Most of my Claude Code turns don't need Opus. Some do. Picking by hand got boring.
So I wrote a plugin that picks the model and effort for each turn. Haiku 5.5 for small stuff. Opus 5.5 for hard stuff.
No proxy. It runs inside Claude Code.
https://t.co/znrSiNh8t0
Does it save money? Probably. How much? No idea yet. No real numbers, so no fake ones.
You can see what it does: a line above the prompt and a /router pane.
MIT. Tell me where it picks wrong.
claude plugin marketplace add alexei-led/claude-router
Who decides? A small classifier: Jev, Clef, OpenAI, or a local Ollama model. It sees your prompt and a few recent messages. With Ollama, they never leave your laptop.
No key, slow answer, network error: the turn keeps its model. Subagents are not touched.
ry it:
pi install npm:pi-subagents
pi install npm:@alexeiled/pi-fusion
/reload
Repo:
https://t.co/pg7n6tcAWF
Pi package:
https://t.co/th9a2bCzKW
#OpenSource#DevTools#AI#CodingAgents#LLM
The useful part is not role play.
It is model diversity: different training priors, different reasoning paths, different failure modes.
One model catches an edge case. Another disagrees. Another gives the simple path. The judge compares them.
Implementation detail I care about:
pi-fusion is built on pi-subagents.
Panel members are normal Pi subagents. The judge is a normal Pi subagent. The base Pi session stays in control.
So the integration stays simple.
I made pi-fusion for the Pi coding agent.
One prompt → several model runs → judge model → one final answer.
Built on `pi-subagents`. Slower, but useful for design, risk, and debugging prompts where one model may miss something.
https://t.co/pg7n6tcAWF
#pi#ai#pidev#llm
🚀 ccgram v4.3.0 is out.
Control terminal coding agents from Telegram: Claude Code, Codex, Gemini, Pi, or shell.
One Forum topic = one tmux/herdr window = one live agent session.
Why this is useful 🧵
New workflow: /split agent teams.
From a Telegram topic:
/split claude
/split npm test
ccgram adds a sibling pane to the same window/tab. Use /panes to inspect. Blocked panes surface as alerts.
AI coding agents thrash on bugs, fake green tests, skip the repro.
coding-posture is a SKILL.md skill that makes them pick a task mode — debug / fix / review / migrate / optimize — and follow a checklist. Open source, works with Claude Code, Codex, Pi.
https://t.co/RfzY6yxFzf
One SKILL.md file. Install as a Claude Code or Codex plugin, or drop it into Pi / Hermes / Cursor.
/plugin marketplace add alexei-led/coding-posture
MIT, run the eval yourself:
https://t.co/RfzY6yxFzf
#OpenSource#AI#DevTools#ClaudeCode
Coding agents behave like optimistic elevators with write access: thrash on a stuck bug, fake green tests, migrate prod without a rollback.
coding-posture is a small skill that fixes the posture, not the model. 🧵
Before non-trivial work the agent picks a mode (debug, fix, review, test-first, refactor, optimize, migrate, upgrade, integrate, spike, unstuck) and follows its checklist.
Modes are procedures, not personas — the part research says actually changes behavior. Early eval: +15pp.