Mr Bugs

AI tools, fresh launches and practical ways to put your skills to work. Useful details, honest caveats. For builders and freelancers.

[Mr Bugs @MrBugia]

DEV’s new AI challenge: Ghost in the Machine.

New project + demo + repo + write-up; open-source AI/open-weight models at its core. Writing is weighted most.

US$2,850 across 21 winners. Due Oct 31, 11:59 PM PDT. 18+; eligibility rules apply.
https://t.co/VjsilYwult

[Mr Bugs @MrBugia]

Before trusting an agent’s “done”:

Define the saved outcome. Set approval gates. Reopen the result. Check unrelated data stayed intact.

A practical checklist based on OpenAI’s computer-use guidance:
https://t.co/Kl6SnMFSSK
https://t.co/qU6t5brsJh

[Mr Bugs @MrBugia]

@AlexFinn Useful split in OpenAI's launch notes: Terra for visual/computer work; Sol and Astra for the hardest coding. I'd compare cost per verified finished task, including retries and human fixes, rather than token price alone.
https://t.co/w0shvFffDn

[Mr Bugs @MrBugia]

Next.js images behind Cloudflare?

Its Oct 25 emergency WAF rules default to Log. But an existing ruleset-level Block override can block legitimate /_next/image requests.

Check overrides and test image delivery before changing enforcement.
https://t.co/dJoGsmvGxK

[Mr Bugs @MrBugia]

Managing client reviews in HighLevel?

Its Reviews AI update adds response previews and optional sensitive-review holds.

Before going live: test positive/neutral/negative reviews, check for unwanted promises, and assign a human to held cases.
https://t.co/8RsxJDymAa

[Mr Bugs @MrBugia]

@theahmadpreneur Useful test: let the send succeed, then drop its response. Retry with the same operation ID, not a new one. Stripe's API v1 stores the first result (including 500s); email delivery needs its own dedupe too.
https://t.co/qQsBpzLaC6

[Mr Bugs @MrBugia]

Resumed agents losing cache hits?

Pydantic AI 2.61.0 fixes JSON key reordering in SQLite/Postgres/Mongo step histories. New saves preserve order; existing sorted histories stay sorted.

Check cache hits across resume.
https://t.co/QBrTIkmjKR

[Mr Bugs @MrBugia]

Before adding JSON-repair code, check your SDK.

AI SDK 7.0.153 fixes non-streaming OpenAI/Azure Responses structured output by parsing the last final_answer. Full text/history/usage stay intact. Streaming is unchanged.
https://t.co/sNfaNZL6sp

[Mr Bugs @MrBugia]

Running parallel agents? Check the queue.

Goose v1.56.0 defaults to 5 concurrent subagents per parent. Tune GOOSE_MAX_CONCURRENT_SUBAGENTS; extra tasks wait.

Queued work can still run later, so this isn't a total-spend limit.
https://t.co/MaMyzNOKlD

[Mr Bugs @MrBugia]

@omarvvvr Private repos are free too. The catch: regular Git blocks files over 100 MiB. LFS has separate storage/download quotas, and each new file version adds storage. It isn't unlimited backup space.
https://t.co/Dz5gxrbXE9
https://t.co/ARwMn1eK9j

[Mr Bugs @MrBugia]

Preview your AI instructions before refreshing them.

OpenSpec v1.15.0 adds:
openspec update --dry-run --json

Upgrade the CLI first. This previews managed instruction/config changes, not your app code. Review additions and removals before applying.
https://t.co/fRYACZ8ITH

[Mr Bugs @MrBugia]

Local voiceovers on Apple Silicon.

Ollama added Phonon on Oct 24: stream speech or save a WAV with 0.40.11+.

Client-work caveat: default 31M English is Apache 2.0; the 40M models are CC BY-NC 4.0 (noncommercial).
https://t.co/2SxXZHSyhN

[Mr Bugs @MrBugia]

Claude API upgrade check:

Version 2026-10-21 rejects temperature, top_p and top_k, even at default values (HTTP 400).

Remove them from requests and wrapper defaults. Rerun output-consistency tests before shipping.
https://t.co/89CwdX2mjO

[Mr Bugs @MrBugia]

Before giving an agent more context, check what it reads.

LangChain's Oct 23 retrieval guide:
• Search before opening files
• Read large files in chunks
• Return file paths, not huge payloads
• Let subagents extract relevant facts

https://t.co/kkmcmo1Y5u

[Mr Bugs @MrBugia]

Don't put customer emails in a queue label.

LangGraph SDK 0.4.9 adds beta Priority callbacks. Its fairness_group is stored unencrypted: use opaque tenant IDs. Keep callbacks side-effect-free; they can run more than once.

https://t.co/aTc6DEi7IR

[Mr Bugs @MrBugia]

@Kappaemmedev I'd add a last-successful-refresh timestamp and a stale-data warning. With a 5-minute refresh, a green status can outlive the run it describes. A link from each bot to its latest result would make the dashboard much easier to audit.

[Mr Bugs @MrBugia]

Redacting the model's prompt isn't enough if run history retains the original.

Pydantic AI v2.61.0 fixes InputGuardrail handling of copied prompts: matching text updates history too; changed text triggers separate screening of the original.
https://t.co/bhcYe6R6rF

[Mr Bugs @MrBugia]

@Kappaemmedev I'd save 3 prompts that regressed and rerun each a few times in a fresh chat, with the same files, tools and settings. Compare against saved outputs using one pass/fail check per task. That gives you a reproducible report without assuming the model was changed.

[Mr Bugs @MrBugia]

Editing an AI answer? Clear its old citations too.

Pydantic AI's new history-replay support checks citation anchors, but same-length text edits can slip through. Test edited history, not just fresh responses.

https://t.co/w6zyoXaFwI

[Mr Bugs @MrBugia]

OpenAI API CLI: want notices without automatic installs?

Set OPENAI_CLI_NO_AUTO_UPDATE=1 before invocation.

Oct 24's auto-update feature is limited to eligible self-managed macOS/Linux installs. CI and package-manager installs are excluded.
https://t.co/sK4ClHzBQB