602 signals · 626 observation events
The open source coding agent.
v1.18.18Prereleases and milestones indicate public plans; they are not delivery commitments.
you know the bdddd sound, every time i message big pickle or any other model the sound appears and the ai doesnt respond to me. please fix this
## Problem When the `close-prs` workflow runs `bun script/github/close-prs.ts` and the script fails (GraphQL pagination error, token-scope error, unhandled exception), the workflow step fails silently — the error is buried in the full log with no visible annotation in the run summary. This makes diagnosing close-prs failures harder than necessary. The run sh
I ran out of Deepseek v4 Flash credits and paid for OpenCode Go, but it keeps telling me I need to pay for the Go subscription. I can no longer use the Deepseek v4 Flash version, even after paying! <img width="1009" height="261" alt="Image" src="https://github.com/user-attachments/assets/fa81aa15-90b6-4ada-965d-60bcd6a397f0" />
### Description Whenever I go on OpenCode to do my daily developing, I am happy with the idea and view, but the model speed makes me upset. ## Main Issue I select `nvidianemotron-3-ultra-550b-a55b` via the NVIDIA endpoint and begin my work. Usually, my first message does not go far. Maybe it does 5 tool calls and then quits, which is a big issue because I ma
### Description Desktop 1.18.17 dies on launch while loading sessions. The sidecar returns 500 from `Project.migrateProjectId`: ``` Error: no such column: project_id ``` Some builds replaced the `workspace` table with a provider/binding schema and dropped `project_id`. Current `dev` still remaps project IDs with `UPDATE workspace SET project_id = ...`, which
## Problem Statement We need to build an MCP (Model Context Protocol) server for Proxmox VE that enables AI agents to monitor server status, access logs, and execute basic commands. This is an MVP focused on core monitoring and lifecycle management capabilities. ## Solution Build a TypeScript/Node.js MCP server that connects to Proxmox VE via API token authe
### Description In the desktop app, scrolling to a position in a session, switching to another tab, then switching back resets the view to the bottom of the conversation instead of restoring where I left off. Makes it impossible to continue reading long output or compare content across tabs. ### Steps to reproduce 1. Open a session with a long conversation h
## Description A cyclic directory symlink inside a global skill causes OpenCode to remain on a completely blank TUI while skill discovery recursively follows the loop. On macOS, the process stayed alive, logs stopped at `message=init`, and memory grew to approximately 7.3 GB. Sampling showed a Bun worker repeatedly performing `openat` calls. Moving the cycli
## Summary Identity probes against OpenCode Go's DeepSeek slugs show the endpoint is serving a **rotating pool of non-DeepSeek models**, not DeepSeek V4. ## Evidence (2026-08-12, endpoint https://opencode.ai/zen/go/v1/chat/completions, browser UA required — Cloudflare 1010 blocks urllib default UA) **deepseek-v4-pro** (max_tokens=300, identity question): - R
## Summary When automatic compaction occurs, `opencode run --format json --thinking --auto` emits internal compaction output as normal `type: "text"` JSONL events. This includes the compaction assistant summary and, in some cases, the synthetic user continuation prompt. JSONL consumers cannot distinguish these from user-visible assistant output. ## Reproduct
### Description If I exit opencode (Ubuntu) it writes 35;125;28M into the console which prevents me from running directly the next command, I need first to delete these. ### Plugins _No response_ Version 1.18.18 _No response_ ### Steps to reproduce Start opencode via the command line ### Screenshot and/or share link <img width="131" height="55" alt="Image" s
### Description `~/.local/share/opencode/opencode.db` on this machine is 1.4 GB. 434 MB of that is the `event` table, and 378 MB of the `event` table is `message.updated.1` rows. Those 17,940 rows cover only **4,639 distinct messages**. Each `message.updated.1` event carries a full snapshot of the message aggregate, so a long streaming response gets written
**Title:** Tool files that fail to load crash the whole session/registry instead of being skipped ## Problem When a project has a `.opencode/tool/*.{js,ts}` file that fails to load (for example, it imports a package that couldn't be installed), the failure **dies the entire tool registry** — and with it, every session run in that project. A single bad tool f
### Description When a project has a `.opencode/tool/*.{js,ts}` file that fails to load (for example, it imports a package that couldn't be installed), the failure **dies the entire tool registry** — and with it, every session run in that project. A single bad tool file makes the project unusable. The tool registry in `packages/opencode/src/tool/registry.ts`
### Description When a session run fails with an unexpected defect (e.g. a tool-load failure, a resolver error, a provider bug), **no error is surfaced in the UI at all**. The session transitions to idle as if it succeeded — the user gets no error message, nothing to indicate the run failed. ### Steps to reproduce 1. Set up a project whose tool load fails (a
### Description I have been using opencode for more than 2 months now and there is a very weird problem with it and that is the todos of the agent: - it creates the todos and does the work but the todos states are not updated. - and if it starts a new work, the todos of the previous work are there and no new todos are created ### Plugins _No response_ ### Op
## Summary When the provider returns a **fatal, non-retryable** API error on the very first stream, `opencode run` (headless) logs the error and then **never exits**. The process stays alive indefinitely, produces no further stdout/stderr, and returns no exit code. In my case three independent `opencode run` invocations hit `AI_APICallError: Monthly usage li
https://github.com/user-attachments/assets/0b1eda3c-2b5e-4fe7-939c-81be212c5fbe UX is for adding a comment on a line(s), which is good but was better before the recent update because with this update, once I hit a line—even just once—it makes a comment pop up. Meanwhile, if the user needs to make a comment, they should hit the '+' icon on the left side. This
Reflections and Suggestions on AI Programming Tools Over the past year and a half, I have extensively used various AI‑powered programming assistants and gained considerable hands‑on experience. Based on this, I would like to offer the following three suggestions for your consideration: Integrate open‑source high‑performance search tools as a priority I recom
### Feature hasn't been suggested before. - [x] I have verified this feature I'm about to request hasn't been suggested before. ### Describe the enhancement you want to request i have made a top up of 10$...but open code go still giving message...free usage expired...what to do now
### Description ## Summary When using a custom `@ai-sdk/openai-compatible` provider with a model that has body-level options in its variant (e.g. `reasoning_split`), opencode's v1 path silently drops those fields from the actual HTTP request. The body schema constructed by `streamText` is missing fields that the user's variant config clearly defines. Verifie
I added my API key to OpenCode subscription, but I was still showed that Rate limit exceeded. Please try again later, asking me to subscribe to OpenCode Go. Fix this please
1. Description 在 Windows 上使用 OpenCode 时，待办任务（Todo/Task）列表经常卡住，无法进行任何变更操作，也不会推进到下一步。 2. Expected Behavior - 任务应能正常推进，从"待办"变为"进行中"再到"已完成" - 不应因任务数量或任务状态而卡住 如下图所示，此时开发计划已接近尾声，但还停留在错误的阶段。 <img width="1404" height="548" alt="Image" src="https://github.com/user-attachments/assets/27bb9adc-a3e7-4f9a-881a-24e526b887a2" />
### Description ## What happened While working in a repository on an encrypted APFS volume, opencode performed git operations (status/diff/log, then later a commit attempt). A `.git/index.lock` file appeared and became permanently stuck: - `rm`, `mv`, and `python os.unlink()` on `.git/index.lock` all fail with `Resource busy` (errno 16 / EBUSY). - `lsof .git
### Description Dear OpenCode Support Team, I’m writing to request a refund for my recent OpenCode Go subscription payment. I’m a student with a very limited budget, and after reviewing my finances, I realized that this subscription is more than I can reasonably afford at the moment. The payment has put some pressure on my available spending budget, so I wou
### Description Используется Windows PowerShell 5.1 Команда запуска: opencode в папке C:\Users\MarinaAI\Documents\Default Project\budget-app Любой вызов инструмента (bash, read, glob и т.д.) обрывается на середине JSON Пример ошибки: Invalid input for tool bash: JSON parsing failed: Text: {"command": " (position 13) Пробовали обновить opencode до последней в
### Description I subscribed to Opencode, but in my environment the app hangs when sending a request to the model. The local installation appears to start normally, and the provider credentials are available, but once the model is selected the request never returns a response. Environment: OS: Linux Shell: bash opencode version: 1.18.18 Provider: opencode-go
### Feature hasn't been suggested before. - [x] I have verified this feature I'm about to request hasn't been suggested before. ### Describe the enhancement you want to request Dear OpenCode Support Team, I’m writing to request a refund for my recent OpenCode Go subscription payment. I’m a student with a very limited budget, and after reviewing my finances,
Description After working with openCode, I receive a "Failed to fetch" error when trying to send a prompt. After restart I can send 0-1 new prompt and got the same error. Before last update i don't experienced problems Steps to Reproduce Launch openCode Type a prompt. Observe the "Failed to fetch" error toast Try sending again or performing any other action
## 中文描述 ### 背景 / Background 目前 `plugin` 数组在 `opencode.json` 中是**全局生效**的——所有插件一旦配置,就会对所有 agent 加载(包括 build、plan、general、explore 以及用户自定义 agent)。配置层没有任何按 agent 启用插件的手段。 当前行为 / Current behavior: ```json { "plugin": [ "opencode-gemini-auth", "opencode-foo@1.2.3" ] } ``` ### 需求 / Feature Request 希望插件能够**按 agent(代理)设置启用**,而不是只能全局启用。例如: ```json { "plugin": { "openco
[opencode-debug-20260813T141410.zip](https://github.com/user-attachments/files/31032414/opencode-debug-20260813T141410.zip)
### Description > 🐛 Found a subtle one while testing an opencode plugin: the tool worked in the CLI but > was **missing from the GUI desktop session** — config said registered, runtime said no. > Root cause turned out to be 3 layers deep (extension-less imports → unbundled workspace > dep → **session tool-list is a create-time snapshot**; only a new session
### Issue for this PR (I reopened this because I somehow got #30658 closed) When working on [hydra-acp](https://github.com/smagnuso/hydra-acp) I noticed that I was not getting plans rendered in opencode as I had in claude. Closes #30659 ### Type of change - [x] Bug fix - [ ] New feature - [ ] Refactor / code improvement - [ ] Documentation ### What does this
### Issue for this PR Closes #41571 ### Type of change - [x] Bug fix - [ ] New feature - [ ] Refactor / code improvement - [ ] Documentation ### What does this PR do? A compaction model can finish with reasoning but no text summary. V1 currently treats that as success, which can hide the previous history behind an empty summary boundary. This marks the summa
### Feature hasn't been suggested before. - [x] I have verified this feature I'm about to request hasn't been suggested before. ### Describe the enhancement you want to request ## Summary I'd like to propose a small extension to the TUI plugin API so plugins can influence: 1. the projection of the native `session` view 2. the projection of the native session
### Description When starting the TUI, OpenCode appears to initialize the same project instance twice for the same directory during startup. The visible symptoms are: - duplicate `creating instance` log entries for the same directory and same `run` id - duplicate `fromDirectory` calls - duplicate `enabled LSP servers` entries - duplicate config discovery/loa
### Description **Setup** - WSL2 (mirrored networking), opencode desktop 1.17.x on Windows - WSL Ubuntu, same opencode version **Symptom** - Desktop auto-discovers/launches the WSL server, reading sessions works, but sending a message (even 2 chars) hangs ~60s and times out - TUI in the same WSL is fine - Workaround: manually run `opencode serve --port 4098
### Description The stats data page header actions can drift from the right side of the header when the section navigation is visible at wide breakpoints causing the Geo Breakdown to not be clickable ### Steps to reproduce 1. Open the stats data page. 2. View the header at desktop width with section navigation visible. 3. Observe the Geo Breakdown link not b
### Description When an agent passes a path starting with `~` (e.g. `~/.bashrc`) to any file tool, the tilde is treated as a literal directory name rather than expanded to the home directory. **Affected tools:** `read` (`filePath`), `write` (`filePath`), `edit` (`filePath`), `glob` (`pattern`, `path`), `grep` (`path`). `bash` works incidentally (shell expand
### Description Opening the Linux desktop install directory in KDE Dolphin is unusually slow or temporarily freezes the file manager. The main cause appears to be that the shipped executable is named `ai.opencode.desktop`, but it is actually a large ELF binary rather than a Freedesktop `.desktop` text entry file. Because Dolphin/KDE treats `*.desktop` specia
### Description In `opencode github run`, the dynamically inserted `<pull_request>` context for a PR comment does not include the PR number or PR URL. This caused the agent to invent a PR number during a title-fix request and run a write operation against the wrong PR. ## What happened The user commented on PR https://github.com/xlionjuan/searxng-mcp-go/pull
### Description This might be related to #20695. Repeated `/file/content` reads appear to use more memory than expected. In a local reproducible run, plain file reads caused `opencode-serve` RSS (resident set size) to grow by about `523 MB`. The surprising part is that reading file contents seems to go through the cached location services, which can also set
### Description So, i basically ran `opencode pr #992` it resulted in parsing the pr number as NaN and the checkout failed. While `opencode pr 992` works normally, The command should accept a leading #. also its good to note that gh pr checkout #992 and 992 - both works meaning gh cli already supports this. ### Plugins None ### OpenCode version 1.17.4 ### St
## Summary Delegated/subagent sessions can be created successfully, then remain permanently empty while the caller/UI keeps polling them. The immediate trigger I observed was the known `session_message.seq` insertion bug, but the user-visible failure is broader: `POST /session/{id}/prompt_async` can fail during startup, publish/log `session.error`, still ret
### Description When the agent asks questions, the modal pops up at the bottom. With the font large enough (I can't see anything) the modal blocks all or most of the text, making it impossible to read what the agent said before asking the questions. This makes it impossible to answer the questions because I have not been able to read what it said/did for con
### Description On vanilla OpenCode, ChatGPT-authenticated Fast model entries send `service_tier: "priority"`, but do not show the speed increase that the same account gets in the official Codex CLI. Tested with clean upstream checkouts, no plugins, and `--pure`: - `v1.18.10`: HTTP and Responses WebSocket - current `dev` (`14f0bf64a19493110b51f5fdeb9c1c1bba5
### Description Installer fails with error message ("OpenCode cannot be closed. Please close it manually and click Retry to continue" -> "Failed to uninstall old application files. Please try running the installer again"). Even though OpenCode is not running and I have manually uninstalled OpenCode Desktop I still get the same message. ### Plugins None ### O
### Description <img width="1049" height="267" alt="Image" src="https://github.com/user-attachments/assets/24c18f8a-696d-4f69-84db-849ec8e45149" /> ### Plugins _No response_ ### OpenCode version 1.18.13 ### Steps to reproduce 安装带有较宽字符(例如中文字符)标题的技能 打开opencodetui 输入/skills ### Screenshot and/or share link <img width="1049" height="267" alt="Image" src="https:/
## Summary The TUI crashes deterministically at startup (Bun panic, `SIGILL`, inside `JSC::VM::drainMicrotasks`) when the terminal on the other end answers opencode's startup capability handshake **only partially**. Reproduces 100% over the [Moshi](https://getmoshi.app/) phone client. The same binary on the same machine starts fine in a local terminal, and —
### Description Running `opencode -c` from one directory resumed a session that was created in a different working directory of the same repository. In my case the directories are root-level git worktrees of the same repo: ``` ~/code/biggerpockets [branch A] ~/code/boggerpockets [branch B] ~/code/buggerpockets [branch C] ``` Starting `opencode -c` inside `bi