have you checked out rsc's codesearch? that's probably the OG of this concept afaik

Booted a Debian 13 desktop VM, ran Firefox, no calls to pidfd_getfd at all. YMMV of course, just mentioning it as the debian codesearch results look daunting

𝙊𝙥𝙚𝙣𝘾𝙤𝙙𝙚 v1.14.42 released. TL;DR: native LLM core groundwork; interactive opencode run; Effect HTTP API replaces Hono paths; Scout/repo research provider retry fixes. 𝗖𝗼𝗿𝗲 & 𝗟𝗟𝗠 • Added a new internal @opencode-ai/llm foundation: schema-first requests/events, typed tools, tool-loop execution, provider route composition, HTTP/WebSocket transports, and recorded provider replay. • First route set covers Anthropic Messages, Bedrock Converse, Gemini, OpenAI Chat, OpenAI Responses, OpenAI-compatible chat, plus provider helpers for OpenAI, Anthropic, Azure, Bedrock, Cloudflare, Google, GitHub Copilot, OpenRouter, xAI. • Added @opencode-ai/http-recorder for cassette-style HTTP/WebSocket replay, including binary Bedrock event-stream handling. 𝗖𝗟𝗜 / 𝗧𝗨𝗜 • opencode run now has --interactive split-footer mode: scrollback stays append-only while the footer handles prompt input, command/model/variant menus, permissions, questions, subagent tabs, history, and interrupts. • The TUI key system moved onto @opentui/keymap with flat command ids, leader-key support, command palette integration, and scoped plugin keymap cleanup. • Added an internal which-key plugin, inactive by default, that can show reachable shortcuts and pending key sequences. • Session picker sorting now uses full updated timestamps, not day buckets. Prompt clear retains substantial drafts in history instead of dropping them. 𝗦𝗲𝗿𝘃𝗲𝗿 & 𝗦𝗗𝗞 • The server route tree is now Effect HttpApi-first; legacy Hono backend/adapter/proxy routes were removed. • HTTP API behavior tightened: gzip/deflate response compression, /doc OpenAPI serving, structured validation errors, fixed CORS Vary: Origin, typed 401 auth challenge headers, validated permission/question ids, and logged instance-disposal failures. • Workspace APIs gained sync-list/status/warp coverage and fixed workspace fence headers for mutating calls. • JS SDK/OpenAPI output was regenerated for the updated Effect HTTP surface. 𝗪𝗼𝗿𝗸𝘀𝗽𝗮𝗰𝗲𝘀 & 𝗦𝗰𝗼𝘂𝘁 • Workspaces now track time_used, can discover existing git worktrees, and start syncing even before they have sessions. • Detached session claims now use the source project, avoiding cross-project sync ownership mistakes. • New experimental Scout path adds codesearch, repo_clone, and repo_overview tools, plus config reference entries for git/local reference subagents. • Websearch now rolls out Exa Parallel provider selection with provider labels surfaced in TUI/Web tool output. 𝗣𝗿𝗼𝘃𝗶𝗱𝗲𝗿𝘀 & 𝗥𝗲𝘁𝗿𝗶𝗲𝘀 • GPT-5 reasoning variants were corrected across OpenAI, Azure, OpenRouter, Cloudflare AI Gateway, and compatible providers, including pro/chat/codex/deep-research edge cases. • Gemini thinking controls were aligned for Gemini 2.5 and Gemini 3 variants, including flash/image model differences. • Anthropic Opus/Sonnet effort handling was updated, and Anthropic/Bedrock reasoning blocks now preserve signed/redacted metadata even when text is empty. • Retry state can now carry provider-specific actions, so Go/free-tier/account-rate-limit failures show actionable dialogs instead of matching brittle strings. 𝗗𝗲𝘀𝗸𝘁𝗼𝗽 & 𝗪𝗲𝗯 • Desktop main process was converted to Effect-style orchestration with deferred startup state, sidecar lifecycle cleanup, update handling split out, system cert/proxy setup retained, and Windows installer switched to silent user-scope NSIS. • Project sidebar now shows a working spinner when any project/workspace session is busy or retrying. • Web/UI shell output normalizes carriage returns, websearch cards show the actual provider, and tool status title measurement is deferred to avoid extra layout work. • npm shim now forwards SIGINT/SIGTERM/SIGHUP to the spawned native binary instead of blocking in spawnSync. Bundle size change macOS arm64 • Total: 100.0 MB -> 100.0 MB ( 32.3 KB) • Bun runtime: 60.4 MB -> 60.4 MB ( 256 B) • CLI/TUI JS: 14.3 MB -> 14.3 MB ( 38.0 KB) • Web UI assets: 17.9 MB -> 17.9 MB ( 1.1 KB) • Native addons: 1.7 MB -> 1.7 MB ( 720 B) • WASM: 5.7 MB -> 5.7 MB (no change) • Bundle metadata: 81.9 KB -> 74.1 KB (-7.8 KB) Linux x64 • Total: 139.6 MB -> 139.7 MB ( 64.0 KB) • Bun runtime: 96.2 MB -> 96.2 MB (no change) • CLI/TUI JS: 14.3 MB -> 14.3 MB ( 38.0 KB) • Web UI assets: 17.9 MB -> 17.9 MB ( 1.1 KB) • Native addons: 5.5 MB -> 5.6 MB ( 25.4 KB) • WASM: 5.7 MB -> 5.7 MB (no change) • Bundle metadata: 70.4 KB -> 70.0 KB (-435 B) Windows x64 • Total: 151.5 MB -> 151.5 MB ( 43.5 KB) • Bun runtime: 111.5 MB -> 111.5 MB (-187 B) • CLI/TUI JS: 14.3 MB -> 14.3 MB ( 38.0 KB) • Web UI assets: 17.9 MB -> 17.9 MB ( 1.1 KB) • Native addons: 2.1 MB -> 2.1 MB ( 4.0 KB) • WASM: 5.7 MB -> 5.7 MB (no change) • Bundle metadata: 66.5 KB -> 67.2 KB ( 672 B) Compare: github.com/anomalyco/opencod…

𝙊𝙥𝙚𝙣𝘾𝙤𝙙𝙚 v1.14.30 released. TL;DR: Effect HTTP API parity workspace routing; desktop path/session identity fixes; scoped session lists; provider/TUI/editor stability fixes. 𝗖𝗼𝗿𝗲 & 𝗦𝗗𝗞 • Expanded the experimental Effect HTTP API backend into route groups/handlers across sessions, files, providers, PTY, sync, TUI, config, and workspaces, while Hono remains the stable default. • Added explicit backend selection/telemetry via OPENCODE_EXPERIMENTAL_HTTPAPI and kept generated OpenAPI/JS SDK parity during the migration. • Reworked workspace routing/proxy boundaries so instance routes resolve directory/workspace context consistently and can proxy remote workspace HTTP/WebSocket traffic safely. • Session listing now supports project-relative path filtering and scope=project, exposed through the SDK. 𝗗𝗲𝘀𝗸𝘁𝗼𝗽 & 𝗪𝗲𝗯 • Fixed missing sessions caused by path key mismatches: workspace/session cache keys now normalize slash style, trailing slashes, and Windows drive roots, with legacy storage migration so existing data survives. • Fixed beta desktop download asset names after the electron → desktop package rename. • Disabled Windows updater code-signature verification where packaging/signing mismatches were blocking updates. 𝗧𝗨𝗜 & 𝗘𝗱𝗶𝘁𝗼𝗿𝘀 • TUI session lists now default to the current project-relative path, with a command-palette toggle to show all project sessions instead. • Paste summaries can now be toggled directly in the TUI and persist independently of config defaults. • Editor context reconnects against the active session directory, so Claude/IDE lock-file selection follows the workspace you are actually using. • Zed selection handling now converts UTF-8 byte offsets correctly, fixing non-ASCII/emoji ranges. • Repeated submits no longer resend the same editor selection context, and invalid custom themes are ignored instead of crashing startup. 𝗣𝗿𝗼𝘃𝗶𝗱𝗲𝗿𝘀 & 𝗧𝗼𝗼𝗹𝘀 • Azure requests now align closer to OpenAI defaults, including store=false and ID stripping for Azure/OpenAI request bodies, fixing reasoning item sequence errors. • Added reasoning variant support for mistral-medium-3.5 and made DeepSeek model matching case-insensitive. • GitHub Copilot model variants now sync from remote capability data instead of preserving stale local variants. • Removed the broken codesearch tool path from agent permissions, registry, CLI/TUI, and UI rendering. • Fixed bash parser tree cleanup, timeout cleanup after rejected promises, and compaction tail_start_id remapping when forking sessions. • System prompt assembly now orders instructions as Global, Project, then Skills. Bundle size change macOS arm64 • Total: 99.6 MB -> 99.7 MB ( 16.1 KB) • Bun runtime: 60.4 MB -> 60.4 MB ( 128 B) • CLI/TUI JS: 14.0 MB -> 14.0 MB ( 26.8 KB) • Web UI assets: 17.8 MB -> 17.8 MB (-1.4 KB) • Native addons: 1.7 MB -> 1.7 MB (no change) • WASM: 5.7 MB -> 5.7 MB (no change) • Bundle metadata: 76.2 KB -> 66.8 KB (-9.4 KB) Linux x64 • Total: 139.3 MB -> 139.3 MB ( 28.0 KB) • Bun runtime: 96.2 MB -> 96.2 MB (no change) • CLI/TUI JS: 14.0 MB -> 14.0 MB ( 26.8 KB) • Web UI assets: 17.8 MB -> 17.8 MB (-1.4 KB) • Native addons: 5.5 MB -> 5.5 MB (no change) • WASM: 5.7 MB -> 5.7 MB (no change) • Bundle metadata: 67.1 KB -> 69.6 KB ( 2.6 KB) Windows x64 • Total: 151.2 MB -> 151.2 MB ( 25.5 KB) • Bun runtime: 111.5 MB -> 111.5 MB ( 64 B) • CLI/TUI JS: 14.0 MB -> 14.0 MB ( 26.8 KB) • Web UI assets: 17.8 MB -> 17.8 MB (-1.4 KB) • Native addons: 2.1 MB -> 2.1 MB (no change) • WASM: 5.7 MB -> 5.7 MB (no change) • Bundle metadata: 65.9 KB -> 65.9 KB (no change) Compare: github.com/anomalyco/opencod…

I had custom software built on top of google's codesearch that efficiently flagged dirty files and re-indexed efficiently in the background while I made changes myself that gave me sub 50ms keystroke search on twitter's entire codebase, all running on my laptop

I tried it and it generated some serious amount of total slop that I had to immediately delete. Not impressed at all so far with agentic. Sonnet 4.5 is LEAGUES better for agentic coding. Very good "ask" or codesearch model and very well behaved model otherwise.

Source Code Search Engine osint.link/#code #osint #search #codesearch #searchengine

deepwiki mcp <-> codesearch (Exa code api) 在使用opencodian的时候突然发送它调用了codesearch工具，并让它解释了一下： codesearch就是针对代码的“websearch”，搜代码库和文档来回答代码相关的问题 codesearch是opencode zen模型自带的tool，由exa提供，似乎才关注到，好奇跟Deepwiki mcp 比何如 前几天才用deepwiki mcp专门做deepwiki-expert-agent，来处理访问公开代码库的问题，比本地agent自己去探索本地clone下来的代码库，质量高很多，可靠性也不错，速度也快很多

ArkTS-CodeSearch: A Open-Source ArkTS Dataset for Code Retrieval Yulong He, Artem Ermakov, Sergey Kovalchuk, Artem Aliev, Dmitry Shalymov arxiv.org/abs/2602.05550 [𝚌𝚜.𝚂𝙴 𝚌𝚜.𝙲𝙻]

have you seen the new bug? it's on buganizer. it's literally on issuetracker. dude it's on monorail. find it on gerrit. you can probably find it on googlesource. check it on luci. look for it on codesearch. it's running on clusterfuzz. just sign up for embark. just bugcrowd it.

at twitter i wrote my own code indexer based on google's codesearch, which had a custom neovim plugin. it was good enough with grok at the time to keep me productive

I have 66 installed but the ones I use the most are Google Drive, Slack, Codesearch (a homemade one that searches our company's entire codebase), Context7, Repomix, Chrome Devtools

Recorded the first demo of Minorag’s new chat mode today. Watching it go from idea → working feature feels surreal. If you’re curious how it looks in action, here’s the video 👇 #buildinpublic #CodeSearch #dotnet #AIEngineering

we have a codesearch tool under experimental flag, gonna release it this week it triggers from your query

Ok did some testing, both on 4.5 haiku, having it build a sign up form in sveltekit with the new remote functions. They just released and the model's can't write them out of the box: with local search: opencode.ai/s/i9vY6wq2 - 55K tokens in main agent - $0.42 for main agent - sub agent tokens: 40K, 24K, 34K, 24K - sub agent costs: $0.10, $0.05, $0.07, $0.03 - $0.67 for entire run - really good final result with web search: opencode.ai/s/defPOAza - 114K tokens - $0.87 for entire run - final output was functional, but rough Right off the bat the agent tried to search for very much the wrong thing in the web search test which definitely hurt it, I had to re-steer it halfway through and then it did search the right thing. I definitely prefer the local search. The search outputs from Exa were kinda all over the place, like in this one opencode.ai/s/defPOAza#msg_a… it gave like 5 different ways of doing the same thing. The search query was also bad, but that's not the search tool's fault. Could see myself using the codesearch as a one off thing for little implementation questions, but for bigger day to day stuff I'm absolutely sticking with the local. Especially with sub agents this works absurdly well

Prompts and code snippets you submit are treated as user content. We keep them only while your account is active or needed to run the service; when no longer needed, we delete them. What's important, you control whether content helps improve the product via the Share&Improve settings. For generation, inputs go to vetted LLM providers, but we don’t send raw repo code to third‑party codesearch services; embeddings run on our infra and the local graph stays on your machine. Data is encrypted in transit and at rest.

Cognition 发布 SWE-grep 和 SWE-grep-mini: 用于多轮快速上下文检索的强化学习模型 核心创新：速度与精度的双突破 SWE-grep 系列专为多轮智能体式代码搜索设计，通过强化学习（RL）优化，支持每轮最多 8 个并行工具调用（如 grep、read 和 glob），将典型 10-20 轮交互压缩至 4 轮（3 轮探索 1 轮总结）。 输出速度对比：SWE-grep-mini 达 2858 tokens/s，SWE-grep 为 682 tokens/s，远超 Sonnet 4.5（69 tokens/s）和 Haiku 4.5（142 tokens/s），实现 20 倍加速。 驱动“Fast Context”子智能体，在 @windsurf 中为用户提供 3 秒内完成的检索，避免中断“流状态”工作流。模型匹敌前沿编码模型的检索精度，同时节省主智能体的上下文预算，减少“污染”风险。 问题驱动：从瓶颈到帕累托前沿 编码智能体常耗费 60% 时间于检索：嵌入式 RAG 快但不准，传统智能体搜索灵活却慢（数分钟）。帖子直击痛点，宣称 SWE-grep “杀死长上下文依赖”，创建“新帕累托前沿”——在 Cognition CodeSearch Eval（128 个智能体搜索任务，34 个仓库）上，SWE-grep-mini 的加权 F1 分数达 0.44，高于 Sonnet 4.5（0.38）和 GPT-4.1（0.34）。 团队研发动机源于真实场景，如 Devin 在大型代码库（如 React 或 PyTorch）中的延迟，目标是 5 秒内检索以维持工程师生产力。 训练与部署：RL 的高效实践 采用策略梯度 RL，奖励函数优先精确率（β=0.5 的 F1 分数），结合序列重要性采样和 Dr. GRPO token 限制，确保稳定性。并行调用从 4 增至 8 个，轮次从 5 减至 3。SWE-grep-mini 通过 RL 蒸馏进一步压缩规模。 而与 @cerebras 的合作，利用其晶圆级硬件加速推理，端到端延迟降至 5 倍快。 部署上，已渐进上线 Windsurf，并在 playground 模拟本地环境对比 Claude Code 和 Cursor CLI。 实证与影响：数据说话 · 博客评估基于内部数据集（真实查询 bug 标签），下游测试 SWE-Bench Verified 显示检索步骤减少、时间缩短，任务成功率持平。 · 直观呈现：F1 分数柱状图中 SWE-grep-mini 领跑，速度条形图凸显其 20 倍优势。 展望与价值 团队视 Fast Context 为“快速智能体”起点，未来扩展至 DeepWiki 和 Devin，优化可变轮次与工具集。引用 Csikszentmihalyi 的“流”理论，强调延迟每超 5 秒，生产力中断概率升 10%。

super excited to release SWE-grep and SWE-grep-mini! SWE-grep-mini achieves extreme inference speeds of >2,800 TPS: 20x faster than Haiku 4.5 while beating Sonnet 4.5, Opus 4.1 & GPT-5 on our CodeSearch eval our vision: make agentic search as fast as embedding search. the SWE-grep models can perform multi-turn agentic search in <3 seconds compared to 20-60 seconds for other frontier models. x.com/cognition/status/19788…

a week after leaving Google, and i still find my muscle memory going to my moma/codesearch/golinks