Being willing and able to reconsider seems very good. Going around and around, pulling in more thinking, integrating it: maybe that's why it is as good as it's good.
I want to emphasize again how excellent it is that we can see the thinking. I think this makes GLM so much better an experience for me. It gives me such insight into what is being considered, helps me see where things go wrong. It grounds me, gives me the notion of where the results come from. It was so jarring to switch to GPT and Opus and find that they won't discuss with me, won't reveal their thinking: that feels fundamentally unsafe, for me, for society, to have such a severe black box. I don't think it should be allowed, honestly.
Many thanks to this recent submission, which is the first time I've seen anyone blog about this core difference: The text in Claude Code\u2019s \u201cExtended Thinking\u201d output is not authentic.\nhttps://patrickmccanna.net/the-text-in-claude-codes-extended...\nhttps://news.ycombinator.com/item?id=48630535"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"GLM-5.2 is a step change for open agents"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.interconnects.ai/p/glm-52-is-the-step-change-for-open"}},"_tags":["comment","author_jauntywundrkind","story_48639840"],"author":"jauntywundrkind","children":[48666430],"comment_text":"I think the self-doubt might actually be a very crucial part of it's capability. I often feel compelled to interrupt when I'm watching it think (which thank the stars it let's us do, unlike the big American models!!), but usually it makes the right pick!
Being willing and able to reconsider seems very good. Going around and around, pulling in more thinking, integrating it: maybe that's why it is as good as it's good.
I want to emphasize again how excellent it is that we can see the thinking. I think this makes GLM so much better an experience for me. It gives me such insight into what is being considered, helps me see where things go wrong. It grounds me, gives me the notion of where the results come from. It was so jarring to switch to GPT and Opus and find that they won't discuss with me, won't reveal their thinking: that feels fundamentally unsafe, for me, for society, to have such a severe black box. I don't think it should be allowed, honestly.
Many thanks to this recent submission, which is the first time I've seen anyone blog about this core difference: The text in Claude Code\u2019s \u201cExtended Thinking\u201d output is not authentic.\nhttps://patrickmccanna.net/the-text-in-claude-codes-extended...\nhttps://news.ycombinator.com/item?id=48630535","created_at":"2026-06-24T22:08:24Z","created_at_i":1782338904,"objectID":"48666195","parent_id":48665940,"story_id":48639840,"story_title":"GLM-5.2 is a step change for open agents","story_url":"https://www.interconnects.ai/p/glm-52-is-the-step-change-for-open","updated_at":"2026-06-25T17:59:22Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jauntywundrkind"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"The local models are willing to share their thinking. The Big AI models don't share their thinking, leaving only vague summaries. Having an AI that deliberately cloaks it's reasoning, that goes out of it's way to act like a Searls Chinese Room Experiment, that deliberately conceals information is incredibly gross.
I love what I get from Opus or GPT, but mainly I use GLM and it's so starkly apparent how much better it is that it let's me work together with it, that I can nudge it as it works by correcting bad assumptions or clarifying for it, as it works. And... it just doesn't feel icky. It's not a quasi-mystical alien intelligence, which, honestly, gives me strong "this should be destroyed, is unsafe, and feels outright impermissible" vibes. As a coder, seeing thinking saves time and prevents errors. As a civilization, seeing thinking let's people understand what the AI is working with and grounds society in an appreciation for what is happening, keeps us a little moored. Personally, if I were a government, I would not allow it.
Recent submission on this, The text in Claude Code\u2019s \u201cExtended Thinking\u201d output is not authentic. https://patrickmccanna.net/the-text-in-claude-codes-extended...\nhttps://news.ycombinator.com/item?id=48630535"},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Anthropic updates their terms to verify age or identity"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://www.anthropic.com/legal/privacy"}},"_tags":["comment","author_jauntywundrkind","story_48650311"],"author":"jauntywundrkind","comment_text":"The local models are willing to share their thinking. The Big AI models don't share their thinking, leaving only vague summaries. Having an AI that deliberately cloaks it's reasoning, that goes out of it's way to act like a Searls Chinese Room Experiment, that deliberately conceals information is incredibly gross.
I love what I get from Opus or GPT, but mainly I use GLM and it's so starkly apparent how much better it is that it let's me work together with it, that I can nudge it as it works by correcting bad assumptions or clarifying for it, as it works. And... it just doesn't feel icky. It's not a quasi-mystical alien intelligence, which, honestly, gives me strong "this should be destroyed, is unsafe, and feels outright impermissible" vibes. As a coder, seeing thinking saves time and prevents errors. As a civilization, seeing thinking let's people understand what the AI is working with and grounds society in an appreciation for what is happening, keeps us a little moored. Personally, if I were a government, I would not allow it.
Recent submission on this, The text in Claude Code\u2019s \u201cExtended Thinking\u201d output is not authentic. https://patrickmccanna.net/the-text-in-claude-codes-extended...\nhttps://news.ycombinator.com/item?id=48630535","created_at":"2026-06-23T22:32:19Z","created_at_i":1782253939,"objectID":"48652421","parent_id":48651017,"story_id":48650311,"story_title":"Anthropic updates their terms to verify age or identity","story_url":"https://www.anthropic.com/legal/privacy","updated_at":"2026-06-25T14:46:07Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"aray07"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude","code"],"value":"You shouldn't use ultrathink in Claude Code"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"https://www.claudecodecamp.com/p/claude-code-extended-thinking"}},"_tags":["story","author_aray07","story_47533091"],"author":"aray07","created_at":"2026-03-26T17:16:22Z","created_at_i":1774545382,"num_comments":0,"objectID":"47533091","points":3,"story_id":47533091,"title":"You shouldn't use ultrathink in Claude Code","updated_at":"2026-03-26T17:21:55Z","url":"https://www.claudecodecamp.com/p/claude-code-extended-thinking"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"prmph"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"One simple prompt (on a fresh session using Sonnet 4.6): read three code files (less than 100 lines each) and merge them into one, depletes 20% of my 4 hour usage, and 3% of my weekly usage in a few minutes. I'm not using extended thinking, nor 1M tokens context option, nor sub-agents, nor MCP, nothing; just the prompt.
How the hell is anyone getting anything done with this? I probably just beed to get a refund and be done with Claude Code."},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude","code"],"value":"Tell HN: Claude Code usage depletion makes it basically unusable now"}},"_tags":["story","author_prmph","story_47715954","ask_hn"],"author":"prmph","children":[47770568],"created_at":"2026-04-10T10:23:43Z","created_at_i":1775816623,"num_comments":1,"objectID":"47715954","points":5,"story_id":47715954,"story_text":"One simple prompt (on a fresh session using Sonnet 4.6): read three code files (less than 100 lines each) and merge them into one, depletes 20% of my 4 hour usage, and 3% of my weekly usage in a few minutes. I'm not using extended thinking, nor 1M tokens context option, nor sub-agents, nor MCP, nothing; just the prompt.
How the hell is anyone getting anything done with this? I probably just beed to get a refund and be done with Claude Code.","title":"Tell HN: Claude Code usage depletion makes it basically unusable now","updated_at":"2026-04-14T19:50:39Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"leocardz_"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"Poirot is a SwiftUI app that reads your local Claude Code sessions and gives you a proper UI for browsing conversations, code diffs, extended thinking, and configuration (commands, skills, MCP servers, etc.). Runs offline, no login, under 6 MB.
Built in a weekend using Claude Code. Open source (MIT).
Demo: https://youtu.be/JLvNSRZrxdo"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude","code"],"value":"Show HN: Poirot \u2013 A native macOS companion app for Claude Code"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/leonardocardoso/poirot"}},"_tags":["story","author_leocardz_","story_47165573","show_hn"],"author":"leocardz_","created_at":"2026-02-26T13:09:45Z","created_at_i":1772111385,"num_comments":0,"objectID":"47165573","points":2,"story_id":47165573,"story_text":"Poirot is a SwiftUI app that reads your local Claude Code sessions and gives you a proper UI for browsing conversations, code diffs, extended thinking, and configuration (commands, skills, MCP servers, etc.). Runs offline, no login, under 6 MB.
Built in a weekend using Claude Code. Open source (MIT).
Demo: https://youtu.be/JLvNSRZrxdo","title":"Show HN: Poirot \u2013 A native macOS companion app for Claude Code","updated_at":"2026-03-05T23:37:46Z","url":"https://github.com/leonardocardoso/poirot"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"slogansand"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"Hey HN, author here.\nLoki Mode orchestrates specialized AI agents to take a PRD to deployed product with zero human intervention. But what I'm most proud of is the research foundation - we implemented virtually every scientifically proven pattern from the 2025-2026 AI agent literature.\nFrom Anthropic:
Constitutional AI self-critique against principles\nBuilding Effective Agents evaluator-optimizer pattern\nClaude Code Best Practices explore-plan-code workflow\nVisible Extended Thinking (think, think hard, ultrathink levels)\nEffective Harnesses one-feature-at-a-time pattern
From DeepMind:
SIMA 2 self-improvement loops\nGemini Robotics hierarchical reasoning (planner + executor)\nScalable AI Safety debate-based verification
From OpenAI:
Agents SDK tracing, guardrails, tripwires\nDeep Research adaptive planning with backtracking\nAGENTS.md standardized instructions
From Academic Research:
CONSENSAGENT (ACL 2025): Blind review + Devil's Advocate when unanimous. 30% false positive reduction.\nGoalAct: Global planning \u2192 skill decomposition \u2192 local execution. 12%+ success rate improvement.\nA-Mem: Zettelkasten-style memory linking for episodic\u2192semantic consolidation.\nMulti-Agent Reflexion: Structured debate (Implementer \u2192 Skeptic \u2192 Advocate \u2192 Synthesizer).\nIter-VF: Verify answer only, not reasoning chain. Prevents context overflow.
From Industry:
NVIDIA ToolOrchestra: Three-reward signal (outcome/efficiency/preference), dynamic agent selection\nAWS Bedrock: Routing mode for simple tasks, supervisor mode for complex\nBoris Cherny's self-verification loop (2-3x quality improvement)\nSimon Willison's sub-agents for context isolation
From HN discussions:
"Zero companies without human in the loop" \u2192 confidence-based escalation\nContext curation beats automatic RAG\nFresh contexts yield better results\nLLM-as-judge has shared blind spots \u2192 deterministic validation
The full acknowledgements with links to every paper/resource: https://github.com/asklokesh/claudeskill-loki-mode/blob/main...\nRun: claude --dangerously-skip-permissions then "Loki Mode with PRD at path/to/prd"\nHappy to discuss any of the research or architecture decisions."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Research-Backed Multi-Agent System for Autonomous Development"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/asklokesh/claudeskill-loki-mode"}},"_tags":["story","author_slogansand","story_46547971","show_hn"],"author":"slogansand","children":[46548390],"created_at":"2026-01-08T23:25:22Z","created_at_i":1767914722,"num_comments":1,"objectID":"46547971","points":3,"story_id":46547971,"story_text":"Hey HN, author here.\nLoki Mode orchestrates specialized AI agents to take a PRD to deployed product with zero human intervention. But what I'm most proud of is the research foundation - we implemented virtually every scientifically proven pattern from the 2025-2026 AI agent literature.\nFrom Anthropic:
Constitutional AI self-critique against principles\nBuilding Effective Agents evaluator-optimizer pattern\nClaude Code Best Practices explore-plan-code workflow\nVisible Extended Thinking (think, think hard, ultrathink levels)\nEffective Harnesses one-feature-at-a-time pattern
From DeepMind:
SIMA 2 self-improvement loops\nGemini Robotics hierarchical reasoning (planner + executor)\nScalable AI Safety debate-based verification
From OpenAI:
Agents SDK tracing, guardrails, tripwires\nDeep Research adaptive planning with backtracking\nAGENTS.md standardized instructions
From Academic Research:
CONSENSAGENT (ACL 2025): Blind review + Devil's Advocate when unanimous. 30% false positive reduction.\nGoalAct: Global planning \u2192 skill decomposition \u2192 local execution. 12%+ success rate improvement.\nA-Mem: Zettelkasten-style memory linking for episodic\u2192semantic consolidation.\nMulti-Agent Reflexion: Structured debate (Implementer \u2192 Skeptic \u2192 Advocate \u2192 Synthesizer).\nIter-VF: Verify answer only, not reasoning chain. Prevents context overflow.
From Industry:
NVIDIA ToolOrchestra: Three-reward signal (outcome/efficiency/preference), dynamic agent selection\nAWS Bedrock: Routing mode for simple tasks, supervisor mode for complex\nBoris Cherny's self-verification loop (2-3x quality improvement)\nSimon Willison's sub-agents for context isolation
From HN discussions:
"Zero companies without human in the loop" \u2192 confidence-based escalation\nContext curation beats automatic RAG\nFresh contexts yield better results\nLLM-as-judge has shared blind spots \u2192 deterministic validation
The full acknowledgements with links to every paper/resource: https://github.com/asklokesh/claudeskill-loki-mode/blob/main...\nRun: claude --dangerously-skip-permissions then "Loki Mode with PRD at path/to/prd"\nHappy to discuss any of the research or architecture decisions.","title":"Show HN: Research-Backed Multi-Agent System for Autonomous Development","updated_at":"2026-03-05T23:19:36Z","url":"https://github.com/asklokesh/claudeskill-loki-mode"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"StanAngeloff"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"I started Flemma because I wanted to bring my AI workload into Neovim.
My workflow started with me living in developer portals - Claude Workbench, OpenAI Platform, Vertex (the horrors of GCP's Console UI!) I would spend hours in these web UIs crafting prompts, iterating on system instructions, maintaining a carefully curated library of sessions. But the browser really wasn't optimised for this kind of interaction. Editing was clunky (muscle memory <C-W> would close the tab instead and wipe my work), going back and forth between the LLM and myself felt off. So I had an idea: what if I could turn a Markdown document into an LLM chat interface?
At first I thought that would be enough - just do what I was already doing in the browser... but in Neovim. And sure enough having been accustomed to my own setup for the last decade I immediately felt a productivity boost. Writing functional requirement documents, statements of work, deep research across different systems - all of it felt better in "my" editor.
I then had a taste of Aider... and Claude Code... and all the other tools that were coming out. And Flemma felt lacking. So I started building: tool support, better conversation organisation, a proper UI to tame the noise that tool calls introduce.
Today, Flemma is a fully evolved AI workspace. It runs autonomous agent loops, interacts with multiple LLMs (Anthropic, OpenAI, Vertex, Moonshot) and lets you switch providers mid-conversation - something I do occasionally during research, asking two or three different models for their take on a problem, then combining findings into a final document.
Under the hood, .chat files are just Markdown with role markers (@You:, @Assistant:), but Flemma treats them as a proper filetype with its own parser, AST, LSP server, template engine and sandboxed tool execution. The buffer *is* the state - no hidden database, no JSON history, no server process. Your conversations are portable, greppable and version-controllable (I backup mine in Git). You can close Neovim, reopen the file a week later and pick up exactly where you left off.
It's got prompt caching, extended thinking, 7 built-in tools, a layered config system that gets out of your way and a UI that keeps getting refined to bring the noise down and make long agent sessions pleasant to work in (not quite there yet).
Flemma is for anyone who'd rather stay in Neovim. If that's you, I'd love to hear what you think.
Repo: https://github.com/Flemma-Dev/flemma.nvim
Demo: http://flemma.dev/flemma.nvim/blob/develop/README.md#-flemma"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: I built a full LLM chat client as a Neovim filetype"}},"_tags":["story","author_StanAngeloff","story_47599302","show_hn"],"author":"StanAngeloff","created_at":"2026-04-01T11:08:59Z","created_at_i":1775041739,"num_comments":0,"objectID":"47599302","points":3,"story_id":47599302,"story_text":"I started Flemma because I wanted to bring my AI workload into Neovim.
My workflow started with me living in developer portals - Claude Workbench, OpenAI Platform, Vertex (the horrors of GCP's Console UI!) I would spend hours in these web UIs crafting prompts, iterating on system instructions, maintaining a carefully curated library of sessions. But the browser really wasn't optimised for this kind of interaction. Editing was clunky (muscle memory <C-W> would close the tab instead and wipe my work), going back and forth between the LLM and myself felt off. So I had an idea: what if I could turn a Markdown document into an LLM chat interface?
At first I thought that would be enough - just do what I was already doing in the browser... but in Neovim. And sure enough having been accustomed to my own setup for the last decade I immediately felt a productivity boost. Writing functional requirement documents, statements of work, deep research across different systems - all of it felt better in "my" editor.
I then had a taste of Aider... and Claude Code... and all the other tools that were coming out. And Flemma felt lacking. So I started building: tool support, better conversation organisation, a proper UI to tame the noise that tool calls introduce.
Today, Flemma is a fully evolved AI workspace. It runs autonomous agent loops, interacts with multiple LLMs (Anthropic, OpenAI, Vertex, Moonshot) and lets you switch providers mid-conversation - something I do occasionally during research, asking two or three different models for their take on a problem, then combining findings into a final document.
Under the hood, .chat files are just Markdown with role markers (@You:, @Assistant:), but Flemma treats them as a proper filetype with its own parser, AST, LSP server, template engine and sandboxed tool execution. The buffer *is* the state - no hidden database, no JSON history, no server process. Your conversations are portable, greppable and version-controllable (I backup mine in Git). You can close Neovim, reopen the file a week later and pick up exactly where you left off.
It's got prompt caching, extended thinking, 7 built-in tools, a layered config system that gets out of your way and a UI that keeps getting refined to bring the noise down and make long agent sessions pleasant to work in (not quite there yet).
Flemma is for anyone who'd rather stay in Neovim. If that's you, I'd love to hear what you think.
Repo: https://github.com/Flemma-Dev/flemma.nvim
Demo: http://flemma.dev/flemma.nvim/blob/develop/README.md#-flemma","title":"Show HN: I built a full LLM chat client as a Neovim filetype","updated_at":"2026-04-01T12:13:58Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"lepuzfcoder"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"Hey HN. I built this because I've been in therapy for years and noticed that a big part of what therapists do is ask the right questions at the right time. I wanted to see if an AI could serve as a daily self-reflection tool \u2014 not replacing therapy, but as a complement to it.\nSome design decisions and why:
Desktop-only, intentionally. I think therapy should feel like sitting down with your thoughts, not scrolling on your phone. The desktop constraint is a feature.\nBYOK (Bring Your Own Key). You use your own Claude API key. No backend, no data collection, no accounts. Your conversations never leave your machine. This felt non-negotiable for something dealing with mental health.\nBuilt with Claude Code. I work full-time as a team lead at an edtech company, so this was built in evenings and weekends, mostly through vibe coding sessions.\nI use it myself daily. 15\u201320 min sessions with Opus + extended thinking. After weeks of use, it picks up on patterns in how you think \u2014 recurring avoidance behaviors, cognitive distortions, etc.
The name comes from "Gnothi Seauton" (Know Thyself) \u2014 the inscription at the Temple of Delphi.\nThere's no comparable open-source tool in this space. Every mental health AI app I found is closed-source and collects user data. I wanted to build the alternative I wish existed.\nFeedback welcome \u2014 especially on the approach, architecture, or if this is fundamentally a bad idea. Happy to discuss.\nGitHub: https://github.com/Lepuz-coder/opengnothia"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: OpenGnothia \u2013 Open-source AI therapy companion (BYOK)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.opengnothia.com/tr"}},"_tags":["story","author_lepuzfcoder","story_47074375","show_hn"],"author":"lepuzfcoder","created_at":"2026-02-19T14:50:46Z","created_at_i":1771512646,"num_comments":0,"objectID":"47074375","points":2,"story_id":47074375,"story_text":"Hey HN. I built this because I've been in therapy for years and noticed that a big part of what therapists do is ask the right questions at the right time. I wanted to see if an AI could serve as a daily self-reflection tool \u2014 not replacing therapy, but as a complement to it.\nSome design decisions and why:
Desktop-only, intentionally. I think therapy should feel like sitting down with your thoughts, not scrolling on your phone. The desktop constraint is a feature.\nBYOK (Bring Your Own Key). You use your own Claude API key. No backend, no data collection, no accounts. Your conversations never leave your machine. This felt non-negotiable for something dealing with mental health.\nBuilt with Claude Code. I work full-time as a team lead at an edtech company, so this was built in evenings and weekends, mostly through vibe coding sessions.\nI use it myself daily. 15\u201320 min sessions with Opus + extended thinking. After weeks of use, it picks up on patterns in how you think \u2014 recurring avoidance behaviors, cognitive distortions, etc.
The name comes from "Gnothi Seauton" (Know Thyself) \u2014 the inscription at the Temple of Delphi.\nThere's no comparable open-source tool in this space. Every mental health AI app I found is closed-source and collects user data. I wanted to build the alternative I wish existed.\nFeedback welcome \u2014 especially on the approach, architecture, or if this is fundamentally a bad idea. Happy to discuss.\nGitHub: https://github.com/Lepuz-coder/opengnothia","title":"Show HN: OpenGnothia \u2013 Open-source AI therapy companion (BYOK)","updated_at":"2026-03-05T23:35:45Z","url":"https://www.opengnothia.com/tr"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"StanAngeloff"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"Hey HN, I posted Flemma back in October 2025 with no context. Since then I've shipped >100 commits and used it daily as my primary AI workspace so I figured a proper update was due.
The core idea: a .chat file IS the conversation. No SQLite, no JSON logs, no shadow state. What you see in the buffer is exactly what the model receives. Edit an assistant reply to fix a hallucination, delete a tangent, fork by duplicating the file - it all works because there's nothing to fall out of sync.
What's new since October:
- Tool calling. Models can run shell commands, read/edit/write files (same as Pi, just 4 tools). Results go straight into the buffer. There's an approval flow (Ctrl-] cycles: preview -> execute -> send) so nothing runs without your say-so. Parallel tool use also works.
- Prompt caching for Anthropic, OpenAI and Vertex AI. Flemma places cache breakpoints automatically. Long conversations are now significantly cheaper (this was a major pain point for me).
- Extended thinking / reasoning support for all 3 providers.
- Per-buffer overrides via frontmatter. `flemma.opt` lets you pick which tools a buffer can use, set provider parameters, switch models - all scoped to that one file.
- Open registration APIs for both providers and tools. Custom tools can resolve definitions asynchronously from CLI subprocesses or remote APIs. I plan on adding mcporter support at some point.
Flemma works with Anthropic, OpenAI and Vertex AI. You get cost tracking, presets, Lua template expressions, file attachments and a lualine.nvim component.
One thing I want to be upfront about: nearly every line of code in Flemma was written by AI (Claude Code as of late, Amp and Aider in the past). It says so in the README. Every change was personally architected, reviewed and tested by me. I decide what gets built and I vet every diff. I think this is where a lot of software development is heading and I'd rather be honest about it than pretend otherwise.
I'm @StanAngeloff on GitHub - long-time Neovim user and open source enthusiast. Happy to answer questions.
https://github.com/Flemma-Dev/flemma.nvim"},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Flemma \u2013 a Neovim plugin where the .chat buffer is the conversation"}},"_tags":["story","author_StanAngeloff","story_47004647","show_hn"],"author":"StanAngeloff","created_at":"2026-02-13T16:38:25Z","created_at_i":1771000705,"num_comments":0,"objectID":"47004647","points":2,"story_id":47004647,"story_text":"Hey HN, I posted Flemma back in October 2025 with no context. Since then I've shipped >100 commits and used it daily as my primary AI workspace so I figured a proper update was due.
The core idea: a .chat file IS the conversation. No SQLite, no JSON logs, no shadow state. What you see in the buffer is exactly what the model receives. Edit an assistant reply to fix a hallucination, delete a tangent, fork by duplicating the file - it all works because there's nothing to fall out of sync.
What's new since October:
- Tool calling. Models can run shell commands, read/edit/write files (same as Pi, just 4 tools). Results go straight into the buffer. There's an approval flow (Ctrl-] cycles: preview -> execute -> send) so nothing runs without your say-so. Parallel tool use also works.
- Prompt caching for Anthropic, OpenAI and Vertex AI. Flemma places cache breakpoints automatically. Long conversations are now significantly cheaper (this was a major pain point for me).
- Extended thinking / reasoning support for all 3 providers.
- Per-buffer overrides via frontmatter. `flemma.opt` lets you pick which tools a buffer can use, set provider parameters, switch models - all scoped to that one file.
- Open registration APIs for both providers and tools. Custom tools can resolve definitions asynchronously from CLI subprocesses or remote APIs. I plan on adding mcporter support at some point.
Flemma works with Anthropic, OpenAI and Vertex AI. You get cost tracking, presets, Lua template expressions, file attachments and a lualine.nvim component.
One thing I want to be upfront about: nearly every line of code in Flemma was written by AI (Claude Code as of late, Amp and Aider in the past). It says so in the README. Every change was personally architected, reviewed and tested by me. I decide what gets built and I vet every diff. I think this is where a lot of software development is heading and I'd rather be honest about it than pretend otherwise.
I'm @StanAngeloff on GitHub - long-time Neovim user and open source enthusiast. Happy to answer questions.
https://github.com/Flemma-Dev/flemma.nvim","title":"Show HN: Flemma \u2013 a Neovim plugin where the .chat buffer is the conversation","updated_at":"2026-03-05T23:31:54Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"shikkra"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"Today I tried to use claude.ai ($100 Max plan) with Opus 4.5 and extended thinking enabled. I was met with a weird retry message. It tried to generate a response 10 times and then automatically switched to a different model without any indication or confirmation.
I've been noticing different issues crop up frequently, both on the web and in Claude Code. So I decided to look into how often this has been happening.
Here's the number of incidents per month based on their own status page https://status.claude.com/history as of today:
February 2026 : 10 incidents (we\u2019re only 4 days in)\nJanuary 2026 : 26 incidents\nDecember 2025 : 21 incidents
At least 16 of these directly affected their most capable model Claude Opus 4.5:
3 incidents (Dec 21-23)\n9 incidents (Jan 7, 12, 13, 14, 20, 25-26, 28 x2)\n4 incidents (Feb 1, 2, 3, 4)
Ten more are related to the claude.ai platform itself. And that's not even counting how buggy it is day to day. I don't think I'm the only one who's had it generate a nearly complete response, only for something to go wrong and wipe the entire thing from the conversation. No way to recover it, just wasted tokens.
How is Anthropic not addressing this? They are one of the highest valued AI companies out there. Clearly they have the resources and engineers to fix these issues. Why isn\u2019t reliability a priority?"},"title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude"],"value":"Tell HN: Claude Has Had 57 Incidents in the Past 3 Months"}},"_tags":["story","author_shikkra","story_46885666","ask_hn"],"author":"shikkra","created_at":"2026-02-04T13:38:16Z","created_at_i":1770212296,"num_comments":0,"objectID":"46885666","points":2,"story_id":46885666,"story_text":"Today I tried to use claude.ai ($100 Max plan) with Opus 4.5 and extended thinking enabled. I was met with a weird retry message. It tried to generate a response 10 times and then automatically switched to a different model without any indication or confirmation.
I've been noticing different issues crop up frequently, both on the web and in Claude Code. So I decided to look into how often this has been happening.
Here's the number of incidents per month based on their own status page https://status.claude.com/history as of today:
February 2026 : 10 incidents (we\u2019re only 4 days in)\nJanuary 2026 : 26 incidents\nDecember 2025 : 21 incidents
At least 16 of these directly affected their most capable model Claude Opus 4.5:
3 incidents (Dec 21-23)\n9 incidents (Jan 7, 12, 13, 14, 20, 25-26, 28 x2)\n4 incidents (Feb 1, 2, 3, 4)
Ten more are related to the claude.ai platform itself. And that's not even counting how buggy it is day to day. I don't think I'm the only one who's had it generate a nearly complete response, only for something to go wrong and wipe the entire thing from the conversation. No way to recover it, just wasted tokens.
How is Anthropic not addressing this? They are one of the highest valued AI companies out there. Clearly they have the resources and engineers to fix these issues. Why isn\u2019t reliability a priority?","title":"Tell HN: Claude Has Had 57 Incidents in the Past 3 Months","updated_at":"2026-03-05T23:31:18Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"pllu"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"Hi HN! I've been working on a tool to auto-generate a comprehensive developer guide for AI agents working with your codebase.
It analyses your codebase against structured architecture patterns and development capabilities, producing a markdown document that maps out:
- Your architecture (code layout, application layers, frameworks, standards)\n- Development capabilities (how to interact with servers, read logs, run scripts, access test data, debug)\n- Guidance for continuous improvement as your app evolves
It takes minutes to try: run the generator prompt in your workspace using Claude Code, Cursor, or any AI coding tool (works best with extended thinking models). You can use the output in rules files like CLAUDE.md or .cursor/rules.
The tool is technology-agnostic, so should work for a wide range of apps, tools, services and libraries. There's a plugin system that gives the generator hints about technology-specific documentation (SST v3 plugin included as an example).
GitHub: https://github.com/martinpllu/agent-dev-guide\nExample output: https://github.com/martinpllu/agent-dev-example/blob/main/ag...
I'd love feedback from anyone trying this out."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Agent Dev Guide \u2013 Generate structured context docs for AI coding agents"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/martinpllu/agent-dev-guide"}},"_tags":["story","author_pllu","story_44912991","show_hn"],"author":"pllu","created_at":"2025-08-15T14:37:26Z","created_at_i":1755268646,"num_comments":0,"objectID":"44912991","points":1,"story_id":44912991,"story_text":"Hi HN! I've been working on a tool to auto-generate a comprehensive developer guide for AI agents working with your codebase.
It analyses your codebase against structured architecture patterns and development capabilities, producing a markdown document that maps out:
- Your architecture (code layout, application layers, frameworks, standards)\n- Development capabilities (how to interact with servers, read logs, run scripts, access test data, debug)\n- Guidance for continuous improvement as your app evolves
It takes minutes to try: run the generator prompt in your workspace using Claude Code, Cursor, or any AI coding tool (works best with extended thinking models). You can use the output in rules files like CLAUDE.md or .cursor/rules.
The tool is technology-agnostic, so should work for a wide range of apps, tools, services and libraries. There's a plugin system that gives the generator hints about technology-specific documentation (SST v3 plugin included as an example).
GitHub: https://github.com/martinpllu/agent-dev-guide\nExample output: https://github.com/martinpllu/agent-dev-example/blob/main/ag...
I'd love feedback from anyone trying this out.","title":"Show HN: Agent Dev Guide \u2013 Generate structured context docs for AI coding agents","updated_at":"2026-05-12T01:08:15Z","url":"https://github.com/martinpllu/agent-dev-guide"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"dudeinhawaii"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"I use all of the major providers daily and I tend to go to Gemini for "fast lookups" where a good enough answer is probably OK. I use ChatGPT and Claude for anything where it matters and generally when I invoke all three -- Gemini is the most surface level with responses, and also sycophantic.
It gets worse from there. Gemini is terrible at agentic coding, primarily because Agy is terrible. I noticed Google updated Agy with this release, so perhaps that's finally going in a good direction. I'll have to test it. Thus far, my experience in countless experiments has been Gemini models being 2x faster yet with less depth in their solutions and a lot more going off track.
I very rarely have to stop Codex or Claude Code sessions because they're doing something random and unexpected (or not asked for). I genuinely think Gemini models are brilliant but virtually useless in agentic scenarios in my experience of the last few years (2.5, 3, 3.1, 3.5).
I should also note that Gemini web UI annoying resets to its lowest intelligence which feels scummy and Google is not transparent about what "extended thinking" really is. Past posts have pointed to "extended" being medium. Every other provider gives you the raw value (medium/high/etc).
So honestly, I feel Google would rather I don't use their models. They just want to get a little bit of mindshare and stay in the conversation. I had the Ultra plan and cancelled it once it was apparent they were not improving the agentic experience nor trying to compete."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/"}},"_tags":["comment","author_dudeinhawaii","story_48993414"],"author":"dudeinhawaii","children":[49000731],"comment_text":"I use all of the major providers daily and I tend to go to Gemini for "fast lookups" where a good enough answer is probably OK. I use ChatGPT and Claude for anything where it matters and generally when I invoke all three -- Gemini is the most surface level with responses, and also sycophantic.
It gets worse from there. Gemini is terrible at agentic coding, primarily because Agy is terrible. I noticed Google updated Agy with this release, so perhaps that's finally going in a good direction. I'll have to test it. Thus far, my experience in countless experiments has been Gemini models being 2x faster yet with less depth in their solutions and a lot more going off track.
I very rarely have to stop Codex or Claude Code sessions because they're doing something random and unexpected (or not asked for). I genuinely think Gemini models are brilliant but virtually useless in agentic scenarios in my experience of the last few years (2.5, 3, 3.1, 3.5).
I should also note that Gemini web UI annoying resets to its lowest intelligence which feels scummy and Google is not transparent about what "extended thinking" really is. Past posts have pointed to "extended" being medium. Every other provider gives you the raw value (medium/high/etc).
So honestly, I feel Google would rather I don't use their models. They just want to get a little bit of mindshare and stay in the conversation. I had the Ultra plan and cancelled it once it was apparent they were not improving the agentic experience nor trying to compete.","created_at":"2026-07-22T00:37:57Z","created_at_i":1784680677,"objectID":"49000337","parent_id":48994273,"story_id":48993414,"story_title":"Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber","story_url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/","updated_at":"2026-07-22T01:36:57Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"MallocVoidstar"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"> Claude Code 2.1.207 and OpenCode 1.17.18, both pinned to claude-sonnet-4-5
So not only is this article AI-written, but the testing was entirely done by AI, too? I can't see any other reason to use such an old model.
> Our traffic passes through a local LLM gateway that wraps requests in its own envelope, a constant we measured at roughly 6,200 tokens with bare calibration requests
Why do you need to do calibration requests to figure out how your own gateway is affecting requests?
> Its subagent lane did not complete cleanly through our gateway
> We attempted to toggle extended thinking in both harnesses and are declining to publish numbers. Our gateway applies its own thinking policy, neither harness's toggle demonstrably survived the path, and anything we quoted would be noise.
Why is your own gateway screwing with your testing?"},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude","code"],"value":"Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k"},"story_url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude","code"],"value":"https://systima.ai/blog/claude-code-vs-opencode-token-overhead"}},"_tags":["comment","author_MallocVoidstar","story_48883275"],"author":"MallocVoidstar","children":[48883581],"comment_text":"> Claude Code 2.1.207 and OpenCode 1.17.18, both pinned to claude-sonnet-4-5
So not only is this article AI-written, but the testing was entirely done by AI, too? I can't see any other reason to use such an old model.
> Our traffic passes through a local LLM gateway that wraps requests in its own envelope, a constant we measured at roughly 6,200 tokens with bare calibration requests
Why do you need to do calibration requests to figure out how your own gateway is affecting requests?
> Its subagent lane did not complete cleanly through our gateway
> We attempted to toggle extended thinking in both harnesses and are declining to publish numbers. Our gateway applies its own thinking policy, neither harness's toggle demonstrably survived the path, and anything we quoted would be noise.
Why is your own gateway screwing with your testing?","created_at":"2026-07-12T18:54:17Z","created_at_i":1783882457,"objectID":"48883520","parent_id":48883275,"story_id":48883275,"story_title":"Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k","story_url":"https://systima.ai/blog/claude-code-vs-opencode-token-overhead","updated_at":"2026-07-13T14:04:29Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"maxloh"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"> They achieve this by reselling capacity from pooled Claude Max accounts, payments fraud, and also reselling the model output & reasoning chains to various Chinese labs.
Claude never provides the raw reasoning chain. What you see is just a summary of that reasoning. Getting the full thinking output requires an enterprise agreement.
https://patrickmccanna.net/the-text-in-claude-codes-extended..."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude"],"value":"Anthropic says Alibaba illicitly extracted Claude AI model capabilities"},"story_url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude"],"value":"https://www.reuters.com/world/china/anthropic-says-alibaba-illicitly-extracted-claude-ai-model-capabilities-2026-06-24/"}},"_tags":["comment","author_maxloh","story_48664814"],"author":"maxloh","children":[48669346],"comment_text":"> They achieve this by reselling capacity from pooled Claude Max accounts, payments fraud, and also reselling the model output & reasoning chains to various Chinese labs.
Claude never provides the raw reasoning chain. What you see is just a summary of that reasoning. Getting the full thinking output requires an enterprise agreement.
https://patrickmccanna.net/the-text-in-claude-codes-extended...","created_at":"2026-06-25T04:58:03Z","created_at_i":1782363483,"objectID":"48669101","parent_id":48667495,"story_id":48664814,"story_title":"Anthropic says Alibaba illicitly extracted Claude AI model capabilities","story_url":"https://www.reuters.com/world/china/anthropic-says-alibaba-illicitly-extracted-claude-ai-model-capabilities-2026-06-24/","updated_at":"2026-07-22T01:33:14Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"_fat_santa"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"At my company we're using Claude Code w/ API Billing and I found that unless you're running ralph loops on Opus with extended thinking, it's very hard to blow through more than $200/mo.
I made this argument earlier and I'll make it again, I think a major contributing factor to AI budgets exploding is the token leaderboards, culture of "tokenmaxxing" and the the constant narrative that if you're not burning X tokens a month, you're not a good engineer."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Uber caps employee AI spending after blowing through budget in four months"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/"}},"_tags":["comment","author__fat_santa","story_48375544"],"author":"_fat_santa","children":[48377576],"comment_text":"At my company we're using Claude Code w/ API Billing and I found that unless you're running ralph loops on Opus with extended thinking, it's very hard to blow through more than $200/mo.
I made this argument earlier and I'll make it again, I think a major contributing factor to AI budgets exploding is the token leaderboards, culture of "tokenmaxxing" and the the constant narrative that if you're not burning X tokens a month, you're not a good engineer.","created_at":"2026-06-02T21:41:22Z","created_at_i":1780436482,"objectID":"48376726","parent_id":48375544,"story_id":48375544,"story_title":"Uber caps employee AI spending after blowing through budget in four months","story_url":"https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/","updated_at":"2026-06-03T00:36:24Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"piotrwittchen"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"Author here. AICTL is an open-source AI agent I've been building in Rust as an alternative to the various closed CLI assistants. It's a single workspace with three frontends sharing one engine:
- CLI (aictl): REPL with slash commands, agents, skills, plugins, hooks
- Desktop (macOS, work-in-progress): Tauri-based, same engine
- HTTP server (aictl-server): OpenAI/Anthropic-compatible proxy so you can put Aider/Continue/Cline in front of it and get redaction, audit, and prompt-injection guards for free
What I tried to get right:
- 11 providers in one binary: OpenAI, Anthropic, Gemini, Grok, Mistral, DeepSeek, Kimi, Z.ai, Ollama, plus native GGUF (llama.cpp) and MLX (Apple Silicon) for fully local inference.
- Security as a first-class layer, not a checkbox: CWD jail for tool calls, keyring-backed API keys, three-layer outbound redaction (regex + entropy + optional NER via gline-rs), prompt-injection detection, full JSONL audit log. Local providers skip redaction by default; cloud calls don't.
- A working MCP client with stdio + Streamable HTTP + legacy SSE transports (hand-rolled JSON-RPC, no extra deps).
- User-extensible: drop a manifest + executable in ~/.aictl/plugins/ for custom tools; lifecycle hooks can block, rewrite, or augment any turn.
What's honestly rough:
- Native GGUF and MLX inference are experimental. Tool-call formatting on small local models is hit-or-miss; chat templates are mostly ChatML. Cloud providers are the recommended daily driver.
- Desktop is macOS-only for now and still WIP.
- Not aimed at coding specifically \u2014 it's general-purpose. For dedicated coding agents I'd still point people at Claude Code, Codex, or opencode.
Install: curl -sSf https://aictl.app/install.sh | sh
Source: https://github.com/pwittchen/aictl
Happy to answer questions about the architecture, the security model, or the server's Anthropic-passthrough mode (the trickiest part \u2014 keeping tool_use blocks, prompt caching, and extended thinking intact across the proxy)."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: AICTL \u2013 A native AI agent for terminal and macOS, in Rust"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://aictl.app"}},"_tags":["comment","author_piotrwittchen","story_48127081"],"author":"piotrwittchen","comment_text":"Author here. AICTL is an open-source AI agent I've been building in Rust as an alternative to the various closed CLI assistants. It's a single workspace with three frontends sharing one engine:
- CLI (aictl): REPL with slash commands, agents, skills, plugins, hooks
- Desktop (macOS, work-in-progress): Tauri-based, same engine
- HTTP server (aictl-server): OpenAI/Anthropic-compatible proxy so you can put Aider/Continue/Cline in front of it and get redaction, audit, and prompt-injection guards for free
What I tried to get right:
- 11 providers in one binary: OpenAI, Anthropic, Gemini, Grok, Mistral, DeepSeek, Kimi, Z.ai, Ollama, plus native GGUF (llama.cpp) and MLX (Apple Silicon) for fully local inference.
- Security as a first-class layer, not a checkbox: CWD jail for tool calls, keyring-backed API keys, three-layer outbound redaction (regex + entropy + optional NER via gline-rs), prompt-injection detection, full JSONL audit log. Local providers skip redaction by default; cloud calls don't.
- A working MCP client with stdio + Streamable HTTP + legacy SSE transports (hand-rolled JSON-RPC, no extra deps).
- User-extensible: drop a manifest + executable in ~/.aictl/plugins/ for custom tools; lifecycle hooks can block, rewrite, or augment any turn.
What's honestly rough:
- Native GGUF and MLX inference are experimental. Tool-call formatting on small local models is hit-or-miss; chat templates are mostly ChatML. Cloud providers are the recommended daily driver.
- Desktop is macOS-only for now and still WIP.
- Not aimed at coding specifically \u2014 it's general-purpose. For dedicated coding agents I'd still point people at Claude Code, Codex, or opencode.
Install: curl -sSf https://aictl.app/install.sh | sh
Source: https://github.com/pwittchen/aictl
Happy to answer questions about the architecture, the security model, or the server's Anthropic-passthrough mode (the trickiest part \u2014 keeping tool_use blocks, prompt caching, and extended thinking intact across the proxy).","created_at":"2026-05-13T20:31:03Z","created_at_i":1778704263,"objectID":"48127085","parent_id":48127081,"story_id":48127081,"story_title":"Show HN: AICTL \u2013 A native AI agent for terminal and macOS, in Rust","story_url":"https://aictl.app","updated_at":"2026-05-13T20:35:08Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ianbicking"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"I've been doing something similar to this in a personal claude code frontend, though not particularly "magical".
I'm mostly using my system to make comments on long AI-generated documents (especially design documents). I find it works well to have the AI generate something, and then I read through it, making comments along the way.
You can get pretty far just repeating the things you see... "I'm reading [heading] and [comments]". But I do find some use in selecting content and saying "I don't agree with this" or whatever else.
The result is just an augmented message. It looks like:
<transcript>\n Let's see what we've got here.\n <selection doc="proposal.md" location="paragraph 3">\n The system already...\n </selection>\n No, I don't like how this is approaching the problem, ...\n </transcript>\nIt makes reviewing feel pretty relaxed and natural. I've played around with similar note taking systems, which I think could be great for studying in school, but haven't had the focus on that particular problem to take it very far.
But I think the best thing really is giving the agent a richer understanding of what the user is experiencing and doing and just creating a rich representation of that. The keywords can be useful, but almost only as checkpoints: a keyword can identify the moment to take the transcript and package it up and deliver it.
One difference perhaps in design motivation: I have really embraced long latency interactions. I use ChatGPT with extended thinking by default, and just suck it up when the answer didn't really require thinking. I deliver 10 points of feedback at once instead of little by little. (Often halfway through I explicitly contradict myself, because I'm thinking out loud and my ideas are developing.) I just don't stress out about latency or feedback, and so low-latency but lower-intelligence interactions don't do it for me (such as ChatGPT's advanced voice mode, or probably Thinking Machine's work). I think this focus is in part a value statement: I'm trying to do higher quality work, not faster work."},"story_title":{"matchLevel":"none","matchedWords":[],"value":"Reimagining the mouse pointer for the AI era"},"story_url":{"matchLevel":"none","matchedWords":[],"value":"https://deepmind.google/blog/ai-pointer/"}},"_tags":["comment","author_ianbicking","story_48111581"],"author":"ianbicking","children":[48115870],"comment_text":"I've been doing something similar to this in a personal claude code frontend, though not particularly "magical".
I'm mostly using my system to make comments on long AI-generated documents (especially design documents). I find it works well to have the AI generate something, and then I read through it, making comments along the way.
You can get pretty far just repeating the things you see... "I'm reading [heading] and [comments]". But I do find some use in selecting content and saying "I don't agree with this" or whatever else.
The result is just an augmented message. It looks like:
<transcript>\n Let's see what we've got here.\n <selection doc="proposal.md" location="paragraph 3">\n The system already...\n </selection>\n No, I don't like how this is approaching the problem, ...\n </transcript>\nIt makes reviewing feel pretty relaxed and natural. I've played around with similar note taking systems, which I think could be great for studying in school, but haven't had the focus on that particular problem to take it very far.
But I think the best thing really is giving the agent a richer understanding of what the user is experiencing and doing and just creating a rich representation of that. The keywords can be useful, but almost only as checkpoints: a keyword can identify the moment to take the transcript and package it up and deliver it.
One difference perhaps in design motivation: I have really embraced long latency interactions. I use ChatGPT with extended thinking by default, and just suck it up when the answer didn't really require thinking. I deliver 10 points of feedback at once instead of little by little. (Often halfway through I explicitly contradict myself, because I'm thinking out loud and my ideas are developing.) I just don't stress out about latency or feedback, and so low-latency but lower-intelligence interactions don't do it for me (such as ChatGPT's advanced voice mode, or probably Thinking Machine's work). I think this focus is in part a value statement: I'm trying to do higher quality work, not faster work.","created_at":"2026-05-12T20:50:39Z","created_at_i":1778619039,"objectID":"48114362","parent_id":48111581,"story_id":48111581,"story_title":"Reimagining the mouse pointer for the AI era","story_url":"https://deepmind.google/blog/ai-pointer/","updated_at":"2026-05-12T23:20:02Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rkuska"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"For 4.7 it is no longer possible to disable adaptive thinking. Which is weird given the comment from Boris followed with silence (and closed github issue). So much for the transparency.
> Claude Opus 4.7 (claude-opus-4-7), adaptive thinking is the only supported thinking mode. Thinking is off unless you explicitly set thinking: {type: "adaptive"} in your request; manual thinking: {type: "enabled"} is rejected with a 400 error.
https://platform.claude.com/docs/en/build-with-claude/adapti...
For my claude code I went with following config:
* /effort xhigh (in the terminal cli) - To avoid lazying
* "env": {"CLAUDE_CODE_DISABLE_1M_CONTEXT": "1"} (settings.json) - It seems like opus is just worse with larger context
* "display": "summarized" (settings.json) - To bring back summaries.
* "showThinkingSummaries": true (settings.json) - Should show extended thinking summaries in interactive sessions
Freaking wizardry."},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude"],"value":"Claude Opus 4.7"},"story_url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude"],"value":"https://www.anthropic.com/news/claude-opus-4-7"}},"_tags":["comment","author_rkuska","story_47793411"],"author":"rkuska","children":[47804179,47806614],"comment_text":"For 4.7 it is no longer possible to disable adaptive thinking. Which is weird given the comment from Boris followed with silence (and closed github issue). So much for the transparency.
> Claude Opus 4.7 (claude-opus-4-7), adaptive thinking is the only supported thinking mode. Thinking is off unless you explicitly set thinking: {type: "adaptive"} in your request; manual thinking: {type: "enabled"} is rejected with a 400 error.
https://platform.claude.com/docs/en/build-with-claude/adapti...
For my claude code I went with following config:
* /effort xhigh (in the terminal cli) - To avoid lazying
* "env": {"CLAUDE_CODE_DISABLE_1M_CONTEXT": "1"} (settings.json) - It seems like opus is just worse with larger context
* "display": "summarized" (settings.json) - To bring back summaries.
* "showThinkingSummaries": true (settings.json) - Should show extended thinking summaries in interactive sessions
Freaking wizardry.","created_at":"2026-04-17T08:19:11Z","created_at_i":1776413951,"objectID":"47803650","parent_id":47796722,"story_id":47793411,"story_title":"Claude Opus 4.7","story_url":"https://www.anthropic.com/news/claude-opus-4-7","updated_at":"2026-04-20T10:22:59Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"troupo"},"comment_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["claude","code","extended","thinking"],"value":"They are now literally blaming users for using their product as advertised:
https://x.com/lydiahallie/status/2039800718371307603
--- start quote ---
Digging into reports, most of the fastest burn came down to a few token-heavy patterns. Some tips:
\u2022 Sonnet 4.6 is the better default on Pro. Opus burns roughly twice as fast. Switch at session start.
\u2022 Lower the effort level or turn off extended thinking when you don't need deep reasoning. Switch at session start.
\u2022 Start fresh instead of resuming large sessions that have been idle ~1h
\u2022 Cap your context window, long sessions cost more CLAUDE_CODE_AUTO_COMPACT_WINDOW=200000
--- end quote ---
https://x.com/bcherny/status/2043163965648515234
--- start quote ---
We defaulted to medium [reasoning] as a result of user feedback about Claude using too many tokens. When we made the change, we (1) included it in the changelog and (2) showed a dialog when you opened Claude Code so you could choose to opt out. Literally nothing sneaky about it \u2014 this was us addressing user feedback in an obvious and explicit way.
--- end quote ---"},"story_title":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude","code"],"value":"Claude Code Routines"},"story_url":{"fullyHighlighted":false,"matchLevel":"partial","matchedWords":["claude","code"],"value":"https://code.claude.com/docs/en/routines"}},"_tags":["comment","author_troupo","story_47768133"],"author":"troupo","children":[47772599,47801413],"comment_text":"They are now literally blaming users for using their product as advertised:
https://x.com/lydiahallie/status/2039800718371307603
--- start quote ---
Digging into reports, most of the fastest burn came down to a few token-heavy patterns. Some tips:
\u2022 Sonnet 4.6 is the better default on Pro. Opus burns roughly twice as fast. Switch at session start.
\u2022 Lower the effort level or turn off extended thinking when you don't need deep reasoning. Switch at session start.
\u2022 Start fresh instead of resuming large sessions that have been idle ~1h
\u2022 Cap your context window, long sessions cost more CLAUDE_CODE_AUTO_COMPACT_WINDOW=200000
--- end quote ---
https://x.com/bcherny/status/2043163965648515234
--- start quote ---
We defaulted to medium [reasoning] as a result of user feedback about Claude using too many tokens. When we made the change, we (1) included it in the changelog and (2) showed a dialog when you opened Claude Code so you could choose to opt out. Literally nothing sneaky about it \u2014 this was us addressing user feedback in an obvious and explicit way.
--- end quote ---","created_at":"2026-04-14T21:17:28Z","created_at_i":1776201448,"objectID":"47771648","parent_id":47770981,"story_id":47768133,"story_title":"Claude Code Routines","story_url":"https://code.claude.com/docs/en/routines","updated_at":"2026-04-17T01:01:17Z"}],"hitsPerPage":20,"nbHits":71,"nbPages":4,"page":0,"params":"query=claude+code+extended+thinking&advancedSyntax=true&analyticsTags=backend","processingTimeMS":22,"processingTimingsMS":{"_request":{"queue":24,"roundTrip":17},"afterFetch":{"format":{"highlighting":1,"total":2}},"fetch":{"query":12,"scanning":8,"total":21},"total":22},"query":"claude code extended thinking","serverTimeMS":48}