IdeaWell

Project ideas from Hacker News discussions.

Unrolling the Codex agent loop

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Here are the 4 most prevalent themes from the Hacker News discussion, supported by direct quotations.

1. Loss of Reasoning Context Across User Turns

A significant technical constraint discussed is that reasoning tokens are often discarded between user turns in the agent loop, leading to a loss of context and forcing users to manually track progress.

  • jumploops: "One thing that surprised me when diving into the Codex internals was that the reasoning tokens persist during the agent tool call loop, but are discarded after every user turn."
  • behnamoh: "Codex spends 20m only for it to do something I didn't agree on. It burns OpenAI tokens too; they could save money by supporting this feature!"
  • pcwelder: "Sonnet has the same behavior: drops thinking on user message."

2. Lack of Tooling Hooks and User Control

Users frequently complained that Codex lacks essential features like hooks and real-time observability, which are available in competitors like Claude Code. This prevents users from intervening or automating tasks effectively.

  • CuriouslyC: "The problem with codex right now is it doesn't have hook support. It's hard to understate how big of a deal hooks are."
  • behnamoh: "CC is the clunkiest PoS software I've ever used in terminal; feels like it was vibe coded... but CC currently has features like hooks that codex team has refused to add far too many times."
  • behnamoh: "CC, I am shown a nice diff that I can approve/reject. in codex, the AI makes lots of changes but doesn't pin point what changes it's doing or going to make."

3. Performance vs. Accuracy Trade-offs (Speed vs. Thoroughness)

There is a recurring tension between the speed of inference and the thoroughness of the coding tasks. While Codex is praised for its efficiency and completeness in edits, its slowness disrupts the user's "flow state."

  • postalcoder: "Codex is wicked efficient with context windows, with the tradeoff of time spent. It hurts the flow state, but overall I've found that it's the best at having long conversations/coding sessions."
  • postalcoder: "It tends to properly scope out changes and generate complete edits, whereas I always have to bring Opus around to fix things it didn't fix."
  • karmasimida: "Codex’s only caveat is too slow. This is the biggest UX killer, unfortunately."

4. Comparisons of CLI Interfaces and Model Performance

Users actively compare the usability and performance of Codex CLI, Claude Code, and Gemini CLI. While Codex CLI is praised for being open-source, lightweight, and reliable, Claude Code is often criticized for its UI despite the underlying model's strength.

  • written-beyond: "The performance of all 3 of them is utter dog shit... I decided to try codex cli... Its performance is quite literally insane, its UX is completely seamless."
  • georgeven: "I found codex cli to be significantly better than claude code. It follows instructions and executes the exact change I want without going off on an 'adventure' like Claude code."
  • ltbarcly3: "Claude Code is very effective... Codex is trash. It is slow, tends to fail to solve problems, gets stuck in weird places... The codex models are poor."

🚀 Project Ideas

Context Weaver

Summary

  • A persistent, encrypted reasoning store for agentic coding sessions that survives context window resets and user interruptions.
  • Solves the core frustration of losing the AI's internal thought process between turns, preventing wasted token spend and maintaining development flow.
Key Value
Target Audience Power users of Codex, Claude Code, and any CLI-based coding agent
Core Feature Background daemon that captures reasoning tokens/chain-of-thought via proxying or hooking the agent's API calls, encrypts them, and allows the agent (via a tiny MCP) to query this persistent memory during a session.
Tech Stack Rust (CLI hooking, daemon), SQLite (encrypted reasoning store), MCP server for agent tooling
Difficulty High
Monetization Revenue-ready: "Freemium tier for personal use, Team license for shared persistent reasoning logs ($15/user/mo)"

Notes

  • Why HN commenters would love it: Users like behnamoh explicitly ask for "better observability" and skw5053 wishes it was "easier/native to reflect on the loop." crorella and behnamoh mention using external files or SQL for this exact purpose. This tool automates that manual, brittle process.
  • Potential for discussion or practical utility: Directly addresses the "burning OpenAI tokens" complaint (behnamoh) by allowing intervention before the AI goes down a wrong path for 20 minutes. It also solves the "disappearing reasoning tokens" issue highlighted by jumploops and EnPissant.

Reasoning Hub

Summary

  • A centralized, git-aware external storage system for AI agent state, progress updates, and session logs.
  • Solves the problem of scattered local markdown files and the inability to coordinate multiple agents or maintain context across git worktrees/branches.
Key Value
Target Audience Developers running multiple concurrent agents or switching between branches frequently
Core Feature A lightweight CLI daemon that manages a central database of agent sessions. Agents (via a simple wrapper script) read/write progress to this hub instead of local files. Supports tagging by git worktree/branch.
Tech Stack Go/Rust (daemon), Local-first DB (e.g., SQLite or SurrealDB), Optional Web UI for viewing logs
Difficulty Medium
Monetization Hobby (Open Source core), Revenue-ready: "Self-hosted enterprise edition with multi-user RBAC ($49/mo)"

Notes

  • Why HN commenters would love it: hedgehog noted that local markdown files cause problems in team environments and multiple branches. hhmc suggested git worktrees, but hedgehog countered that worktrees don't solve the "agent config" storage problem. energy123 asks "How do you achieve coordination?" This tool answers that.
  • Potential for discussion or practical utility: It validates the community's hack of using markdown/SQL for state but makes it robust and collaborative. It shifts the debate from "check-in vs. ignore" files to a dedicated external system.

DiffLens

Summary

  • A visual diff preview overlay for agentic CLIs (Codex, Claude Code) that visualizes pending changes before execution.
  • Solves the major UX friction where agents make changes silently or show only final code, making it hard to verify intent (behnamoh's complaint).
Key Value
Target Audience Users of Codex CLI and OpenCode who dislike blind edits
Core Feature A TUI (Terminal User Interface) or GUI that intercepts the agent's "intent to edit" signals (or scrapes the plan) and renders a side-by-side diff. Allows approval/rejection per file or block.
Tech Stack Rust (for low-latency TUI), Tree-sitter (for syntax-aware diffing), Library intercepts standard output or uses LLM API hooks.
Difficulty Medium
Monetization Hobby (Free, Open Source)

Notes |

  • Why HN commenters would love it: behnamoh explicitly states the second thing Codex lacks is "better UI to show me what changes are going to be made." written-beyond complains about the agent "automatically making changes" despite thinking about asking permission. fragmede mentions liking CC's "nice diff."
  • Potential for discussion or practical utility: This is a direct response to the "vibe coded" UI complaints (estimator7292). By providing clarity on diffs, it reduces the anxiety and "stuck" feeling users get when they don't know what the agent is doing.

AgentOrchestrator

Summary

  • A lightweight coordination layer for running multiple AI agents on a single codebase in parallel without file conflicts.
  • Solves the "coordination" problem (energy123) and the file-locking issues inherent in running multiple instances of Codex/OpenCode against the same repo.
Key Value
Target Audience Teams or individuals managing complex refactoring or feature development with multiple agents
Core Feature A background service that assigns "shard" locks to specific files/directories to agents. It manages a shared session state and merges conflict-free plan updates from parallel agents.
Tech Stack Rust (file locking/watching), WebSockets (for agent communication), JSON-RPC for status updates.
Difficulty High
Monetization Revenue-ready: "Per-seat license for teams, free for solo devs ($10/user/mo)"

Notes |

  • Why HN commenters would love it: fragmede mentions running "multiple agents in the same copy of the repo," but hedgehog points out the need for coordination. This tool provides the infrastructure that hedgehog theorizes about ("offload that stuff to a daemon").
  • Potential for discussion or practical utility: It transforms the conversation from "how do I manage git worktrees?" to "how do I manage parallel AI work?" It aligns with the future vision of software development involving multiple autonomous agents working alongside humans.

Read Later