The three open-source AI commits that shipped on April 25-26, 2026 and made my Mac agent better at seeing itself

Saturday, April 25

The analytics patch that exposed 96.4% of self-reported user feedback as silent log uploads

The fastest way to ruin a feedback funnel is to fire the same PostHog event for two genuinely different things. Fazm has a floating action bar with an exclamationmark.triangle button labelled Report an issue. The intent is one-click log upload, no form. The implementation, until April 25, called FeedbackWindow.sendSilently(), which fired Feedback Opened and Feedback Submitted (length 0) in the same tick. From the funnel's point of view a silent log upload looked indistinguishable from a real form-based complaint.

What the commit body actually says, verbatim: 133 of 138 Feedback Submitted events in the last 30 days had length 0. Almost the entire stream of inbound feedback was the floating-bar button, not the modal. Anyone reading the dashboard before April 25 would have concluded that users were submitting empty complaints, when in fact users were uploading logs through a completely separate code path.

The fix is two characters of intent and a property bag. The Swift call sites for the floating bar pass source: "silent"; the modal-form code path passes source: "modal" (the default). Now the funnel can be filtered to real form-based feedback and the dashboard stops lying.

// Desktop/Sources/PostHogManager.swift, before commit 2fbc891c
func feedbackOpened() {
    track("Feedback Opened")
}

func feedbackSubmitted(feedbackLength: Int) {
    track("Feedback Submitted", properties: [
        "feedback_length": feedbackLength
    ])
}

This is the boring kind of open-source AI work that no roundup ever covers and that every product eventually has to do. Telemetry hygiene is half the surface area of any agent that shipped a year ago. Most of the user-feedback decisions a consumer AI app makes downstream of release week are made on dashboards. If the dashboard is wrong, the decisions are wrong.

Sunday, April 26 at 14:12 PT

v2.4.2 went out, mostly papering over the seams of a busy week

The release itself is a CHANGELOG.json bump (commit bdf57afd) and a Codemagic build trigger. What it carried into users' hands is small but intentional. Read these the way you would read a diff in a long-running open-source repo: each line is the shape of a complaint someone filed.

The third bullet is the one that connects to the wider news cycle for that weekend. The Custom API Endpoint accepts any Anthropic-compatible base URL. Until v2.4.2 the help text implied that meant a corporate gateway or a GitHub Copilot proxy. Naming local LLM bridges turns the field into the official integration point for whatever Hugging Face checkpoint you happen to be running on your own hardware. Point ANTHROPIC_BASE_URL at a local serving layer, restart the app, and the same accessibility-tree tool surface rides along behind whichever model you point at it.

The Smart/Opus mapping fix is a smaller story but a familiar one. When a backend renames a model ID, every client that stored the old key has to translate at read time or the preference goes stale. The fix is a one-line lookup that rewrites the persisted opus key to the new ACP ID before comparing against the live available-model list. Boring plumbing, the kind that exists in every cross-version client that wants to survive its own backend.

Sunday, April 26 at 16:02 PT

The six-line schema doc that taught the agent how to read its own observer_activity log

Two hours after the release went out, commit be29a3c7 landed. The diff is six lines of natural language inside a tool description string in acp-bridge/src/fazm-tools-stdio.ts. No new code paths. No new endpoints. Just words.

The SQL tool runs against the local fazm.db SQLite database. One of its tables is observer_activity, which logs every Card-style nudge the observer surfaced (a meeting reminder, a task suggestion, a stale-tab notice). The table's columns are id, type, content, status, userResponse, createdAt, actedAt. The catch: the human-meaningful fields (the card's title, body, the option labels the user could click) live JSON-encoded inside content, not as siblings. Before April 26, when a user asked what did I dismiss yesterday, the agent would write SELECT body FROM observer_activity WHERE ... and bounce off a column-not-found error. After April 26 the tool description tells the model the truth.

// acp-bridge/src/fazm-tools-stdio.ts, before commit be29a3c7
description: `Run SQL on the local fazm.db database.
Supports: SELECT, INSERT, UPDATE, DELETE.
SELECT auto-limits to 200 rows. UPDATE/DELETE require WHERE.
DROP/ALTER/CREATE blocked.
Use for: app usage stats, time queries, task management,
aggregations, anything structured.`,

What does this look like at runtime, end to end? When the user asks what reminders fired this morning, the bridge sends the updated tool description with the SQL request. The model writes a SELECT json_extract(content, '$.title') query. The bridge runs it against fazm.db, gets back rows, and pipes them back through the tool result. The model now has the real card titles, not opaque JSON blobs.

This is the smallest possible kind of code change with the largest possible kind of effect on agent behaviour. It is the difference between a model that bluffs through a schema and a model that reads the schema and quotes it back. None of this looks like AI work in the way the news cycle defines AI work. All of it is what the news cycle's output runs on, once it reaches a real machine.

Why the boring layer matters more than the headline layer

Every weekend in late April 2026 produced a wave of new checkpoints, new fine-tunes, new arXiv preprints. None of them showed up on a Mac without a client that knew how to talk to them. The client work is rarely flashy. It is help text on a settings field, a property name on an analytics event, a schema-notes paragraph in a tool description. It is the work that decides whether the model on the other end of your local API endpoint can actually answer the question you asked.

The pattern across these three commits is the same: be honest about what your software actually sees. The PostHog patch stopped the dashboard from claiming users had submitted 138 empty complaints when they had really uploaded 133 logs. The v2.4.2 help text stopped pretending the Custom API Endpoint was only for corporate proxies. The schema-notes paragraph stopped letting the agent guess at table layouts. None of those is a model release. All of them are open-source AI updates in the only sense that matters once the model has finished downloading.

The repo is at github.com/m13v/fazm. The four files referenced in this piece are Desktop/Sources/PostHogManager.swift, Desktop/Sources/FeedbackView.swift, acp-bridge/src/fazm-tools-stdio.ts, and the top of CHANGELOG.json. Every claim in this article points at one of them.