part of the execution contract.

Verified contribution ledger →

Fail-closed execution integrity for streamed LLM tool calls.

0.4.2 is published on npm with provider terminal-state handling, incremental argument evidence, schema validation, identity correlation, one-shot execution decisions, and AI SDK execution guards.

npm install prefix-safe-json@0.4.2

The release path is independently auditable: the published npm tarball is reproducible byte-for-byte from the tagged source, SLSA provenance is verified against the exact package/version bundle npm authenticated, and the verifier fails closed when release identity cannot be established.

AI SDK v5/v6/v7 · streaming JSON · tool calling · schema validation · reproducible release · fail closed

Executable compliance testing for AI coding instructions. Turns CLAUDE.md, AGENTS.md, Cursor rules, and similar repository instructions into sandboxed scenarios with scored reports, CI integration, SARIF output, provider comparison, and regression-oriented evidence.

AI agents · developer tooling · sandbox testing · CI · SARIF

PR-scoped mutation testing for JavaScript / TypeScript. Uses StrykerJS to focus mutation testing on changed lines and turn surviving mutants into actionable review signals instead of broad, expensive mutation runs.

mutation testing · StrykerJS · TypeScript · GitHub Actions · CI quality gates

• Node.js — nodejs/node#64954: recursive readdir with Buffer encoding across callback, sync, promises, and withFileTypes paths.
• Trendyol Baklava — Trendyol/baklava#1220: fix a resize-listener reference leak in bl-pagination and prove cleanup with a regression test.
• Vercel AI SDK docs — vercel/ai#18770: distinguish truncation from malformed JSON before jsonrepair in the cookbook flow.
• Sandbase Harness — sandbaseai/sandbase-harness#73: open prefix-safe-json@0.4.2 integration pilot for confirmation-required tool execution; not an adoption claim unless merged.

• Reproduce against current upstream before proposing a fix.
• Prefer regression tests that fail on the old behavior.
• Test the real execution boundary when side effects are involved.
• Compare patch failures against a clean baseline before calling them regressions.
• Keep security and reliability claims scoped to what the evidence proves.
• Fail closed where ambiguous state can trigger an irreversible action.

Backend systems · AI agent/tool execution safety · developer infrastructure · testing systems · OSS reliability

looks valid ───────────────► safe to execute
not the same thing