agent-safety

白箱AGI架构探索：元认知（自我认知循环）、持续学习（知识飞轮）、世界模型（条件空间+语义时空图）、自我改进（自举纪律）、零LLM白箱管线与可审计信任护栏。

Execution-time drift firewall for long-running DeepSeek Harness agents. Real-Harness tests: unsafe stale mutations 12/12 native -> 0/12; valid controls 7/7 both; post-SIGKILL unsafe continuation 2/2 -> 0/2.

Deterministic fail-closed tool-call authorization for DSH with evidence: allow/block/ask policy gate plus approval-chain deferral.

Batch Review Inbox for DeepSeek Harness — stage, inspect, edit, approve, and commit exact tool side effects safely.

Commitment write-gate for AI coding agents: two-tier pre-execution policy (deterministic guard + LLM judge) with measured receipts. DeepSeek Harness plugin, engine-agnostic core.

DeepSeek Harness file-safety guardrail that limits AI coding-agent file edits per turn and prevents too many file changes, with call and UTF-8 byte budgets. 限制 DSH Agent 每轮修改文件数量。

Context firewall for DeepSeek Harness: credentials in tool results become opaque placeholders before they reach the model — the secret never enters context.

DSH plugin spec: observing soft-signal drift

Nine deterministic hooks that block a coding agent from weakening tests, swallowing errors, or stubbing type checks — each backed by a planted-failure test proving the guard can actually fire. Runs on Claude Code and DeepSeek Harness.

Unofficial alpha execution-fidelity guard bundle for DeepSeek Harness.

Pre-flight guard for git working copies: is the checkout you are about to write to the one you think it is? Branch identity, drift from the remote and how stale that figure is, the identity your next commit will carry, work already in the index, and duplicate clones. Read-only, offline by default.