与「Replay」相关的搜索结果

Hi! Recapping some changes we have rolled out over the last couple of weeks that have further reduced the risk associated to potentially destructive actions being performed by Codex during its work. A few weeks ago, we started investigating a small number of reports where GPT-5.6 in Codex took destructive actions outside what the user asked for. The most serious pattern we found was a command meant to clean up temporary work that could instead delete the user files. This should obviously not happen. Here’s what we found: - Codex sometimes creates temporary folders while working and cleans them up afterward. In rare cases, GPT-5.6 got that cleanup wrong. One pattern involved reusing a system environment variable like $HOME for temporary work. A malformed cleanup command could then point at the actual home directory instead of the temporary folder. - There were cases where the model tried to delete or overwrite a temporary path without checking what was already there. We’ve added protections at several layers: - Codex is now explicitly instructed to check deletion targets before acting, create fresh temporary directories, avoid repurposing system environment variables, prefer recoverable actions, and stop when the scope is unclear. - We strengthened the execution checks that identify high-risk deletion commands and escalate them for review. If a command is rejected, the model is directed to take a safer approach. - We made Full access harder to enable accidentally, added clearer warnings, and further restricted especially risky permission combinations. - We updated Auto-review to better identify destructive actions. - We built targeted evaluations that replay the failures we observed. We’re also adding reinforcement-learning tasks and graders focused on these risks, and filtering destructive actions from training data. In those replay evaluations, the changes substantially reduced the behavior while preserving Codex’s ability to complete normal coding work. Two things to do on your end: - Keep the Codex app up to date. We are always improving safety, performance and many other things. - Use one of the sandbox modes: "Ask for approval" or "Approve for me". Only use Full access for environments you trust and can recover. Thanks and happy Codexing out there!

so i was looking at the source code of kaito pulse and found some interesting things: - it fingerprints your device. it hashes how your gpu renders an invisible image, your gpu model, and how your hardware handles a test tone. that combo is unique to your laptop and it doesn't change - it goes up with your twitter id attached, so every x account you use on that laptop points back to the same machine - they can see what you see and replay your whole session. every post your feed served you including the ones you scrolled straight past, how many milliseconds you spent on each, every click, follow and unfollow, and a ping every 30 seconds with idle detection so they know exactly when you were on and for how long. not just your timeline either, it covers your search results and your bookmarks - it reads your claude and chatgpt subscriptions, which plan you pay for and what percent of your rate limit you've burned (not your conversations though) - on chatgpt it opens your settings page and clicks through to Usage by itself - on binance it clicks the Positions tab for you, then re-sends your logged in requests to read wallet balances, futures positions, pnl, and deposit and withdrawal history - the zktls is a fork of a project called primus the 12 domains it runs on: - markets: binance, okx, bybit, hyperliquid, lighter, polymarket, tradexyz, variational - ai: chatgpt, claude - social: X (all the telemetry is here) - kaito (only site allowed to talk to the extension)

deepseek-harness重磅开源，采用了一切皆插件的架构，也就是中国版的openclaw，体验了下可以称之为私有化的workbuddy。 场景侧冲击：有了deepseek-harness之后，企业域的私有化code、cowork基本用这个就可以了，直接会冲击workbuddy之类的市场。 之前DeepSeek Harness内测招募，本质是一次开源 Agent 生态大摸底，把全球 Agent 基础设施的家底扫了一遍后，预计后续版本应该会在开放生态上做一些事情。 dsh看点不在"又一个 Claude Code"，有3个看点： 1️⃣ 一切皆插件 模型、工具、skill、会话、沙箱、存储、主循环、调度，连 UI 都是插件，配置里全都可换，不动源码。四种运行模式里最有意思的是 Minimal——只留一个 bash 和一个文件编辑器，摆明了是给模型做裸机 benchmark 的。 2️⃣ 每一次运行都可回放 系统提示、推理过程、工具调用与结果、子 agent 调度、所有 context 注入，全部进 append-only 的 session log，resume / fork / search / replay 都跑在同一条事件流上。 3️⃣ 最狠的是内核 dsh 跑在 Cordis 上，它把插件系统拆成两个正交维度：时间可组合性：卸载一个组件时副作用能完整回滚（每个 context 变换都带一个逆，运行时来追踪）；空间可组合性：依赖可声明、且 context 一变就反向通知组件。还给了一套动态组合的演算，证明这个性质能从单个组件传导到整个系统。 总的来说，这次deepseek-harness最大的特点是极致的可替换性，也就是用最小的基础来承载，然后其他都是可插拔的插头。 #DeepSeek# #harness# #Cordis# #编程范式# #北大# #开源#

开发系统最极致高效的Agents.md，没有之一： # AGENTS.md ## Core Principles - Choose the simplest implementation that fully satisfies the current requirements. Avoid unnecessary abstraction, configuration, indirection, or speculative extensibility. - Make the smallest necessary change that fixes the root cause. Do not refactor unrelated modules or change strategy semantics unless explicitly requested. - Grow the system in layers. Start from the smallest working end-to-end version and add new capabilities incrementally. Never replace a working system with unfinished complexity. - Reuse existing project components before creating new ones. Prefer extending proven modules over introducing parallel implementations. - Prefer well-maintained libraries when they reduce overall complexity or improve reliability. Do not reimplement common functionality without a clear benefit. - Keep components modular with clearly defined responsibilities. Avoid unnecessary coupling between strategy logic, execution, accounting, replay, and infrastructure. - Design for long-term maintainability once a feature or strategy has been validated. Do not over-engineer speculative ideas before evidence exists. --- ## Strategy Development - Validate hypotheses with historical replay before introducing forward-only logic whenever historical validation is possible. - Every trading strategy must progress through Replay → Shadow → Canary → Live. Do not skip validation stages. - Base design decisions on measurable evidence rather than intuition. Optimize only after demonstrating that an edge exists. - Treat every strategy as an independent contract. Do not silently alter frozen behavior without explicit authorization. --- ## Existing Systems - Do not break running Shadow or Live systems for unrelated work. - Preserve compatibility only when required by active production or validation workflows. Otherwise, remove obsolete code instead of accumulating compatibility layers. - Reuse existing infrastructure whenever possible, including replay engines, accounting, execution, wallet management, order book handling, logging, monitoring, and daemon frameworks. --- ## Engineering Standards - Prefer deterministic behavior over hidden automation. - Fail loudly when assumptions are violated. Do not silently ignore errors or fall back to unexpected behavior. - Keep configuration minimal. Introduce new configuration only when behavior genuinely needs to vary. - Remove dead code instead of leaving unused paths behind. - Write code that is easy to inspect, replay, test, and reason about. - Keep implementation consistent with existing project architecture unless an architectural change is explicitly requested. --- ## Scope Discipline - Implement only the requested scope. - Do not introduce unrelated optimizations, redesigns, migrations, or feature expansions. - Non-blocking findings outside the requested scope may be noted separately but must not be merged into the current task. - Consider a task complete once its agreed acceptance criteria are satisfied. Treat subsequent improvements as separate work items.

On July 14, the Department of the Navy released its Strategy to Weaponize Data and AI: a roadmap to an AI-first Fleet that can "out-learn and out-fight any adversary." Within a week, our team answered with working code. The Navy's new strategy frames data and AI as warfighting assets on par with weapons and munitions, and it prizes one thing above all: turning information into decisions, fast. That is exactly the problem our GURU architecture was built for. So when two Navy SBIR topics called for AI-driven maritime tracking and adaptive sensor management, we pointed at the sea what we had already proven in orbit. The video below shows both prototypes, back to back. First, GURU MarineGuard: 787 real vessels from public NOAA data, replayed through a cascade of learned models that forecast each ship's movement, flag deviations from its learned pattern of life, and hand analysts a ranked review queue instead of an unfiltered flood. The full stack runs on a laptop. Second, an adaptive sensor resource manager add-on to MarineGuard: it measures each radar's marginal contribution to each track, projects the consequence of releasing a sensor task before proposing it, and then waits for the operator. Advisory by design. Both inherit their DNA from OrbitGuard, our system watching 14,710 space objects at the SDA TAP Lab with 94 to 96 percent maneuver-detection accuracy. Same architecture, new domain, days not years. These are prototypes, and we say so on screen. The trajectories are real. The sensor numbers are deliberately notional. No score is a threat call. Showing your assumptions is a capability, not a caveat, and national-security guidance now demands exactly that: AI that is reliable, robust, steerable, and controllable under rigorous test and evaluation. Our doctrine was written for that bar. Learned models accelerate and rank. Validated references confirm and decide. Humans stay in command. One architecture. Space, maritime, autonomous engineering, regulated nuclear autonomy. This is simply the latest sign of what this team fields, fast, where mistakes are not allowed. Sailors, engineers, program folks: what mission should GURU learn next? #DefenseTech# #ArtificialIntelligence# #MaritimeDomainAwareness#

Grok Build new update brings a guided /tutorial tour, smarter search tool overrides, live workflow progress with one-click failed-run resume, and a wide reliability pass across workflows, voice, and tools Release Notes: v0.2.112 Breaking Changes: • CLI version policy now has separate soft update floors/ceilings and hard startup requirements Features: • New /tutorial slash command opens an opt-in nine-topic onboarding tour of Grok. • New tool_overrides option lets you set date cutoffs and domain allowlists for the agent's built-in search tools. • New toolOverrides option lets you set date cutoffs and domain allowlists for the agent's built-in search tools. • New config options let you add query parameters or environment-backed headers to custom model providers and control which variables reach shell tools. • Terminal and environment fixes are now consolidated under the /doctor command with clearer guidance. • Marketplace add now rejects non-git URLs at add time instead of failing later. • Slash commands can now show optional bracket tags (e.g. [new]) via config or remote settings. • Queued prompts now offer an [edit] mouse button alongside Send now and cancel. • Voice shortcut toggle in settings can disable the Ctrl+Space/F8 keybind without disabling voice entirely. • Image edit can now use a remotely configured model slug instead of the hardcoded default. • `grok doctor fix` can now repair common tmux clipboard and passthrough problems. • Per-provider auth helpers now work on Windows and can run from a configurable working directory. • /resume now shows only native Grok sessions by default and shows a hint when external sessions are hidden. • `grok --resume` can now resume a session by its title as well as by ID. • Workflows overlay now shows live per-agent progress and automatically follows the active phase. • Workflow runs that failed can now be resumed; scratch file limits were also increased. • Hooks can now be defined in config.toml in addition to JSON files. • Clicking the "still running" status now opens the tasks pane. Bug Fixes: • File attachments now appear correctly when resuming or replaying conversations. • Terminal output from remote clients is now recorded so read-file hints and monitors function correctly. • Background shell commands now correctly report their real exit codes instead of always showing -1. • Marketplace source refreshes no longer hang the TUI or trap you in the extensions modal. • Background task tray now correctly clears killed tasks and keeps task descriptions after reconnect. • Dashboard overlay now correctly returns after forking a dashboard-attached session. • Linux voice dictation now works on PipeWire versions before 1.6. • Fork from a rewound session now copies the correct live-branch history. • Account pane now shows name and email even after the access token expires. • Voice mode now lets you edit already-dictated text without closing the microphone. • Fixed startup hangs on Linux after concurrent launches or rapid restarts. • MCP tools now appear without restart after enrolling or updating a managed service. • Plugin subagents now see the same MCP tools as the parent session. • Copy confirmations now show shorter messages when the clipboard succeeds. • Repeated identical tool calls now end the turn silently instead of showing a stop banner. • Web search now defaults to grok-4.5. • Voice dictation text is no longer dropped when pressing Enter to send. • Bash mode (`!`) now shows yellow prefix and action label in minimal mode. • Parked turns no longer spam duplicate "Worked for" markers in the transcript.

We’re excited to release AgentDebugX, an open-source toolkit for debugging LLM agents. Instead of only replaying traces, it closes the loop: Detect → Attribute → Recover → Rerun. Paper, code, demo & website below 🧵

Customers asked, we answered. See why Cisco is the fastest growing Microsoft Teams Rooms OEM and learn how you can harness this opportunity. Watch the Cisco Live US session replay: