All

What's Changed release: v0.11.1 by @github-actions[bot] in https://github.com/aliou/senzu/pull/15 Full Changelog: https://github.com/aliou/senzu/compare/v0.11.0...v0.11.1

Minor Changes fff80de: Add correlated prompt lifecycle events and report active approval prompts to Herdr.

Okay, but what can I actually do with that token?

Even after building OIDC into orbs, it wasn’t immediately obvious to me how far we could push an OIDC-based dev env. So here are a few ideas to get the juices flowing, drawn from both internal usage and feedback from our early partners.

1. Configure workload identity federation in GCP and grant access to your production logs and metrics from all orbs in your workspace.

I’ve been banging my head around how to give the agent the credentials necessary for whichever task it needs to do in the sandboxI had decided without much conviction to inject encrypted secrets and have them be decrypted on sandbox boot with the caveats that I would spend more time researching a better solution later. .

To optimize for scale, we decided that everything taking O(sandboxes) or O(nodes) load must be horizontally scalable by default, the sandbox creation path should be as simple as possible, and everything else should be secondary. The solution we came to is notably different than existing systems. We completely dispensed with any sort of central coordination, and traded global consistency for scalability and performance everywhere on the critical path for running and creating sandboxes.

I have nowhere near the same requirements for my personal sandbox setup but it still super interesting to see how what is needed to reach massive scale.

Release v0.12.0

I think that’s a great summary of the things we’ve done to make agents more productive in our codebase: don’t make them guess.

I’ve been trying to articulate to friends and colleagues why I never seem to have my agent screw up that much and it boils down to this: not letting them guess. Whether it’s via skills or instructions, but also toolinge.g. plugins for your favorite linter that explain why something is prohibited. and scripts, having everything in front of your agent decreases its chances of doing things wrong.

Of course it’s hard work to keep all of this up to date, but I find it way more rewarding than running a random npx skills addNo shade! and crossing finger that it will solve all my problems.

My initial instinct was to go full near-airgap: remove the default route, run an SNI-based TLS proxy on the host, and only allow traffic to api.telegram.org and openrouter.ai.

I built this with unbound returning fake DNS records pointing at the host bridge IP, nginx stream with ssl_preread doing SNI-based forwarding, nftables dropping everything in the FORWARD chain.

It worked, technically. But it was brittle. The proxy needed careful handling of dynamic IPs, and Openclaw’s error messages when DNS returned unexpected results were… unhelpful.

For this demo, I just said screw it and allowed everything (with a log). The VM has full internet access, but every DNS query and every new connection is visible from the host.

Continuing on my research around microvm, I discovered microvm.nix and yet another way of monitoring and/or locking down networking for vms that will be used by agents.

Minor Changes b4f5b84: Update Synthetic model catalog with Kimi K2.7 Code and remove delisted models.

Agents naturally gravitate toward the smallest stable abstraction. So, in an agent-first world, the most important question facing product teams is not what features and why, but instead how do I design the right primitive ?

Building great developer tools (e.g. primitives) has always been an exercise in exquisite taste and deep domain expertise because it requires balancing a set of competing, and often contradictory, forces: namely the tension between expressivity and simplicity. — The primitive is the product

Will be super interesting to see how companies adopt this and it will truly be a differentiator in adoption.

A VM that can reach the host’s local network is a VM that can reach things it has no business reaching. So QEMU doesn’t run in the host’s network namespace at all. We unshare into fresh user, net and mount namespaces first. Inside that namespace a small wrapper bind-mounts a resolv.conf pointing at 127.0.0.1 so that QEMU’s built-in slirp DNS isn’t used, then installs blackhole routes for every special-use IP range (RFC 6890, so private networks, link-local, loopback, etc.) before it execs QEMU. slirp4netns then provides the namespace’s outbound internet connection, with --disable-host-loopback, sandbox and seccomp all on. QEMU runs inside that namespace, and the guest’s network card is attached to QEMU’s own built-in user-mode networking. So every packet from the guest takes two hops: guest → QEMU’s slirp → the namespace’s slirp4netns → the internet. The guest never sees the host’s network and the host’s network never sees the guest. All of this is done without needing any privileges!

I’ve been reading a lot about microVMs and Firecracker lately as I work on async agent side projectMore on that soon hopefully. and I find it super interesting how people tackel the same UX problems (in this case: locking down the network).

Release v0.2.0

Minor Changes 61e66dc: Filter disabled gateway providers out of proxy and dedicated flows. /api/providers lists every provider regardless of its disabled flag, so disabled providers and their models leaked into the proxy settings/onboarding checklist, the dedicated provider list…

this is a huge reason why i use pi.

i absolutely do not want my harness regularly changing behavior out from under me, including system prompt changes, on top of an already stochastic llm

Imagine I have /skill-1 and /skill-2, both with disable-model-invocation: true I tell Claude Code: “/skill-1, /skill-2 do XYZ”

Which of the skills will be invoked?

I’ll give you a clue - the behaviour was one thing 3 weeks ago, another thing 1 week ago, and back to the original today.

Matt Pocock (@mattpocockuk) on Twitter

I’ve been playing around making my own harnessPi remains my main harness, but trying to think about how I work in this agent-looping world. recently, and also noticed how the behaviour of many Claude features seemingly seem to change every time I spin it up to test something. For my harness, I just decided on a simple principle: trying to keep the primitives (i.e. skills and mid-turn messages affect the agentMost of the time i want immediate steering, but I realized I do use follow-up quite a lot when leaving agents do their thing, especially when running in sandboxes and/or building prototypes where the result itself doesn’t matter much for me. ) consistent with how I expect agent to react.

introducing tau τ — an educational agent harness that teaches you how to build agent harnesses

This is so cool. I truly think we’re moving towards a world where writing/making a great agent harness ends up becoming accessible to most people because the knowledge is just readily available.

Minor Changes e86681a: Add a pythonToPython3 rule that rewrites bare python command invocations to python3 when enabled.

Application scope and user experience: Sun Microsystems co-founder Bill Joy famously said: “No matter who you are, most of the smartest people work for someone else.” That applies with particular force to AI. The range of what agents can do — and the variety of industries, workflows, and real-world contexts they touch — is so vast that no single company can build the best solution for all of them. We see a near-infinite opportunity to rewrite the world using agentic AI as foundational infrastructure, and it will require many companies with focused effort.

Thinking a lot about this lately, especially around my own use. I want from using Anthropic models and tools exclusively a year ago to barely using them and instead using different models in different harnesses depending on the task at hand.

It’s very likely that I’m part of the minority of AI users in this case currently but I wonder how long this will become more common ground amongst the majority of users.

• Add connectors extension and restructure shared config
• Tolerate the admin-only /aperture/config endpoint so non-admin Aperture grants can use the extension

Durable Objects now supports a new us jurisdiction. Create namespaces restricted to the US to keep compute and storage local.

i feel like i have a aluminium hatit’s called a tinfoil hat dummass , but it seems like everyone is preparing for us/non-us tiered system around llm access.

• Add a “decline and stop” option to the permission-gate prompt
• Update AGENTS.md to reflect the current three-extension layout and remove the unused, broken test harness

This article will sound like a schizo ramble, so prepare. To start off I want to state my true beliefs on the matter:

1. LLMs are not conscious and won’t ever be.
2. LLMs are inaccurate statistical machines.
3. Open Source AI is good but must get better.
4. I have no hatred, or malice for any closed source lab
5. I respect what I see as my peers at OpenAI, Anthropic, xAI etc.

You might think, well if that’s the case what’s the problem? These models aren’t intelligent, they can’t really replace people, can they?

Yeah that’s my shit.

• Add optional message field to Migration interface for user-facing migration notifications

Release v0.10.0

been playing with exe.dev this weekend and was curious how they made ssh [vm-name].exe.dev just work, and it’s pretty cool: https://blog.exe.dev/ssh-host-header

• Fix high CPU in sesame watch caused by repeated full-directory re-indexing and redundant file I/O

we landed on a pretty good workflow for doing parallel work in OpenCode

this demo is with git worktrees but i also preview an alternative we’re working on at the end

this will be in 1.6.0

— @thdxr

Rift looks interesting.

Take a look, I really like how this is explained.

• Library changes

I like the idea of Claude’s workflow thingy but since it’s writing code, it keeps failing with these kind of errors:

In addition, the UI of Claude feels a bit overwhelming to me.

There’s also some strong validations around making the workflow deterministic and I wonder if it could still be handled? I’m also curious on why they’re not giving guidance to the model about this and/or linting workflow code to have this be caught before the harness runs the workflow.

What's Changed release: v0.2.0 by @github-actions[bot] in https://github.com/aliou/granola-cli/pull/2 Full Changelog: https://github.com/aliou/granola-cli/compare/v0.1.0...v0.2.0

What's Changed release: v0.1.0 by @github-actions[bot] in https://github.com/aliou/granola-cli/pull/1 New Contributors @github-actions[bot] made their first contribution in https://github.com/aliou/granola-cli/pull/1 Full Changelog: https://github.com/aliou/granola-cli/commits/v…

Missed that claude/opus has a fast mode so I had Pi make a quick extensionDon’t look too closely at that horrendous code that Kimi with reasoning disabled produced. to play with it. Below, a comparison with the vanilla Opus.

• Add syn:large:text, syn:small:text, syn:large:vision, syn:small:vision alias models

Closing my open tabs and finally read Mikker’s post. Borrowed his review workflow but wanted to have it in a Ghostty split (thanks to AppleScript).

I then had Kimi and GLM figure it out and make me an extension.

Release v0.1.0

Been using variations of this prompt every time I notice a gap in my knowledge. Choosing a starting point that I broadly understand and then throwing stuff I literally don’t know about works super well for me.

New TIL: I figured out how to use my LLM CLI tool in a shebang line, which means you can write executable scripts in English, or hook up more complex scripts with a snippet of YAML template

— @simonw

It’s becoming a meme at this point, but here’s a Pi extension to do the same:

Also watch GLM-5.1 completely overthink while trying to speed up the video and add a fast-forward overlay.

• Prevent crash on /new with running background processes

Current status.

• Migrate Pi core package dependencies from @mariozechner/* to @earendil-works/* namespace
• Truncate Linkup search, answer, and fetch result blocks, with full content saved to temp files when truncated

Currently cleaning up a backlog of issues and PRs in a client’s repo and was tired of alt-tabbing between Pi and my browser, so made a quick Pi extension to autocomplete GitHub PRs and issues.

Never going to use this ever again, but might be useful to other people so sharing it as a gist.

• Use event-driven Synthetic quota updates without polling
• Avoid stale contexts in status update

Was about to make a Pi extension to communicate between sessions, but actually only needed Amp’s read thread tool.

• Add setting to control Synthetic-proxied models. New installs default to Synthetic-hosted models only, while existing configs keep proxied models enabled

Can’t help but re-read this every time I write “just” somewhere.

• Initial release of pi-neuralwatt — Neuralwatt inference API provider with energy transparency

Banger.

• Extract pure core functions from Pi glue.
• Add per-provider gateway model checking.
• Add gateway model checking to settings UI
• Rewrite setup wizard with Wizard + FuzzyMultiSelector

Needed to scaffold a project this morning with an AGENTS.md and liked how Claude Code’s new /init looked, so had Pi figure out how it works and make its own version.

Another one-time use extension.

Source

• Rename the package from @aliou/pi-extension-dev to @aliou/pi-dev-kit

The models are getting better and better which makes it easy to forget to not delegate your thinking.

• Redesign web search tool UI to match read_url pattern

What's Changed release: v0.5.0 by @github-actions[bot] in https://github.com/aliou/preview-markdown/pull/6 Full Changelog: https://github.com/aliou/preview-markdown/compare/v0.4.1...v0.5.0

Finally got around to reading Linear’s Agent interaction guidelines and did a quick implementation with Pi. It was pretty simple to set up and I kinda like the UX!

Take a look at the code on GitHub.

Minor Changes 9231b94: Add a new no-ts-import-extension plugin that disallows .ts extensions in import and re-export paths.

Had to update Ghostty to fix that annoying auto-scroll bug but didn’t want to lose my open tabs. Asked Pi/GPT-5.4 to dump and restore my tabs and splits. It mostly worked and I’m never going to reuse this.

• Add optional warning-level notifications for rewritten commands so demos and debugging make toolchain rewrites visible.

Got absolutely screwed by Opus thinking it was running on Linux instead of macOS while debugging some CLI stuff with tmux.

But the session of Opus figuring out what happened is pretty fascinating:

Also created a small tool that uses BM25 to find content in previous sessions. It’s super useful to just say “I was running a Pi session at the same time, could it be related?” This is completely borrowed from Amp.

• Replace deep boolean with depth enum on web-answer tool for consistency with web-search

• Add “/synthetic:quotas” command to display API usage quotas
• Add Synthetic web search tool

Very meta: wanted to use Kimi K2.5 via Synthetic, so I had Pi set up a one-off extension that creates the provider and defines its model. Then reloaded Pi and asked Kimi to create the Synthetic provider.

Release pi-linkup@0.1.0

Another one, but this one probably more for me than other people: guardrails!

Fun extension I made for Pi: processes. Let your agent handle running processes in the background and be notified when they finish / die etc.

Ported by beloved jellybeans-mono for Pi, both dark and light!

What's Changed release: v0.1.0 by @github-actions[bot] in https://github.com/aliou/fastmail-cli/pull/1 New Contributors @github-actions[bot] made their first contribution in https://github.com/aliou/fastmail-cli/pull/1 Full Changelog: https://github.com/aliou/fastmail-cli/commit…

What's Changed release: v0.2.0 by @github-actions[bot] in https://github.com/aliou/preview-markdown/pull/1 New Contributors @github-actions[bot] made their first contribution in https://github.com/aliou/preview-markdown/pull/1 Full Changelog: https://github.com/aliou/preview-mar…

Full Changelog: https://github.com/aliou/preview-markdown/commits/v0.1.0

I needed to quickly investigate some data and used the following pattern to do so. Using the model of an e-commerce store as example, here’s what I did.

The Setup

First, setup a cube.js defining the dbType:

module.exports = {
  dbType: 'duckdb',
}

Then, define the driverFactory, and configure the access to the GCS bucket:

const driverFactory = async () => {
  const { DuckDBDriver } = require('@cubejs-backend/duckdb-driver');

  return new DuckDBDriver({
    database: ':memory:',
    initSql: `
      INSTALL httpfs;
      LOAD httpfs;

      CREATE SECRET (
        TYPE gcs,
        KEY_ID 'YOUR_KEY_ID',
        SECRET 'YOUR_SECRET_ID'
      );
    `,
  });
}

Finally, add to initSql the commands to setup our views and access our data:

CREATE VIEW orders AS
SELECT * FROM read_parquet('gs://tmp-bucket/orders/*.parquet');

CREATE VIEW products AS
SELECT * FROM read_parquet('gs://tmp-bucket/products/*.parquet');

CREATE VIEW line_items AS
SELECT * FROM read_parquet('gs://tmp-bucket/line_items/*.parquet');

CREATE VIEW users AS
SELECT * FROM read_parquet('gs://tmp-bucket/users/*.parquet');

CREATE VIEW inventory_transactions AS
SELECT * FROM read_parquet('gs://tmp-bucket/inventory/*.parquet');

Key Implementation Details

Views, Not Tables: I specifically used CREATE VIEW instead of CREATE TABLE AS to avoid loading all the data into memory. Views just store the query definition - DuckDB streams the Parquet data chunk by chunk when you actually query it. With tables, DuckDB would copy everything into its internal format, eating up memory or disk space.

Zero Infrastructure: The :memory: database means no provisioning. Each Cube process gets a fresh DuckDB instance that reads directly from existing Parquet files.

Direct Access: No ETL pipeline, no data loading. The read_parquet() function streams data directly from GCS, and DuckDB’s columnar engine only fetches the columns needed for each query.

Wildcard Patterns: The /*.parquet pattern lets you treat multiple files as a single table without any preprocessing. Perfect for data that’s already partitioned by date or other dimensions.

Performance Characteristics

In this example, you could run complex queries across orders, products, line items, and inventory - joining across millions of rows within seconds. What would normally require hours of data pipeline setup can be done immediately.

DuckDB’s aggressive caching means the first query might take a few seconds to fetch metadata, but subsequent queries run pretty much instantly even on larger datasets.

Trade-offs

Benefits:

• Zero data movement—queries run directly against GCS files
• No infrastructure to manage
• Instant setup for ad-hoc analysis
• Reuses existing Parquet files from other processes

Drawbacks:

• :memory: means state vanishes on restart
• Not suitable for production workloads
• Performance degrades on datasets larger than available RAM

Notes

This pattern works for any quick data investigation. Got Parquet files? Throw them in GCS, point Cube at them with this config, and you’re analyzing data in minutes. For more permanent setups, swap :memory: for a file path to persist the DuckDB catalog between restarts.

Your data lake is already an OLAP database. You just need DuckDB to unlock it.

Continuing my work on cube-records, I wanted to support joined field names like orders.total or users.email - matching exactly how the underlying Cube names them. Instead of manually defining these combinations, TypeScript’s template literal types generate them automatically.

The Pattern

Here’s a simplified version of the Cube schema:

// User defines their schema
interface Schema {
  users: {
    fields: { id: string; email: string };
    joins: ['orders'];
  };
  orders: {
    fields: { total: number; status: string };
    joins: [];
  };
}

Then, we use a template literal type to defined our joined fields:

// Template literal magic
type JoinedFields<T extends keyof Schema> =
  Schema[T]['joins'][number] extends infer Join
    ? Join extends keyof Schema
      ? `${Join}.${keyof Schema[Join]['fields'] & string}`
      : never
    : never;

Finally, we can also use the type to create our own custom type.

// TypeScript now knows these fields exist:
type UserJoinedFields = JoinedFields<'users'>;
// Result: 'orders.total' | 'orders.status'

The infer keyword extracts each join name, then template literals compose the dot-notation paths.

Usage

With the schema defined above and our new type, we can define a Query type like the following:

type Query<T extends keyof Schema> = {
  model: T;
  fields: (Schema[T]['fields'] & JoinedFields<T>)[];
}

And in our code, we can declare a query serenely thanks to the type system preventing us from writing a query with invalid fields:

const query: Query = {
  model: 'users',
  fields: [
    'email',          // ✓ users field
    'orders.total',   // ✓ joined field
    'orders.status',  // ✓ joined field
    'invalid.field'   // ✗ Type error
  ]
};

The real implementation has more conditionals for safety, but the core idea remains the same: extract the join names, then use template literals to compose the field paths.

Trade-offs

Benefits:

• Zero maintenance—add a field to the schema, get autocompletion everywhere
• Catches typos at compile time instead of runtime
• Scales to hundreds of field combinations without manual work

Drawbacks:

• Type computation can slow down in massive schemas
• Error messages become cryptic when deeply nested
• Only works with predictable string patterns

Notes

This pattern shines when you control the schema and need to generate predictable string patterns. It’s less suitable for user-defined schemas or when you need runtime validation anyway. The sweet spot is internal APIs where you want compiler-enforced consistency between your types and runtime behavior.

What you don’t write is as important as what you do. No manual string unions. No keeping field lists in sync. Add a new join relationship, and TypeScript immediately knows about every possible field combination.

When Ghostty came out, I wrote this blog post which contained a few screenshots. I forgot how I managed to have both light and dark mode work seamlessly in Eleventy, so I figured I’d write it down for future me or anyone that sees this.

The Pattern

I use an Eleventy shortcode that conditionally renders different images based on theme preference. The basic version wraps images in figure tags with captions:

config.addShortcode('image', (src: string, alt: string, withDarkMode = false) => {
  const lastDotIndex = src.lastIndexOf('.');
  const basePath = src.slice(0, lastDotIndex);
  const extension = src.slice(lastDotIndex);

  const lightPath = `${basePath}-light${extension}`;
  const darkPath = `${basePath}-dark${extension}`;

  return `<figure>${
    withDarkMode
      ? `<img class="not-dark:hidden" src="${darkPath}" alt="${alt}">` +
        `<img class="dark:hidden" src="${lightPath}" alt="${alt}">`
      : `<img src="${src}" alt="${alt}">`
  }<figcaption>${alt}</figcaption></figure>`;
});

Smooth Transitions

The images would swap instantly when toggling themes, which felt jarring. Adding Tailwind’s transition utilities creates a cross-fade effect:

withDarkMode
  ? `<img class="not-dark:hidden not-dark:opacity-0 dark:opacity-100 starting:dark:opacity-0 transition-discrete transition-opacity duration-1000" src="${darkPath}" alt="${alt}">` +
    `<img class="dark:hidden dark:opacity-0 not-dark:opacity-100 starting:not-dark:opacity-0 transition-discrete transition-opacity duration-1000" src="${lightPath}" alt="${alt}">`
  : `<img src="${src}" alt="${alt}">`

The starting: modifiers prevent the fade animation on initial page load, which I borrowed from how Tailwind’s own docs handle theme switching.

Usage

Now {% raw %}{% image "ghostty-screenshot.png" "Terminal with ligatures" true %}{% endraw %} automatically looks for ghostty-screenshot-light.png and ghostty-screenshot-dark.png, then smoothly transitions between them when you toggle themes.

Trade-offs

Benefits:

• Smooth 1-second cross-fade between theme variants
• Both images preload—no flicker on theme switch
• Simple to maintain with clear naming conventions

Drawbacks:

• Downloads both variants even if user never switches themes
• Requires maintaining two versions of every screenshot

Notes

This approach works well for static sites where you control every image. For dynamic content or high-traffic sites, you’d want lazy loading for the hidden variant or CSS filters for automatic adjustments.

References

Figured how to create custom shortcode after reading this article by Anh.

While working on @general-dexterity/cube-records, I needed a way to defineOr use codegen with another library. my domain-specific Cube models in a way that would provide type-safe autocompletion. The solution was ended up pretty simple: global interface augmentation.

The Problem

The lib defines an empty global interface that users can augment with their own cube definitions. However, after publishing, I discovered that tsup was optimizing the empty interface into a type alias during the build:

// What we write
export interface CubeRecordMap {}

// What tsup outputs with dts: true
export type CubeRecordMap = {}

// Now augmentation fails
declare global {
  interface CubeRecordMap { // Error: can't augment a type alias
    users: { /* ... */ }
  }
}

The Solution

Adding a dummy property prevents this optimization:

export interface CubeRecordMap {
  __empty: {
    measures: {};
    dimensions: {};
    joins: [];
  };
}

That __empty property isn’t arbitrary - it ensures the interface remains augmentable through the entire build pipelineThis is specifically a build tool optimization when generating declaration files, not a TypeScript language limitation. .

Implementation

Users can now extend the interface in their projects:

declare global {
  interface CubeRecordMap {
    users: {
      measures: { count: { type: number } };
      dimensions: {
        id: { type: string };
        email: { type: string };
      };
      joins: readonly ['orders'];
    };
    // ... other models and views
  }
}

The library extracts type-safe cube names and fields:

type CubeRecordName = keyof CubeRecordMap;
type CubeRecordMeasure<T extends CubeRecordName> =
  keyof CubeRecordMap[T]['measures'] & string;

Trade-offs

Benefits:

• Interface stays augmentable through the build pipeline
• Zero runtime overhead - types exist only at compile time
• Incremental cube definitions with immediate IDE support

Drawbacks:

• The interface includes a dummy property
• Requires some documentation to explain the pattern

Notes

This isn’t just for my personal benefitsEven though it was pretty fun to try to figure this out. , this pattern provides domain-specific autocompletion without runtime cost. Instead of searching through Cube models for field names, TypeScript provides instant feedback as you type.

While writing my Ghostty blog post, I wanted syntax highlighting for the configuration snippets. Since Shiki doesn’t support Ghostty’s config format, I added a custom language definition.

const ghostty = {
  name: 'ghostty',
  scopeName: 'source.ghostty',
  fileTypes: ['ghostty'],
  repository: {},
  patterns: [
    {
      include: '#strings',
    },
    {
      name: 'comment.line.number-sign.ghostty',
      match: '^\\s*#.*$',
    },
    {
      match: '\\b(font-family|font-feature)\\b',
      name: 'keyword.other.ghostty',
    },
  ],
};

I only defined patterns for comments (#) and the font keywords I was actually using. The grammar follows TextMate rules but doesn’t try to implement the full Ghostty spec - just enough for my examples.

highlighter = await createHighlighter({
  themes: ['github-dark', 'github-light'],
  langs: ['ruby', 'javascript', 'json', 'elixir', 'typescript', ghostty],
});