Skip to content

Credential, concurrency and recovery experiments

These are first-party controlled experiments against pinned upstream source, not independent QA or whole-product tests. The public package includes the original probe harnesses, two saved rounds, immutable source URLs and SHA-256 identities.

SuiteObservationBoundary
OpenHandsClaude file callbacks 4→0 and child credential-blob variables 0→4; Codex 4→1 and 0→3Original extracted methods and startup slice; fixture provider identity and file callbacks; real Python child with synthetic values; no ACP handshake
AG2Two-loop local peaks 2/1→1/1, aggregate 3→2Original admission method; real threads/loops and waiting synthetic subtasks; five trials per configuration per round
OrcaDivergent-index requests: 200→1 at 10ms and 10→3 at 16sComplete recovery/lookup source modules with store/time/PTY fixtures; no Electron or production crash reproduction
PaperclipKill request returns before observed exitOne real harmless Node child; no descendant or tool-effect claim

Public package · Pinned source identities · Detailed method and controls in Chinese

Reproduction

Fetch this package, run node fetch-pinned.mjs, then run the relevant probe with a fresh output number:

text
python probe-ag2.py 3
python probe-openhands.py 3
node probe-orca.mjs 3
node probe-paperclip.mjs 3

The measured interpreters were Python 3.10 and Node 24.16.0. The fetcher verifies the exact source hashes before execution. Upstream source is downloaded from its original repository, not redistributed here. Matplotlib and Pillow are required only for chart generation.

Saved AG2, OpenHands, Orca and Paperclip observations each have a matching second-round file. Do not replace historical rounds while reproducing. Fixed versions and constructed conditions define the supported scope; a current PR may subsequently change.

Last updated: