genji1024/dotfiles
Watch
1
0
Fork
You've already forked dotfiles
 0
Code Issues Pull requests Wiki Activity Actions

Migrate dotfiles from GitHub #2

Merged
genji1024 merged 260 commits from migrate/from-github into main 2026-08-20 16:44:30 +09:00
Conversation 1 Commits 260 Files changed 132 +3776
bot-genji1024 commented 2026-08-20 14:18:30 +09:00
Collaborator
Copy link

Closes #1

実施内容

GitHub 上の dotfiles リポジトリを本リポジトリ（Forgejo）へ移行しました。全 260 コミットについて、Issue #1 の要件どおり以下を実施しています。

  • author を genji1024 <[email protected]> に変更
  • コミットメッセージを Conventional Commits（<type>(scope): <description>）に正規化
  • body の無いコミットは headline のみ維持（body の新規追加は 0 件）

変更内容

  • main ブランチ: 空 init コミット 1 件（chore: initialize empty repository）を新規作成
  • migrate/from-github ブランチ: 移行後の全履歴（260 コミット + 空 init、main の子として接続）
  • マージコミット（旧 "Merge pull request #2"）は両親を保持して再現（全履歴保持のため）

検証結果（ローカルで実施）

  • コミット数: 260 → 261（+空 init）で一致
  • author/committer: 移行対象 260 件すべて genji1024 <[email protected]>（空 init は bot 身元）
  • 内容不変: tree ハッシュが書き換え前後で 260/260 件完全一致（マージ含む）
  • メッセージ形式: 全行 72 字以内・bullet 間空行なし・body 無しコミットへの body 追加 0 件
  • トポロジ: 元履歴と同型（脱落・順序変更ゼロ、マージ 1 件のみ）

動作確認手順

本 PR は履歴の移行のためビルドは不要です。Forgejo 上で以下を確認できます。

  1. migrate/from-github のコミット一覧・メッセージ・author を確認
  2. 全コミットの author が genji1024 <[email protected]> であることを確認
  3. マージコミットが 1 件、両親つきで残っていることを確認

備考

  • 移行ソース: Issue 記載の github.com/genji1024/dotfiles ではなく、ローカル clone の g-ohara/dotfiles（同一人物と判断）を使用しました。差異があればご指摘ください。
  • headline が 50 字を超えるコミットが 97/260 件あります（元から 79 件・正規化で 18 件）。規則は「約 50 字以内」のため許容と判断しました。
  • メッセージ正規化は 65 件のみ変更（195 件は author のみ変更）。意図が不明だった 1 件（旧 fix #1）のみ diff を確認して正規化しました。
Closes #1 ## 実施内容 GitHub 上の dotfiles リポジトリを本リポジトリ（Forgejo）へ移行しました。全 260 コミットについて、Issue #1 の要件どおり以下を実施しています。 - author を `genji1024 <[email protected]>` に変更 - コミットメッセージを Conventional Commits（`<type>(scope): <description>`）に正規化 - body の無いコミットは headline のみ維持（body の新規追加は 0 件） ## 変更内容 - `main` ブランチ: 空 init コミット 1 件（`chore: initialize empty repository`）を新規作成 - `migrate/from-github` ブランチ: 移行後の全履歴（260 コミット + 空 init、`main` の子として接続） - マージコミット（旧 "Merge pull request #2"）は両親を保持して再現（全履歴保持のため） ## 検証結果（ローカルで実施） - [x] コミット数: 260 → 261（+空 init）で一致 - [x] author/committer: 移行対象 260 件すべて `genji1024 <[email protected]>`（空 init は bot 身元） - [x] 内容不変: tree ハッシュが書き換え前後で 260/260 件完全一致（マージ含む） - [x] メッセージ形式: 全行 72 字以内・bullet 間空行なし・body 無しコミットへの body 追加 0 件 - [x] トポロジ: 元履歴と同型（脱落・順序変更ゼロ、マージ 1 件のみ） ## 動作確認手順 本 PR は履歴の移行のためビルドは不要です。Forgejo 上で以下を確認できます。 1. `migrate/from-github` のコミット一覧・メッセージ・author を確認 2. 全コミットの author が `genji1024 <[email protected]>` であることを確認 3. マージコミットが 1 件、両親つきで残っていることを確認 ## 備考 - 移行ソース: Issue 記載の `github.com/genji1024/dotfiles` ではなく、ローカル clone の `g-ohara/dotfiles`（同一人物と判断）を使用しました。差異があればご指摘ください。 - headline が 50 字を超えるコミットが 97/260 件あります（元から 79 件・正規化で 18 件）。規則は「約 50 字以内」のため許容と判断しました。 - メッセージ正規化は 65 件のみ変更（195 件は author のみ変更）。意図が不明だった 1 件（旧 `fix #1`）のみ diff を確認して正規化しました。
- merge pull request #2 (branch 1-switch-initvim-to-initlua)
- fix #1 by moving the settings in init.vim to init.lua
- Remove `scripts/system/` and the 2 scripts in it.
- Remove `scripts/apps/` and move the 3 scripts in it to `scripts/`.
- Add instructions for creating a new user and configuring sudo access.
- Add Docker installation and setup instructions to README.
- Add instructions for exporting secret keys and importing them.
- Add Dockerfile for OpenCode
  - Based on node:26.3.0-trixie-slim
  - Install opencode-ai, difit and GitLab CLI
- Add compose service with port 4966
  - Mount config, cache, state, and agents dirs
- Add entrypoint script for running via compose
  - Support host-level directory creation
- Add skill management files
  - `.skill-lock.json` to track installed remote skills
  - tracks find-skills, grill-me, difit and difit-review
  - `.gitignore` to keep `original/` as the only tracked dir
- Add custom skills for agent workflows
  - commit-message: commit message generation
  - doc-writing: Japanese documentation formatting
  - general: baseline session rules
  - gitlab-cli: GitLab CLI operations
  - mr-review: review comment resolution
  - project-context: project-specific guidelines
  - session-lessons: reusable lesson extraction
- Add cmake Dockerfile based on ubuntu:22.04
  - Install build dependencies and cmake 4.3.3 via tarball
- Register cmake service in compose.yaml with shared mounts
- Add shell wrapper for convenient docker compose invocation
- Add glab Dockerfile based on ubuntu:24.04
  - Install glab 1.95.0 via deb package
- Register glab service in compose.yaml with config volume mount
- Add shell wrapper for convenient docker compose invocation
- Add marksman Dockerfile based on alpine:3.23.4
  - Install marksman 2026-02-08 via GitHub releases
- Register marksman service in compose.yaml with version arg
- Add shell wrapper for convenient docker compose invocation
- Switch neovim from apk package to source build
  - Clone stable branch and build with RelWithDebInfo
  - Add build dependencies (cmake, ninja, coreutils)
- Add tree-sitter-cli as multi-stage build dependency
  - Build from rust:1.95.0-alpine3.22 via cargo-binstall
  - Copy binary into neovim final image
- Update runtime dependencies: add ripgrep, drop neovim pkg
- Configure neovim clipboard to use OSC 52 protocol
  - Enables copy/paste across SSH and tmux sessions
- Add tmux config with set-clipboard on
  - Allows tmux to forward OSC 52 sequences
- Add opencode.jsonc with bash/webfetch/websearch/skill permissions
- Update .gitignore to allow tracking opencode config
  - Fix root-level ignore pattern from `*` to `/*`
- Delete `SKILL.md` as it is no longer needed
- Add prettier service using `jauderho/prettier:3.8.3-alpine` image
- Add shell wrapper with healthcheck for dependency validation
- Add YAML language server service to compose
- Add shell wrapper with healthcheck for dependency validation
- Check if rootless docker socket exists before setting context
- Fall back to default context when rootless socket is absent
- Fix rootless docker detection in profile
  - Socket file is absent at profile load time
  - Check systemctl for enabled service instead
- Fall back to default context when rootless is unavailable
- Pin neovim build to a specific version tag
  - Replace stable branch tracking with v0.12.3 tag
  - Pass version via build arg for traceability
- Stop hard-coding an 80-character column ruler globally
- Add hermes-agent service to docker compose
  - Mount .agents and .hermes configs via bind volumes
  - Pass agent and GitLab MCP versions as build args
- Add Dockerfile based on node:26-alpine
  - Install hermes-agent from official installer
  - Add GitLab MCP server for repository integration
- Add shell wrapper for convenient compose invocation
- Pass `--fallback-style=none` to clangd to suppress formatting
  - Avoids applying LLVM default style when no config exists
  - Respects project-level `.clang-format` when present
- Add environment skill describing opencode's Docker container
  - Records Debian trixie base, host Arch kernel, and 1:1 bind mounts
  - Specifies unavailable tools (pacman, docker, systemd, GUI/X11)
  - Lists pre-installed binaries (node, npm, glab, difit)
- Update general skill to autoload environment
  - Required before running any shell command
- Add gitlab-mcp image and compose service (@zereight/[email protected])
  - via Streamable HTTP on port 3002
  - Bearer auth via STREAMABLE_HTTP_AUTH_TOKEN
  - MCP_ALLOWED_HOSTS and MCP_SERVER_URL permit non-loopback Host
  - No host port publish; reachable at http://gitlab-mcp:3002/mcp
- Make opencode and hermes-agent depend_on gitlab-mcp
  - Pass STREAMABLE_HTTP_AUTH_TOKEN for clients to send as Bearer header
- Drop bundled mcp-gitlab npm install from hermes-agent Dockerfile
- Drop glab CLI from opencode Dockerfile
  - switch opencode service from `extends: glab` to `extends: common`
- Remove glab service, Dockerfile, and shell wrapper
- Remove gitlab-cli skill
  - update environment, general, and mr-review skills for MCP GitLab
- Add opencode.jsonc and .hermes/config.yaml mcp.gitlab entries
- Add .dockerscripts/.env.example and .gitignore for secret handling
- Add mcp.github to opencode.jsonc and .hermes/config.yaml
  - points at GitHub hosted server (api.githubcopilot.com/mcp/)
  - PAT bearer auth; oauth auto-detection disabled
- Pass GITHUB_PERSONAL_ACCESS_TOKEN from .env to opencode/hermes-agent
  - lets client configs interpolate the token
- Extend .env.example with GitHub PAT placeholder and scope guidance
- No new container: GitHub hosts the server (unlike gitlab-mcp)
- Export ~/.supabase/bin so the Supabase CLI is in interactive shells
- Track frontend-design skill from anthropics/skills in the lock file
- Persist the list of agents last selected during skill sync
- Serialize toolset lists as JSON strings to match agent config schema
- Enable memory and user profile features
- Drop stale inline documentation comments superseded by agent docs
- Record onboarding prompts as already seen
- Build a claude-code image from an Alpine Node base
  - installs bash, git, and Claude Code via npm
- Add a claude-code compose service depending on gitlab-mcp
  - bind-mounts ~/.agents, ~/.claude, and ~/.claude.json
  - keeps config and skills persistent across runs
- Add a `claude` wrapper script running via `docker compose run --rm`
- Map port 3000 to host so node-tsserver servers are reachable locally
- Install git, make, and python3 for node-gyp native addon builds
- Install python3, python3-pip, and make in the opencode image
  - enables building native dependencies and running Python tooling
- Add Ubuntu-based image with Qt6, Mesa/Vulkan, and a JRE
- installs PrismLauncher from its official APT repository
- Add a thin wrapper around `docker compose run --rm`
- invokes any compose.yaml service ad hoc, like the `claude` wrapper
- Build a github-mcp image from GitHub's distroless compiled binary
  - runs under mcp-proxy (github-mcp-server lacks HTTP transport)
- Expose a Streamable HTTP endpoint on port 3003
  - serve /mcp path, mirroring the gitlab-mcp container
- Build the github-mcp image with pinned version pins
  - pins github-mcp-server binary and mcp-proxy wrapper
- Pass GITHUB_PERSONAL_ACCESS_TOKEN only to this container
  - centralizes the credential instead of copying it to every client
- Repoint opencode/hermes-agent github MCP at the github-mcp container
  - http://github-mcp:3003/mcp instead of GitHub's hosted endpoint
- Drop per-client GITHUB_PERSONAL_ACCESS_TOKEN header and compose env
  - only the github-mcp container needs the token now
- Add github-mcp to depends_on for both services
- Update the .env.example comment to reflect the new consumer
- Add project-scoped .mcp.json pointing Claude Code at the containers
  - gitlab-mcp and github-mcp, mirroring opencode/hermes-agent configs
- Add an orchestrator primary agent on glm-5.2
  - can only invoke the task tool
  - reading, commands, and GitLab/GitHub MCP calls all denied
- Add a worker subagent on deepseek-v4-flash with full tool access
- Keep the expensive model on planning/synthesis only
  - keeps autonomous runs within usage limits
  - cheap model does the token-heavy work
- The orchestrator's permission block denied "skill"
  - blocked loading github-autonomous-engineer and other skills
  - it had no way to learn its task instructions
- Loading a skill only pulls its SKILL.md into context
  - reads no project code and calls no MCP tools
  - keeps the orchestrator thin
- bind mount fails if these paths don't exist on host
- github-mcp and gitlab-mcp gain curl-based healthchecks
- claude-code, hermes-agent, opencode wait on healthy MCP
- avoids clients talking to an MCP server before it's ready
- mcp-proxy base image needed curl for the check to run
- claude-code talks to GitHub via github-mcp over MCP
- it never called the GitHub API directly
- host symlinks under ~/.agents or ~/repos use absolute paths
- translating to /root broke them inside the container
- mount both trees at the same path so they resolve
- Replace single worker with role-scoped subagents
  - implementer: edit+bash allowed, git commit/push and MCP denied
  - verifier: bash allowed for tests only, edit and MCP denied
  - committer: bash limited to git add/commit/status/diff/log
  - pusher: bash limited to git push/status/log/branch
  - vcs-mcp: gitlab_*/github_* allowed, edit and bash denied
- Keep the pipeline fully autonomous
  - drop "ask" permissions in favor of allow/deny throughout
- Route orchestrator delegation through the new pipeline
  - rewrite description/prompt to list the 5 subagents in order
- Disable unused default subagents
  - general: disable true
  - explore: disable true
  - scout: disable true
- Keep build and plan primary agents enabled
- Give OpenCode a way to build and test Docker images safely
  - new docker-mcp/{Dockerfile,entrypoint.sh,server.py,requirements.txt}
  - privileged DinD container isolates docker access from opencode
  - exposes build_docker_image and run_test_container over SSE
- Wire docker-mcp into the compose stack
  - opencode depends_on docker-mcp and shares /workspace with it
  - opencode env gets DOCKER_MCP_URL pointing at the sse endpoint
  - launch script mkdir's ~/workspace before compose up
- Scope docker_* tool access to the verifier role only
  - register docker-mcp in opencode.jsonc's mcp block
  - allow docker_* only for verifier; deny it on every other agent
- Fix -32001 timeouts during docker image builds
  - remote MCP clients default to a 5s timeout
  - docker-mcp's own build timeout is 300s, far longer
  - set the docker MCP entry's timeout to match, 300s
- Let build_docker_image accept per-build --build-arg values
  - new optional build_args mapping on the tool signature
  - each entry rendered as --build-arg before the build context path
- Needed to build subpath-deploy variants without hand-editing ARGs
- Add lifecycle tools beyond the existing throwaway run_test_container
  - start_container launches a named container in the background
  - container_logs and exec_in_container inspect/drive it while alive
  - stop_container force-removes it when the caller is done
- No automatic reaper; docker-mcp restarts wipe nested dind state anyway
- Let docker_cp move files between /workspace and a running container
  - direction picks to_container or from_container
  - workspace-side path stays bounded to /workspace like other tools
- Pairs with start_container for injecting config into a live container
- Let the agent see what's already on the docker-mcp daemon
  - list_images wraps docker images
  - list_containers wraps docker ps, showing stopped ones by default
  - inspect_object wraps docker inspect for an image or container
- Needed to check for stray images/containers left by earlier runs
- Add create_network/remove_network wrapping docker network create/rm
- Let run_test_container and start_container join a network
  - pass the same network name to reach containers by name
  - enables nginx-in-front-of-app style reverse proxy verification
- Add ports mapping to start_container, rendered as -p flags
  - nested daemon shares its netns; port lands on docker-mcp's address
  - reachable as docker-mcp:<host_port> from sibling services
- Not added to run_test_container; it exits before anything can dial it
- Switch docker-mcp off the SSE transport to streamable-http
  - serve the endpoint under /mcp via streamable_http_path
  - point client url, DOCKER_MCP_URL, and healthcheck at /mcp
- Require mcp>=1.9.0 for streamable-http transport support
- Move the catch-all "*" to the top of each bash block
  - applies to the read, commit, and push agent definitions
- Use shorthand passthrough for vars already in the environment
  - XDG_RUNTIME_DIR, DOCKER_CONTEXT, STREAMABLE_HTTP_AUTH_TOKEN
- Forward GITHUB/GITLAB personal access tokens to the mcp service
- Insert three phase leads between orchestrator and workers
  - planner delegates to issue-reader and code-explorer
  - builder delegates to coder and verifier
  - integrator delegates to committer, pusher, and pr-manager
- Run leads on deepseek-v4-pro, keeping leaves on v4-flash
- Scope permission.task per level so each tier calls only its own
  - leads coordinate only; direct read/edit/bash/MCP all denied
  - leaves set task deny to pin them as the bottom layer
- Rename implementer to coder and vcs-mcp to pr-manager
- Extract all 11 agent definitions from opencode.jsonc
  - into separate Markdown files under agents/
  - frontmatter carries mode/model/description/permission
  - lead prompts move to body; leaves stay body-less
  - config keeps only mcp, permission, and disable toggles
- Set subagent depth to 2 so leads can delegate to leaf agents
  - two-level nested delegation was silently blocked before
- Add `question: deny` to the orchestrator and every agent below
  - covers the primary, three middle leads, and seven leaves
  - stops the whole hierarchy from prompting the user mid-task
- Add a sessions MCP server over opencode session history
  - Serve list, full-text search, and get across past sessions
  - Read the opencode SQLite database read-only via FastMCP
  - Fall back to immutable reads when the WAL is unavailable
- Wire the server into the container stack
  - Add the sessions-mcp compose service with a healthcheck
  - Mount the opencode data directory read-only into it
  - Register the endpoint and gate opencode startup on it
- Scope access to preserve the delegation hierarchy
  - Grant the sessions tools to issue-reader alone
  - Deny them for every coordinator and other leaf
- Deny all tools by default, then allow only what each needs
  - Start every agent's permission block with a deny-all rule
  - Rely on last-match-wins so the explicit allows take effect
  - Drop the per-tool deny lines the default now covers
- Preserve each agent's existing effective capabilities
  - Keep coordinators limited to skill and their delegations
  - Keep leaves scoped to their own tools and MCP servers
  - Let coder and verifier still search via allowed bash
- Stop new tools from leaking in without an explicit grant
  - A newly added MCP server is denied everywhere by default
  - Only the agent meant to use it needs a single allow line
- Spell out that subagents share no conversation memory
  - Each task call is a fresh subagent's entire context
  - Forbid vague back-references like "as above"
- Tell each lead exactly what to paste into delegated prompts
  - orchestrator: requirement, then each lead's full prior output
  - planner: a self-contained ask for issue-reader/code-explorer
  - builder: requirement/files, and verifier's raw failure output
  - integrator: the actual commit message and PR title/description
- Add skill: allow to every lead and leaf
  - Three leads: planner, builder, integrator
  - Seven leaves: issue-reader through pr-manager
- Restore what the whitelist refactor had dropped
  - None allowed skill explicitly, so it defaulted to deny
  - Matches orchestrator, which already allowed it
- Add a forgejo-mcp server ahead of the GitHub/GitLab migration
  - Wrap the community forgejo-mcp npm package's HTTP mode
  - Talk to any self-hosted Forgejo/Gitea instance over its API
  - Guard the endpoint with the shared streamable-http bearer token
- Wire the server into the container stack
  - Add the forgejo-mcp compose service with a healthcheck
  - Gate opencode startup on the new service being healthy
  - Register the endpoint in opencode.jsonc alongside gitlab/github
- Document the new required environment variables
  - FORGEJO_URL, the target instance's base URL
  - FORGEJO_PERSONAL_ACCESS_TOKEN, the API token to call it with
- Whitelist forgejo MCP for the GitLab/GitHub agents
  - issue-reader: read-only context gathering for the planner
  - pr-manager: issue/PR operations for the integrator
- Update both descriptions to mention Forgejo alongside GitLab/GitHub
- Shared StreamableHTTPServerTransport rejected 2nd+ clients
  - error: "Server already initialized"
- Replace with per-session transport and server, keyed by session ID
- Verify single and concurrent multi-client MCP connections
- Build context is ./dockerfiles, not the forgejo-mcp subdir
- COPY forgejo-mcp/http.js so the patch file is found during build
- Verified with a docker build mirroring the compose context layout
- COPY overwrote http.js with mode 0644
  - broke the forgejo-mcp-http symlink with "Permission denied"
- Add --chmod=0755 so the patched file stays executable
- Verified the binary launches in a container from the fixed image
- Add gitpush-mcp: FastMCP server holding GitHub/GitLab/Forgejo tokens
  - same pattern as docker-mcp/sessions-mcp
  - exposes a single push(repo_path, remote, branch) tool
  - resolves the right token by matching the remote hostname
  - authenticates via a short-lived GIT_ASKPASS helper
  - token never touches argv or .git/config; scrubbed from output
  - always a plain push; force-push intentionally unsupported
- Remove PAT env vars from the opencode service in compose.yaml
  - tokens now reach only gitpush-mcp; no agent can read them via bash
- Wire gitpush-mcp into compose.yaml and opencode.jsonc
  - repos bind mount, healthcheck, depends_on, and mcp.gitpush
- Update pusher: replace bash `git push*` with `gitpush_*` MCP access
  - git status/log/branch remain allowed for read-only checks
- Add role/scope/output-contract prompts to leaf agents
  - code-explorer, coder, committer, issue-reader, pr-manager, verifier
  - they had frontmatter and permissions but no instructions
  - style matches the middle-lead agents (builder/planner/integrator)
- Reinforce each agent's permission boundary in prose too
  - defense in depth: prompt states it, permission block enforces it
  - add an explicit output-format expectation for the calling lead
  - e.g. verifier returns full failure output verbatim, not a paraphrase
genji1024 approved these changes 2026-08-20 16:43:17 +09:00
genji1024 left a comment
Copy link

LGTM!

LGTM!
genji1024 deleted branch migrate/from-github 2026-08-20 16:44:30 +09:00
genji1024 referenced this pull request from a commit 2026-08-20 16:44:32 +09:00
merge: switch init.vim to init.lua
genji1024 referenced this pull request from a commit 2026-08-20 16:44:32 +09:00
merge: migrate dotfiles from GitHub
Sign in to join this conversation.
Reviewers
No reviewers
genji1024
Labels
Clear labels   
enhancement
No labels
enhancement
Milestone
Clear milestone
No items
No milestone
Assignees
Clear assignees
No assignees
bot-genji1024
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
genji1024/dotfiles!2
No description provided.