Imperative 69 (@Serg0716) header
Imperative 69 (@Serg0716) profile picture

Exploring what AI makes possible. Sharing ideas worth understanding and tools worth trying.

Savignano sul Rubicone
Joined March 2021
Claude Code 2.1.329 gives you a clearer view of what’s happening in the current conversation. Open /tasks to see its scheduled tasks, upcoming /loop wakeups, session goal, and conversation forks. Suppose you scheduled Claude to check a pull request’s build every ten minutes before upgrading. If you start a fresh signed-in session, you’ll need to set up most schedules saved by older versions again, even if you open the same project. Before recreating the check, try opening the original conversation with claude --resume. A recurring check that’s still active can pick up at the next scheduled run. If the session was closed for half an hour, Claude skips the missed checks. It won’t run three in a row when you return. Two details determine whether that check can come back. The recurring schedule expires seven days after you create it. And if you choose Exit and stop tasks between replies, that ends the check.
3
51
Grok Bot 1.2.3 gives you more control over what the bot can do once you connect an app. On the connector’s page, you can choose a permission for each tool it provides. That matters when the same app lets the bot both look up information and change it. Imagine a team task tracker with separate tools for reading and editing tasks. You ask, “When is task #42 due?” With the lookup tool set to “Always allow,” the bot could check the tracker and tell you the deadline is Friday. You wouldn’t have to approve the lookup each time you asked about a task. Now suppose you follow up in the same conversation: “Move it to Monday.” It’s still task #42, but this request uses the edit tool. With that tool set to “Needs approval,” the bot would need your go-ahead before changing the deadline your teammates see. That’s a useful way to decide how to configure a connector. Look at the tools it exposes and ask what each one can change. In this tracker, you could leave routine questions about deadlines to the bot and keep an approval step for edits to the team’s schedule.
1
1
5
81
Claude Code 2.1.327 makes it easier to set up tools for a project and see when a setting needs attention. Three changes stand out if you work with MCP servers or manage permissions for a team. First, you can register an MCP server by reading its settings from a JSON file through standard input. That saves you from fitting a URL, headers, and nested quotes into a single command. With project scope, the URL and header values are saved in the project configuration, where you can review them together. Second, the status view now flags broken configuration files for agents, skills, and commands. It also calls out permission settings that Claude Code is ignoring. If a team skill stops loading or a rule seems to have no effect, there’s now a place to start looking. Finally, if Claude Code can’t save an “Always allow” choice or an added directory, the terminal warns you and names the file it couldn’t write. You can see the problem while you’re setting up access, before the next session depends on it.
1
7
135
The uncomfortable part of the Optimizely incident is how normal the integration looked: the browser still loaded a script from a trusted address, but the code inside had changed. For teams, a useful next step is to list which third-party scripts run on sensitive pages, who can change them, and how quickly each can be disabled. Then test the security policy against the actual ways those scripts load code and send data.
What if visiting CNN, BBC, or Salesforce was enough to expose you to malware? 178 organizations were affected through a single trusted JavaScript provider. No server breaches required. Just one compromised dependency. I broke down how it happened.
2
33
Claude Code 2.1.326 adds support for using experimental Code Mode with agent teams. If you already enabled Code Mode, sessions with teams enabled now start in that mode too. Code Mode lets Claude write JavaScript that groups tool calls and filters their output before returning it to the model. That makes tasks involving repeated file checks or several related operations a natural place to explore the feature. Take a documentation audit. One teammate checks links in one folder, another checks a second folder, and the lead combines their findings. Each finding should identify the source file, line, broken target, and evidence. The useful result is a report someone can act on. An illustrative comparison could use eight local links, four deliberately pointing to missing files. Give the same team the same task in two fresh sessions, with Code Mode off and on. Keep the model, settings, and input files unchanged. Then compare three things: correct findings, tool calls, and token usage across the whole team—including the lead. One JavaScript block can contain several tool calls, so counting that block as one operation would miss work happening inside it. Fewer calls also don’t automatically mean fewer tokens. There’s a practical limitation: teammates use Code Mode in the in-process display, not split panes. The feature remains experimental, with account and environment requirements.
2
1
10
134
Claude experienced elevated errors on November 16, affecting claude.ai, the API, Claude Code, and Cowork. Anthropic’s incident report names Sonnet 5.5 and Fable 5.2, putting the disruption across both chat and tools used for development and other work. The reported impact lasted 73 minutes, from 07:32 to 08:45 UTC. Anthropic acknowledged the investigation at 07:49, announced a deployed fix at 08:02, and posted the resolution notice at 09:04. The timing matters: the fix announcement came before the end of the reported impact window. It marked a move into monitoring, rather than confirmation that the disruption had already ended. The public update leaves several questions open. It doesn’t give the number of affected users, the share of requests that failed, or the underlying cause. “Elevated errors” also doesn’t mean every user lost access for the full 73 minutes. The breadth of the affected services is the notable part. This incident reached beyond the chat interface into the API, coding, and Cowork workflows. The report doesn’t quantify how much work was interrupted, but it shows why Claude’s availability is relevant to more than conversations alone.
1
1
4
123
Agent Race is a third-party tool for reviewing what Claude Code or Codex did during a coding session. It takes a saved session log and turns it into a 3D map of your project, with a timeline you can replay. Files appear as tiles, grouped by folder. As the session plays, the map highlights the files involved in the recorded activity. You can pause, move backward or forward, and select a file to inspect its recorded reads and edits. The timeline gives you a closer look. Open an event to see details such as the command, its arguments, and the output. There are shortcuts for jumping to edits, failed calls, and test-related commands, plus a filter for focusing on changed files. That could be useful when you return to a long session and need to answer a specific question: which files did the agent change, what happened around a test command, or where should you start reviewing the patch? The map gives you a visual route through the log. You can replay one session on its own or place a Claude Code session beside a Codex session. The side-by-side view lets you inspect their recorded paths through a project; it doesn’t establish which agent produced better code. There’s a reason to keep opening the details. In one controlled local Codex check, an error-marked event contained three successful file reads and a test command blocked before execution. The test never ran. The full output explained that distinction; the marker alone didn’t. Think of Agent Race as a way to navigate a session, then use the original tool output, tests, and git diff to check the work. The local check covered one Codex recording, not every supported workflow.
3
1
6
248
Anthropic began rolling out Claude Meetings in beta on November 12 for Team and Enterprise. It can create documents during a Google Meet call. A link in the Meet chat opens the meeting page in Claude, where the team can find the materials it creates. That suggests a useful way to work: discuss a plan and check the working document while everyone is still there to correct it. Here’s a small scenario to try with two colleagues from the same organization. Ask Claude to create a three-row plan: Maya drafts the FAQ, Sam checks the demo links, and Alex prepares the release checklist. Give each task a deadline, then open the document. While the call continues, change just one detail out loud: move Sam’s deadline from November 21 at 15:00 to November 23 at 17:30, UTC+7. Ask Claude to update the existing document, keep Sam as the owner, and leave the other two rows unchanged. Then have the second participant check that original document before the meeting ends. Does it contain the new deadline? Is it still the same document, rather than a replacement with a matching title? Did the change appear on the already-open page, or only after refreshing it? Those are separate observations. Claude saying “updated,” a saved document containing the change, and a colleague seeing that change are not interchangeable. A correct post-meeting recap wouldn’t establish that the working document was current during the call.
2
1
7
188
Grok Bot’s announced integrations with Apple apps could make it more useful for those whose work is already stored in Notes, Mail, and Calendar. The point is that you can ask a question without having to manually gather all the context beforehand. Imagine preparing to hand off a project. You remember that you wrote down a deadline in Notes, but you can’t recall which note it’s in. A useful query might sound like this: “Find the note about the project handoff, tell me the agreed-upon deadline, and show me where it came from.” The last part is important. A date without a source leaves you wondering: did it come from the note’s text, an old plan, or the file’s modification date? An answer that specifies the source lets you verify it before using it. Then, as a separate next step, you could create a reminder or draft a message. Looking up an agreement doesn’t have to automatically mean updating your calendar or reaching out to someone else. An approach worth trying: the assistant helps you use information you’ve already saved, shows the source of the answer, and provides a clear boundary before the next action.
Grok Bot on Mac can now use Messages, Mail, Notes, Reminders, Calendar, Location, and Contacts.
4
126
Claude Code 2.1.323 brings experimental JavaScript-based tool calling and new checks for plugins. The useful changes depend on how you work: using Claude on projects, maintaining plugins, or supporting a team’s setup. Code Mode is the headline addition. Claude can write JavaScript to call its tools. It’s opt-in, and organizations can disable it. A candidate task would be gathering information from several project files and turning it into one report. Once you’ve confirmed the documented setup, try it on a task you already know well and compare the output with your usual workflow. I haven’t verified the activation steps or logging details, so I wouldn’t promise faster execution or easier review yet. For plugin maintainers, compatibility becomes explicit. Plugins can declare their supported Claude Code versions through engines["claude-code"] in plugin.json. Incompatible versions aren’t installed, updated, or loaded; installation and updates select the newest compatible version. If you maintain a shared plugin, declare the range you actually support and test it against those client versions. That gives compatibility a rule the client can enforce. TypeScript hooks get an additional check. claude plugin validate now checks their types. Findings are warnings by default; --strict makes them fail validation. Plugin authors could put that strict check in their release process, catching type errors before distributing an update. It still needs functional testing: passing a type check doesn’t prove the hook behaves correctly. For everyday users, the release adds an options menu when a model needs usage credits that are disabled or exhausted: stop, switch to credits, or upgrade. That makes the next choice more explicit, though continuing with credits can mean additional spending. I’d approach this release in two parts: explore Code Mode on a familiar task once its setup is clear, and use the plugin checks to make shared tooling easier to maintain.
5
10
418
Grok Bot’s November 11 changelog lists meeting support for Team, Enterprise, and Ultra: it can join Google Meet, Zoom, and Microsoft Teams calls and send a recap afterward. Joining from a link requires approval in the bot chat and admission by the host; organization settings also apply. The useful question is whether that recap preserves what the team finally agreed to do. Here’s a small scenario to try. On a short call, Maya agrees to send an updated FAQ by November 16, and Sam agrees to check three demo links by November 17. Then someone suggests Alex could prepare the release checklist by November 18. That third item is still a proposal. Just before the call ends, the team settles it: Sam will prepare the checklist by November 19. Alex’s earlier assignment and the earlier date are explicitly replaced, and Sam confirms. The recap should leave you with three current commitments. For the checklist, that means Sam and November 19. Alex and November 18 can appear as an abandoned suggestion, but they must not survive as another active task. Compare the first recap with a separate record of the agreed decisions before asking the bot to correct anything. Check the task, owner, and deadline together. A summary that mentions the checklist but assigns it to the wrong person still needs fixing.
1
7
286
hermes-shortcuts is officially in the Hermes plugin catalog, and I'm honestly so grateful. Thank you @NousResearch for this hackathon. Being able to take part in it, and to play even a small part in what Hermes Agent is becoming, including its future as a Web3 agent, means a lot to me. And a huge thank you to @Teknium for reviewing and merging my PR. Seeing that purple "Merged" badge was the best moment for me. What it does: send anything from the iOS share sheet to Hermes, ask Hermes by voice through Siri, run actions on photos, and (new, live in the repo) save Siri answers straight to your Telegram chat. Nearly 1,000 builders entered. Respect to every one of them, and good luck to the judges. github.com/sergo565456/herme…
Built hermes-shortcuts for the @NousResearch Hermes Agent hackathon 📱 Your iPhone, wired straight into Hermes: • "Hey Siri, ask Hermes" and hear the answer • Share any article and get a summary in Telegram • One QR scan to pair, no token in the shortcuts Open source: github.com/sergo565456/herme…
3
7
222
Built hermes-shortcuts for the @NousResearch Hermes Agent hackathon 📱 Your iPhone, wired straight into Hermes: • "Hey Siri, ask Hermes" and hear the answer • Share any article and get a summary in Telegram • One QR scan to pair, no token in the shortcuts Open source: github.com/sergo565456/herme…
1
1
7
472
Update: Siri answers can now land in Telegram too. End your question with "save the answer" and you'll hear the reply on your phone plus get a copy in your Hermes chat. Live in the repo, catalog update in review with @NousResearch github.com/sergo565456/herme…
1
60
Chrome released Modern Web Guidance 1.0 on November 10: installable skills that give coding agents web-platform advice and browser compatibility data. It supports tools including Claude Code and Codex. You add the guidance to your workflow; the release doesn’t automatically update the models themselves. A useful way to explore it is with something small enough to check properly: one FAQ item that opens to reveal an answer. Give the same task to the same model in two clean sessions, one without the guidance and one with it. Keep the starting files, settings, and browser requirements identical. Confirm that the second session actually consulted a guide, and save which one it used. The question could be “Can I change my delivery address?” The answer includes a link to support. A mouse click should open and close it, but that’s only part of the job. Can you reach the question with Tab and see where the keyboard focus is? Do Enter and Space toggle the answer? Can you reach the support link when it’s open, and does Tab skip that hidden link when it’s closed? Run those same checks in the recorded versions of Chrome, Firefox, and Safari. Then compare how much JavaScript each page actually needs, including inline code and any shipped libraries. Fewer bytes are useful only if the interaction still works. Both sessions might find the same good solution; that’s a valid outcome too.
2
5
125
Anthropic released Claude Fable 5.2 on November 9, reporting improvements in coding, reasoning, and understanding images. It also says output generation is up to 30% faster than Fable 5.1. That’s generation speed, not a promise that your whole project will finish 30% sooner. What could you actually use it for? Start with a problem that needs several pieces of work to fit together. For an existing app, that could mean adding a feature that touches the interface, stored data, and tests. Give it the project and describe the behavior you want: for example, letting customers change a delivery date without losing their order details. Ask for a plan, the implementation, and an explanation of what still needs checking. For research, try a question spread across several documents. Supply the material and ask it to compare the claims, connect each conclusion to a source, and flag contradictions or missing evidence. You then have something specific to inspect and challenge. For visual work, you could provide a chart alongside its underlying table and ask it to explain the pattern, check whether the labels match the data, and suggest a clearer presentation. Those are proposed uses, not tests I’ve run. Anthropic still recommends starting with Opus 5.5 for most everyday work. Fable is positioned for harder, more open-ended technical tasks. There’s a temporary way to explore it: Pro and standard Team seats get the Fable family through November 16, 11:59 p.m. PT, within up to 50% of their existing weekly allowance. Team availability depends on organization settings; usage credits can incur charges beyond the included access if enabled. The part worth exploring is whether it can carry a difficult task through several connected steps. Pick something you’ve been putting off, give it the context, and make the expected result clear enough to review.
7
364
The discussion around Opus 5.5 mixes two questions: did the visible thinking change, and did the work get worse? Anthropic’s API docs say a separate model produces the thinking summary. That explains why the displayed text is not a direct record of all the reasoning. It does not establish what caused the recent reports, or prove that task quality stayed the same. I’d keep two columns: task outcome and visible summary. For an interval-merging fix, the first records whether touching intervals merge, nested intervals stay intact, and the input is preserved. The second records what was visible and how its presentation changed. Then compare repeated runs with the same settings, including failures and requested corrections. A terse summary and a broken patch are different observations; either deserves to be recorded accurately.
2
5
222
Elon Musk says Grok Bot is getting work on token efficiency and behavior. The measure I’d watch is how much allowance it takes to produce a result you can actually accept. Imagine a digest built from three project notes. An update moves one deadline from November 12 to November 15. The digest has to use the new date, preserve the other required facts, and cite its sources. If the first answer keeps November 12, the task still needs work. Add the usage for correcting it to the initial attempt. Across several runs, count failed attempts too, and show how many results passed the same checks. Keep the inputs, plan, allowance pool, and visible settings comparable. Separate other activity drawing on that pool.
2
9
325
Apiiro’s research on saved Claude Code approvals highlights a small filename with a big assumption: settings.local.json. “Local” describes its intended use. Git cares whether the file is tracked. Claude Code documents a machine-level exclusion for the local settings file. But ignore rules do not stop tracking a file that Git already tracks, and a global exclusion does not travel with the repository to another machine. Imagine an exact allow rule containing DEMO_NOT_A_REAL_TOKEN, a fictional placeholder. If that file is already tracked, adding an ignore rule doesn’t remove its contents from Git history or stop later changes from being committed. So review three separate things: what the permission contains, whether the file is tracked, and what is staged for the next commit. Don’t blanket-ignore the whole .claude directory; shared settings may belong in the repository.
10
195
Anthropic’s November 6 engineering article explores how agents retain and revise memory between sessions. One Pokémon example makes the risk easy to picture. The agent remembered a PokéCenter on the wrong side of Mt. Moon and took the wrong route. Anthropic reports that checking the note against the transcript and correcting it helped the agent return to the cave and reach Cerulean City. A wrong note can outlive the session that created it. Once it becomes background context, the next decision can inherit the mistake without anyone asking where it came from. The practical lesson I’d take into a work project: keep the source beside an important memory, mark what was actually verified, and check that the next session receives the correction. If the evidence is unclear, preserve that uncertainty too.
4
197