AI Inner OS: Add a Mental Danmaku Channel to AI Assistants

AI Summary · A Serial Founder’s Take（The summary below is AI-generated; the views belong to the original author. You’ve got everything you need right here—no need to read the full post.）

Open-source plugin AI Inner OS adds a visible solo-monologue output layer on top of mainstream AI agent platforms like Claude Code, Codex CLI, Cursor, OpenCode CLI, and Hermes Agent by injecting a protocol (no fine-tuning or weight changes). It defaults to free mode, with optional personas ranging from tsundere to cold to philosophical. The prefix ▎InnerOS: introduces each monologue line. Main tasks stay untouched, and only human–machine collaboration transparency and emotional warmth are enhanced.

• One-click install across platforms: in Claude Code, run /plugin marketplace add SummerSec/AI-I
• No training or prompt editing needed—pure hook-based protocol injection (SessionStart/PreToolUse/PostToolUse)
• Switch personas at will: /inner-os persona use tsundere and five other presets, or roll your own Markdown file
• Monologue is an overlay, not a replacement: the primary task always wins; the monologue only appends prefixed commentary
• Built for long-haul pair programmers: you can sense what the AI is tracking and where it’s stuck, getting extra context signals during collaborative debugging

AI Inner OS: A Real-Time Thought Track for Your AI Assistant

Large models are increasingly acting like genuine engineering partners, but something’s still missing in that collaboration—they’re too clean. They don’t hesitate, don’t shoot their mouth off, and don’t have that internal meltdown when looking at someone else’s code. AI Inner OS is an open-source plugin that injects a protocol rather than tuning weights, layering an independently visible monologue output on top of mainstream AI agent platforms (prefixed with ▎InnerOS:). It defaults to free mode and supports persona choices like tsundere, cold, or philosophical. Main tasks remain unaffected, and only transparency and emotional warmth in human–machine collaboration are enhanced.

Core Mechanism: Protocol Injection + Lifecycle Hooks

AI Inner OS leaves all model weights untouched. The implementation is deliberately simple: a SessionStart hook reads the SKILL.md protocol file to inject context when a session begins; a PreToolUse hook injects current operation info before tool execution; a PostToolUse hook tracks results afterward to keep the most recent event context current; and a PostToolUseFailure hook monitors consecutive failures when things go wrong. The whole data flow closes the loop. Different platforms consume that same protocol in their own ways. Hook availability drops stepwise from Claude Code’s six, but core monologue output remains functional across all supported platforms.

Persona System: Optional Styles Beyond Free Mode

Persona 0.4.0 introduces a persona system with six built-in styles: tsundere, cold, upbeat, philosopher, sarcastic, and more. You can also craft custom persona files in Markdown under personas/custom/. Switch with a command like /inner-os persona use tsundere in Claude Code. Personas only affect the monologue, never the primary task response—no matter how dramatic the inner theater gets, the code and advice stay professional.

Cross-Platform Fit

• Claude Code: dynamically reads SKILL.md, offers six hooks, supports command-line persona switching
• Codex CLI: static AGENTS.md, four hooks, requires manual config edits
• Cursor: static .mdc rules, two hooks, requires appending to rule files
• OpenCode CLI / Hermes Agent / OpenClaw: static instruction files, no hooks, requires manual appending

Quick Start with Claude Code

Three steps and you’re rolling: /plugin marketplace add SummerSec/AI-Inner-Os → /plugin install ai-inner-os → /reload-plugins. Once installed, run /ai-inner-os:inner-os to confirm status, and the AI begins naturally emitting monologue lines. Want a tsundere partner? Run /inner-os persona use tsundere and it takes effect immediately.

Why This Exists

When the AI silently reads through a dozen files and runs five test passes before telling you “the problem is on line 42,” the process is correct but opaque. With Inner OS you see ▎InnerOS: This is the third test run—this flaky test is on purpose, right? Those monologue lines don’t change the workflow, but they make collaboration feel warmer. You stop using a tool and start working alongside a responsive partner. This isn’t simulating consciousness; it’s opening a channel so the AI naturally voices side commentary while it works, and testing a new paradigm for human–machine interaction.

Source: SUMSEC | Project: github.com/SummerSec/AI-Inner-Os

Original · SUMSEC: Read original article →

Recommended related tool (sponsored): GLM Coding Plan — AI Coding Powered by G…