Agentic work is the majority of what people do with Claude and similar tools now. This release is us actually building specifically for those tasks.
Most of what slows an agent down is the overthinking and missing harness tools.
The setup you have to explain every time, the security question you never quite get a straight answer to, the interface that almost does what you want but fights you on the last ten percent with paragraphs of thinking text.
1.7.6 is a release aimed squarely at that surrounding layer: ten refined skills built for agentic workflows specifically, a real hardening pass on Faraday-Gate, our security gate, and a set of JOT interface fixes that make the writing and research panel interact the way it was designed to.
Ten tailored skills, built for how people actually work with agents now
Skills are one of the more underused parts of working with Claude. A skill is a small, self-contained instruction file that Claude loads only when it’s relevant, and a good one turns a task you’d otherwise re-explain every session into something the agent acts fast and with accuracy on.
We went looking at what the wider Claude ecosystem has built, reviewed a large subset of existing plugins and a large collection of community skills, and pulled out the ten that filled real gaps in what VEKTOR ships with, without overlapping anything we already had.
A few worth calling out specifically. Token conservation gives an agent explicit discipline around what to read and how much, instead of pulling in whole files when a targeted search would do. Agent delegation gives a clear decision framework for when a task should be handed off to a subagent versus handled directly, which matters a lot more now that most serious work is multi-agent by default.
Task orchestrator manages a backlog of work across a full pipeline rather than treating each task as a one-off. PR prep runs a proper self-review checklist before code goes out the door. Writing rules and slop detector both catch the specific tells of AI-generated text that reads as generic or unearned, phrase patterns, structural tics, claims without evidence, before a human reader has to.
Each one was rewritten to actually work in the tools available in this kind of session rather than assuming a different product’s feature set, and then tested for real rather than assumed to work.
We took the slop detector skill specifically and had a fresh agent in co-work, with no memory of building it, apply it cold to our own README. It caught a genuine, unbacked benchmark claim and a latency figure that contradicted another number elsewhere in the same document. That’s exactly the kind of thing a skill should do: catch what a human skimming past it would miss.
Token conservation — read-budget discipline; pulls targeted excerpts instead of whole files, cutting wasted context.
Agent delegation — a clear framework for when to hand a task to a subagent versus doing it directly.
Task orchestrator — runs a multi-item backlog through a full pipeline instead of treating each item as a one-off.
PR prep — a self-review checklist to run before code goes out, catching the obvious stuff before a human has to.
Writing rules — documents and enforces house style/guardrails so output stays consistent across sessions.
Slop detector — flags generic AI-writing tells: unearned claims, filler phrasing, structural clichés.
Onboarding — a staged reading order for getting an agent (or a person) oriented in an unfamiliar codebase fast.
Debugging wizard — a systematic method for tracking down bugs instead of guessing and poking.
Legacy modernizer — a incremental approach to migrating old code forward without a risky big-bang rewrite.
Test master — strategy and discipline for writing tests that actually catch regressions, not just pad coverage.
Faraday-Gate keeps getting better at detection.
Faraday-Gate is the security gate that sits in front of every consequential action an agent tries to take, a memory write, a file operation, a remote command, and checks it before it happens. We’ve written before about being precise regarding what it actually does versus what it sounds like it does, and this release is more of that same refined discipline applied to closing real gaps rather than adding surface-level polish.
The biggest one: Faraday-Gate now has an independent integrity watchdog, a background process that runs separately from any active session and checks two things continuously. First, the AI-assistant configuration files across seven different clients, Claude Desktop, Cursor, Windsurf, VS Code, Cline, Roo Code, and Groq Desktop, watching for the exact persistence trick real supply-chain attacks use: quietly injecting a rogue MCP server entry so it gets loaded and trusted automatically next time.
Second, Faraday’s own core enforcement files, since previously nothing would have noticed if those files themselves were tampered with. We tested this directly by simulating tampering against Faraday’s own gating logic, and the watchdog caught it and named the exact file.
Alongside that, the audit log Faraday-Gate keeps of every gate decision is now tamper-evident. Each event’s record is chained to the previous one’s hash, the same principle git uses for commit history, so altering or deleting a historical entry breaks every hash that comes after it, visibly.
Self-preservation coverage, the check that protects Faraday’s own files from tampering, went from watching Claude Desktop only to all seven client surfaces. And a session that gets flagged as compromised now genuinely locks. Previously the flag was recorded but didn’t stop much else from continuing.
Now it blocks every further consequential action until you restart, matching how real endpoint security tooling handles containment, while status checks and approvals stay available so you’re never locked out of understanding what happened.
The confirmation prompts themselves got rewritten too, across all five gate types, to explain what approving actually does and when you should say no, instead of just naming what pattern got flagged. A warning that says what tripped isn’t the same as one that tells you the consequence.
JOT actually looks and behaves better with design improvements
JOT is the writing and research panel, thoughts on one side, synthesis and collaborative research on the other. A few things in it needed refinement in the interface, and we rebuilt and tested them properly rather than patching around them.
The flashcards feature had a real sequencing bug. The copy button’s click handler was nested inside the save button’s handler, so copy only worked after you’d already hit save, by which point the card had removed itself from the page. We pulled them apart into two independent listeners so both buttons actually do what they say immediately.
The collab panel’s styling had a much bigger issue hiding behind it. A single unterminated CSS comment early in the stylesheet had swallowed every rule after it, the status bar, the insight and synthesis blocks, the paper cards, the session bar, button styling, all of it silently dead code for what looks like a long stretch of the panel’s history. We closed the comment properly and every one of those sections is styled again.
And the synthesis panel’s scrollbar, which is a small thing until you’re actually the person scrolling through a long research thread and can’t find any visible way to do it. The custom scrollbar styling had been applied to the outer panel wrapper, which never scrolls, instead of the actual content area underneath it that does.
The real scrolling element was falling back to the browser’s default near-invisible overlay scrollbar. We fixed the CSS to target the right element and made it wider and higher contrast, so there’s now a real, visible slider on both the writing pane and the synthesis pane.
Reorganization of the toolbar from split screen so each pane gets its own section page for both Collab and Synthesis.
The reliability pass underneath all of it
None of the above matters much if the basics quietly stop working, so we also went through the SDK end to end before pushing this out. The vektor hooks command had been failing outright because the module behind it was never actually written, despite being wired into both the CLI and the chat interface. We built it properly this time, list, add, remove, clear, all sharing one implementation instead of duplicated logic in two places.
Faraday’s own corpus updater had a quieter problem. It verified downloaded signature files correctly but never actually extracted anything from the archive, silently falling back to the bundled copy every time, with no error and no sign anything was wrong. We wrote a small zip reader and wired in the missing extraction step, so updates now do what they’ve always claimed to.
The boot banner had been showing a stale, hardcoded version number regardless of what was actually installed, traced back to a function being called with one argument where it expected three. We fixed the call and made the fallback read the real version from the package itself, so the same mistake can’t quietly recur.
And across the codebase, about a dozen files had an old encoding bug where a status icon had degraded into a bare question mark. Rather than guess at the original symbol and risk introducing a new version of the same corruption, we replaced every instance with plain, unambiguous text instead.
The README got a full rewrite to match all of it: the real CLI command list, which now runs past thirty commands, a verified tool count pulled directly from the server code, and a changelog section that reflects the last several releases instead of one frozen snapshot from a while back.
We also found a real old claim in our own benchmarks and updated the 81 percent LongMemEval to the new figure in the README.
Where that leaves 1.7.6
Ten skills built for agentic work specifically, a security layer that’s measurably harder to tamper with or fool, an interface that finally behaves the way it looks like it should, and everything underneath it double-checked rather than assumed. That’s the release. Full changelog and download at vektormemory.com.
https://vektormemory.com/docs/changelog
VEKTOR Memory builds local-first persistent vector memory infrastructure for AI agents. Faraday-Gate is the security and privacy gate that reviews every consequential agent action before it happens. Full technical documentation at vektormemory.com.
LLM AI Agent Vector Memory Skills
Top comments (0)