New AI projects on Hugging Face or GitHub, May 28 2026. One repo changed a single model ID and broke every long chat, then fixed it the same evening.

Neither platform publishes a dated release list. Models, datasets, papers, and repositories surface continuously, ordered by a rolling trending score rather than by calendar date. The closest thing to a true daily record is a tagged release or a commit log, where each change carries an author, a message, and a timestamp. Two records that genuinely carry 2026-05-28: huggingface_hub v1.17.0 was uploaded that day with cross-repo file copies, ssh-to-Spaces, and a security fix that writes token files at 0o600 instead of world-readable defaults; and the open-source macOS agent Fazm pushed 54 commits and cut five patch releases (v2.9.48 through v2.9.52). Fazm's standout change was a one-line default-model bump from claude-opus-4-7 to claude-opus-4-8 in ShortcutSettings.swift, plus the bug fix that bump forced: commit e4cb4f09, which stops poisoned thinking-block history from returning HTTP 400 on the next message.

Because the new default, claude-opus-4-8, is an extended-thinking model, and Fazm runs it over the Agent Client Protocol bridge at version 0.37.0. Extended-thinking models emit signed thinking blocks. If a turn is cancelled or stalls mid-thinking, the SDK leaves an assistant message whose thinking block was never finalized with a signature. That unsigned block lives in the session history, in memory and on disk. The very next prompt replays it, and the Anthropic API rejects the entire request with `400 ... thinking ... blocks ... cannot be modified`. The plain session-resume retry just replays the same poison and 400s again, which is what 'it destroyed the whole conversation' looked like to a user. The model ID was one line; the blast radius was every long chat that had ever been interrupted.

A regex test against the error message and a reroute. The detector is `isThinkingBlockPoison`, which matches either /thinking[\s\S]{0,40}blocks[\s\S]{0,80}cannot be modified/i or /blocks in the latest assistant message cannot be modified/i. When that matches and a sessionId exists and retry depth is under MAX_QUERY_RETRIES, the bridge unregisters the cached session, clears the pending tools and timers, sets msg.resume to undefined, records the stuck session id, and calls handleQuery again at depth + 1. Because msg.resume is now undefined, the retry goes through session/new and replays priorContext (plain user and assistant text, no signed thinking blocks) instead of resuming the poisoned session. The fix is +55 lines, -5 lines, all in acp-bridge/src/index.ts.

Because priorContext is plain text. A session resume reloads the full message history including the signed thinking blocks, and the unsigned, mid-cancel block among them is what the API refuses to accept. priorContext is the conversation re-expressed as ordinary user and assistant turns with no thinking blocks at all, so there is nothing for the API to reject. The trade-off is that the freshly started session loses the model's private reasoning trace from earlier turns, but it keeps every visible message. To the user the conversation continues; to the model it is a clean session seeded with what was said, not what was thought.

The bridge dependency @agentclientprotocol/claude-agent-acp moved three times on May 28. It went to 0.37.0 at 11:59 PDT (commit b92b19b0), was downgraded to 0.29.2 at 18:48 (commit 1b0f9086) when the new version started surfacing the thinking-block rejections, then went back to 0.37.0 at 19:08 to 19:10 (commits 931775c0 and b860a170) once the poison handler in e4cb4f09 landed at 19:09. The downgrade-then-re-upgrade is the visible shape of a maintainer isolating whether the failure was the protocol version or the session state. It was the session state; with the handler in place, 0.37.0 was safe to keep.

Most of the other 54 commits were performance work. v2.9.47 and v2.9.49 chased an idle CPU storm: four withAnimation(.repeatForever) patterns (an update-available pulse, an update-button pulse, an upgrade-Claude pulse, and a hang-flash) were keeping a SwiftUI animation transaction open forever, recomputing the chat scroll view every display cycle and pegging the main thread. The fix disabled those animations and added automatic stack-sampling: when the top thread sustains over 85 percent CPU for a minute, the app shells out to /usr/bin/sample and writes the hot-stack trace to the log. v2.9.49 also turned session recording and screen observation off by default. v2.9.48 started reporting Fazm-paid LLM usage with the real model id and real Gemini token counts. v2.9.51 fixed a forever-pulsing Update Available indicator and the optional-sign-in experiment. v2.9.52 closed the day with the chat-stuck-loading fix and the referral Copy Link button.

It is a real model id, and the change was more than a label. The diff in ShortcutSettings.swift swaps the Smart picker's default id from claude-opus-4-7 to claude-opus-4-8 and updates the normalizeModelId function so the legacy 'default' alias and any prior-pinned Opus version (claude-opus-4-5, claude-opus-4-7) all resolve to the new canonical id. That normalization matters because the bridge no longer advertises the bare 'default' alias; without the mapping, an old pinned id would silently fall back to whatever the SDK treats as default, which is not guaranteed to be Opus. So the one-line default change is backed by a migration path for every previously stored model preference.

Three artifacts, in order. Open the project's CHANGELOG.json or CHANGELOG.md at the root and read the entries dated 2026-05-28. Open the commit log with `git log --since=2026-05-28 --until=2026-05-29 --no-merges` and read the messages, ignoring the release-changelog commits. Then pick one commit whose subject sounds load-bearing (here, 'Fix 400 errors from poisoned thinking-block session history') and run `git show <hash>` to read the diff and, more importantly, the comment the author left for future readers. The README is marketing; the diff is the truth. A model-version bump that looks like a one-line change is often the most dangerous kind, because the line is small and the consequences are not.

github.com/mediar-ai/fazm is the public repo. The poison handler lives in acp-bridge/src/index.ts inside handleQuery, in the block that defines isThinkingBlockPoison (commit e4cb4f09). The default-model change is in Desktop/Sources/FloatingControlBar/ShortcutSettings.swift, in defaultModels and normalizeModelId (commit ac5b4f81). The ACP version moves are in acp-bridge/package.json. Release notes are in CHANGELOG.json at the repo root. For the Hugging Face side, the v1.17.0 notes are at github.com/huggingface/huggingface_hub/releases.