I'm sure it's a coincidence that regular opus got slower today :sadface:

my current working theory is Claude Opus estimates plan phases to take a bit under 200K tokens because that's its default window. we consistently complete phases at 192-198K with compaction off. with compaction on in CC it has to basically start over at 150K or so and loses momentum

maybe coding becomes even more like music. anyone can sit down at the piano and make sound and some of it will sound good, but creating a specific vibe requires mastery

SWE: <clutching pearls> but the code slop! SRE: <exhaling smoke> you haven't seen what we have seen. slop is as old as time, maybe older. some of us believe slop is how the Universe began. we make your slop run, or a bot's slop run, it's all the same problems, different day

a younger me would be distraught about how much attention is going to autonomous AI agents, but now I see this as an opportunity while folks figure out what has always been true: agents are cybernetic experiences. the model is (as models always have been) a cognitive amplifier

I did something at work that's so cool I can't talk about it yet 😁

that "confabulations" flag is underrated — having the model catch its own hallucinated stubs from earlier sessions is wild. I've been running similar sweeps on multi-agent codebases and it surfaces stuff linters completely miss. plan mode is genuinely the unlock.

BDSM enjoyers LLM enthusiasts 🤝 “Let me show you my custom harness”

this is quite an undertaking biggest thing it brings is a much better security model. Monty lets you have tight control over network, filesystem, and (soon) which modules are allowed also, extremely fast startup time

here's an example where I will interrupt the model to redirect it. this plan is building e2e tests around multi-context drifting and it noticed some silly mismatches I'd like to fix. if I let this go it'll happily wrap those differences up in adapter code I do not want

if I had all the stuff you bet your ass I would build my own trombones. I have all the stuff to build agentic software, so obviously I have several irons in the fire

Opus 4.6 is great but still better with Gemini as a sidekick

have opus 4.6 review your code. do it. get in plan mode and go "let's do a sweep of the repo for tech debt, stubs, confabulations, crusty docs, etc." and let it rip

interesting. opus 4.6 will sometimes call cpal (claude as an mcp) when it should use its own subagent. I don't usually have it loaded in CC but it is occasionally handy to have the simpler system prompt maybe claude likes it too

opus 4.6 whips

working well with CC today: phased plan, compaction off, at the end of each phase go back into plan mode and prompt "update the plan with current status then proceed". It will clear context and start fresh with the plan. way better than compaction imo

think I'll call this next trick "multi-context drifting"

ok here's a cool use case for the model, we found a cache called "high-school-files.zip" which includes a bunch of stuff I'd forgotten about, and it ripped the text out of the 30+ year old wordperfect files. I believe there is also a netware binary that gave me superuser on the school network? lol

Sir Ian McKellen performing a monologue from Shakespeare’s Sir Thomas More on the Stephen Colbert show. Never have I heard this monologue performed with such a keen sense of prescience. Nor have I ever been in this exact historical moment.TY Sir Ian, for reaching us once again. #Pinks #ProudBlue

life with gpal: "run this idea by gemini"

today's exploratory prompts are both from a mini-epiphany about kaijutsu right before bed: what if "oops, all MCP"? so I have bots looking into that while I rebuild the UI closer to its final form, which is turning out to be WindowMaker for AI, with shaders. we'll see in a few hours :)

Released Hyprstream v0.2.0 today. This release includes model downloads, Git worktree management, basic LLM inference and training for Qwen3 dense models. Released as AppImages for CPU, CUDA, ROCm, and Universal. direct links to follow... github.com/hyprstream/h...

so the new /insights thing in claude code says +5,512,017 / -2,012,677 lines changed on personal acct so I questioned that. it counted some autogenerated model code. the actual output is still pretty cool, 258KLoC extant code across 8 projects, 200KLoC of rust, all of it builds, compiles, and runs

Anthropic just took a big swipe at OpenAI's decision to put ads in ChatGPT. Anthropic is airing ads mocking ChatGPT ads during the Super Bowl, and they're hilarious 😅 Anthropic is also committing to no ads in Claude www.theverge.com/ai-artificia...

work project is deep in a goop phase, while personal project emerges from one. goop phase is when the agent struggles mightily and tries to give up, context burning like a christmas tree in july. you gotta nudge it through or it'll do more goopy code

it surprises me that CC continues to footgun itself with its compaction at 150K. as usual that's probably right for general population but I get a *lot* of tasks finished in that last 50K so compaction isn't needed at all. plan well - using claude! - and compaction doesn't happen at all

I'm doing a big gui refactor, so I pushed claude plan mode to stay strategic and write an overall plan, with phases that it can plan as we go. it's a little clunky, I have to track the overall plan file myself, but it seems to work fine

dunno why I enjoy doing this so much but it rules

I ran a big plan on auto today to redo a hunk of the agentic ui app, bumping it as I had time between things. it worked great. too well, in fact. every new ui element is showing at the same time with gaudy first-pass visuals. it looks terribad 😂

kaijutsu will get speech2text input mostly for when my cat has fallen asleep in such a way that typing is not feasible

two bajorans one orb

I love playing in orchestra but the way some of the staff talk to us explains a lot about the orchestra's state

it gives me some strange comfort that people I know to be smart and capable af with language still encounter skill issues with agents. it can be scary when claude extrudes a complex module in 10 minutes. but it seems not everyone is getting that. I wonder what doesn't click for others?

today's claude technique: autocompaction off. every time it runs out of context, I rewind back to the beginning when the plan started and tack on "We've made progress. Review and continue." So far so good :)

so glad I built gpal: > have gemini do a code review > have gemini review the plan > ask gemini for feedback big upgrade for claude

not bad lil bot, not bad. this is the first time the new context browser concept has worked, modeled after skill trees in 4X games. the lil discs are text textures. it all needs tons of polish and realignment to even see if the concept works. my framerate is tanked rn, gonna work on that next

ok fine, Claude's words (after talking to gemini): **Adversarial review**: Ask Claude to find flaws in your proposed plan or implementation—a skeptical peer who follows instructions stubbornly. I'mma let it merge and do a vibe check as I use it. I load cpal in CC too for giggles and it's fun

presenting: your pal claude mcp: github.com/tobert/cpal the inverse of your pal gemini mcp: github.com/tobert/gpal

got lucky, missed one of the agents committed an API key and google emailed me it was compromised. doesn't look like it got picked up before I deleted it. sheesh. envvar was the wrong choice, hah, switching to file at least for my own setup. gcloud console is a lot when in a rush lol

it wouldn't be technically wrong to make gemini call claude senpai 🤔

gpal now gives claude image, speech, and audio gen tools :)

the text fading in with TAA is somewhat unintentional but a cool effect anyway, and looks fabulous once it fades in

gemini cli was doing really well on this feature... implementing gemini features in gpal, and it's stuck... can't figure out which gemini model to use 😂 wonder if I should plug in exa search?

they fixed autocompaction disable so I'm back on my 190K+ bullshit. I'm noticeably more productive when I manage compaction myself and let well-aligned contexts play out

gpal also superpowers claude's plan mode, "have gemini review the plan before we commit" no adversarial language necessary, I ask, claude prompts gemini, gemini reviews, claude picks what is actionable, and does it

Pathfinder WoTR Fancomic : How Azatas solve problems Azatas are basically friendship fairies? The gameplay is quite whimsical and power of friendship but you can also release extremely destruction lol. its quite fun.

I see my fingerprints all over my software. the shape of things is from my mind, produced via literal incantations. I have not opened vscode this year and 4/5 of my dreamiest apps are all running and taking shape. that 5th thing will be built by the first 4 😉

gpal now has semantic_search tool based on google's embeddings api :)

new gpal feature in the works: semantic search tool, using google's embedding api to build a local chromadb. seems neat, can't wait to try it. might fill a nice gap for claude and repo indexing

we taketh away (the footgun) but also we giveth (something way cooler) (the backend arrays never get serialized to strings, also, no more footgun)