Hugging Face new models or GitHub trending AI, May 30 2026. The shipment worth reading was a session that got stuck on “Compacting...” forever, and the six checks that freed it.

What the log actually showed

The bug had a single, reproducible signature. A pop-out chat in session 0ea67ea0 resumed after a prior heavy turn had unwound a live subagent mid-flight. From then on, every resume of that session did the same four things and then stopped. This is the trace, condensed from fazm.log on 2026-05-30.

Read the fourth line carefully. The prompt resolved with stop reason end_turn in zero milliseconds, with zero output tokens and zero assistant text. That combination is not a model deciding to stay quiet. It is the Claude Code SDK underneath short-circuiting the turn without ever calling the model, and then reporting the dead turn as an ordinary success. The placeholder chunk had already streamed to the window, so the UI sat on “Compacting...” with nothing coming after it.

The hard part: an empty success that is real, versus one that is poison

You cannot just retry every empty end_turn. Per the Agent Client Protocol, end_turn is a success: a model is allowed to finish a turn without saying anything. Retrying that case is wrong, and worse, the retry replays prior context, which can destroy live tool, browser, and subagent state. So the recovery has to fire only on the one signature that proves the model was never called. That is six conditions, all true at once, in acp-bridge/src/index.ts.

Each condition is carrying weight. Remove any one and the guard either misses the bug or starts eating legitimate empty answers.

The two that do the real separating are promptDurationMs < 50 and outputTokens === 0. A genuine empty turn that actually ran the model needs a real API round-trip, which takes longer than 50ms and bills at least one token. A short-circuit resolves in zero and bills nothing. The clock and the meter are what make the distinction provable rather than a guess.

Why the same recovery was deleted two weeks earlier

This is the part the commit log tells you and a trending page never could: the May 30 fix is a re-add. A blanket version of empty-turn recovery existed before and was removed on 2026-05-13 (commit fad9af4f) for two good reasons. It was lossy, because replaying prior context tears down whatever MCP, browser, or subagent state is still live. And it was out-of-spec, because it retried a case the protocol defines as success.

So what changed to make it safe again? Nothing about the danger; only the precision of when it fires. At a 0ms, zero-token resolve, the session is already inert. The heavy turn that mattered ended long ago, and its subagent and tool state went with it. There is nothing live left to destroy by forcing a fresh session and replaying the context, because the thing the blanket version used to break is no longer running. The narrow signature is not a smaller version of the old fix. It is the one window in which the old fix's failure mode cannot occur.

The arc reads cleanly across three dates. On May 13 the blanket recovery came out. On May 29, v2.9.54 shipped a first attempt at the stuck-on-“Compacting...” symptom. On May 30, v2.9.57 added the signature-gated recovery for the resume short-circuit the earlier attempt missed. Three moves on one bug, each visible only because someone read the diff, not the leaderboard.

The rest of May 30, briefly

The stuck-session recovery was the headline, but it shared the day with five other changes across v2.9.56 and v2.9.57. They are worth a line each because they show where an open-source desktop agent actually spends its time.

• Pop-out New Chat purge. A new chat opened from a popped-out window sometimes appeared already in a thinking or spinner state after a stuck turn. A helper now purges the stale streaming state on new-chat handoff (roughly +43 lines in ChatProvider.swift).
• Phantom stall indicator cleared. A “tool not responding” indicator could reappear on the next prompt and show a healthy session as hung for thousands of seconds. A clearStallIndicator helper resets that stale state on stream error or timeout.
• Knowledge graph went local. The Memory page had a dead no-op call to a remote knowledge-graph API. It was removed so a refresh re-reads local storage and recovers a stale database handle instead of doing nothing.
• Gemini failover released. The transparent Claude-to-Gemini failover built the evening before reached users in v2.9.56, so a dead or rate-limited built-in key continues the chat instead of erroring.
• Honest analytics. Empty or dropped replies are no longer counted as successful answers, which is the measurement that made the stuck-session bug visible in the first place.

How to read any project's real May 30 yourself

The method generalizes past Fazm. A trending feed tells you what existed at the end of a day; the diff tells you what someone decided, and why. Three artifacts, in order:

1. Open the project's CHANGELOG.json or CHANGELOG.md at the root and read the entries dated 2026-05-30.
2. Run git log --since=2026-05-30 --until=2026-05-31 --no-merges and read the messages, skipping the release-changelog commits.
3. Pick one commit whose subject sounds load-bearing (here, “Add recovery for poisoned resume compaction stalls”) and run git show 3e4fb0cd to read the diff and the comment the author left for the next reader.

The comment block above the guard is the whole story in one place: the field-case session id, the deleted blanket version, and the reasoning for why this narrow re-add is safe. That is the kind of record a leaderboard cannot hold.

Questions about May 30, 2026