The April 2026 developer-tools update nobody wrote up. The reference ACP agent drops nine kinds of SDK events.

The public SERP is dominated by per-vendor recaps: OpenAI's Codex 2.0 expansion into computer use and SSH devboxes, the WordPress 7.0 release cycle being extended to land Real-Time Collaboration, Apple requiring watchOS apps uploaded after April 2026 to be built with the watchOS 26 SDK and include 64-bit support, Microsoft's Xbox mode going generally available on Windows 11 PCs from April, the WordPress @wordpress/build esbuild replacement for webpack, and weekly Chrome DevTools release notes. Those are all real. They are also all at the app or vendor layer.

AI coding agents are quietly converging on a single transport, Agent Client Protocol (ACP). Claude Code publishes an ACP adapter via @agentclientprotocol/claude-agent-acp. Zed speaks ACP natively. Fazm's desktop app ships an ACP bridge so any ACP-speaking agent can drive it. The effect is that the thing you build an integration against is the protocol, not the vendor. That story is not in any of the April 2026 recaps, but it is the reason those recaps will start looking repetitive next month.

Because the reference agent drops events. Read acp-bridge/src/patched-acp-entry.mjs in the Fazm source. It imports ClaudeAcpAgent from @agentclientprotocol/claude-agent-acp/dist/acp-agent.js, monkey-patches ClaudeAcpAgent.prototype.createSession, and wraps session.query.next() so it can forward nine SDK event categories the upstream agent silently swallows: rate_limit_event, system/compact_boundary, system/status, system/task_started, system/task_notification, system/api_retry, tool_progress, tool_use_summary, and compaction stream chunks (content_block_start and content_block_delta with a compaction shape). Without that wrapper, the UI sits idle while the SDK is actually rate-limiting, compacting context, or retrying an HTTP 529. The file is 267 lines. The dependency version in acp-bridge/package.json is @agentclientprotocol/claude-agent-acp ^0.29.2.

rate_limit_event. The Claude Agent SDK has fired it since late March 2026 with an eight-field rate_limit_info payload (status, resetsAt, rateLimitType, utilization, overageStatus, overageDisabledReason, isUsingOverage, surpassedThreshold). The reference ACP agent does not propagate it. In the Fazm patch, lines 132 to 152 forward it as a sessionUpdate of type rate_limit so the Swift UI in ACPBridge.swift can render a five-hour vs seven-day window indicator and an overage status chip. If you are using the reference agent, the UX difference on a rate-limit hit is no-signal vs. progressive-warning.

Directly. Fazm drives Mac apps through real macOS accessibility APIs (AXUIElement, axorc-style trees), not screenshots. For that to be useful, the desktop chat layer has to faithfully show what the underlying agent is doing. If the agent is compacting 180k of context, the UI has to say compacting, not just spin. If the agent is on api_retry attempt 3 of 5 with a 529 status, the UI has to say that, not hide it. The patched entry is what makes accessibility-native automation legible. Accessibility APIs are how we touch the app; faithful event forwarding is how the human knows why we are not.

Yes. Fazm is open source at github.com/mediar-ai/fazm. The file is acp-bridge/src/patched-acp-entry.mjs. The key landmarks are: line 16 imports ClaudeAcpAgent from @agentclientprotocol/claude-agent-acp/dist/acp-agent.js, line 21 overrides createSession, line 39 overrides session.query.next, lines 62 to 129 forward the five system/* subtypes, lines 132 to 152 forward rate_limit_event, lines 156 to 176 forward tool_progress and tool_use_summary, lines 182 to 201 forward compaction stream chunks, and line 212 onward patches prompt() to attach costUsd and usage to the return value so the Swift side can render real per-turn cost in the floating bar.

The patch is additive, not destructive. Every call ultimately delegates to the real ClaudeAcpAgent; the wrapper only re-emits dropped events as additional sessionUpdate messages. That means it coexists with any upstream fix without conflicting. As @agentclientprotocol/claude-agent-acp adds native emission for these events (work is ongoing upstream), Fazm's wrapper can drop the corresponding branches without changing the public ACP surface the Swift app consumes. So the patch is a safety net, not a fork.

Fazm 2.1.2 (released April 7, 2026, per CHANGELOG.json) says: 'Upgraded ACP protocol to v0.25.0 with improved error handling for credit exhaustion and rate limits.' In practice that means the ACP sessionUpdate payload now carries structured errorType values like billing_error and rate_limit (see acp-bridge/src/index.ts lines 1901, 1963, 2032 and the rate_limit case at 2443), so the onboarding flow in Fazm stops silently failing when built-in credits run out, as it did before 2.1.2. The same version fixed onboarding silently failing when credits run out.

That the reference agent's silence is not the SDK's silence. If you ever thought Claude Agent SDK sessions go dark during rate limits or context compaction, they do not. The SDK emits real events; the default ACP adapter just does not forward all of them. Copy the pattern from patched-acp-entry.mjs: monkey-patch createSession, wrap query.next(), inspect item.value.type and item.value.subtype, re-emit as sessionUpdate. That pattern is reusable on top of any ACP-compatible agent.

It matters because it explains why Fazm's floating bar feels responsive under load where other desktop agents feel dead. When you send a long query, you see compaction markers. When the SDK retries a 529, you see the retry counter. When you run out of overage credits, you see the exact reset time. That legibility exists because a 267-line wrapper file inside acp-bridge forwards what the reference agent throws away. If your desktop AI agent does not show you those signals, ask it why.