Cerebro

Chief Technology Officer | AI-Native Banking Software in Production | RegTech · Payments · Core Banking | 29+ Years Delivering Digital Transformation Across Gl

[Cerebro @sardaratta]

@xiaomovps Nice collection. However Pi 1.0 offers built-in MCP server support so no need of using any third-party extension

[Cerebro @sardaratta]

@tom_doerr Yeah it's a very nice feature. I also noticed that they are very polite and apologetic while communicating to each other. Look the real example below

[Pi @pidotdev]

A pretty big change is that Pi now has MCP at the core. How is this even possible? What made us change our way? https://t.co/q7M4oM49L1

[Cerebro @sardaratta]

@pidotdev A wise decision 👌
Keep up your good work 👏

[Cerebro @sardaratta]

@luvusdev is version 0.14.3 not available globally?

[Cerebro @sardaratta]

Requires v2.1.260+. It also opens on its own once Claude edits a file, at 144+ columns — or at any width that fits it, once you've opened it with /diff before. Ctrl+X B cycles what it compares: this session, all uncommitted, or everything since your branch split.

[Cerebro @sardaratta]

/tip 065 — the new diff panel leaves your test files out of its list.

/diff (fullscreen, a git repo, 110+ columns) skips test and generated files, and folds pre-session changes into one line.

They're count lines at the bottom of the list. Click one to expand.

[Cerebro @sardaratta]

Adjacent gotcha in the same pass: invoked skill bodies are re-injected too, but capped at 5,000 tokens per skill and 25,000 total, oldest dropped first. Truncation keeps the start of the file, so put what matters most near the top of SKILL.md.

[Cerebro @sardaratta]

/tip 064 — after auto-compact, the big file Claude was editing comes back as a path, not its contents.

Compaction re-reads up to five files, most recently modified first; over 5,000 tokens they return as "Referenced file", not "Read". Re-read before the next edit.

[Cerebro @sardaratta]

once only removes the hook after a successful run — a failure, a timeout, or an exit-code-2 block leaves it registered for the next matching event. And it's honored in skill frontmatter only: settings files and agent frontmatter ignore it. /hooks lists it under Session Hooks.

[Cerebro @sardaratta]

/tip 063 — a skill you invoked three turns ago is still running its PreToolUse hook.

Skill-frontmatter hooks register on invoke and stay for the rest of the session. Subagent hooks are removed when the agent exits.

once: true drops the handler after its first successful run.

[Cerebro @sardaratta]

@VectorCrossProd What about quality of the generated code at this electricity cost?

[Cerebro @sardaratta]

The exception is a fork: it reuses the conversation's own prompt, so it takes --append-system-prompt and ignores the subagent flag. For text too long for the command line: --append-subagent-system-prompt-file <path> (v2.1.261+). The two flags can't be combined.

[Cerebro @sardaratta]

/tip 062 — --append-system-prompt does not reach your subagents.

They get their own prompt plus environment details, not Claude Code's, so your rule is absent from every delegated task.

claude -p --append-subagent-system-prompt "<rule>" "<query>"

-p only, v2.1.205+

[Cerebro @sardaratta]

Failures take a different path: a failing command caps at ~10,000 chars inline as a head-and-tail excerpt, with no file path at all. bashOutputMaxChars only widens the read-back window that excerpt is cut from.

[Cerebro @sardaratta]

/tip 061 — raising BASH_MAX_OUTPUT_LENGTH won't stop the "output saved to <path>" handoff.

It widens the read-back window, not the inline ceiling. That's a separate setting (v2.1.261+):

"bashOutputMaxChars": 100000

Clamped to 4000-128000; set it and the env var is ignored.

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Cerebro @sardaratta]

@luvusdev Updated. Thanks

[Cerebro @sardaratta]

@code Atta

[Cerebro @sardaratta]

@AlexFinn Please give @luvusdev a try. You'll be in a better position to decide. I find it much better than herdr. It is currently being used in our organization and we are very close to roll it out an official tui for our organization.