The one thing

LiteLLM 1.103.0 deprecates the `litellm-proxy` command. Scripts that call it should switch to `lite`.

The release deprecates the litellm-proxy entrypoint in favour of lite, and renames lite autoroute up/down to start/stop, keeping the old names as deprecated aliases.¹ According to the pull request, litellm-proxy now prints one deprecation line on stderr and otherwise runs as before.² Neither page gives a removal date.

The same release removes a duplicate user-budget hook that returned 429 on zero-cost models.¹

VERDICT · BREAKING — Nothing fails today. Replace litellm-proxy with lite in scripts, Dockerfiles and CI now, and check any job that treats stderr output as a failure.

Shipped

Breaking

worktrunk 0.80.0 runs commit hooks on a detached HEAD during a squash

When a commit hook or signing stops the commit in wt merge or wt step squash, the branch is now left untouched instead of reset to the merge base. The cost is that git commit hooks now run on a detached HEAD during a squash.³ Separately, the branch field in wt config state logs --format=json now holds the real branch name, such as feature/x, instead of the sanitised directory name.³

MATTERS TO · Anyone running worktrunk to manage agent worktrees with commit hooks that read the current branch, or scripts that parse its hook logs.

nofeed.dev/issues/2026-09-28/worktrunk-0-80-0-squash-hooks/

Library API changes in the same release

Repository::capture_refs_with_ahead_behind is removed; git::parse_porcelain_z and Repository::changed_files now use bytes; GitError gains StaleWorktreeHoldsWork.³ This matters only if you import worktrunk as a Rust crate.

The conversation

The claim — including ours

Hacker News read Fireworks' Ember-1 as a cost story rather than a model story.

From the floor

tukHelix, Hacker Newsreader-posted, unverified
Welcomed a cheaper open model but said a model-training team inside an API provider made them less sure about using that provider.
netvarun, Hacker Newsreader-posted, unverified
Said that after a recent price cut, GPT-6 Sol beats Kimi K3 on both quality and cost in their internal tests, and quoted $2/$10 against $3/$15.

Our take

Ember-1 is Kimi K3 trained to reason less. Fireworks reports 34.5% fewer total tokens at a score of 0.753 against K3's 0.750, on its own table and in A/B tests with two unnamed customers.⁴ That is a cost-per-task claim, the measure we argued for in No. 39, but the vendor measured it. We did not read a price page, so the per-token figures above stay a field report.

Hacker News was swept in full: 24 stories in the window, 7 above 80 points, with comments. Vendor changelogs and status pages (12 feeds) ran complete with no entries in the window. arXiv, 48 hours: 200 papers fetched, none kept. The X layer ran on background only: the newest roster digest (100 accounts) and Reddit digest were collected 27 September 05:30 UTC, about 24.5 hours old. The ai-dev-x keyword file holds a placeholder instead of posts, and the Reddit fallback search returned no posts.

Skip this

Ember-1 benchmark results Published 23 September, outside the window; the headline table is Fireworks scoring its own model.

OpenAI Codex 0.158.0 Drops GPT-5.4 from the bundled catalog with a migration prompt and adds MCP OAuth client secrets. Nothing breaks.

Unsealed briefs in Authors Guild v. OpenAI 610 points, but it is a court filing. Nothing a developer installs or pins changes.

There are no "rogue" AI agents An essay about who is liable. No new incident or measurement.

TinyAIArena A Show HN spectator game with no benchmark method.

Everything we saw

18 candidates scanned · 3 used in this issue — the rest, with the reason each one was left out

The receipt. Ranking is only trustworthy if the discarded pile is visible, so here it is: everything the collectors surfaced in the window, with its signal and what we did with it.

Item	Source	Signal	Call
BerriAI/litellm v1.103.0	GitHub Releases	Deprecates litellm-proxy entrypoint	Led the issue
max-sixty/worktrunk v0.80.0	GitHub Releases	Three items marked Breaking	Printed as Breaking
Introducing Ember-1 (HN, 412 points)	Hacker News	Blog dated 23 September	Used in the conversation
openai/codex rust-v0.158.0	GitHub Releases	GPT-5.4 removed from bundled catalogs; approval default for elevated commands	No breaking section; watching
sst/opencode v1.18.33; vercel/ai 7.0.118, workflow 2.0.49, workflow-harness 1.0.128	GitHub Releases	Patch releases	No decision changes
BuilderIO/agent-native v0.1.271, clips 0.1.319 and 0.1.320	GitHub Releases	Patch cadence	No decision changes
pacifio/atlas alpha-0.3.4, Atlas Teams Early	GitHub Releases	Empty release notes; new to the watchlist today	Nothing to evaluate
Law firms and AI discounts (NYT); SNL on Anthropic; chat templates and self-reference	Hacker News	105 to 181 points	Off the beat
Vendor changelogs and status pages	12 feeds	Sweep complete, zero entries	Quiet weekend
ai-dev-x keyword digest, 2026-09-27	X digest	File present, contents a placeholder	Source unreachable

EOF

End of feed. That is everything from the window worth your time.
Explore more news, analysis and practical guides on NO FEED.

Sources — each with the date it was read01 LiteLLM v1.103.0 release notes (2026-09-28) · 02 BerriAI/litellm PR #41673: deprecate the litellm-proxy entrypoint (2026-09-28) · 03 worktrunk 0.80.0 release notes (2026-09-28) · 04 Fireworks: Introducing Ember-1 (published 23 September 2026) (2026-09-28)